Experimenting with P2MT’s MeshRanger2-o mesh routers. 

Context in source publication

Context 1

... to the multihop throughput degradation problem, and a number of network planning and communication algorithms were proposed based on the multiradio architecture in the literature. However, our experimental results indicate that it is not the case for TCP connections in 802.11n WMNs. For instance, we found through real-life experiments that the multihop TCP throughput degrades for more than 70 percent after traversing five hops over an 802.11n mesh network even under some nearly perfect conditions. The bandwidth-delay product [15] is a useful quantity for analyzing network performance. It represents the number of bytes necessary to fill a TCP path. This quantity also implies that the TCP throughput is limited inversely by the Round-Trip Time (RTT) or delay of the communication path given the receiver advertised window size, RWIN (which represents the amount of data that the receiver can accept without acknowledging the sender), as follows: In this paper, we demonstrate through experimental results that using multiple radios and channels solely is not enough to improve the multihop TCP throughput in an 802.11n-based WMN. We also verify through control experiments that the multihop TCP throughput drop in MRMC 802.11n mesh networks is primarily due to the large RTT of multihop wireless communication path and the high bandwidth of 802.11n, resulting in a large bandwidth-delay product, where protocol tuning or other remedies are required for achieving the peak throughput. This also explains why we found in our experiments that multihop (e.g., five hop) TCP throughput of 802.11a is similar or comparable to that of 802.11n. To illustrate how (1) limits the multihop TCP throughput of 802.11n, let us consider an example. Assume RWIN 1⁄4 16 KB 1⁄4 16 ; 384 bytes and packet size of 1,500 bytes. Then, the RWIN 1⁄4 16 ; 384 = 1 ; 500 % 11 packets. We found from our experiment that the average measured TCP data rate of 802.11n is about 80 Mbps for a single-hop wireless link. Hence, the time needed to transmit one packet at the first hop is 1 ; 500 Ã 8 = 80 s 1⁄4 0 : 15 ms. It will take 11 Ã 0 : 15 ms 1⁄4 1 : 65 ms before the source transmits all the 11 packets in a TCP window. If at this time, the ACK for the first packet still has not returned from the destination that is multiple hops away, then the TCP source will be idling, and the bandwidth at the first hop will be wasted. Hence, the RTT for the communication path must be smaller than 1.65 ms to have no bottleneck. According to our experiments, the RTT for a two-hop wireless path is more than 2 ms, and this is the reason why multihop TCP throughput for 802.11n degrades starting from the second hop (more about the results in Section 3). On the other hand, for 802.11a, the average measured TCP data rate is only about 20 Mbps. Therefore, the critical RTT that limits the throughput is larger ( 1 ; 500 Ã 8 Ã 11 = 20 s 1⁄4 6 : 6 ms). For instance, the five- hop RTT for a wireless path is about 5 ms according to our experiments. Thus, the TCP throughput for MRMC 802.11a can be maintained after traversing five hops in the network. As a quick remedy to the problem, we attempt to examine the aggregated throughput in a WMN with multiple simultaneously transmitting TCP streams. Our experimental results indicate that the use of multiple parallel TCP connections between the transmitter and receiver that are multiple hops away can better utilize the wireless bandwidth and boost the aggregated throughput. Therefore, TCP tuning techniques such as the use of parallel streams and dynamic adjustment of the advertised window based on the measured behavior need to be enabled in commercial wireless networking products (e.g., via firm- ware upgrade to enable the window scaling option) to fulfill the stringent bandwidth requirement for various real-time applications of wireless mesh network nowadays. The rest of the paper is organized as follows: Section 2 describes the experimental setup. The experimental results and discussion of the potential causes of multihop TCP throughput drop are presented in Section 3. In Section 4, we attempt to use parallel TCP streams for the transmission and measure the network aggregated throughput. Section 5 discusses proposals and research efforts on the TCP protocol that should be implemented into future wireless commercial products. Finally, Section 6 concludes the paper. Let us consider two experiments in Sections 2 and 3. One is for measuring the multihop TCP throughput for 802.11n and 802.11a networks (up to five hops). The other one is a control experiment that replaces the wireless links gradually with wired links (up to three links wired) to demonstrate and verify that the large round-trip delay for wireless path is the key factor that limits the multihop TCP throughput performance. Six mesh nodes are used in the experiments. The mesh nodes we used are the MeshRanger2-o mesh routers provided by P2 Mobile Technologies Limited [16] as shown in Fig. 1. Each of the MeshRanger2-o has two dual-band WiFi radio interfaces, in which the Atheros AR7161 chipset is used as the interface controller, capable of operating in either the 2.4- or 5-GHz frequency bands. The six mesh nodes are aligned in a chain topology (with uniform separation of 1.5 m) in an open indoor environment and operating in nonoverlapping 5-GHz channels as illustrated in Fig. 2. Note that when the high throughput (HT) operation of 802.11n is enabled, a 40-MHz-wide channel is used. So, the channel assignment needs to take this into account and needs to avoid channel overlapping and radio interference introduced by the “secondary” channel used in the HT operation. All of the nodes use a transmit power of 17 dBm (50 mW) and are equipped with two pairs of omnidirectional antennas with a gain of 5 dBi. Note that the results for 802.11n with 20 MHz-wide channel and 802.11n with single antenna are not presented here, since we found that these scenarios also suffer from the same multihop throughput degradation problem as 802.11n with MIMO and HT operation, and we have verified that it is not the 802.11n features (such as the 40-MHz-wide channel and MIMO) that undermine the multihop throughput performance. We measure the multihop data rate for 802.11a and 802.11n (with 40-MHz-wide channels used in the latter case) with RTS/CTS disabled. Specifically, the network tool, Iperf [17], is used to measure the TCP and UDP throughput for connections with different number of hops, and packets originating from node 1 are transmitted to other nodes in the linear network for the measurements. At the beginning of the experiments, the throughput for every wireless link is examined to ensure that it is working properly. For instance, we found that the normal single-hop TCP throughput for 802.11n is about 80 to 85 Mbps, while that for 802.11a is about 20 to 25 Mbps. In the experiments, the TCP advertised window size is set as 16 KB, while for UDP transmission, the offered UDP bandwidth is 100 Mbps. The ping command is used to get a rough estimation of the RTT for communication paths. According to the experimental results in [18], TCP RTT is slightly smaller than ping RTT for paths destined to light-loaded host. For paths destined to high-loaded host, it depends on an expand coefficient which is the ratio of the mean RTT to the minimum RTT. If the coefficient is small (e.g., less than 20), TCP and Ping RTT are basically the same. Otherwise, TCP RTT appears to be a bit larger. The list of experimental parameters is summar- ized in Table 1. Fig. 3 plots the TCP and UDP multihop throughput for single-channel 802.11n (channel 149), MRMC 802.11n, and MRMC 802.11a. First of all, we can see from the figure that both the TCP and UDP throughput of single-channel 802.11n (green-dotted lines) drop drastically when traversing the network, which is primarily due to radio interference and collision. For its TCP throughput, it is more than halved for every increase of hops. With the use of multiple radios and channels, the TCP and UDP throughput of 802.11a, and the UDP throughput of 802.11n can be sustained over five hops (a fluctuation of less than 10 percent could be due to the processing of routing overheads or variation of the wireless link quality and is considered to be normal). However, the TCP throughput of MRMC 802.11n (blue-solid line) still drops significantly (over 70 percent) after five hops. This result is disappoint- ing, especially to researchers that propose the use of multiple radios and channels in wireless mesh networks for boosting the overall network capacity. It is also surprising to note that the TCP throughput of 802.11a (red-dashed line) is almost equal to that of 802.11n after five hops. In general, there are several potential reasons to account for the multihop TCP throughput drop in MRMC 802.11n mesh network that are worth ...