AutoPlex: inter-session multiplexing congestion control for large-scale live video services

Full text

AutoPlex: Inter-Session Multiplexing Congestion Control for
Large-Scale Live Video Services
Bo Wu
Tencent Technologies
Tong Li∗
Renmin University of China
Cheng Luo
Tencent Technologies
Changkui Ouyang
Tencent Technologies
Xinle Du
Tsinghua University
Fuyu Wang
Tencent Technologies
ABSTRACT
The increasingly obvious advances in live video services introduce
the urgent need for enhancing network transmission performance,
especially by designing an ecient congestion control (CC) scheme.
Unfortunately, the previous rule-based CC methods cannot adapt
well to various network conditions and statuses while machine-
learning-powered CC paradigms always suer from non-trivial
system overhead and unstable eects.
In this paper, we rst conduct a large-scale network measure-
ment for 800+ million live video streams, and nd that QoS metrics
of better-performed sessions show similarity in the same user group.
We then propose AutoPlex, an inter-session multiplexing CC frame-
work that makes full use of this similarity and automatically adjusts
CC parameters (i.e., pacing rate and congestion window size). Au-
toPlex supports user-dened policies that can act as standards to
learn QoS features of better-performed sessions. We implement the
proposed AutoPlex prototype based on QUIC protocol and BBR
algorithm, and conduct experiments in the real live CDN proxy. The
experimental results demonstrate the potentials of AutoPlex for the
transmission optimization of live video applications, in which the
average (or 90th-percentile) retransmission ratio can be reduced by
24%
⇠
27% (or 32%
⇠
40%) while the average value of goodput/rtt is
promoted by 14% ⇠32%.
CCS CONCEPTS
•Networks →Transport protocols;
KEYWORDS
Congestion Control; Network Measurement
ACM Reference Format:
Bo Wu, Tong Li, Cheng Luo, Changkui Ouyang, Xinle Du, and Fuyu Wang.
2022. AutoPlex: Inter-Session Multiplexing Congestion Control for Large-
Scale Live Video Services. In ACM SIGCOMM 2022 Workshop on Network-
Application Integration (NAI ’22), August 22, 2022, Amsterdam, Netherlands.
ACM, New York, NY, USA, 6 pages. https://doi.org/10.1145/3538401.3546596
∗Tong Li is the corresponding author (tong.li@ruc.edu.cn).
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for prot or commercial advantage and that copies bear this notice and the full citation
on the rst page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specic permission and/or a
fee. Request permissions from permissions@acm.org.
NAI ’22, August 22, 2022, Amsterdam, Netherlands
©2022 Association for Computing Machinery.
ACM ISBN 978-1-4503-9395-9/22/08. . . $15.00
https://doi.org/10.1145/3538401.3546596
1 INTRODUCTION
In the past few years, live video services (LVS) such as TikTok Live,
YouTube Live, and Twitch have experienced rapid growth, where
the weekly coverage of live streaming worldwide has reached 30.4%
in Q3 of 2021 [
1
]. However, the real network conditions vary with
regions, resulting in vastly dierent network QoS among these
regions. In particular, poor network QoS has negative impacts on
both LVS consumers and CDN vendors. For example, users might
complain or stop using some LVS as their deteriorated QoE (e.g.,
video freezing) while CDN vendors are also unable to accept their
cost increases, e.g., caused by higher retransmission ratios.
The existing mechanisms mainly focus on designing, optimizing
congestion control (CC) algorithms for enhancing transmission per-
formance of LVS trac. On the one hand, rule-based CC schemes
(e.g., Reno [
2
], Cubic [
3
] and BBR [
4
]) can introduce better perfor-
mance in some specic network status or services. However, the
way they set CC parameters (e.g., maximum pacing rate) as xed val-
ues obviously cannot adapt to dierent network conditions, which
is the main reason for the huge QoS dierences between regions.
For example, 10Mbps maximum pacing rate might incur lower re-
transmission ratio in 10Mbps bottleneck bandwidth (BtlBW) than
in 1Mbps BtlBW. On the other hand, machine learning (ML) enables
more exible CC parameter conguration for dierent network
conditions [
5
–
7
]. However, the way they construct an ML model
for each user group (UG) cannot address the issue that a large num-
ber of UGs exist in a large-scale network. For example, over 18000
UGs (grouped based on "state-city-ASN" rule) can be obtained in
Brazil, in which 27.3% of national LVS sessions will occupy at least
100 ML models to achieve trac transmission optimization (see
§3.2). That is actually unable to be deployed due to the non-trival
memory or computation overhead caused by numerous ML models.
Therefore, a deployable CC scheme is highly required for adaptive
parameter congurations based on dierent network conditions.
In this paper, we make large-scale network measurements for
LVS trac transmissions, in which we can nd the same CC param-
eter conguration can bring dierent network QoS performance
among various UGs (see §3.3). Besides, we also learn the better-
performed sessions with lower retransmission ratios (retran_ratios)
or larger goodput/rtt value
1
have similar QoS features in a same
UG (see §3.4). Based on the above measurement and analysis, a
novel CC framework named AutoPlex is proposed, which can au-
tomatically multiplex QoS metrics of better-performed sessions in
the last period and makes adaptive CC parameter congurations
for next-period LVS trac. Take BBR as an example, maximun
1
The value of goodput/rtt is usually leveraged to evaluate the CC performance [
8
],
which is used as a target for LVS transmission optimization.

NAI ’22, August 22, 2022, Amsterdam, Netherlands Wu. et al.
pacing rate (max_pacing_rate) and maximum congestion window
(max_cwnd) can be congured based on little-changed QoS val-
ues such as goodput, maximum inight size (max_in�ight) and
max_cwnd that better-performed LVS sessions introduce in the
last period. Compared to rule-based CC methods, AutoPlex en-
ables adaptive parameter congurations for dierent UGs that
has various network conditions without incurring heavyweight
ML overhead. Meanwhile, AutoPlex supports user-dened policies
to achieve directional optimization for some specic QoS metric,
which act as selection criteria to extract better-performed LVS
sessions. For example, the QoS metrics of LVS trac with lower
retran_ratio can be leveraged to congure CC parameters if the pol-
icy is dened to reduce retran_ratio. We implement our proposed
AutoPlex and partially deploy it in live CDN proxy that is based
on QUIC protocol and BBR algorithm for LVS trac transmission.
The experimental results demonstrate the potential of AutoPlex
for LVS application. Concretely, the average (or 90th-percentile)
retran_ratio can be reduced by 24%
⇠
27% (or 32%
⇠
40%) while the
average value of goodput/rtt is promoted by by 14% ⇠32%.
The remainder of this paper is organized as follows. §2 intro-
duces the related work about congestion control. §3 introduces the
background and motivation. Then, the design details of AutoPlex
are depicted in §4, and experimental evaluation is described in §5.
Finally, we conclude this article in §6.
2 RELATED WORK
Reno [
2
] is the most well known scheme that proposes the key
concepts of parameters in CC such as congestion window (cwnd),
inight size, initial window, etc. Many TCP variants [
3
,
4
,
9
–
14
] are
aimed at specic networks and modify or improve the settings of
these parameters, based on their specic insights or assumptions.
For example, Vegas [
9
] sets cwnd based on RTT changes in networks
with stable delay. BIC [
10
] uses binary search to set cwnd, and
Cubic [
3
] uses a cubic function. Verus [
11
] and Sprout [
12
] are
designed specically for wireless networks to better t the frequent
changes in bandwidth. BBR [
4
] estimates the maximum bandwidth
to set pacing rate and computes the bandwidth and delay product
(bdp) to set cwnd. TACK [
15
] computes bdp and the minimum
RTT to set ACK frequency rather than cwnd in wireless network
cases. Although these schemes can improve TCP’s adaptability to
dierent network scenarios, for a specic scheme, its underlying
logic and way of setting CC parameters remain inexible.
Machine learning (ML) enables more exible CC parameter con-
guration for dierent network conditions [
5
–
7
]. We roughly di-
vide the ML-powered schemes into intra-session learning-based
and inter-session learning-based schemes.
Intra-session learning-based schemes learn the network con-
ditions (as well as status) in a same session and adjust the CC
parameters in real time. For example, Vivace [
16
] and PCC [
5
]
adjusts sending rates in real time, and determines the size of the
increment according to a performance utility function gradient.
RemyCC [
17
] iteratively searches for a state-action mapping table
to maximize an objective function. However, this intra-session esti-
mation still suer from inexibility since it relies on the stability
and predictability of the underlying network.
Table 1: User group (UG) amount and session ratio with 13.56
million live video trac in Brazil.
Grouping Rules UG Num. Session Ratio
Top 10 Top 50 Top 100
State-City-ASN 18620 11.8% 22.2% 27.3%
State-ASN 10326 16.7% 27.7% 32.9%
ASN 8329 30.9% 35.7% 38.3%
In the contrast, inter-session learning-based schemes learn the
network conditions across dierent sessions and adjust CC parame-
ters in the context of the same UG. For example, Indigo [
8
] uses an
LSTM model to train across a wide range of scenarios, guided by
reaching the optimal operating point of the network. Orca [
18
] com-
bines a learning model and rule-based scheme to ensure available
recoveries from wrong equilibrium. However, as discussed before,
the way they construct an ML model for each UG cannot address
the issue that a large number of UGs exist in a large-scale network
(see §3.2). In this paper, we propose AutoPlex to improve the CC
adaptability while reduces the computation and memory overhead
of inter-session learning-based schemes.
3 MOTIVATION
3.1 LVS Demands Flexible CC Schemes
It is well-studied that one-size-ts-all scheme does not exist in
CC [
19
]. Specically, for the worldwide LVS, covering a wide va-
riety of network conditions in various regions and countries, no
single scheme can adequately prevail. Take a famous live platform
as an example, the average delay in Malaysia is 86ms, while in
Hainan Province of China it is 36ms. Besides, the average available
bandwidth in Hainan is 7.3 Mbps for each LVS streams compared
to 3.2Mbps in Turkey. Moreover, LVS streams incur average 5.2%
packet loss rate in Turkey and 3.8% in Brazil, respectively. For a
same region such as Brazil, the average, 50th- and 90th-percentile
delays (available bandwidths) are 52ms (3.9Mbps), 27ms (5.4Mbps)
and 169ms (15.2Mbps), respectively2.
As discussed above, rule-based CC schemes (e.g., Reno, CUBIC,
Sprout, BBR, etc.) show inexibility in the underlying logic and way
of setting CC parameters. It is expected that dierent choices of
CC parameter settings to fare dierently in dierent contexts (e.g.,
UGs). Recently, it has been demonstrated that ML-powered schemes
show great potential to improve the exibility of CC [5–8, 17, 18].
3.2
UG-based and ML-powered Schemes Do Not
Scale Well
To pursue a more stable and precise control eect, many existing
researches focus on dividing user groups (UGs) based on client’s
region and ISP/AS number (ASN) information and provide some spe-
cic parameters for each UG [
20
][
6
]. However, non-trival memory
or computation overhead incurred by ML models can be the bottle-
neck factor of real deployment [
21
], especially when the amount
of user group becomes very large. Table 1 shows the amount of
2
The data mentioned above is all based on our large-scale network measurements for
LVS trac transmission.

AutoPlex: Inter-Session Multiplexing Congestion Control NAI ’22, August 22, 2022, Amsterdam, Netherlands
Avg. 50th 75th 90th
Metrics
0
2%
4%
6%
8%
10%
12%
Retran ratio
UG 1
UG 2
UG 3
(a) Retransmission ratio
Avg. 50th 75th 90th
Metrics
0
20
40
60
80
100
120
SRTT (ms)
UG 1
UG 2
UG 3
(b) Smooth RT T
Avg. 50th 75th 90th
Metrics
0
2
4
6
8
10
12
14
Goodput (Mbps)
UG 1
UG 2
UG 3
(c) Goodput
Figure 1: The QoS dierences of LVS streams that belong to three dierent UGs.
UG under dierent grouping rules in Brazil. For more ne-grained
grouping rules (e.g., state-city-ASN), more UGs can be obtained
while with the most coarse-grained rule (i.e., ASN)
3
, UG number
can also reach 8329. Meanwhile, we also count the top 10, 50, and
100 UGs based on their session amounts and nd that leverag-
ing 100 ML models for top 100 UGs can only cover 27.3% ~38.3%
steams of entire LVS sessions. Compared to their limited benets,
the overhead introduced by ML models is unacceptable in complex
networks, especially with a large number of UGs. This reveals that
UG-based and ML-powered schemes do not scale well.
3.3
LVS Session Performance Varies in Dierent
UGs
We make large-scale network measurements for 3 months and
record session data of 800+ million LVS streams, which contains
client IP address and some QoS metrics, e.g., minimum, maximum
and smooth RTT (denoted by min_rtt,max_rtt and srtt, respec-
tively), goodput, max_cwnd,max_in�ight,retran_ratio, etc. We use
"state-city-ASN" as a grouping rule and construct <IP, UG> mapping
table to extract QoS values for every UG.
Figure 1 shows the dierences of retran_ratio,srtt and goodput
between dierent UGs, which is depicted from the perspective of
average, 50th-, 75th and 90th-percentile values. We can learn LVS
sessions in dierent UGs perform dierently in terms of the above
three QoS metrics. For example, the LVS sessions of UG 2 result in
a smaller and more stable retran_ratio while a lower latency (evalu-
ated by srtt) is actually introduced in UG 1 (instead of in UG 2 and
3) from the perspective of average, 50th-, 75th- and 90th-percentile
values. Meanwhile, UG 3 always holds a smaller goodput than the
other two UGs, which might be caused by its higher retran_ratio
and smaller goodput. Therefore, adaptively conguring a series of
CC parameters for each UG will be highly required to meet dierent
network conditions and optimize LVS transmissions.
3.4 Better-Performed Sessions Perform
Similarly in the Same UG
To explore potential associations of session QoS within a same UG,
we randomly extract six-day LVS trac data of UG 1. The time
3
In this paper, we regard ASN (instead of ISP) as the basic for user grouping, as ASN
can better reect the session similarity in a same UG. With user grouping, each user
or client IP address can be exactly divided into a UG.
interval Tis set to 15 minutes, where some LVS trac with spe-
cic ranges of retran_ratio and goodput/srtt is extracted as better-
or worse-performed LVS sessions (dened as below). Figure 2
shows that better-performed sessions extremely outperform worse-
performed streams in terms of retran_ratios (
⇠
0.03% vs.
⇠
2.9%) and
goodput/srtt (⇠872 Mbps/s vs. ⇠53Mbps/s) on average.
•
Better-performed sessions are dened as the LVS streams
with 5th- to 20th-percentile retran_ratios or descending good-
put/srtt in all last-period sessions.
•
Worse-performed sessions are dened as the LVS streams
with 75th- to 90th-percentile retran_ratios or descending
goodput/srtt in all last-period sessions.
03/25 03/26 03/27 03/28 03/29 03/30 03/31
0
2.5%
5.0%
7.5%
10.0%
12.5%
15.0%
17.5%
Average retran_ratio
All LVS sessions
Better-performed sessions
Worse-performed sessions
(a) Retransmission ratio
03/25 03/26 03/27 03/28 03/29 03/30 03/31
0
5
10
15
20
25
Average goodput/srtt (×10 2Mbps/s)
All LVS traffic
Better-performed sessions
Worse-performed sessions
(b) Goodput/srtt
Figure 2: Better- and worse-performed LVS sessions.
To explore the similarity between better-performed LVS sessions
in a same UG, we analyze the changes of goodput, max_in�ight
and max_cwnd in two adjacent time periods. Figure 3(a) depicts the
cumulative goodput change rates of better- and worse-performed
sessions in a same UG. We nd that better-performed sessions has
little-changed goodput in two adjacent periods, whose average
change rate is 25%
⇠
30% of that of worse-performed sessions.
The cumulative changes of max_cwnd and max_in�ight for better-
and worse-performed session is shown in Figure 3(b) and 3(c), in
which the curve slope represents the change rate compared to last-
period values. We can learn the better-performed LVS sessions
can introduce smoother max_cwnd and max_in�ight changes than
those worse-performed trac. For example, 3.0
⇥⇠
5.1
⇥
changes
of max_cwnd and 1.3
⇥⇠
3.0
⇥
changes of max_in�ight are caused
by worse-performed sessions compared to better-performed LVS
streams. Therefore, better-performed LVS sessions in a same UG
perform similarly with little-changed QoS values, which can be
leveraged to automatically congure CC parameters for next period.

NAI ’22, August 22, 2022, Amsterdam, Netherlands Wu. et al.
03/25 03/26 03/27 03/28 03/29 03/30 03/31
0
2
4
6
8
Cumulative gooput change rate (×10 4)
Better-performed sessions (retran_ratio)
Worse-performed sessions (retran_ratio)
Better-performed sessions (goodput/srtt)
Worse-performed sessions (goodput/srtt)
(a) Cumulative goodput change rates
03/25 03/26 03/27 03/28 03/29 03/30 03/31
0
1
2
3
4
5
Cumulative max_cwnd changes (×10 4)
Better-performed sessions (retran_ratio)
Worse-performed sessions (retran_ratio)
Better-performed sessions (goodput/srtt)
Worse-performed sessions (goodput/srtt)
(b) Cumulative max_cwnd changes
03/25 03/26 03/27 03/28 03/29 03/30 03/31
0
50
100
150
200
250
300
Cumulative max_inflight changes (Mb)
Better-performed sessions (retran_ratio)
Worse-performed sessions (retran_ratio)
Better-performed sessions (goodput/srtt)
Worse-performed sessions (goodput/srtt)
(c) Cumulative max_in�ight changes
Figure 3: The QoS similarities of better-performed LVS session in a same UG.
4 THE AUTOPLEX DESIGN
4.1 AutoPlex Framework
AutoPlex enables to enhance LVS transmission performance by
automatically multiplexing inter-session QoS similarities of better-
performed LVS sessions in the last period, which will be used to
congure next-period CC parameters. In particular, the maximum
value of cwnd and pacing_rate are all congured as little-changed
based on the QoS values of latest better-performed sessions. To
meet diverse requirements, AutoPlex framework supports user-
dened QoS policies (e.g., lowering retran_ratio or promoting the
value of goodput/srtt), which will be followed to learn similar QoS
metrics of better-performed LVS sessions. Figure 4 depicts AutoPlex
framework that contains three modules for measurement, decision
and execution. Note that these modules can be deployed on a same
CDN proxy server or on dierent servers.
Data
Collection
User
Grouping
QoS of
Each UG
Measurement Module Decision Module
User-defined
QoS targets
User-defined
QoS targets
data
collection
user
grouping
QoS of
each UG
Measurement Module Decision Module
parameters of 'better' sessions
user-defined
QoS policies
parameters of each UG
Execution Module
New LVS
request <IP, UG> <UG, Parameters>
UG_ID max_cwnd
max_pacing_rate
New LVS
response
Figure 4: The framework of AutoPlex.
Measurement module can perform data collection of all LVS
sessions and get last-period QoS values of each UG by looking up
the pre-established <IP, UG> mapping table. With user-dened QoS
policies, decision module obtains better-performed LVS sessions
and records their similar QoS values (e.g., goodput,max_cwnd and
max_in�ight) that will be utilized for next-period CC parameter
congurations. When receiving a new LVS request, execution mod-
ule extracts remote IP and gains UG identier UG_ID by inquiring
our pre-established <IP, UG> table. With <UG, parameters> map-
ping table, CC parameters (i.e., max_cwnd and max_pacing_rate) of
this UG will be obtained and then be congured for the response
of this LVS session. In AutoPlex, measurement module counts QoS
values of each UG for every time interval of T(e.g., T = 30 mins),
Algorithm 1 CC parameter conguration pseudo code.
1: function CC  ()
2: Require:
3: %DBA ,&D6
4: Compute:
5: &14CC4A ⇢G CA02C (&D6,%
DBA )
6: 6>>3? DC ,<0 G _8=5 ;86⌘C,<0G_2F=3 <40= (&14C C4A )
7: ?028=6_A0C 4_;0BC =6>>3 ?D C
8: <0G_2F=3_;0BC =<0G _8=5 ;86⌘C+<0G_2F=3
2
9: Operate:
10: ?028=6_A0C 4_2?C =?028=6_608= ⇤<0G _1F
11: ?028=6_A0C 4 =5(?028=6_A0C 4 _2?C,?028=6_A0C4 _;0BC)
12: 8=5 ;86⌘C_20? =5(8=5 ;86⌘C_20?,<0G_2F=3_;0BC)
13: end function
which is utilized by decision module for maintaining latest <UG,
parameters> mapping table in execution module.
4.2 Design Details
AutoPlex focuses on achieving automated congurations for CC pa-
rameters by multiplexing similar QoS values of those LVS sessions
that performed better in the last period. In this paper, more ne-
grained grouping rule (i.e., state-city-ASN) is leveraged to classify
LVS consumers. Algorithm 1 shows the pseudo code for CC pa-
rameter congurations, where user-dened policies
%DBA
and each
UG’s QoS values
&D6
(i.e., retran_ratio,goodput,srtt,max_in�ight
and max_cwnd) act as inputs for parameter computations.
Better-performed session extraction. AutoPlex enables deci-
sion module to extract similar QoS values (denoted by
&14CC4A
) of
better-performed LVS sessions in each UG. In Algorithm 1, function
⇢GCA 02C(·)
denotes sorting
&D6
according to
%DBA
-related metric
(e.g., retran_ratio) and then obtains
&14CC4A
from
8
to
9
percentile
values of the sorted
&D6
. For example,
8=
5and
9=
20 are set to
extract
&14CC4A
from
&D6
(ascended by retran_ratio) in §3.3. Then,
the average values of goodput,max_in�ight and max_cwnd can be
obtained while taking the mean operation for &14CC4A .
CC parameter conguration. With goodput,max_in�ight and
max_cwnd of better-performed sessions, AutoPlex rst computes
pacing_rate_last and max_cwnd_last, which reects average pac-
ing_rate and our desired max_cwnd, respectively, in the last period.

AutoPlex: Inter-Session Multiplexing Congestion Control NAI ’22, August 22, 2022, Amsterdam, Netherlands
f1 f2 f3 f4
f1 f2 f3 f4
path
Internet
云服务器
终端用户
音视频报文
音视频报文
消息确认报文 消息确认报文
CDN servers
UG 1 UG 2 UG 3 UG 4
CDN proxy
Internet
Figure 5: The experimental evaluation.
Then pacing rate and inight cap can be congured for next-period
LVS trac transmissions, which will be elaborated in §4.3.
4.3 AutoPlex-based BBR Implementation
In this section, we give an example of how BBR makes use of the
two outputs of AutoPlex, i.e., pacing_rate_last and max_cwnd_last,
to adjust its pacing rate and inght cap.
Adjust pacing rate. A BBR sender controls its pacing_rate with
the help of pacing and an estimated maximum bandwidth max_bw,
where the estimate for max_bw is based on a windowed maximum
lter of the delivery rate that the receiver experiences. The pac-
ing rate is varied in an eight-phase cycle using a pacing_gain of
5/4,3/4,1,1,1,1,1,1, where each phase lasts for an RTT. The BBR
implementation based on AutoPlex computes the pacing rate (i.e.,
pacing_rate_cpt) as line 11 of Algorithm 1 shows. In this paper,
5(·)
is a customizable function according to the optimization goal. For
lower retran_ratio or higher goodput/srtt, we set
5(·)
as a function
that takes the minimum value when
01
and takes the average
value when 0>1as follows.
5(0,1)=(001
0+1
20>1(1)
Adjust inight cap. Although BBR does not use a congestion
window or ACK clocking to control the amount of inight data,
it uses an inight data limit (e.g., 2bdp), which we call it inght
cap (denoted by in�ight_cap) in this paper. The BBR implemen-
tation based on AutoPlex computes the inight cap as line 12 in
Algorithm 1 shows. This enables little-changed maximum values
of in�ight_cap, especially when the computed in�ight_cap exceeds
last-period max_cwnd_last.
Discussion. In real-world deployment, we can activate AutoPlex
just for some selected LVS trac, e.g., high-priority sessions from
VIP users. In this case, low-priority sessions are leveraged to explore
better QoS metrics, which will be multiplexed by high-priority ses-
sions. This partial deployment strategy enhances the convergence
of AutoPlex. Note that, even though the same CC parameters are
congured, better- and worse-performed LVS sessions in a same
UG can be still identied as changing network conditions and status.
UG 1 UG 2 UG 3 UG 4
0
1%
2%
3%
4%
Avgerage retran_ratio
Baseline
AutoPlex
(a) Average retran_ratio
UG 1 UG 2 UG 3 UG 4
0
0.2%
0.4%
0.6%
0.8%
50th-percentile retran_ratio
Baseline
AutoPlex
(b) 50th-percentile retran_ratio
UG 1 UG 2 UG 3 UG 4
0
1%
2%
3%
4%
75th-percentile retran_ratio
Baseline
AutoPlex
(c) 75th-percentile retran_ratio
UG 1 UG 2 UG 3 UG 4
0
2%
4%
6%
8%
10%
12%
90th-percentile retran_ratio
Baseline
AutoPlex
(d) 90th-percentile retran_ratio
Figure 6: Retransmission ratios of AutoPlex.
Therefore, the idea of multiplexing inter-session QoS metrics for
CC parameter congurations can be always achieved over time.
5 EXPERIMENTAL EVALUATION
In this section, we implement AutoPlex prototype and partially
deploy it in the real live CDN proxy (as Figure 5 shows) that is de-
veloped based on Nginx [
22
], which enables measurement, decision
and execution modules of AutoPlex
4
. We randomly select four UGs
(i.e., UG 1
⇠
UG 4) to evaluate AutoPlex using the following metrics:
retran_ratio and gooodput/srtt, which can be set as user-dened
policies. The time period is set to 30 mins and LVS sessions in our
evaluation are all based on QUIC protocol and BBRv1 scheme (as
baseline) for transmission.
5.1 Retransmission Ratio
In AutoPlex framework, users can dene their own preference as an
optimization goal (e.g., lowering retran_ratio) for LVS trac trans-
mission. In this section, we select better-performed sessions with
5th- to 20th-percentile retran_ratio for CC parameter conguration
of next-period LVS trac.
Figure 6 depicts average, 50th-, 75th- and 90th-percentile re-
tran_ratio changes of dierent UGs, respectively, when deploy-
ing AutoPlex prototype. We can learn AutoPlex can introduce
lower retran_ratio, where the average, 75th- and 90th-percentile
retran_ratios are reduced by 24%
⇠
27%, 28%
⇠
41% and 32%
⇠
44%, respectively, compared to the existing BBR scheme. Note that
50th-percentile retran_ratio does not gain a obvious optimization,
which always keeps close to optimal values (i.e., 0.37%
⇠
0.70%)
that are dierent to be further optimized. Meanwhile, AutoPlex
brings dierentiated reduction ratios for dierent UGs to optimize
their retran_ratios. This is because traditional xed CC parameter
settings for entire networks can introduce various baseline values
for each UG while AutoPlex can adjust CC parameters based on
each UG’s network conditions.
4
In this paper, we meanly focus on the LVS transmission optimization from CDN proxy
to user clients. Then video streaming interaction is beyond the scope of our research.

NAI ’22, August 22, 2022, Amsterdam, Netherlands Wu. et al.
UG 1 UG 2 UG 3 UG 4
0
100
200
300
400
500
Avgerage goodput/srtt (Mbps/s)
Baseline
AutoPlex
(a) Average goodput/srtt
UG 1 UG 2 UG 3 UG 4
0
100
200
300
400
50th-percentile goodput/srtt (Mbps/s)
Baseline
AutoPlex
(b) 50th-percentile goodput/srtt
UG 1 UG 2 UG 3 UG 4
0
100
200
300
400
500
600
700
75th-percentile goodput/srtt (Mbps/s)
Baseline
AutoPlex
(c) 75th-percentile goodput/srtt
UG 1 UG 2 UG 3 UG 4
0
200
400
600
800
1000
1200
90th-percentile goodput/srtt (Mbps/s)
Baseline
AutoPlex
(d) 90th-percentile goodput/srtt
Figure 7: Goodput/srtt values of AutoPlex.
5.2 Ratio of Goodput to SRTT
AutoPlex enables to enhance goodput/srtt by multiplexing QoS val-
uess of better-performed LVS sessions to congure max_cwnd and
max_pacing_rate for the next LVS sessions. Note that the sessions
with 80th- to 95th-percentile values of goodput/srtt in the last period
are extracted as better-performed LVS sessions.
Figure 7 shows the values of goodput/srtt in dierent UGs. We
can know AutoPlex can achieve a better optimization for the value
of goodput/srtt, in which the average, 50th- and 75th-percentile
values are all signicantly enhanced (i.e., by 14%
⇠
32%, 44%
⇠
128%, and 19%
⇠
56%, respectively). By contrast, the 90th-percentile
values have not been improved obviously. This might be because
these values are already close to the optimal and have limited room
to be further optimized. AutoPlex can introduce dierentiated op-
timization results for various UGs. For example, 32% and 14% im-
provements of average goodput/srtt are incurred for UG 1 and UG
3, respectively. For the metric of goodput/srtt, AutoPlex can achieve
obvious optimization for those poor performances (e.g., 50th- and
75th-percentile values) while keeping plausible improvements for
already-better performances.
6 CONCLUSION AND FUTURE WORK
In this paper, we take a rst step to design, analyze, implement
and evaluate inter-session multiplexing congestion control frame-
work named AutoPlex, which is based on our performed large-scale
network measurements for live video streams. AutoPlex enables
adaptive CC parameter congurations by multiplexing QoS values
of better-performed LVS trac in the last period. Besides, user-
dened policies can also be supported in AutoPlex, which act as
standards to learn QoS features of better-performed LVS trac. We
complete experimental evaluation by implementing and deploying
AutoPlex in the real live CDN proxy, whose results demonstrate
the huge potentials that AutoPlex introduces for optimizing LVS
trac transmission.
In the future, machine learning such as DRL can be leveraged to
select better-performed LVS sessions based on the QoS performance
gains of prior session selection. Besides, exploring more parameter
settings [
23
,
24
] by multiplexing inter-session QoS values is also a
natural extension of AutoPlex.
ACKNOWLEDGMENTS
We would like to thank the NAI reviewers and Dr. Kai Gao for their
extensive and valuable feedback. We also thank Senzhen Liu for his
kind comments during our performing experimental evaluation.
REFERENCES
[1]
Max Wilbert. Top 22 live streaming platforms: Everything you need to know.
https://www.dacast.com/blog/, 2022.
[2]
Van Jacobson. Congestion avoidance and control. In ACM SIGCOMM computer
communication review, volume 18, pages 314–329. ACM, 1988.
[3]
Sangtae Ha, Injong Rhee, and Lisong Xu. Cubic: a new tcp-friendly high-speed
tcp variant. ACM SIGOPS operating systems review, 42(5):64–74, 2008.
[4]
Neal Cardwell, Yuchung Cheng, C Stephen Gunn, Soheil Hassas Yeganeh, and
Van Jacobson. BBR: Congestion-based congestion control: Measuring bottleneck
bandwidth and round-trip propagation time. ACM Queue, 14(5):20–53, 2016.
[5]
Mo Dong, Qingxi Li, Doron Zarchy, P Brighten Godfrey, and Michael Schapira.
PCC: Re-architecting congestion control for consistent high performance. In
USENIX NSDI, pages 395–408, 2015.
[6]
Xiaohui Nie, Youjian Zhao, Zhihan Li, Guo Chen, Kaixin Sui, Jiyang Zhang, Zijie
Ye, and Dan Pei. Dynamic tcp initial windows and congestion control schemes
through reinforcement learning. IEEE JSAC, 37(6):1231–1247, 2019.
[7]
Xu Li, Feilong Tang, Jiacheng Liu, Laurence T Yang, Luoyi Fu, and Long Chen.
AUTO: Adaptive congestion control based on multi-objective reinforcement
learning for the satellite-ground integrated network. In USENIX ATC, 2021.
[8]
Francis Y Yan, Jestin Ma, Greg D Hill, Deepti Raghavan, Riad S Wahby, Philip
Levis, and Keith Winstein. Pantheon: the training ground for internet congestion-
control research. In USENIX ATC, pages 731–743, 2018.
[9]
Lawrence S Brakmo, Sean W O’Malley, and Larry L Peterson. TCP Vegas: New
techniques for congestion detection and avoidance, volume 24. ACM, 1994.
[10]
Lisong Xu, Khaled Harfoush, and Injong Rhee. Binary increase congestion control
(bic) for fast long-distance networks. In IEEE Infocom, volume 4, pages 2514–2524,
2004.
[11]
Yasir Zaki, Thomas Pötsch, Jay Chen, Lakshminarayanan Subramanian, and
Carmelita Görg. Adaptive congestion control for unpredictable cellular networks.
In ACM SIGCOMM, pages 509–522, 2015.
[12]
Keith Winstein, Anirudh Sivaraman, and Hari Balakrishnan. Stochastic forecasts
achieve high throughput and low delay over cellular networks. In USENIX NSDI,
pages 459–472, 2013.
[13]
Venkat Arun and Hari Balakrishnan. Copa: Practical delay-based congestion
control for the internet. In USENIX NSDI, pages 329–342, 2018.
[14]
Tong Li, Kai Zheng, and Ke Xu. Acknowledgment on demand for transport
control. IEEE Internet Computing, 25(2):109–115, 2021.
[15]
Tong Li, Kai Zheng, Ke Xu, Rahul Arvind Jadhav, Tao Xiong, Keith Winstein, and
Kun Tan. Tack: Improving wireless transport performance by taming acknowl-
edgments. In ACM SIGCOMM, pages 15–30, 2020.
[16]
Mo Dong, Tong Meng, DoronZarchy, Engin Arslan, Yossi Gilad, Brighten Godfrey,
and Michael Schapira.
{
PCC
}
vivace: Online-learning congestion control. In
USENIX NSDI, pages 343–356, 2018.
[17]
Keith Winstein and Hari Balakrishnan. Tcp ex machina: Computer-generated
congestion control. ACM SIGCOMM Computer Communication Review, 43(4):123–
134, 2013.
[18]
Soheil Abbasloo, Chen-Yu Yen, and H Jonathan Chao. Classic meets modern: A
pragmatic learning-based congestion control for the internet. In ACM SIGCOMM,
2020.
[19]
Michael Schapira and Keith Winstein. Congestion-control throwdown. In ACM
Hotnets, pages 122–128, 2017.
[20]
Junchen Jiang, Shijie Sun, Vyas Sekar, and Hui Zhang. Pytheas: Enabling
data-driven quality of experience optimization using Group-BasedExploration-
Exploitation. In USENIX NSDI, 2017.
[21]
Minsik Cho and Daniel Brand. MEC: Memory-ecient convolution for deep
neural network. In International Conference on Machine Learning, pages 815–824.
PMLR, 2017.
[22]
F5 Networks Inc. NGNIX: Advanced load balancer, web server, & reverse proxy.
https://www.nginx.com, 2022.
[23]
Tong Li, Kai Zheng, Ke Xu, Rahul Arvind Jadhav, Tao Xiong, Keith Winstein,
and Kun Tan. Revisiting acknowledgment mechanism for transport control:
Modeling, analysis, and implementation. IEEE/ACM TON, 29(6):2678–2692, 2021.
[24]
Hui Xie and Li Tong. Revisiting loss recovery for high-speed transmission. In
IEEE WCNC, pages 1–6, 2022.