ArticlePDF Available

Towards Understanding and Enhancing Association and Long Sleep in Low-Power WiFi IoT Systems

June 2021
IEEE Transactions on Green Communications and Networking PP(99):1-1

June 2021
PP(99):1-1

DOI:10.1109/TGCN.2021.3085908

Authors:

Jaykumar Sheth

Santa Clara University

Behnam Dezfouli

Santa Clara University

Enhancing the energy efficiency of WiFi IoT stations introduces unique challenges compared to 802.15.4 and BLE. The four essential operations performed to ensure connectivity between stations and the access point in a WiFi network are association, periodic beacon reception, maintaining association, and station wake up. Understanding and enhancing these operations are essential for building energy-efficient and dependable IoT systems. However, it is unclear how the software and hardware configuration of station and access point, concurrent traffic, power management, and security protocols affect the reliability and energy efficiency of these operations. In this paper, first, we present a thorough analysis of the association cost of WPA2 and WPA3 and mitigate the effect of key computation on association overhead. Second, we prove that increasing listen interval to reduce beacon reception wake-up duration may negatively impact energy efficiency. We identify the primary causes of this problem subject to link quality estimation algorithm and beacon delay. Third, we show that maintaining association by relying on access-point-based polling is not reliable. In particular, we confirm the wake-up delay of low-power stations is highly affected by factors such as channel utilization and beacon listen interval. We also confirm that key renewal aggravates the chance of disassociation.

Impact of DL and UL traffic on beacon transmission delay. (a) and (b): voice AC. (c) and (d): video AC. UL traffic has a higher effect on beacon transmission delay because it prevents the AP from channel access during TxOP.

…

Figures - uploaded by Behnam Dezfouli

Content may be subject to copyright.

Content uploaded by Behnam Dezfouli

Content may be subject to copyright.

IEEE TRANSACTIONS ON GREEN COMMUNICATIONS AND NETWORKING, 2021 1

Towards Understanding and Enhancing Association and Long

Sleep in Low-Power WiFi IoT Systems

Vikram K. Ramanna∗†, Jaykumar Sheth∗, Simon Liu∗, and Behnam Dezfouli∗

∗Internet of Things Research Lab, Computer Science and Engineering, Santa Clara University, USA

†Inﬁneon Technologies, San Jose, USA

{vramanna, jsheth, sliu3, bdezfouli}@scu.edu

Abstract—Enhancing the energy efﬁciency of WiFi IoT stations intro-

duces unique challenges compared to 802.15.4 and BLE. The four

essential operations performed to ensure connectivity between stations

and the access point in a WiFi network are association,periodic beacon

reception,maintaining association, and station wake up. Understand-

ing and enhancing these operations are essential for building energy-

efﬁcient and dependable IoT systems. However, it is unclear how the

software and hardware conﬁguration of station and access point, con-

current trafﬁc, power management, and security protocols affect the

reliability and energy efﬁciency of these operations. In this paper, ﬁrst,

we present a thorough analysis of the association cost of WPA2 and

WPA3 and mitigate the effect of key computation on association over-

head. Second, we prove that increasing listen interval to reduce beacon

reception wake-up duration may negatively impact energy efﬁciency.

We identify the primary causes of this problem subject to link quality

estimation algorithm and beacon delay. Third, we show that maintaining

association by relying on access-point-based polling is not reliable.

In particular, we conﬁrm the wake-up delay of low-power stations is

highly affected by factors such as channel utilization and beacon listen

interval. We also conﬁrm that key renewal aggravates the chance of

disassociation.

Index Terms—Empirical Evaluation, Energy Efﬁciency, 802.11, WPA2,

WPA3, Interference, Delay, Reliability.

1 INTRODUCTION

A WiFi network (a.k.a., 802.11 network) is composed of

two components: stations and Access Point (AP). Stations

connect to the AP to communicate with other stations and

the Internet. WiFi is an appealing technology for IoT connec-

tivity [1]–[4] considering multiple reasons: First, large-scale

deployment of WiFi APs provides a ready infrastructure

for IoT connectivity in license-free bands. Second, the high

data rate of this standard, compared to low-power tech-

nologies such as 802.15.4 and Bluetooth Low Energy (BLE),

facilitates the development of applications such as medical

monitoring, video streaming, process control, and robotics

[5]–[7]. Third, the energy consumption of WiFi stations has

been signiﬁcantly reduced during recent years. Speciﬁcally,

existing studies show the higher energy efﬁciency of WiFi’s

physical layer compared to BLE and LTE [3], [4].

Improving the energy efﬁciency of Internet of Things

(IoT) systems is important from two perspectives: First,

the impact of the energy consumption of these stations on

global Information and Communications Technology (ICT)

energy footprint is increasing [8]. Second, many of these

stations (e.g., smart locks, security cameras) run on bat-

tery or energy harvesting mechanisms. Thereby, resource-

constrained stations need to put their transceiver into sleep

mode aggressively to reduce energy consumption whenever

no communication is taking place.

Regardless of the application type, all WiFi IoT systems

need to perform the following operations to ensure AP-

station connectivity: association,periodic reception of beacon

packets,maintaining association, and station wake up. Associ-

ation refers to the process of exchanging station and AP

information (e.g., supported data rates) and establishing

keys for secure communication. Once associated, a low-

power IoT station needs to periodically wake up and receive

beacon packets sent by the AP. Beacon reception serves

three primary purposes: ﬁrst, the AP uses beacon packets

to inform the stations about their buffered packets, second,

the stations can measure their link quality to the AP, and

third, stations sync up their clock with the AP. Maintaining

association is performed by both sides—stations and the

AP. If a station misses a particular number of beacons,

it may initiate the roaming process or retry to reassociate

with the AP. On the other hand, if the AP does not hear

from a station for a particular time duration, it may try

to poll the station or simply disassociates the station. The

periodic wake up of stations and the need to maintain

association may cause unreliable or delayed station wake-

up. Despite the importance of these operations, existing

studies primarily focus on the effects of trafﬁc on energy

efﬁciency [1], [9], [10], or evaluate performance at a high

level [11], [12], and unfortunately, much less attention has

been paid to the impact of the above operations on system

dependability and energy efﬁciency.

In this paper, we present an empirical study of associa-

tion, beacon reception, maintaining association, and station

wake up delay, with focus on parameters including energy

efﬁciency, reliability, and delay. We also propose methods to

improve the performance of these operations. Speciﬁcally,

the contributions of this paper are as follows.

Association Overhead. The process of associating a sta-

tion with an AP impose a non-negligible delay and energy

consumption burden on the station. Through empirical eval-

uations, ﬁrst, we show that instead of exhaustive channel

sweeping, using speciﬁc AP probing can reduce association

IEEE TRANSACTIONS ON GREEN COMMUNICATIONS AND NETWORKING, 2021 2

overhead by about 40%. However, using this method causes

a higher number of association failures, especially when

the key generation duration increases or channel utilization

escalates. Second, we reveal the high overhead of IP assign-

ment even when a static IP is used. Speciﬁcally, our results

conﬁrm the higher IP assignment overhead of ThreadX with

NetXDuo stack compared to FreeRTOS with LwIP stack.

Third, we compare WPA2 and WPA3 and conﬁrm their rel-

ative performance depends on the processor and ﬁrmware

implementation. Fourth, to reduce the overhead of WPA2

key generation, we ofﬂoad the hashing function to the

application processor. We then compare the performance of

WPA2 with key ofﬂoading against WPA3 with key caching.

Beacon Reception. To reduce the overhead of beacon

reception, a straightforward approach is to increase the sta-

tion’s listen interval. However, we substantiate that employ-

ing this method may considerably increase the overhead

of beacon reception in terms of awake duration, which

translates to higher energy consumption. We identify three

reasons for this behavior: First, link quality estimation algo-

rithms (implemented in station’s ﬁrmware) exhibit various

levels of sensitivity to beacon loss. Second, we signify that

the Transmit Opportunity (TxOP) reserved by voice and

video ﬂows can cause considerable beacon transmission

delay. Third, as beacon loss increases, the awake time of

the station in advance of beacon reception may increase.

Maintaining Association. To avoid the overhead of re-

association, it is essential to maintain association with mini-

mum overhead. This is particularly challenging because APs

maintain a per-station inactivity timer to ensure the station

is alive and within the communication range. Although APs,

by default, poll inactive stations periodically, we prove the

unreliability of this method. In particular, this may lead

to what we call a ‘disassociation-unaware station’, which

means the station is unaware of its disassociation by the

AP. We also reveal key renewal as a time-critical operation

that may result in disassociation. We present solutions to

remedy these problems.

Wake-up Delay. Ensuring short, reliable wake-up delays

prevents cases such as polling failure and key renewal, and

it is also important for applications where the station must

be woken up by user or cloud applications. We consider

the two widely-used power-saving methods of WiFi, namely

Power Save Mode (PSM) and Automatic Power Save Deliv-

ery (APSD), and quantify wake up delay. Our main ﬁndings

are as follows. First, uplink trafﬁc has a higher effect on

prolonging wake up duration. Second, although APSD is

accepted as a more efﬁcient method to reduce AP-station

delay, the wake up delay of this method is higher than that

of PSM.

Segregating the overall system operation into the above

fundamental components allows for extending the pre-

sented studies and methods to a wide range of applica-

tions, hardware, and software platforms. This segregation

also allows us to tackle the challenges of holistic system

evaluation considering the large number of combinations of

conﬁguration parameters.

The rest of this paper is organized as follows: Sec-

tion 2 overviews the association process and power-saving

mechanisms. Section 3 presents our research methodology.

Association overhead is studied in Section 4. In Section 5,

Probe Request

Probe Response

(PRB

<latexit sha1_base64="V+6TZUu5jRA6RSd2E2mXgbxPPjE=">AAAB7nicbVBNSwMxEJ3Ur1q/qh69BItQL2VXKnos9eKxFvsB7VKyabYNzWaXJCuUpT/CiwdFvPp7vPlvTNs9aOuDgcd7M8zM82PBtXGcb5Tb2Nza3snvFvb2Dw6PiscnbR0lirIWjUSkuj7RTHDJWoYbwbqxYiT0Bev4k7u533liSvNIPpppzLyQjCQPOCXGSp1yo1kfNC8HxZJTcRbA68TNSAkyNAbFr/4woknIpKGCaN1zndh4KVGGU8FmhX6iWUzohIxYz1JJQqa9dHHuDF9YZYiDSNmSBi/U3xMpCbWehr7tDIkZ61VvLv7n9RIT3Hopl3FimKTLRUEisInw/Hc85IpRI6aWEKq4vRXTMVGEGptQwYbgrr68TtpXFbdauX6olmr1LI48nME5lMGFG6jBPTSgBRQm8Ayv8IZi9ILe0ceyNYeymVP4A/T5A/bKjq4=</latexit>

(PRB

<latexit sha1_base64="rDFiYjS0fPPGC1s3lq2JKFMmUvM=">AAAB7nicbVBNSwMxEJ3Ur1q/qh69BItQL2VXKnos9eKxFfsB7VKyabYNzWaXJCuUpT/CiwdFvPp7vPlvTNs9aOuDgcd7M8zM82PBtXGcb5Tb2Nza3snvFvb2Dw6PiscnbR0lirIWjUSkuj7RTHDJWoYbwbqxYiT0Bev4k7u533liSvNIPpppzLyQjCQPOCXGSp1y46E+aF4OiiWn4iyA14mbkRJkaAyKX/1hRJOQSUMF0brnOrHxUqIMp4LNCv1Es5jQCRmxnqWShEx76eLcGb6wyhAHkbIlDV6ovydSEmo9DX3bGRIz1qveXPzP6yUmuPVSLuPEMEmXi4JEYBPh+e94yBWjRkwtIVRxeyumY6IINTahgg3BXX15nbSvKm61ct2slmr1LI48nME5lMGFG6jBPTSgBRQm8Ayv8IZi9ILe0ceyNYeymVP4A/T5A/VFjq0=</latexit>

Auth Request

(AUT HQ)

<latexit sha1_base64="mJNeLjgPB/h6gWE+HIXFSQ21+TQ=">AAAB73icbVBNT8JAEJ3iF+IX6tHLRmKCF9IajB5RLxwhoUACDdkuW9iw3dbdrQlp+BNePGiMV/+ON/+NC/Sg4EsmeXlvJjPz/JgzpW3728ptbG5t7+R3C3v7B4dHxeOTtooSSahLIh7Jro8V5UxQVzPNaTeWFIc+px1/8jD3O09UKhaJlp7G1AvxSLCAEayN1C3fua36oHk5KJbsir0AWidORkqQoTEofvWHEUlCKjThWKmeY8faS7HUjHA6K/QTRWNMJnhEe4YKHFLlpYt7Z+jCKEMURNKU0Gih/p5IcajUNPRNZ4j1WK16c/E/r5fo4NZLmYgTTQVZLgoSjnSE5s+jIZOUaD41BBPJzK2IjLHERJuICiYEZ/XlddK+qjjVynWzWqrdZ3Hk4QzOoQwO3EAN6tAAFwhweIZXeLMerRfr3fpYtuasbOYU/sD6/AGSXY8F</latexit>

Auth Response

(AUT HR)

<latexit sha1_base64="GlMCmO4W9MM/3IILG/OL5uMlnQo=">AAAB73icbVBNT8JAEJ3iF+IX6tHLRmKCF9IajB5RLxzRUCCBhmyXLWzYbuvu1oQ0/AkvHjTGq3/Hm//GBXpQ8CWTvLw3k5l5fsyZ0rb9beXW1jc2t/LbhZ3dvf2D4uFRS0WJJNQlEY9kx8eKciaoq5nmtBNLikOf07Y/vpv57ScqFYtEU09i6oV4KFjACNZG6pRv3Ga9/3DeL5bsij0HWiVORkqQodEvfvUGEUlCKjThWKmuY8faS7HUjHA6LfQSRWNMxnhIu4YKHFLlpfN7p+jMKAMURNKU0Giu/p5IcajUJPRNZ4j1SC17M/E/r5vo4NpLmYgTTQVZLAoSjnSEZs+jAZOUaD4xBBPJzK2IjLDERJuICiYEZ/nlVdK6qDjVyuV9tVS7zeLIwwmcQhkcuIIa1KEBLhDg8Ayv8GY9Wi/Wu/WxaM1Z2cwx/IH1+QOT4o8G</latexit>

Assoc Request

Assoc Response

(ASCR)

<latexit sha1_base64="NbxjjQ9wtKeq5Lm5w0PBfbThOOE=">AAAB7nicbVDLTgJBEOz1ifhCPXqZSEzwQnYNRo8oF4/44JHAhswOszBhdnYyM2tCNnyEFw8a49Xv8ebfOMAeFKykk0pVd7q7AsmZNq777aysrq1vbOa28ts7u3v7hYPDpo4TRWiDxDxW7QBrypmgDcMMp22pKI4CTlvBqDb1W09UaRaLRzOW1I/wQLCQEWys1CpdP9R692e9QtEtuzOgZeJlpAgZ6r3CV7cfkySiwhCOte54rjR+ipVhhNNJvptoKjEZ4QHtWCpwRLWfzs6doFOr9FEYK1vCoJn6eyLFkdbjKLCdETZDvehNxf+8TmLCKz9lQiaGCjJfFCYcmRhNf0d9pigxfGwJJorZWxEZYoWJsQnlbQje4svLpHle9irli7tKsXqTxZGDYziBEnhwCVW4hTo0gMAInuEV3hzpvDjvzse8dcXJZo7gD5zPH+LSjqE=</latexit>

(ASCQ)

<latexit sha1_base64="ri6E0/bXdJidIV14A22z59BU81Q=">AAAB7nicbVBNSwMxEJ31s9avqkcvwSLUS9mVih6rvXhs0X5Au5Rsmm1Ds9mQZIWy9Ed48aCIV3+PN/+NabsHbX0w8Hhvhpl5geRMG9f9dtbWNza3tnM7+d29/YPDwtFxS8eJIrRJYh6rToA15UzQpmGG045UFEcBp+1gXJv57SeqNIvFo5lI6kd4KFjICDZWapduH2r9xkW/UHTL7hxolXgZKUKGer/w1RvEJImoMIRjrbueK42fYmUY4XSa7yWaSkzGeEi7lgocUe2n83On6NwqAxTGypYwaK7+nkhxpPUkCmxnhM1IL3sz8T+vm5jwxk+ZkImhgiwWhQlHJkaz39GAKUoMn1iCiWL2VkRGWGFibEJ5G4K3/PIqaV2WvUr5qlEpVu+yOHJwCmdQAg+uoQr3UIcmEBjDM7zCmyOdF+fd+Vi0rjnZzAn8gfP5A+FNjqA=</latexit>

EAP Msg 1

EAP Msg 2

(EAP1)

<latexit sha1_base64="q1FYd0CsMj+LtzZmWz3vMMwXKAQ=">AAAB7nicbVDLSgNBEOz1GeMr6tHLYBDiJexKRI9RETxGMA9IljA76U2GzM4uM7NCCPkILx4U8er3ePNvnCR70MSChqKqm+6uIBFcG9f9dlZW19Y3NnNb+e2d3b39wsFhQ8epYlhnsYhVK6AaBZdYN9wIbCUKaRQIbAbD26nffEKleSwfzShBP6J9yUPOqLFSs3R3Xet6Z91C0S27M5Bl4mWkCBlq3cJXpxezNEJpmKBatz03Mf6YKsOZwEm+k2pMKBvSPrYtlTRC7Y9n507IqVV6JIyVLWnITP09MaaR1qMosJ0RNQO96E3F/7x2asIrf8xlkhqUbL4oTAUxMZn+TnpcITNiZAllittbCRtQRZmxCeVtCN7iy8ukcV72KuWLh0qxepPFkYNjOIESeHAJVbiHGtSBwRCe4RXenMR5cd6dj3nripPNHMEfOJ8/rxyOfw==</latexit>

(EAP2)

<latexit sha1_base64="svBSmVN830jb6JDFsBKFuam7PPI=">AAAB7nicbVBNSwMxEJ2tX7V+VT16CRahXspuqeixKoLHCvYD2qVk02wbmmSXJCuUpT/CiwdFvPp7vPlvTNs9aOuDgcd7M8zMC2LOtHHdbye3tr6xuZXfLuzs7u0fFA+PWjpKFKFNEvFIdQKsKWeSNg0znHZiRbEIOG0H49uZ336iSrNIPppJTH2Bh5KFjGBjpXb57rrRr573iyW34s6BVomXkRJkaPSLX71BRBJBpSEca9313Nj4KVaGEU6nhV6iaYzJGA9p11KJBdV+Oj93is6sMkBhpGxJg+bq74kUC60nIrCdApuRXvZm4n9eNzHhlZ8yGSeGSrJYFCYcmQjNfkcDpigxfGIJJorZWxEZYYWJsQkVbAje8surpFWteLXKxUOtVL/J4sjDCZxCGTy4hDrcQwOaQGAMz/AKb07svDjvzseiNedkM8fwB87nD7ChjoA=</latexit>

EAP Msg 3

EAP Msg 4

(EAP3)

<latexit sha1_base64="tUrRBmdvDSxi8lgVTMHLeW6X+JI=">AAAB7nicbVBNSwMxEJ2tX7V+VT16CRahXsquVvRYFcFjBfsB7VKyabYNzWZDkhXK0h/hxYMiXv093vw3pu0etPXBwOO9GWbmBZIzbVz328mtrK6tb+Q3C1vbO7t7xf2Dpo4TRWiDxDxW7QBrypmgDcMMp22pKI4CTlvB6Hbqt56o0iwWj2YsqR/hgWAhI9hYqVW+u673zk97xZJbcWdAy8TLSAky1HvFr24/JklEhSEca93xXGn8FCvDCKeTQjfRVGIywgPasVTgiGo/nZ07QSdW6aMwVraEQTP190SKI63HUWA7I2yGetGbiv95ncSEV37KhEwMFWS+KEw4MjGa/o76TFFi+NgSTBSztyIyxAoTYxMq2BC8xZeXSfOs4lUrFw/VUu0miyMPR3AMZfDgEmpwD3VoAIERPMMrvDnSeXHenY95a87JZg7hD5zPH7ImjoE=</latexit>

(EAP4)

<latexit sha1_base64="AqeOd3smQ2I5mZhWGMFDcyoF+bQ=">AAAB7nicbVDLSgNBEOz1GeMr6tHLYBDiJexKRI9RETxGMA9IljA7mU2GzM4uM71CCPkILx4U8er3ePNvnCR70MSChqKqm+6uIJHCoOt+Oyura+sbm7mt/PbO7t5+4eCwYeJUM15nsYx1K6CGS6F4HQVK3ko0p1EgeTMY3k795hPXRsTqEUcJ9yPaVyIUjKKVmqW761q3ctYtFN2yOwNZJl5GipCh1i18dXoxSyOukElqTNtzE/THVKNgkk/yndTwhLIh7fO2pYpG3Pjj2bkTcmqVHgljbUshmam/J8Y0MmYUBbYzojgwi95U/M9rpxhe+WOhkhS5YvNFYSoJxmT6O+kJzRnKkSWUaWFvJWxANWVoE8rbELzFl5dJ47zsVcoXD5Vi9SaLIwfHcAIl8OASqnAPNagDgyE8wyu8OYnz4rw7H/PWFSebOYI/cD5/ALOrjoI=</latexit>

Auth Commit 1

Auth Commit 2

Auth Conﬁrm 1

Auth Conﬁrm 2

(AUT HCMT 1)

<latexit sha1_base64="JyF8/U23lcwcSctoGGrd9HGc45Y=">AAAB9HicbVBNT8JAEJ3iF+IX6tHLRmKCF9IajB5RLlxMMKFAAg3ZLgts2G7r7paENPwOLx40xqs/xpv/xgV6UPAlk7y8N5OZeX7EmdK2/W1lNja3tneyu7m9/YPDo/zxSVOFsSTUJSEPZdvHinImqKuZ5rQdSYoDn9OWP67O/daESsVC0dDTiHoBHgo2YARrI3nFO7dR6yXVh4Yzu+zlC3bJXgCtEyclBUhR7+W/uv2QxAEVmnCsVMexI+0lWGpGOJ3lurGiESZjPKQdQwUOqPKSxdEzdGGUPhqE0pTQaKH+nkhwoNQ08E1ngPVIrXpz8T+vE+vBrZcwEcWaCrJcNIg50iGaJ4D6TFKi+dQQTCQztyIywhITbXLKmRCc1ZfXSfOq5JRL14/lQuU+jSMLZ3AORXDgBipQgzq4QOAJnuEV3qyJ9WK9Wx/L1oyVzpzCH1ifP/vUkPM=</latexit>

(AUT HCMT 2)

<latexit sha1_base64="rjT5HCrWHXq9doN8xpSnm6plPQg=">AAAB9HicbVBNS8NAEJ34WetX1aOXxSLUS0lKRY/VXnoRKvQL2lA22027dLOJu5tCCf0dXjwo4tUf481/47bNQVsfDDzem2FmnhdxprRtf1sbm1vbO7uZvez+weHRce7ktKXCWBLaJCEPZcfDinImaFMzzWknkhQHHqdtb1yd++0JlYqFoqGnEXUDPBTMZwRrI7mFu2aj1k+qD43S7Kqfy9tFewG0TpyU5CFFvZ/76g1CEgdUaMKxUl3HjrSbYKkZ4XSW7cWKRpiM8ZB2DRU4oMpNFkfP0KVRBsgPpSmh0UL9PZHgQKlp4JnOAOuRWvXm4n9eN9b+rZswEcWaCrJc5Mcc6RDNE0ADJinRfGoIJpKZWxEZYYmJNjllTQjO6svrpFUqOuXi9WM5X7lP48jAOVxAARy4gQrUoA5NIPAEz/AKb9bEerHerY9l64aVzpzBH1ifP/1akPQ=</latexit>

(AUT HCNF 1)

<latexit sha1_base64="6zH0sA9ADQQEkEBM0Y2EJtvLcTo=">AAAB9HicbVBdSwJBFL3bp9mX1WMvQxLYi+yGUY+WED6FgauCLjI7jjo4O7vNzAqy+Dt66aGIXvsxvfVvGnUfSjtw4XDOvdx7jx9xprRtf1tr6xubW9uZnezu3v7BYe7ouKHCWBLqkpCHsuVjRTkT1NVMc9qKJMWBz2nTH1VmfnNMpWKhqOtJRL0ADwTrM4K1kbzCrVuvdpPKw70zvejm8nbRngOtEicleUhR6+a+Or2QxAEVmnCsVNuxI+0lWGpGOJ1mO7GiESYjPKBtQwUOqPKS+dFTdG6UHuqH0pTQaK7+nkhwoNQk8E1ngPVQLXsz8T+vHev+jZcwEcWaCrJY1I850iGaJYB6TFKi+cQQTCQztyIyxBITbXLKmhCc5ZdXSeOy6JSKV4+lfPkujSMDp3AGBXDgGspQhRq4QOAJnuEV3qyx9WK9Wx+L1jUrnTmBP7A+fwDn+pDm</latexit>

(AUT HCNF 2)

<latexit sha1_base64="2X0FpovhGsyaibmVMguxdvgr6ro=">AAAB9HicbVBdSwJBFL1rX2ZfVo+9DElgL7IrRj1aQvgUBq4KusjsOKuDsx/NzAqy+Dt66aGIXvsxvfVvGnUfSjtw4XDOvdx7jxtxJpVpfhuZjc2t7Z3sbm5v/+DwKH980pJhLAi1SchD0XGxpJwF1FZMcdqJBMW+y2nbHdfmfntChWRh0FTTiDo+HgbMYwQrLTnFW7tZ7ye1h/vy7LKfL5glcwG0TqyUFCBFo5//6g1CEvs0UIRjKbuWGSknwUIxwuks14sljTAZ4yHtahpgn0onWRw9QxdaGSAvFLoChRbq74kE+1JOfVd3+liN5Ko3F//zurHybpyEBVGsaECWi7yYIxWieQJowAQlik81wUQwfSsiIywwUTqnnA7BWn15nbTKJatSunqsFKp3aRxZOINzKIIF11CFOjTABgJP8Ayv8GZMjBfj3fhYtmaMdOYU/sD4/AHpgJDn</latexit>

(a) WPA2 (b) WPA3

Station StationAP AP

Probe Request

Probe Response

(PRB

Assoc Request

Assoc Response

(ASCR)

(ASCQ)

EAP Msg 1

EAP Msg 2

(EAP1)

(EAP2)

EAP Msg 3

EAP Msg 4

(EAP3)

(EAP4)

Step 1

Step 2

Step 3

Step 4

Step 1

Step 2

Step 3

Step 4

Step 5

IP Assignment

Fig. 1: Association process performed by (a) WPA2 (b) WPA3.

we study the impact of listen interval on energy efﬁciency.

Mechanisms for preventing reassociation are presented in

Section 6. A quantitative study of wake up delay is given

in Section 7. We overview the related work in Section 8 and

conclude the paper in Section 9.

2 BACKG RO UND

In this section, we provide an overview of the association

process as well as the mechanisms that enable stations

to leverage power saving via switching off their wireless

transceiver.

2.1 Association Process

The association process is composed of the following steps,

as demonstrated in Figure 1. Step 1: The station scans for

nearby APs by sending probe requests on all the channels.

Among the APs whose Service Set Identiﬁer (SSID) matches

with that programmed in the station, the compatible AP

with the highest signal strength is selected. Step 2: The

station authenticates with the AP. With WPA2, this primar-

ily means the two sides share information such as their

MAC addresses to generate Pairwise Master Key (PMK).

With WPA3, the station and AP share information (e.g.,

calculated scalar and point) that allow them to perform the

Simultaneous Authentication of Equals (SAE) handshake,

a method that has been originally designed for 802.11s

mesh networks. The SAE handshake, which is similar to

Difﬁe–Hellman (DH) key exchange with authentication,

is based on a zero-knowledge proof algorithm known as

Dragonﬂy. Step 3: The station sends an association request

to the AP and awaits an association response. The primary

result of this step is assigning an Association ID (AID)

to the station. Step 4: A 4-way key exchange happens to

generate Pairwise Transition Keys (PTK) and Group Tem-

poral Key (GTK), which are used for unicast and broadcast

communication, respectively. Step 5: If no static IP has been

conﬁgured, the station requests for IP address using DHCP.

2.2 Energy Saving Mechanisms

The 802.11 standard offers various power-saving methods.

The Power Save Mode (PSM) enables the stations to wake

up periodically at each Beacon Interval (BI). The BI value

used by commercial APs is 102.4 ms. Before transitioning

into the sleep mode, the station informs the AP about

IEEE TRANSACTIONS ON GREEN COMMUNICATIONS AND NETWORKING, 2021 3

Host 1

Access Point 1

(AP1)

hostapd

AP Monitor

Station

(CYW/BCM)

Collector

Ethernet Connection

Wireless

Connection

Serial

Connection

…

Smartphone1

Other IoT

Stations

Access Point 2

(AP2)

Smartphone2

Host 2

Sniﬀer

Pinger

Fig. 2: The testbed used in this work. As we will explain in the

subsequent sections, various experiments may employ subsets of this

testbed’s components.

its decision via a successful packet exchange where the

power management bit (in the MAC header) is set. Every

BI instance, the AP broadcasts a beacon message which, in

addition to other information, conveys to the stations if the

AP has buffered packets for the stations. This is performed

by setting the AID bit of the station in the Trafﬁc Indication

Map (TIM) ﬁeld. In this case, the station stays awake, sends

a PS-Poll message to the AP, and waits for downlink packet

transmission.

With PSM, stations need to retrieve downlink packets

every BI. Since the delay of this mechanism may not be

acceptable for interactive applications, the Automatic Power

Save Delivery (APSD) mechanism has been introduced.

With APSD, a station can poll the AP anytime by sending a

Null packet. The APSD mechanism, which is part of the

802.11e amendment (and is available in 802.11n/ac/ax),

also introduces the concept of Access Category (AC) to

differentiate the priority of voice, video, best-effort, and

background trafﬁc types. From left to right, the priority of

channel access reduces.

To deliver multicast and broadcast packets such as Ad-

dress Resolution Protocol (ARP), the AP needs to ensure all

the stations are awake. These packets are delivered after the

transmission of a beacon packet including the Delivery Traf-

ﬁc Indication Message (DTIM) element. The DTIM period is

conﬁgured as a multiple of beacon interval in the range of 1

to 255.

3 METHODOLOGY

Figure 2 shows the testbed used in this paper. We explain

the components and operation of this testbed as follows.

3.0.1 IoT stations

We use multiple types of stations in this paper. However, the

main two low-power WiFi transceivers used in our studies

are: CYW43907 [13], and BCM4343W [14]. CYW43907 (hence-

forth as the CYW station) includes two ARM-Cortex R4 pro-

cessors. BCM4343W (henceforth as the BCM station) includes

two ARM-Cortex M3 processors. In both systems, one core

is dedicated to the wireless subsystem and the other core

comprises the application subsystem. These two transceivers

are used by various low-power IoT development boards as

well as products such as Amazon Tap, LG Urbane Watch,

and Samsung Gear S2. The operating systems supported by

these boards are FreeRTOS [15], [16] and ThreadX [17], [18]

and the networking stacks are LwIP [19] (for FreeRTOS) and

NetXDuo [20] (for ThreadX). Unless otherwise mentioned,

ThreadX and NetXDuo are, respectively, the operating sys-

tem and network stack used by these stations.

3.0.2 AP

Unless otherwise mentioned, the transceiver used by the

APs is Atheros AR9462. AP functionality is handled by

hostapd, which is a user-space daemon used by most

commercial APs. We use the 802.11n standard as the com-

munication protocol between stations and AP. The APs

operate in the 2.4 GHz band. The APs and stations support

PSM and APSD.

3.0.3 Channel utilization

We access driver’s counters to measure the channel utiliza-

tion consumed by Downlink (DL)/Uplink (UL) and interfer-

ence trafﬁc. Since we use Atheros transceiver, we monitor

AR_CCCNT register value, which stores the time elapsed

since the start time of the transceiver, and AR_RCCNT reg-

ister value, which stores the activity duration sensed on

the channel. We refer to the value of the ﬁrst and second

registers as Tand B, and compute channel utilization

during an interval t1to t2as (Bt2−Bt1)/(Tt2−Tt1).

3.0.4 Controlled concurrent channel access scenarios

To measure the effect of concurrent trafﬁc, channel access,

and link unreliability, we use the following cases: pres-

ence of DL/UL trafﬁc, and presence of interference. In

the DL/UL case, AP1 is associated with the CYW/BCM

station and Smartphone1. AP1 and Smartphone1 continu-

ously exchange DL and/or UL trafﬁc. In the interference

case, AP1 is associated only with CYW/BCM; also, AP2

and Smartphone2 (associated with it) are operating in the

vicinity (about three meters) and exchange DL and UL

trafﬁc. AP1 and AP2 operate on the same channel. Host1

and Host2 are used to generate DL trafﬁc from AP1 and

AP2 to Smartphone1 and Smartphone2, respectively. The

Pinger is used for generating ping packets towards the

station under test. The distance between stations and their

associated AP is about two meters.

Figures 3 (a) and (b) characterize CYW-to-AP and AP-

to-CYW station link quality by demonstrating the average

number of retransmissions per packet. We ran each exper-

iment 10 times for 1000 packets sent per round. In other

words, for each round of 1000 packets, we compute packet

retransmission rate by dividing the total number of retrans-

missions by 1000. We observe that DL, UL, and interference

almost equally affect station-to-AP link. In contrast, for AP-

to-station communication, the effect of DL trafﬁc is almost

negligible because it causes packet transmission delay in-

stead of packet loss. Note that the maximum number of

MAC layer retransmissions per data packet is 7 with most

wireless transceivers.

3.0.5 Power measurement

The energy measurement tool is EMPIOT [21]. This plat-

form’s basic sampling rate is 500 Ksps, which are then

averaged and streamed to the control software at 1 Ksps.

Its maximum accuracy error is 4% compared to the existing

high-end commercial products.

4 ASSOCIATION OVERHEAD

In this section, we empirically evaluate the overhead of

association process and present methods to alleviate it.

IEEE TRANSACTIONS ON GREEN COMMUNICATIONS AND NETWORKING, 2021 4

Station to AP Transmission AP to Station Transmission

(a) (b)

Fig. 3: Average number of retransmissions per data packet for (a)

station-to-AP transmission, and (b) AP-to-station transmission. X-axis

values refer to the level of channel utilization by Downlink (DL), Uplink

(UL), and interference.

Fig. 4: The duration ((a) and (b)) and energy consumption ((c) and

(d)) of various association methods using the CYW station. The PMK

ofﬂoading method is referred to as PMKO, and the speciﬁc probing

method is denoted as SP.

4.1 Effect of Probing, Key Generation, Operating Sys-

tem, and Network Stack on Association

In this section, we study the impact of various param-

eters on association. Channel utilization level in all the

experiments is less than 10%. The station used in these

experiments is CYW. With regard to the testbed demon-

strated in Figure 2, only the CYW station and AP1 are the

active components of these experiments. We use the WPA2

protocol in this section. Also, it is worth noting that WPA2

is still the most widely used authentication protocol on IoT

stations, including Raspberry Pi (RPi) (3B+ and 4), Ring

Camera, Nest Camera, and Amazon Echo. Figure 4 presents

the results.

4.1.1 Probing

The probing process requires the station to send multiple

probe packets on each channel, and wait long enough (dwell

time) for the reception of responses. This process increases

the energy consumption of the station and elongates the

association process.

Speciﬁc Probing (SP). In this method, we reduce asso-

ciation overhead by sending a probe request to a particular

AP on a channel known to the station. We implement SP by

using driver interfaces that allow us to program the channel

and SSID of a particular AP in the station’s ﬁrmware. When

using SP, the probe request is still a broadcast packet,

however, the packet is sent on the programmed channel

only and includes the SSID of the target AP. As Figure 4

shows, for FreeRTOS, the SP method reduces the duration

and energy consumption by 39.66% and 43.58% on average,

respectively, compared to a regular association. These values

are 21.73% and 21.67% for ThreadX.

4.1.2 PMK Computation

During the 4-way handshake of WPA2, a shared se-

cret key called PMK is generated. Using SHA-1 algo-

rithm, PMK is derived from Pre-Shared Key (PSK), SSID,

and SSID length using the Password-Based Key Deriva-

tion Function 2 (PBKDF2) hashing algorithm as follows:

PBKDF2(HMAC-SHA1, PSK, SSID, 4096, 256). Here,

a 256-bit PMK is generated from 4096 iterations of the

hashing method. This function is implemented in the wire-

less ﬁrmware, and therefore, it is run by the transceiver’s

processor. However, PMK generation is a process-intensive

operation and increases the energy consumption and dura-

tion of association.

PMK Ofﬂoading (PMKO). In this method, we ofﬂoad

the calculation of PMK to the application processor. It is

worth noting that considering the complexity of 802.11

compared to standards such as 802.15.4, a separate processor

is used in the wireless subsystem of WiFi stations. For ex-

ample, both CYW and BCM stations include two processors:

one for wireless connectivity, and one for running applica-

tion threads. Therefore, the PMKO method can leverage the

application processor, which is usually at least as powerful

as the wireless subsystem’s processor. We implemented

PMKO on the CYW station by computing PMK in the

application processor and then passing it to the transceiver.1

As soon as the station boots up, the application processor

computes PMK and makes it available to be used during

the association process.

The results in Figure 4 demonstrate the effect of PMKO

on association overhead. On average, using PMKO with

FreeRTOS reduces delay and energy consumption by 18.87%

and 23.38%, respectively. These values are 4.01% and 5.39%

for ThreadX. Although ThreadX shows higher association

overhead compared to FreeRTOS, the difference is not

caused by the WPA2 method. We observed that the higher

overhead is caused by IP assignment processing and the

packet exchanges made to inform the AP (and other nodes)

about IP assignment. Therefore, the IP assignment overhead

of ThreadX overshadows the beneﬁts achieved by using

PMKO.

1. The implementation is available at the following link:

https://github.com/SIOTLAB/Low-Power-WiFi-Association.git

IEEE TRANSACTIONS ON GREEN COMMUNICATIONS AND NETWORKING, 2021 5

4.1.3 IP Assignment

Although using static IP allocation prevents the need to

communicate with a DHCP server, the results in Figure 4

show the considerable impact of static IP allocation when

using ThreadX w/ NetXDuo, compared to a marginal effect

with FreeRTOS w/ LwIP. With NetXDuo, the association

process includes the transmission of Gratuitous ARP pack-

ets after IP assignment. The station broadcasts these ARP

packets to announce its IP to the entire network. This extra

step is not performed by LwIP stack.

4.2 WPA2 vs WPA3

In this section, we compare the overhead of WPA2 and

WPA3 while varying channel utilization level by controlling

the amount of UL and DL trafﬁc exchanged between AP1

and Smartphone1 (§3). The AC of this trafﬁc is voice. Also,

considering the performance beneﬁts of SP (§4.1.1), we use

this method for all the experiments. Compared to the exper-

iments demonstrated in the previous section, and in order to

make the observations independent of the operating system

and protocol stack used, in this section we merely focus on

the ﬁrst four stages of association process, as demonstrated

in Figure 1. Therefore, the overhead of IP assignment has

been excluded from all the results of this section.

4.2.1 WPA2 vs WPA3

Similar to WPA2, a PMK must be calculated during Step 2

of WPA3 association. However, this calculation is different

than that of WPA2; this is because the main vulnerability

of WPA2 is its heavy reliance on using PSK to compute

PMK. In contrast, WPA3 uses zero-knowledge proof for

PMK generation to ensure no elements of the PSK are

exchanged between the station and AP. This calculation

introduces a heavy processing load due to hashing and DH

key generation. Figure 5 compares the overhead of WPA2

and WPA3. We ﬁrst discuss the default operation of the

two algorithms, demonstrated simply as WPA2 and WPA3

in the legends of this ﬁgure. With CYW, the duration and

energy consumption of WPA3 are lower than that of WPA2,

however, the overhead of WPA3 is higher than WPA2 for

BCM station. Since we use a similar ﬁrmware and driver

on the CYW and BCM stations, their main difference is the

processor used in their wireless subsystems. CYW includes

a Cortex-R4 processor which implements the ARMv7-R

architecture with an eight-stage pipeline and includes data

and instruction caches. BCM’s wireless subsystem includes

a Cortex-M3 processor, which implements the ARMv7-M

architecture with a three-stage pipeline and without cache

memory. We conjecture that the performance differences are

caused by the wireless subsystem’s processor type. To verify

this, we chose additional off-the-shelf stations that employ

Cortex-M3 (CYW43364 [22]) and Cortex-R4 (CYW43455 [23],

RPi) in their wireless subsystem. Since the default ﬁrmware

and driver of the RPi do not support WPA3, we complied

a Full MAC (F-MAC) Linux driver and ﬂashed a new

ﬁrmware to this station to enable WPA3 association. There-

fore, the driver and ﬁrmware used for RPi’s CYW43455 are

different than those used by CYW43455. Figure 6 presents

the PMK calculation and total association duration of these

10% 50% 80% 95%

Channel Utilization

(a)

0.0

0.2

0.4

0.6

0.8

1.0

1.2

1.4

1.6

Duration [sec]

CYW

WPA2

WPA3

WPA2+PMKO

WPA3+PMKC

10% 50% 80% 95%

Channel Utilization

(b)

Duration [sec]

BCM

WPA2

WPA3

WPA2+PMKO

WPA3+PMKC

10% 50% 80% 95%

Channel Utilization

(c)

0.00

0.25

0.50

0.75

1.00

1.25

1.50

1.75

2.00

Energy Consumption [J]

CYW

WPA2

WPA3

WPA2+PMKO

WPA3+PMKC

10% 50% 80% 95%

Channel Utilization

(d)

0.0

0.5

1.0

1.5

2.0

2.5

3.0

Energy Consumption [J]

BCM

WPA2

WPA3

WPA2+PMKO

WPA3+PMKC

Fig. 5: The duration ((a) and (b)) and energy consumption ((c) and (d))

of WPA2 and WPA3 association using CYW and BCM stations. Channel

utilization is generated by controlling the amount of UL and DL trafﬁc

between AP1 and Smartphone1. All the association scenarios use SP.

0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0

Duration [sec]

Samsung G20 WPA3

Samsung G20 WPA2

iPhone XS WPA3

iPhone XS WPA2

Rpi3b+ WPA3

Rpi3b+ WPA2

CYW43364 WPA3(M3)

CYW43364 WPA2(M3)

CYW43455 WPA3(R4)

CYW43455 WPA2(R4)

BCM WPA3(M3)

BCM WPA2(M3)

CYW WPA3(R4)

CYW WPA2(R4)

PMK Calculation

Association Duration

Excluding PMK Calculation Duration

Fig. 6: PMK computation and total association duration for stations

that use Cortex-M3 (BCM4343 [14], CYW43364 [22]) and Cortex-R4

(CYW43907 [13], CYW43455 [23], RPi3B+) in their wireless subsystem.

stations. These results were collected when the channel uti-

lization was no more than 10%. To compute PMK calculation

duration, we used the Sniffer (cf. Figure 2) and measured

the time interval between the reception of probe response

packet (P RBR) and the transmission of ﬁrst authentication

message (AU T HCM T 1). The results conﬁrm that the ratio of

WPA3 to WPA2 PMK calculation duration is less than one

for Cortex-R4 stations; whereas, the ratio is greater than one

for Cortex-M3 stations. The results also reveal the longer

duration of WPA3 compared to WPA2 for Apple iPhone

XS and Samsung Galaxy phones; however, we can cannot

conﬁrm the processor and driver type of these stations.

IEEE TRANSACTIONS ON GREEN COMMUNICATIONS AND NETWORKING, 2021 6

(a) (b)

(e) (f)

CYW | WPA2 | w/o PMK Oﬄoading

[ms]

CYW | WPA2 | w/ PMK Oﬄoading BCM | WPA2 | w/ PMK Oﬄoading

CYW | WPA3 | w/o PMK Caching BCM | WPA3 | w/o PMK Caching

(g)

BCM | WPA3 | w/ PMK Caching

[ms]

[ms] [ms]

(h)

Current Consumption [mA]

CYW | WPA3 | w/ PMK Caching

115 mA 102 mA

BCM | WPA2 | w/o PMK Oﬄoading

[ms]

135 mA

110 mA 138 mA

127 mA

136 mA

104 mA

Authentication Duration = 923 ms Authentication Duration = 2176 ms

Authentication Duration = 92 ms Authentication Duration = 101 ms

Authentication Duration = 820 ms Authentication Duration = 4000 ms

Authentication Duration= 120 ms

Authentication Duration = 110 ms

Fig. 7: Energy consumption of CYW and BCM stations during the

association process using WPA2 and WPA3. The shaded parts refer

to PMK calculation.

4.2.2 PMK Caching for WPA3

Although PMK ofﬂoading is not possible with WPA3, we

can expedite the association process by using PMK Caching

(PMKC), which is part of the standard. If the station and

AP possess PMK caches, they can skip SAE and proceed

with the 4-way handshake. With regard to Figure 1, using

PMKC reduces the number of packets exchanged during

Step 2 to two packets. Speciﬁcally, the station uses the

previously-computed PMK by sending an authentication

message with the authentication algorithm set to open.

The AP then checks the validity of PMKID, which is a

unique key identiﬁer maintained by the AP on a per-station

basis. If the PMKID is valid, the AP responds with an

authentication commit response packet and the association

process proceeds with association request. Figure 7 presents

the power consumption trace of CYW and BCM stations

and compares how PMKO and PMKC affect WPA2 and

WPA3, respectively. These results were collected when the

channel utilization was no more than 10%. As the results

show, both PMKO and PMKC eliminate the long ﬂat part

of the graph where PMK computation is occurring. For

example, on the CYW station, using PMKC reduces the total

association duration of WPA3 from 820 ms to 110 ms, and on

the BCM station, the drop is from 4000 ms to 120 ms. These

results also show the higher power consumption during

key calculation. Referring back to Figure 5, the duration

and energy consumption of WPA2 and WPA3 are very

similar when PMKO and PMKC are used. This is because,

in addition to excluding the heavy processing load of key

calculation, both WPA2 and WPA3 need to exchange eight

packets to perform association.

4.3 Channel Utilization

The results in Figure 5 conﬁrm that as the channel utilization

escalates, the duration and energy consumption of associa-

tion process increase. This is because intensifying channel

utilization increases packet collision rate, the number of

retransmissions, and packet exchange duration.

50% 75% 95%

Channel Utilization

(a)

100

Association Success Rate (%)

CYW

50% 75% 95%

Channel Utilization

(b)

100

Association Success Rate (%)

BCM

WPA2 WPA2+PMKO WPA3 WPA3+PMKC

Fig. 8: Association success rate versus channel utilization level for

WPA2 and WPA3. These results exclude the failures caused by missing

probe request and response packets.

To quantify how packet loss and delay affect associa-

tion reliability, we measured the association success rate

of WPA2 and WPA3 in the presence of various levels of

channel utilization. Figure 8 presents these results. A suc-

cessful association requires completion of the four steps

explained in §2.1, and it is important to note that the success

probability of these steps vary. Probe request is not followed

by layer-2 acknowledgments; it is simply transmitted for a

given number of times, depending on the station and AP

conﬁguration. For example, the CYW43907 ﬁrmware sends

three probe request packets. Hence, if the station or AP miss

these packets, the whole association process fails. For the

results in Figure 8, we have excluded the failures occurring

during Step 1, and instead, focused on the rest of the steps.

In contrast with probe request, the following packets require

layer-2 acknowledgment: probe response, authentication

(Step 2), association (Step 3), and 4-way handshake (Step

4). Nevertheless, the number of allowed retransmissions is

implementation speciﬁc. For example, with CYW43907, the

maximum number of retransmissions per packet of Step 2,

Step 3, and Step 4 are 3, 7, and 7, respectively. Therefore,

intensifying channel utilization causes lower success rate.

As Figure 8 shows, for channel utilization rates higher

than 50%, the success rate of WPA3 is higher than WPA2 for

the CYW station, while the success rate of WPA2 is higher

than WPA3 for the BCM station. By analyzing hostapd,

we identiﬁed a timer is started for the station once a probe

request is received. We also observed that this timer is only

used if the station employs the SP method. In other words,

by receiving a speciﬁc probe packet from a station, the AP is

informed that the station is obliged to associate with the AP.

If the authentication packet is not received from the station

before the timer expiry, the AP sends a disassociation packet

to the station and fails the association process. As explained

earlier, the CYW and BCM stations demonstrate shorter

key computation delays for WPA3 and WPA2, respectively;

thereby, a shorter duration reduces the chance of failure.

On the other hand, during high channel utilization, the

packet delivery delays caused by retransmissions further

exacerbate the chance of timer expiry.

IEEE TRANSACTIONS ON GREEN COMMUNICATIONS AND NETWORKING, 2021 7

1 5 10 20

Listen Interval Coefﬁcient (τ)

(a)

100

200

300

400

Duration [ms]

CYW |Normal

1 5 10 20

Listen Interval Coefﬁcient (τ)

(b)

100

200

300

400

Duration [ms]

CYW |DL

1 5 10 20

Listen Interval Coefﬁcient (τ)

(c)

100

200

300

400

Duration [ms]

CYW |Interference

1 5 10 20

Listen Interval Coefﬁcient (τ)

(d)

Duration [ms]

BCM |Normal

1 5 10 20

Listen Interval Coefﬁcient (τ)

(e)

Duration [ms]

BCM |DL

1 5 10 20

Listen Interval Coefﬁcient (τ)

(f)

Duration [ms]

BCM |Interference

1 5 10 20

Listen Interval Coefﬁcient (τ)

(g)

0.0

2.5

5.0

7.5

10.0

12.5

15.0

Duration [ms]

CYW43364 |Normal

1 5 10 20

Listen Interval Coefﬁcient (τ)

(h)

100

120

Duration [ms]

CYW43455 |Normal

Fig. 9: Per-beacon awake duration. (a)-(c): Awake time of CYW station

with normal (a), DL (b), and interference (c) trafﬁc. (d)-(f): Awake time

of BCM station with normal (d), DL (e), and interference (f) trafﬁc.

(g): Awake time of CYW43364 with normal trafﬁc. (h): Awake time of

CYW43455 with normal tarfﬁc. Cross marks represent the mean values.

Outliers are not demonstrated. These results show that the duration

of receiving each beacon is affected by increasing listen interval and

channel utilization.

5 BEACON RECEPTION OVERHEAD

Each AP transmits beacons every Target Beacon Transmis-

sion Time (TBTT), which is normally set to 102.4 ms. How-

ever, the default wake-up interval causes waste of energy

in two types of applications: (i) if the station-AP trafﬁc

pattern is primarily uplink, or (ii) if the delayed delivery of

downlink packets to the station can be tolerated. To reduce

the number of periodic wake-ups, we modiﬁed the driver to

specify the wake-up duration. Speciﬁcally, the listen interval

coefﬁcient (denoted as τ) speciﬁes wake-up duration as a

multiple of Beacon Interval (BI). For example, when τ= 10,

the station wakes up every 10 ×102.4ms.

We study beacon reception overhead in three scenarios:

(i) normal: where the overall channel utilization does not

exceed 10%; (ii) DL: where AP1 transmits voice trafﬁc to

Smartphone1 and consumes 95% of channel capacity; (iii)

interference: where AP2 and Smartphone2 exchange bidirec-

tional voice trafﬁc and consume 95% of channel capacity.

Figure 9 presents the results. For all the stations, in-

creasing the listen interval results in higher awake duration.

However, the increases in the awake duration of CYW

and CYW43455 are considerably higher compared to BCM

and CYW43364. Also, comparing Figures 9(a)-(c) with (d)-

(f) conﬁrm the higher sensitivity of the CYW station to

DL/UL trafﬁc and interference. The underlying causes of

these observations are threefold: the link quality estimation

algorithm, beacon transmission delay, and variations of the

wake-up timer of the station. We detail these causes as

1 5 10 20

Listen Interval Coefﬁcient (τ)

2000

4000

6000

8000

10000

Duration [ms]

CYW, AR9462

BCM, AR9462

CYW, QCA9565

BCM, QCA9565

Fig. 10: Inter-beacon awake duration

for CYW and BCM stations in the

presence of 95% channel utilization.

AR9462 and QCA9565 refer to the

wireless card used in AP1.

Voice TXOP (1.504 ms)

Time to Send

Beacon Packet

Delayed Beacon

Transmission

Video TXOP (3.008 ms)

Time to Send

Beacon Packet

Delayed Beacon

Transmission

(a)

(b)

Fig. 11: The presence of a ﬂow

belonging to the voice AC can

result in delayed beacon trans-

mission.

follows.

5.0.1 Link Quality Estimation

Beacon loss happens due to multiple factors such as packet

collision, low signal strength, and beacon delay. Manage-

ment frames, such as beacons, are more prone to collisions

because they are not preceded by RTS/CTS. To maintain

its association with the AP, each station needs to receive

a certain number of beacons per second, depending on

the link quality estimation algorithm implemented in the

ﬁrmware. When the beacon interval is one, even if the sta-

tion does not receive a beacon at some beacon instances, the

station can simply switch back to sleep mode and look for

beacon during the next beacon instance. By increasing the

listen interval, the tolerance against beacon loss is reduced

due to the need for time synchronization and link quality

estimation. In this case, the station stays awake for a longer

duration to look for incoming packets. To demonstrate this

behavior, we measured the awake duration between bea-

con reception instances in the presence of 95% interference

trafﬁc. We measured these intervals by accessing driver’s

counters that are updated per beacon reception, and then

correlating the counter values with power traces to measure

awake duration. In these results, to ensure the variations of

awake duration are not caused by the Time Synchronization

Function (TSF) of AP’s wireless transceiver, we ran the ex-

periments using two different wireless transceivers: AR9462

and QCA9565. Figure 10 shows the results. Regardless of the

AP’s transceiver, we observe that the CYW station is more

sensitive to beacon loss, and the sensitivity increases versus

listen interval value. Also, the awake duration between

two beacon receptions may be as short as one BI or span

multiple BIs. For example, by tracking power utilization, we

observed that there exists cases where the station stays in

awake mode after the reception of a beacon until receiving

the next beacon, and this results in 102.4 ms awake duration

between beacons. Although a higher sensitivity to beacon loss is

desired to facilitate handoff and avoid disconnection from the AP,

these results show its adverse effects on beacon reception overhead.

5.0.2 Beacon Transmission Delay

The presence of concurrent trafﬁc, especially voice and

video, causes variations of inter-beacon transmission inter-

vals and increases awake duration. Cisco estimates 79% of

all mobile trafﬁc will be video by 2022 [24]. As per the

802.11e amendment, the voice and video ACs can reserve

Transmit Opportunity (TxOP) of size 1.504 ms and 3.008

IEEE TRANSACTIONS ON GREEN COMMUNICATIONS AND NETWORKING, 2021 8

Inter-Beacon Interval [s]

Channel Utilization [%]

Inter-Beacon Interval [s] Inter-Beacon Interval [s]

Voice | UL Video | UL

(a) (b) (c)

Voice | DL Video | DL

Inter-Beacon Interval [s]

(d)

Fig. 12: Impact of DL and UL trafﬁc on beacon transmission delay. (a)

and (b): voice AC. (c) and (d): video AC. UL trafﬁc has a higher effect

on beacon transmission delay because it prevents the AP from channel

access during TxOP.

ms, respectively. During a TxOP, no other station in the

network commences a transmission. This is demonstrated

in Figure 11. Hence, although at each TBTT the AP is ready

for the transmission of a beacon, it would not be allowed

to send it if a station has reserved a TxOP. To demonstrate

this behavior in a real-world scenario, we measured channel

utilization and correlated it with beacon transmission delay.

Figure 12 displays the impact of UL and DL trafﬁc on beacon

delay. These results conﬁrm the increase of inter-beacon

intervals as channel utilization is intensiﬁed. Consider three

beacon transmission time instances, tk,tk+1, and tk+2 . As-

sume tkand tk+2 are sent at their designated time. If tk+1 is

delayed, then tk+1 −tk>102.4ms and tk+2 −tk+1 <102.4

ms. If tk+2 is delayed, then tk+2 −tk+1 can be shorter,

longer, or equal to 102.4 ms. The results also demonstrate

a lower beacon transmission variation with DL trafﬁc. This

is because beacons are prioritized over all data and man-

agement frames in the AP’s wireless driver. For example,

the software queues inside Atheros drivers (e.g., ath9k) are

organized as follows: a queue for beacon transmission, a

queue for management, multicast, and broadcast packets,

and eight queues for the 802.1p trafﬁc priorities (which are

then mapped to four ACs). The queue assigned to beacon

transmission has the highest priority; thereby, the beacons

are not affected signiﬁcantly by the DL trafﬁc sent by the

AP.

5.0.3 Wake-up Timer

The timestamp inside beacons contains the value of the TSF

timer when the ﬁrst bit of the timestamp is handed to the

physical layer (PHY) from the MAC-PHY interface and also

compensates the transmission delay from the PHY to the

antenna. The receiving station sets the value of its TSF timer

to the timestamp present inside the beacon. To investigate if

CYW can wake up on time to receive beacons, we measured

the awake time of this station before each beacon instance.

Our results show that CYW’s beacon loss rate increases

versus listen interval because it cannot properly adjust its

timer to wake up on time, even in scenarios where channel

utilization is around 10%. We also noticed that this duration

almost doubles for each step of increasing listen interval

coefﬁcient (we do not include the results due to space limita-

tion); thereby conﬁrming the impact of time synchronization

on awake duration.

The studies presented in this section clarify that in-

creasing listen interval cannot be simply used as a method to

improve energy efﬁciency because transceivers exhibit differ-

ent overheads considering listen interval duration, beacon

transmission delay, and interference.

As the value of τincreases, the chance of receiving

multicast and broadcast packets drops. This is because the

DTIM value on commercial APs is usually between 1 to

3, which means multicast and broadcast packets are trans-

mitted every 1 to 3 BIs. Various solutions can be used to

address this problem when τ > 1. For example, reliable

delivery of multicast and broadcast packets can be achieved

by converting these packets into unicast packets. Also, the

AP may prevent the transmission of unnecessary packets

such as ARP. For ARP packets, a proxy implemented on the

AP can be leveraged to reply to queries, without having to

involve stations.

6 MAINTAINING ASSOCIATION

Maintaining association is essential to ensure: (i) stations

spending most of their time in sleep mode can communicate

with the AP without having to reassociate, and (ii) the AP

can wake up and deliver downlink packets to the station.

In this section, we identify the underlying challenges of

maintaining association and present solutions to address

them.

Referring back to Figure 2, the testbed conﬁgurations

used for this experiment are as follows. AP Monitor is a

user-space daemon running on AP1 to monitor association

status, connected time, and idle time of the station. AP

Monitor continuously transmits its status to the Collector

via an Ethernet connection. The Sniffer is used to capture

the packets exchanged between AP1 and the station. The

sniffer reports to the Collector via an Ethernet connection.

The Collector, via a serial connection with the station, gath-

ers the number of received beacons by the station during

the current association period. The Pinger is used to ping

the station. AP1, Sniffer, and Pinger report their statistical

information to the Collector via Ethernet connections.

For the results in §6.1 and §6.2, in addition to the IoT

station under test, there are 12 other IoT stations (such as

cameras, thermostat, and Amazon Echos/Dots) associated

with the AP, as Figure 2 shows.

6.1 Probing or Keep Alive Packets?

Associated stations need to periodically communicate with

the AP to renew their inactivity timer maintained by the AP.

Referring to hostapd operation, if a station does not com-

municate with the AP for ap_max_inactivity duration,

the AP either disassociates the station immediately, or sends

a poll frame to the station and maintains the association if a

response is received. This conﬁguration is available via the

skip_inactivity_poll ﬂag.

In the ﬁrst experiment, we evaluate the effect of listen

interval on maintaining association. Since modifying the

duration of ap_max_inactivity on commercial APs is

not possible, we use its default value of 300 seconds in our

experiments. The skip_inactivity_poll is disabled.

The ﬁrst and second row of Figure 13 present the results

when using τ= 1 and τ= 20, respectively.

Figures 13(a)-(c) show that the AP can successfully poll

the station every 300 seconds until 5000 seconds. At this

point, as we can observe in Figure 13(b), the number of

beacon packets that include the AID of the station is sud-

denly increased to 9. A closer look revealed an interesting

IEEE TRANSACTIONS ON GREEN COMMUNICATIONS AND NETWORKING, 2021 9

(a) (b) (c)

(d) (e) (f)

⌧=1

<latexit sha1_base64="h6iX/NRjHbYDxefGBYs4qEOHag8=">AAAB73icbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0YtQ9OKxgv2ANpTNdtMu3Wzi7kQooX/CiwdFvPp3vPlv3LY5aOuDgcd7M8zMCxIpDLrut1NYWV1b3yhulra2d3b3yvsHTROnmvEGi2Ws2wE1XArFGyhQ8naiOY0CyVvB6Hbqt564NiJWDzhOuB/RgRKhYBSt1O4iTck18Xrlilt1ZyDLxMtJBXLUe+Wvbj9macQVMkmN6Xhugn5GNQom+aTUTQ1PKBvRAe9YqmjEjZ/N7p2QE6v0SRhrWwrJTP09kdHImHEU2M6I4tAselPxP6+TYnjlZ0IlKXLF5ovCVBKMyfR50heaM5RjSyjTwt5K2JBqytBGVLIheIsvL5PmWdU7r17cn1dqN3kcRTiCYzgFDy6hBndQhwYwkPAMr/DmPDovzrvzMW8tOPnMIfyB8/kDxtKPJg==</latexit>

⌧=1

⌧= 20

<latexit sha1_base64="vVWbL88uZMrYF67xxkGfSLRAdLk=">AAAB8HicbVDLSgNBEOyNrxhfUY9eBoPgKeyGiF6EoBePEcxDkiXMTmaTITO7y0yvEEK+wosHRbz6Od78GyfJHjSxoKGo6qa7K0ikMOi6305ubX1jcyu/XdjZ3ds/KB4eNU2casYbLJaxbgfUcCki3kCBkrcTzakKJG8Fo9uZ33ri2og4esBxwn1FB5EIBaNopccu0pRck4rbK5bcsjsHWSVeRkqQod4rfnX7MUsVj5BJakzHcxP0J1SjYJJPC93U8ISyER3wjqURVdz4k/nBU3JmlT4JY20rQjJXf09MqDJmrALbqSgOzbI3E//zOimGV/5EREmKPGKLRWEqCcZk9j3pC80ZyrEllGlhbyVsSDVlaDMq2BC85ZdXSbNS9qrli/tqqXaTxZGHEziFc/DgEmpwB3VoAAMFz/AKb452Xpx352PRmnOymWP4A+fzBzhsj2E=</latexit>

⌧= 20

Station unaware

of disassociation

Station

reassociates

with AP

The AP no longer keeps

track of the inactive time

of the station

Successful

polling

Unsuccessful

polling

Unsuccessful

polling

Fig. 13: AP-based polling and the impact of listen interval. The listen

interval (τ) in sub-ﬁgures (a)-(c) is 1. The listen interval (τ) in sub-

ﬁgures (d)-(f) is 20. (a) and (d): Number of beacons received by the

station (blue curves) as well as the communications between the AP

and station (red bars). (b) and (e): Number of AID-matched beacons

(indicating buffered packets for the station) per 30 seconds intervals.

behavior. After the ﬁrst beacon packet, the station needs to

send two PS-Poll packets to receive an ACK from the AP.

The AP then sends six Null packets to the station, but no

ACK is received; then, the collision avoidance mechanism

(RTS/CTS) is used to communicate with the station, and

20 RTS packets are sent. Meanwhile, hostapd’s probing

timer expires and the station is disassociated by the AP.

Fortunately, link reliability improves shortly and the dis-

association packet is conveyed to the station after the next

beacon instance; thereby, since the station is informed about

the disassociation event, it reassociates with the AP. This

scenario clearly shows the impact of link unreliability and

delayed station wake up on disassociation. Therefore, even

when the listen interval matches the beacon period, the

internal timer of hostapd may expire before the station is

woken up to reply to the probe message.

Figures 13(d)-(f) present the same experiment with τ=

20. Two polling events at 300 and 600 seconds are performed

successfully. The third polling, however, fails because the

AP does not hear back PS-Poll from the station. We ob-

served that AP sends 35 beacons that include the station’s

AID. Since no response is received, the AP infers that the

station is no longer in range; thereby, the AP does not

send a disassociation packet to the station. Therefore, as

Figure 13(d) shows the trend of receiving beacons from

the AP, from the station’s point of view, the station is still

connected to the AP. Adisassociation-unaware station does not

make any attempt to re-associate with the AP as long as it

receives beacon packets. For event-based applications that

trigger uplink trafﬁc (e.g., video streaming when motion

is detected, reporting temperature when a threshold is

passed), this behavior may require the station to re-associate

whenever inter-event interval is longer than 300 seconds.

For applications that require downlink communication with

stations, the disassociation from the AP prevents the user

or cloud platform from communicating with the station.

Based on these discussions, we conclude that maintaining

association by relying on the AP to send probes is unreliable

because the station may be left unaware of disassociation event.

Keep-Alive Packets. The above problem with AP-

Fig. 14: Maintaining association by using station-side keep-alive pack-

ets. Listen interval coefﬁcient (τ) is 20. (a) Number of beacons received

by the station (blue curve) as well as the communication instances

between the AP and station (red bars). (b) Inactivity time of the station

from the AP point of view. (c) Round-Trip Time (RTT) of pinging. (d)

Number of unicast packets exchanged between the station and AP per

30-second intervals. (e) Number of retransmissions between the station

and AP per 30-second intervals. (f) Number of AID-matched beacons

sent by the AP per 30-second intervals. The dotted lines in (c), (d), (e),

and (f) indicate ping transmission instances.

triggered probing can be eliminated by using station-

generated keep-alive packets. A keep-alive message can be

a UDP, TCP, Null or ARP packet. Normally, such packets are

generated by the protocol stack running on the application

subsystem. However, to reduce the burden of packet gen-

eration and reduce energy consumption, we ofﬂoad keep-

alive packet generation to the wireless subsystem; thereby

avoiding the need to wake up the application processor.

To validate the effectiveness of transceiver-generated keep-

alive packets, we utilized driver APIs to set up Null packet

generation. Also, to verify successful downlink packet de-

livery, the Pinger (see Figure 2) pings the station every 450

seconds. Figure 14 shows the results for τ= 20. In terms

of connectivity, Figures 14(a) and (b) conﬁrm the reliabil-

ity of maintaining association by the station and AP. The

inactivity timer kept by the AP for the station is renewed

every 150 seconds, which corresponds to a Null keep-alive

packet or the ping packet. Figure 14(c) conﬁrms that ping

packets are always delivered to the station, although some

RTT instances are as long as 8 seconds. Comparing Figures

14(c) and (f) reveals the relationship between the number of

AID-matched beacons and ping delay. Speciﬁcally, the main

cause of communication delay is the number of beacons sent

by the AP until a successful packet exchange is achieved.

The number of beacons sent by the AP to wake up the

station is, on average, 36. The impact of packet loss can be

observed in Figures 14(d) and (e). Figure 14(d) shows that

packet retransmissions are necessary for both pinging and

keep-alive delivery.

6.2 Key Renewal

In the previous sections, we showed the importance of pe-

riodic communication with AP to renew its inactivity timer.

In this section, we demonstrate that key renewal is also a

time-critical operation and may result in disassociation.

With both WPA2 and WPA3, each station is assigned

two keys: a PTK for unicast communication, and a GTK

for multicast and broadcast communication. Both keys are

periodically renewed based on the timer values deﬁned in

IEEE TRANSACTIONS ON GREEN COMMUNICATIONS AND NETWORKING, 2021 10

(a) (b) (c)

(d) (e) (f)

wpa_group_update_count=4

wpa_group_update_count=32 wpa_group_update_count=32 wpa_group_update_count=32

wpa_group_update_count=4 wpa_group_update_count=4

Unsuccessful

key renewal

Station

disassociated

from the AP

Fig. 15: Impact of key renewal on association durability. Listen interval

coefﬁcient (τ) is 20. In both sub-ﬁgures (i) and (ii): (a) Number of bea-

cons received by the station (blue lines) as well as the communications

between the AP and station (red bars). (b) Inactivity time of the station

from the AP point of view. (c) Connected time of the station from the

AP point of view.

hostapd’s conﬁguration. The GTK is also renewed when-

ever a station leaves the network. This renewal is necessary

to prevent the station from receiving multicast or broadcast

packets of the network that the station no longer belongs to.

This means mobility and disassociation of any station may

require the AP to communicate with all stations.

In the experiments of this section, we use Smartphone1

(see Figure 2) to trigger GTK renewal. This smartphone

has been programmed to leave the network every 450 sec-

onds and then reassociate after 3 seconds. The ﬁrst row of

Figure 15 shows the results with the default conﬁguration

of hostapd. As it can be observed, the GTK renewal is

performed successfully until around 4000 seconds. At this

time, due to link unreliability, the AP fails to renew the

key and disassociates the station. However, since the station

does not receive the disassociation packet, the result is a

disassociation-unaware station.

Customized hostapd conﬁguration. To address the

aforementioned problem, we modify hostapd’s conﬁg-

uration to increase the number of key renewal re-

tries. To leverage this feature, we increase the value of

wpa_group_update_count to 32, compared to its default

value of 4. The results in the second row of Figure 15

conﬁrm that with the new value, the AP can successfully

update the key, even in the presence of long listen interval

(τ= 20). The main takeaway is that key renewal may result

in the disassociation of IoT stations when the listen interval is

long or when the communication link is unreliable. And ﬁxing

this problem requires customizing the conﬁguration values of

hostapd.

7 WAKE UP DE LAY

Ensuring reliable and short station wake up delay prevents

cases such as polling failure (§6.1) and key renewal failure

(§6.2). Such assurance is also important for applications

where timely communication with stations is desired; for

example, when sending an actuation command to a station

or waking up a surveillance camera by the user to start

video streaming. In this section, we consider the PSM and

APSD power saving-methods and measure the wake up

(b)

(a)

UL | LI = 1 UL | LI = 5 UL | LI = 10 UL | LI = 20

DL | LI = 1 DL | LI = 5 DL | LI = 10 DL | LI = 20

50% Channel Utilization 95% Channel Utilization

10% Channel Utilization

50% Channel Utilization 95% Channel Utilization

10% Channel Utilization

[s]

[s] [s]

[s]



<latexit sha1_base64="q6eUG44rqqw6pCEbWiVdTPDj104=">AAAB63icbVBNS8NAEJ3Ur1q/qh69LBbBU0m00B4LXjxWsB/QhrLZbpqlu5uwuxFK6F/w4kERr/4hb/4bN20O2vpg4PHeDDPzgoQzbVz32yltbe/s7pX3KweHR8cn1dOzno5TRWiXxDxWgwBrypmkXcMMp4NEUSwCTvvB7C73+09UaRbLRzNPqC/wVLKQEWxyaZREbFytuXV3CbRJvILUoEBnXP0aTWKSCioN4Vjroecmxs+wMoxwuqiMUk0TTGZ4SoeWSiyo9rPlrQt0ZZUJCmNlSxq0VH9PZFhoPReB7RTYRHrdy8X/vGFqwpafMZmkhkqyWhSmHJkY5Y+jCVOUGD63BBPF7K2IRFhhYmw8FRuCt/7yJund1L1G/fahUWu3ijjKcAGXcA0eNKEN99CBLhCI4Ble4c0Rzovz7nysWktOMXMOf+B8/gASeo47</latexit>

<latexit sha1_base64="K1tnIo6r6AG537Nis8ZQilSWin0=">AAAB7nicbVBNSwMxEJ3Ur1q/qh69BIvgqexqwR4LXjxWsB/QLiWbZtvQbDYk2UJZ+iO8eFDEq7/Hm//GtN2Dtj4YeLw3w8y8UAlurOd9o8LW9s7uXnG/dHB4dHxSPj1rmyTVlLVoIhLdDYlhgkvWstwK1lWakTgUrBNO7hd+Z8q04Yl8sjPFgpiMJI84JdZJnf6UaDXmg3LFq3pL4E3i56QCOZqD8ld/mNA0ZtJSQYzp+Z6yQUa05VSweamfGqYInZAR6zkqScxMkC3PneMrpwxxlGhX0uKl+nsiI7Exszh0nTGxY7PuLcT/vF5qo3qQcalSyyRdLYpSgW2CF7/jIdeMWjFzhFDN3a2Yjokm1LqESi4Ef/3lTdK+qfq16u1jrdKo53EU4QIu4Rp8uIMGPEATWkBhAs/wCm9IoRf0jj5WrQWUz5zDH6DPH3o8j6I=</latexit>

= 307 ms

= 308 ms



= 563 ms

= 605 ms



= 1076 ms

= 1640 ms



= 307 ms

= 308 ms



= 563 ms

= 605 ms



= 1076 ms

= 1640 ms

Packets delivered

immediately after

TIM bit set

Packets delivered

immediately after

TIM bit set

Fig. 16: ECDF of wake up delay for PSM. The vertical red line represents

listen interval (τ) and the horizontal red line represents the 50th

percentile. The notation φdenotes the theoretical expected wake up

delay and the notation ϕis the 50th percentile of empirical wake-up

delay for 10% channel utilization.

UL | LI = 1 UL | LI = 5 UL | LI = 10 UL | LI = 20

DL | LI = 1 DL | LI = 5 DL | LI = 10 DL | LI = 20

(b)

(a)

50% Channel Utilization 95% Channel Utilization

10% Channel Utilization

50% Channel Utilization 95% Channel Utilization

10% Channel Utilization

[s]

[s] [s]

[s]



= 307 ms

= 308 ms



= 563 ms

= 820 ms



= 1076 ms

= 2907 ms



= 307 ms

= 308 ms



= 563 ms

= 820 ms



= 1076 ms

= 2907 ms

Packets delivered

immediately after

TIM bit set

Packets delivered

immediately after

TIM bit set

Fig. 17: ECDF of wake up delay for APSD. The vertical red line

represents listen interval (τ) and the horizontal red line represents the

50th percentile. The notation φdenotes the theoretical expected wake

up delay and the notation ϕis the 50th percentile of empirical wake-up

delay for 10% channel utilization.

delay of stations. Using PSM, when the station receives a

beacon including its AID, it sends a PS-Poll packet to the AP

to trigger downlink delivery; whereas, using APSD, a Null

packet is sent to the AP. The Pinger (§3) sends a ping packet

to the CYW station every tseconds, where tis uniformly

chosen from the range [1,15] seconds. We measure wake

up delay as the interval between the transmission of ﬁrst

beacon including the station’s AID until the transmission of

a PS-Poll or Null packet by the station. Figures 16 and 17

present the results for PSM and APSD, respectively.

The time instances the AP receives the ping packet from

the Pinger follow a continuous uniform distribution. For a

station with listen interval τ, we denote the beacon instances

during the listen interval as [tk, ..., tk+τ]. When a ping

packet arrives at the AP during interval [tk, tk+1), the AP

needs to send τbeacon packets before the next wake-up

instance of the station. Similarly, if the packet arrives at the

AP during interval [tk+1, tk+2 ), the AP needs to send τ−1

beacon packets. Therefore, the expected number of beacons

sent until station wake-up is 1

τ(τ+ (τ−1) + ... + 2 + 1) =

IEEE TRANSACTIONS ON GREEN COMMUNICATIONS AND NETWORKING, 2021 11

(τ+ 1)/2, when τ > 1. The expected wake up delay (φ),

demonstrated by the dashed vertical lines in Figures 16 and

17, is computed as 102.4×(τ+ 1)/2ms, for τ > 1. The

maximum wake up delay is τ×102.4ms, demonstrated by

the solid vertical lines in Figures 16 and 17. As the results

show, the observed values of median wake up delay (ϕ)

increase beyond the theoretical expected values (φ) in the

presence of background trafﬁc and also when τ > 5. For ex-

ample, consider using APSD and UL trafﬁc. For τ= 10, the

expected (φ) and maximum delays are 563 ms and 1024 ms,

respectively; however, the empirical results show median

values (ϕ) of 820 ms and 2162 ms for 10% and 95% trafﬁc,

respectively. We observed a similar behavior in Figure 14,

where, although the expected and maximum number of

beacons are 10.5 and 20, the average number of beacons

sent was 36. These prolonged wake up delays are caused

by missing and delayed transmission of beacon, PS-Poll,

and Null packets. Assuming the probability of successful

beacon reception is p, the expected value of the number of

beacons required to achieve success is τ×1/p when listen

interval τis enforced. Assuming uplink success probability

q, the expected value of the number of packet transmissions

(beacon and PS-Poll/Null) adds up to τ×1/p + 1/q.

For both power-saving methods, and compared to DL

trafﬁc, we observe UL trafﬁc has a considerably higher effect

on prolonging wake up delay. For example, for 95% channel

utilization, the 90th percentile wake up delay of PSM with

τ= 5 increases from 775 ms to 1437 ms when changing the

trafﬁc type from DL to UL. Aligned with our observations

in Figure 3, with DL trafﬁc, the packets sent by AP1 cause

collision with the PS-Poll or Null packet sent by the station.

Whereas, in the UL scenario, both the beacons sent by

AP1 and the PS-Poll or Null packet sent by the station are

susceptible to collide with Smartphone1’s trafﬁc.

In all the results, tail is increased as channel utilization

intensiﬁes. Also, we observe the higher effect of channel

utilization on prolonging APSD wake up delay, compared

to PSM. For example, for τ= 10 and 10% DL channel

utilization, the 50th percentile wake up delay increases by

27% when switching from PSM to APSD; and this increase

is almost 2x for 95% DL channel utilization. Since the

priority of beacon packets is not affected by the power-

saving method used, we looked into the type and priority

of the packet sent by the station in response to an AID-

matched beacon. We noted that PS-Poll is categorized as a

management packet, whereas, Null packets are data pack-

ets. Speciﬁcally, PS-Poll packets are always sent using the

base rate, and Null packets are transmitted at the normal

data rates. For example, in our results, the rate of PS-Poll

packet was 6 Mbps, while the rate of Null packet was at

least 24 Mbps. Also, management packets are prioritized

over all data packets in the driver’s software-queues, and

these packets are transmitted alongside voice AC packets in

the driver’s hardware queues. Although Null data packets

are prioritized over packets belonging to video, best-effort,

and background ACs, these packets are treated as regular

voice AC packets, and thereby, contend with other voice

packets.

The key takeaways are: Wake-up delays are lower with

PSM compared to APSD, in the presence of concurrent trafﬁc.

However, PSM requires sending a PS-Poll packet for retrieving

each packet from the AP, whereas, APSD utilizes only a single

Null frame to retrieve all the queued packets. Hence, the overall

energy consumption when using PSM can still be higher when

more than one packet are queued for downlink delivery at the AP

during each listen interval, on average. Also, the effect of DL

trafﬁc on wake-up delay is lower than that of UL trafﬁc.

8 RE LATED WORK

The suitability of using WiFi for IoT and industrial appli-

cations has been studied in recent works. Luvisotto et al.

[6] discuss the salient features of WiFi’s physical layer to

address the requirements of industrial communication sys-

tems, compared to cellular and 802.15 networks. Similarly,

Tramarin et al. [7] discussed the suitability of 802.11n in

industrial applications. Abedi et al. [3] report the physical

layer energy consumption range of WiFi and BLE are 10-100

nJ/bit and 275-300 nJ/bit, respectively; thereby conﬁrming

the higher efﬁciency of WiFi.

Seneviratne et al. [12] present the loss of DHCP packets

as the main cause of association failure and delay. Pei et

al. [11] conducted a large-scale empirical measurement and

revealed that up to 45% of the users experience association

failure, and 15% of successful associations are longer than

ﬁve seconds. They also demonstrate that the scan process

comprises about 47% of the association duration. Tozlu

et al. [25] evaluate the impact of various conﬁgurations

on the energy efﬁciency of WiFi stations, and show that

WPA2/AES-PSK achieves the highest security-performance

tradeoff. Montori et al. [26] studied the association time

of three authentication schemes, namely, Open, WEP, and

WPA2, and measured their effect on the discharge rate of

three battery types. Abedi et al. [3] show that the overall

energy efﬁciency of WiFi is lower than BLE due to the cost of

association and maintaining association. They propose Wi-

LE, which eliminates these costs by placing data in beacon

packets. Chen and Qiao [27] propose a handoff mechanism

to eliminate the need to dwell on each scanned channel

to receive the response messages. Once a station sends a

probe message, it switches back to its current channel and

delegates the task of collecting information from nearby

APs to the currently-associated AP. None of the existing

works present a detailed study of the various constituents

of association process considering different hardware and

software components and variable trafﬁc loads. Also, WPA3

has been neglected in existing studies and none of them

utilize PMK ofﬂoading or caching to reduce association

costs.

Vasudevan et al. [28] propose a passive mechanism to

measure the potential bandwidth of APs. Assuming bea-

con packet transmissions are not prioritized, stations can

compute the buffering delay of nearby APs by monitoring

their beaconing delay. Their proposed solution assumes the

stations are always listening to the beacon packets and

leverage beacon transmission delays to associate with the

lowest-delay AP nearby. Molina et al. [29] evaluated beacon

transmission jitter in high and normal channel utilization

scenarios. Their results show the inter-quartile range (IQR)

of beacon jitter increases monotonically versus channel

load. Stations can detect channel saturation by using the

Kolgomorov-Smirnov (KS) test and comparing beacon jitters

IEEE TRANSACTIONS ON GREEN COMMUNICATIONS AND NETWORKING, 2021 12

with a reference distribution representing the non-saturated

scenario. In sum, the existing works did not study the

adverse effects of channel load and beacon delay on various

station types that leverage non-standard listen intervals

(τ > 1) to enhance their energy efﬁciency.

Sheth and Dezfouli [1] proposed a packet scheduling

mechanism to reduce the idle time of stations implementing

Adaptive PSM. Peck et al. [30] state that either high delay

or high energy is acceptable from a user’s point of view.

Based on the variations of station-server communication

delay, the sleep duration is dynamically adjusted to satisfy

the desired trade-off. Jang et al. [31] proposed an adaptive

tail time adjustment mechanism by measuring inter-packet

intervals. Once a stream of packets arrives at a station, inter-

packet arrival delay is predicted using a moving average.

Primarily designed for VoIP trafﬁc, Liu et al. [32] propose

a mechanism to reduce contention among stations using

APSD to request packet delivery from the AP. Chen et

al. [33] proposed M-PSM for scenarios where the load of

stations is light and delay-insensitive. The ultimate goal

of M-PSM is to prioritize communication during intervals

that the link quality to the AP allows utilizing higher

communication rates. A per-station DTIM allocation mech-

anism has been proposed in [10]. Instead of maintaining

one DTIM for all the stations, the AP enables the stations

to request their own desired DTIM period. Khorov et al.

[34] present an overview of periodic reservation methods

available in 802.11s, 802.11ad, 802.11ax, and 802.11be. They

propose solutions for addressing the challenges of estab-

lishing communication reservation periods for exchanging

real-time multimedia trafﬁc between stations in the pres-

ence of interference. The aforementioned works introduce

enhancements to the power saving methods of 802.11 to

tailor them for various application scenarios. In contrast, in

this work, we followed an application-agnostic, extendable

analysis of the fundamental operation of power saving

methods considering various hardware platforms and trafﬁc

scenarios.

9 CONCLUSION

Association, periodic beacon reception, maintaining asso-

ciation, and station wake-up are fundamental operations

to ensure connectivity in WiFi networks. In this paper, we

performed an empirical evaluation of these operations using

various hardware and software platforms and identiﬁed

the underlying causes of energy inefﬁciency, delay, and

unreliability. In addition to identifying the challenges of

building WiFi-based, low-power IoT systems, this research

also highlighted the importance of ﬁrmware customization,

AP conﬁguration, and transceiver choice, based on the ap-

plication at hand and environmental parameters.

REFERENCES

[1] J. Sheth and B. Dezfouli, “Enhancing the energy-efﬁciency and

timeliness of iot communication in wiﬁ networks,” IEEE Internet

of Things Journal, vol. 6, no. 5, pp. 9085–9097, 2019.

[2] H. Pirayesh, P. K. Sangdeh, and H. Zeng, “EE-IoT: An Energy-

Efﬁcient IoT Communication Scheme for WLANs,” in IEEE INFO-

COM, 2019, pp. 361–369.

[3] A. Abedi, O. Abari, and T. Brecht, “Wi-LE: Can WiFi Replace

Bluetooth?” in 18th ACM HotNets, 2019, pp. 117–124.

[4] J. Huang, F. Qian, A. Gerber, Z. M. Mao, S. Sen, and O. Spatscheck,

“A close examination of performance and power characteristics of

4G LTE networks,” in 10th ACM MobiSys. ACM, 2012, pp. 225–

238.

[5] C.-C. Li, V. K. Ramanna, D. Webber, C. Hunter, T. Hack, and

B. Dezfouli, “Sensiﬁ: A wireless sensing system for ultra-high-rate

applications,” arXiv preprint arXiv:2012.14635, 2020.

[6] M. Luvisotto, Z. Pang, and D. Dzung, “Ultra high performance

wireless control for critical applications: Challenges and direc-

tions,” IEEE Transactions on Industrial Informatics, vol. 13, no. 3,

pp. 1448–1459, 2017.

[7] F. Tramarin, S. Vitturi, M. Luvisotto, and A. Zanella, “On the use

of IEEE 802.11n for industrial communications,” IEEE Transactions

on Industrial Informatics, vol. 12, no. 5, pp. 1877–1886, 2015.

[8] C. Gray, R. Ayre, K. Hinton, and L. Campbell, “‘smart’ is not free:

Energy consumption of consumer home automation systems,”

IEEE Transactions on Consumer Electronics, vol. 66, no. 1, pp. 87–

95, 2019.

[9] A. Vinhas, V. Bernardo, M. Pascoal Curado, and T. Braun, “Per-

formance analysis and comparison between legacy-PSM and U-

APSD,” 2013.

[10] J. Wang and T. Nagy, “Maintaining delivery trafﬁc indication

message (DTIM) periods on a per-wireless client device basis,”

Patent, Aug. 23, 2011, US Patent 8,005,032.

[11] C. Pei, Z. Wang, Y. Zhao, Z. Wang, Y. Meng, D. Pei, Y. Peng,

W. Tang, and X. Qu, “Why it takes so long to connect to a WiFi

access point,” in IEEE INFOCOM, 2017, pp. 1–9.

[12] S. Seneviratne, A. Seneviratne, P. Mohapatra, and P.-U. Tournoux,

“Characterizing WiFi connection and its impact on mobile users:

practical insights,” in 8th ACM WinTech. ACM, 2013, pp. 81–88.

[13] Cypress Semiconductor. CYW43907: IEEE 802.11 a/b/g/n SoC

with an Embedded Applications Processor. [Online]. Available:

http://www.cypress.com/ﬁle/298236/download

[14] BCM4343W: 802.11b/g/n WLAN, Bluetooth

and BLE SoC Module. [Online]. Avail-

able: https://products.avnet.com/opasdata/d120001/medias/

docus/138/AES-BCM4343W-M1- G data sheet v2 3.pdf

[15] Amazon Inc. (2020) FreeRTOS: Real-time operating system for

microcontrollers. [Online]. Available: https://www.freertos.org

[16] Amazon Inc. (2020) FreeRTOS AWS Reference Integrations.

[Online]. Available: https://github.com/aws/amazon-freertos

[17] Microsoft Inc. (2020) Azure RTOS ThreadX documentation.

[Online]. Available: https://docs.microsoft.com/en-us/azure/

rtos/threadx/

[18] Microsoft Inc. (2020) Azure RTOS ThreadX. [Online]. Available:

https://github.com/azure-rtos/threadx

[19] Free Software Foundation Inc. (2020) LwIP: A Lightweight

TCP/IP stack. [Online]. Available: https://savannah.nongnu.

org/projects/lwip/

[20] Microsoft Inc. (2020) Azure RTOS NetX Duo documentation.

[Online]. Available: https://docs.microsoft.com/en-us/azure/

rtos/netx-duo/

[21] B. Dezfouli, I. Amirtharaj, and C.-C. Li, “EMPIOT: An energy

measurement platform for wireless IoT devices,” Journal of Network

and Computer Applications, vol. 121, pp. 135–148, 2018.

[22] CYW43364 Single-Chip IEEE 802.11 b/g/n

MAC/Baseband/Radio. [Online]. Available: https://www.

cypress.com/ﬁle/298001/download

[23] CYW43455 Single-Chip 5G WiFi IEEE 802.11n/ac

MAC/Baseband/ Radio with Integrated Bluetooth 5.0. [Online].

Available: https://www.cypress.com/ﬁle/358916/download

[24] Cisco, “Cisco visual networking index: global mobile

data trafﬁc forecast update, 2017–2022,” 2019. [Online].

Available: https://s3.amazonaws.com/media.mediapost.com/

uploads/CiscoForecast.pdf

[25] S. Tozlu, M. Senel, W. Mao, and A. Keshavarzian, “Wi-Fi enabled

sensors for internet of things: A practical approach,” IEEE Com-

munications Magazine, vol. 50, no. 6, pp. 134–143, 2012.

[26] F. Montori, R. Contigiani, and L. Bedogni, “Is WiFi suitable for

energy efﬁcient IoT deployments? A performance study,” in IEEE

3rd International Forum on Research and Technologies for Society and

Industry (RTSI). IEEE, 2017, pp. 1–5.

[27] X. Chen and D. Qiao, “Hand: Fast handoff with null dwell time

for ieee 802.11 networks,” in IEEE INFOCOM, 2010, pp. 1–9.

[28] S. Vasudevan, K. Papagiannaki, C. Diot, J. Kurose, and D. Towsley,

“Facilitating access point selection in IEEE 802.11 wireless net-

IEEE TRANSACTIONS ON GREEN COMMUNICATIONS AND NETWORKING, 2021 13

works,” in 5th ACM SIGCOMM IMC. Usenix Association, 2005,

pp. 26–26.

[29] L. Molina, A. Blanc, N. Montavont, and L. Simi´

c, “Identifying

channel saturation in WiFi networks via passive monitoring of

IEEE 802.11 beacon jitter,” in 15th ACM MobiWac. ACM, 2017,

pp. 63–70.

[30] B. Peck and D. Qiao, “A practical PSM scheme for varying server

delay,” IEEE Transactions on Vehicular Technology, vol. 64, no. 1, pp.

303–314, 2015.

[31] S. Y. Jang, B. Shin, and D. Lee, “An adaptive tail time adjust-

ment scheme based on inter-packet arrival time for IEEE 802.11

WLAN,” in IEEE International Conference on Communications (ICC).

IEEE, 2016, pp. 1–6.

[32] L. Liu, X. Cao, Y. Cheng, and Z. Niu, “Energy-efﬁcient sleep

scheduling for delay-constrained applications over WLANs,”

IEEE Transactions on Vehicular Technology, vol. 63, no. 5, pp. 2048–

2058, 2014.

[33] X. Chen, S. Jin, and D. Qiao, “M-PSM: Mobility-aware power save

mode for IEEE 802.11 WLANs,” in 31st ICDCS. IEEE, 2011, pp.

77–86.

[34] E. Khorov, A. Lyakhov, A. Ivanov, and I. F. Akyildiz, “Modeling of

Real-Time Multimedia Streaming in Wi-Fi Networks With Periodic

Reservations,” IEEE Access, vol. 8, pp. 55 633–55 653, 2020.

FLIP: A Framework for Leveraging eBPF to Augment WiFi Access Points and Investigate Network Performance

Conference Paper

Full-text available

Nov 2021

Wireless Security Protocols WPA3: A Systematic Literature Review

Article

Full-text available

Jan 2023

The size of wireless networks and the number of wireless devices are growing daily. A crucial part of wireless security involves preventing unauthorized access by using wireless security protocols in order to protect the data in wireless networks. In 2018, Wi-Fi Protected Access 3 (WPA3) was ratified to protect the data in devices bearing the Wi-Fi trademark. WPA3 has many security improvements over previous wireless security protocols, by providing a better encryption method and key sharing. In this paper, a Systematic Literature Review (SLR) was conducted to analyze three aspects of WPA3 protocol: the reasons behind the release of WPA3, the encryption methods and mode of operation in this protocol, and the attacks that remain penetrating WPA3. In this review, thirty-six articles were identified as the selected research articles, written between 2018 and 2023, focusing mainly on WPA3. After the analysis of the selected articles, the encryption methods and modes of operation were presented in the SLR. In addition, the vulnerabilities that the WPA3 protocol solved and the ones that remain unsolved were discussed. This study concluded that WPA3 excels over its predecessors by providing more security and reliability to wireless networks. The result of this SLR of WPA3 proposes two methods that seek to increase the security level of WPA3 networks, which has been discussed in the discussion section.

A Framework for Managing Device Association and Offloading the Transport Layer’s Security Overhead of WiFi Device to Access Points

Article

Full-text available

Sep 2021
SENSORS-BASEL

Considering the resource constraints of Internet of Things (IoT) stations, establishing secure communication between stations and remote servers imposes a significant overhead on these stations in terms of energy cost and processing load. This overhead, in particular, is considerable in networks providing high communication rates and frequent data exchange, such as those relying on the IEEE 802.11 (WiFi) standard. This paper proposes a framework for offloading the processing overhead of secure communication protocols to WiFi access points (APs) in deployments where multiple APs exist. Within this framework, the main problem is finding the AP with sufficient computation and communication capacities to ensure secure and efficient transmissions for the stations associated with that AP. Based on the data-driven profiles obtained from empirical measurements, the proposed framework offloads most heavy security computations from the stations to the APs. We model the association problem as an optimization process with a multi-objective function. The goal is to achieve maximum network throughput via the minimum number of APs while satisfying the security requirements and the APs’ computation and communication capacities. The optimization problem is solved using genetic algorithms (GAs) with constraints extracted from a physical testbed. Experimental results demonstrate the practicality and feasibility of our comprehensive framework in terms of task and energy efficiency as well as security.

Shallow Sleep Control Management for Low-Power Rockets

Chapter

Full-text available

Oct 2023

In this paper, a control method for shallow sleep management of low-power rockets is studied, and three sleep modes are verified for shallow sleep, and the optimal adaptation of shallow sleep of rockets with minimum power consumption of 16 watts is confirmed In different application environments, multi-mode operation with low power consumption is realized, so as to achieve the goal of long life under the requirements of long-term power up. The low-power mode can make the system run more stably, help reduce system losses, effectively reduce the volume and weight of power supply equipment, and improve the carrying capacity of the rocket. In order to achieve the purpose of low power consumption, based on the principle of “high and low matching”, the performance matching of chips of different performance levels at their optimal working points is analyzed, and the low-power chip architecture design under different computing power requirements is realized, and experimental verification is carried out.

The Optimization of Solar Photovoltaic System for Rural Off-grid Villages

Article

Full-text available

Jan 2022
P IEEE

This work studies improvements that can be made to the efficiency of solar PV systems such that they can be well utilized especially in non-grid connected rural villages to relieve the combustion engine thereby saving diesel as well as providing much needed rest time for the engine parts. Four efficiency factors were worked upon which are solar tracking, the digitization of the solar panel by providing them more intelligence, an improvement of the battery maintenance system and the digitization of the battery bank. To fully accomplish the digitization of the solar panel and battery bank AI is required and this is still being worked upon. An MPPT or Maximum Power Point Tracking system can still be utilized to provide a maximum power transfer to batteries. Solar PV output especially in equatorial Malaysia has a dismal historical record mainly due to the heavy cloud cover which forms intermittent shadows on the panels. This intermittency and high heat shortens lifespans of the battery banks and reduces the efficiency of the panels respectively.

Optimal Transmission Scheduling in Data-Intensive Audio Sensor Networks

Conference Paper

Dec 2023

Simulation of Fault Diagnosis Model for Managing Aeronautical Multivariate Heterogeneous Inputs

Conference Paper

Mar 2023

Leveraging Frame Aggregation in Wi-Fi IoT Networks for Low-Rate DDoS Attack Detection

Chapter

Dec 2022

The proliferation of smart home Internet of Things (IoT) devices is demonstrated by their prominence in people’s lives. However, the resource-constraint essence of these devices introduces various security flaws. One significant attack is the Low-rate Distributed Denial of Service (LR-DDoS) attack, which aims to disrupt the functionalities of the smart home IoT devices in a stealthy way by sending limited malicious traffic to the victim device. This paper proposes a novel set of features based on the 802.11 frame aggregation scheme to detect LR-DDoS attacks. We demonstrate that by conveying the characteristics of subframes during frame aggregation, we can uniquely embody the IoT device’s benign traffic and malicious traffic in smart home networks. Compared to existing works which primarily focus on LR-DDoS attacks launched against data centers, to the best of our knowledge, this paper is the first work focusing on detecting such attacks against smart home IoT devices. We validate the effectiveness of the proposed features using the commercial off-the-shelf smart home IoT devices and by adopting various machine learning algorithms. Empirical results show that adopting the proposed features with the Random Forest achieves a 0.98 accuracy in distinguishing between benign and LR-DDoS attack traffic. KeywordsInternet of Things (IoT) securityDDoSLow-rate attacksSmart homeMachine learning

A 40-nm Low-Power WiFi SoC With Clock Gating and Power Management Strategy

Article

Aug 2022

With the emerging of Internet of Things (IoT) industry, applications like smart power plugs, security ID tags, home automation and wearable electronic devices all make the demand for low-power WiFi chips impendency. In this paper, a low-power 2.4 GHz 802.11b/g/n WiFi system-on-chip (SoC) is designed and implemented with 40-nm CMOS process, with area of 8.1 mm². The low-power SoC integrates 32-bit microcontroller, 802.11b/g/n WiFi baseband, 2.4 GHz RF transceiver, ample memory space, ADC, 6-channel PWM, flexible I/O interfaces, multi-stage power management module, etc.. It has several sleep modes with extremely low leakage current as 0.8 mA/12 µA in light/deep sleep mode and 0.4 µA in Shutdown mode to reduce the power consumption. High performance is demonstrated, including Pout (-28 dB/-30 dB EVM) of 20.1 dBm/19.1 dBm and RX sensitivity of -76 dBm/-74 dBm meanwhile the total current of 148.5 mA/146.5 mA (TX) for 54 Mbps OFDM/HT20 MCS7.

Sensifi: A Wireless Sensing System for Ultra-High-Rate Applications

Article

Full-text available

Jun 2021

Wireless Sensor Networks (WSNs) are being used in various applications such as structural health monitoring and industrial control. Since energy efficiency is one of the major design factors, the existing WSNs primarily rely on low-power, low-rate wireless technologies such as 802.15.4 and Bluetooth. In this paper, by proposing , we strive to tackle the challenges of developing ultra-high-rate WSNs based on the 802.11 (WiFi) standard. As an illustrative structural health monitoring application, we consider spacecraft vibration test and identify system design requirements and challenges. Our main contributions are as follows. First, we propose packet encoding methods to reduce the overhead of assigning accurate timestamps to samples. Second, we propose energy efficiency methods to enhance the system’s lifetime. Third, to enhance sampling rate and mitigate sampling rate instability, we reduce the overhead of processing outgoing packets through the network stack. Fourth, we study and reduce the delay of processing time synchronization packets through the network stack. Fifth, we propose a low-power node design particularly targeting vibration monitoring. Sixth, we use our node design to empirically evaluate energy efficiency, sampling rate, and data rate. We leave large-scale evaluations as future work.

Modeling of Real-Time Multimedia Streaming in Wi-Fi Networks With Periodic Reservations

Article

Full-text available

Mar 2020

An important problem of modern Wi-Fis is the interferences caused by hidden stations active in the same area, or in multihop communications. All these issues significantly degrade the efficiency of the random channel access methods. Recent standardization and research activities are focused on solving coordination problems between various Wi-Fi devices. For example, the ongoing development of Wi-Fi 7 includes a coordinated schedule between the access points as a candidate solution. Consequently, Wi-Fi has many deterministic channel access mechanisms, which schedule channel time in a periodic manner well in advance and, thus, are utilized for streaming QoS sensitive data. However, both random traffic intensity and error-prone nature of the wireless channel complicate choosing such reservation parameters, i.e., the duration and the period of the reserved time intervals, that satisfy QoS requirements while minimizing channel time consumption. This paper introduces a general mathematical framework to solve the problem of choosing appropriate reservations parameters. The comparison of the analytical and simulation results show the high accuracy of the proposed framework. Finally, the paper gives an example of how to use the developed framework to maximize the network capacity.

Enhancing the Energy-Efficiency and Timeliness of IoT Communication in WiFi Networks

Article

Full-text available

Aug 2019

Increasing the number of IoT stations or regularstations escalates downlink channel access contention and queu-ing delay, which in turn result in higher energy consumption andlonger communication delays with IoT stations. To remedy thisproblem, this paper presentsWiotap, an enhanced WiFi accesspoint that implements a downlink packet scheduling mechanism.In addition to assigning higher priority to IoT traffic comparedto regular traffic, the scheduling algorithm computes per-packetpriorities to arbitrate the contention between the transmissionof IoT packets. This algorithm employs a least-laxity first (LLF)scheme that assigns priorities based on the remaining wake-uptime of the destination stations. We used simulation to showthe scalability of the proposed system. Our results show thatWiotap achieves 37% improvement regarding the duty cycle ofIoT stations compared to a regular access point. In addition, wedeveloped a testbed to confirm the implementation correctnessand the performance benefits of Wiotap in a network with fourIoT stations and regular traffic. For the edge and cloud scenarios,our empirical evaluations show up to 44% and 38% improvementin energy and 52% and 41% improvement in delay, respectively.

EMPIOT: An energy measurement platform for wireless IoT devices

Article

Full-text available

Apr 2018

Profiling and minimizing the energy consumption of IoT devices is an essential step towards employing IoT in various application domains. In this paper we propose EMPIOT, an accurate, low-cost, easy to build, and flexible, power measurement platform. We present the hardware and software components of this platform, and study the effect of various design parameters on accuracy. In particular, we analyze the effect of driver, bus speed, input voltage, and buffering mechanism, on sampling rate, measurement accuracy and processing demand. These extensive experimental studies enable us to configure the system in order to achieve its highest performance. We also propose a novel calibration technique and report the calibration parameters under various settings. Using five different IoT devices performing four types of workloads, we evaluate the performance of EMPIOT against the ground truth obtained from high-accuracy devices. Our results show that, for very low-power devices that utilize 802.15.4 wireless standard, measurement error is less than 4%. In addition, for 802.11-based devices that generate short and high power spikes, error is less than 3%.

Identifying Channel Saturation in Wi-Fi Networks via Passive Monitoring of IEEE 802.11 Beacon Jitter

Conference Paper

Full-text available

Nov 2017

Every day large numbers of users connect to IEEE 802.11 networks in order to access the Internet and all sorts of services. However, due to their unplanned and unregulated nature, and the lack of admission control and Quality of Service Guarantees, these wireless networks can experience traffic demand that exceeds the network capacity. In this case, if a device tries to send more traffic, or if a new device joins the network, the aggregate throughput does not necessarily increase. In this paper we show that it is possible for IEEE 802.11 stations to detect a saturated channel by passively monitoring the beacon frames. Access points (AP) send beacon frames periodically and encode them using the strongest modulation and coding scheme, so that even stations far away from the sending APs can decode them correctly. When sending beacons, APs sense the channel first and, if it is busy, delay sending the frame, resulting in unevenly spaced beacon frames, whenever other transmitters are active. We present several experiments, under varying traffic loads, and analyze the distribution of the beacon jitter, whose variance increases as the offered load increases. We show that it is possible to determine, with an acceptable error rate, whether a channel is saturated by comparing the distribution of the beacon jitter with a reference distribution corresponding to a saturated channel.

Is WiFi suitable for energy efficient IoT deployments? A performance study

Conference Paper

Full-text available

Sep 2017

‘Smart’ Is Not Free: Energy Consumption of Consumer Home Automation Systems

Article

Dec 2019

The proliferation of consumer Home Automation Systems (HAS), which is increasingly based on the Internet of Things (IoT) architecture, comes with added costs for the additional electrical energy required to power the automation interfaces and the standby energy consumption required to maintain connectivity and/or “smartness”. In this paper, we use a bottom-up approach to develop novel system-level energy consumption models for consumer HAS devices and quantify the energy consumption for a typical HAS. We then assess the potential impact on global Information and Communications Technology (ICT) energy use. We show that, on average, HAS may consume over one-third of the annual energy used in a mid-sized home, with non-trivial impact on the global ICT energy footprint.

Wi-LE: Can WiFi Replace Bluetooth?

Conference Paper

Nov 2019

Despite the ubiquity of WiFi devices, Bluetooth is widely used for communication in low-power, low data-rate devices. This is because Bluetooth consumes much less power than WiFi which results in longer battery life. The higher power consumption of WiFi devices is due to overheads from either establishing or maintaining connections with the access point. Surprisingly, Bluetooth devices require nearly three times as much energy to transmit a bit of data at the physical layer than WiFi devices. In this paper, we propose Wi-LE a WiFi-compatible communication system that avoids the power hungry process of establishing or maintaining a connection. We implement and evaluate Wi-LE using an off-the-shelf WiFi module. Our results show that Wi-LE has power consumption similar to that of Bluetooth Low Energy (BLE). This demonstrates the potential for Wi-LE to be used in place of BLE.

EE-IoT: An Energy-Efficient IoT Communication Scheme for WLANs

Conference Paper

Apr 2019

Why it takes so long to connect to a WiFi access point

Conference Paper

May 2017

Towards Understanding and Enhancing Association and Long Sleep in Low-Power WiFi IoT Systems

Abstract and Figures

Recommended publications

Empirical Study and Enhancement of Association and Long Sleep in 802.11 IoT Systems

EAPS: Edge-Assisted Predictive Sleep Scheduling for 802.11 IoT Stations

EAPS: Edge-Assisted Predictive Sleep Scheduling for 802.11 IoT Stations

FLIP: A Framework for Leveraging eBPF to Augment WiFi Access Points and Investigate Network Performa...