Virginia Tech ECPE 6504: Wireless Networks and Mobile Computing Analysis of TCP Performance Over Asymmetric Kaustubh S. Phanse (kphanse@vt.edu)
Outline Project Goal Notions of Asymmetry in Wireless Networks and Related TCP Performance Issues Review of Certain Proposed Solutions Simulation Models and Results Conclusions and Scope for Future Work 2
Project Goal Study various notions of asymmetry typically found in wireless or hybrid (wireline-wireless) networks Analyze the implications of these different types of asymmetries on TCP performance Gain insight into certain solutions proposed to mitigate the effect of network asymmetry on TCP performance Use simulations to endorse the theoretical study and gain a further insight into behavior of TCP when used over asymmetric wireless networks 3
Notions of Asymmetry in Wireless Networks and Related TCP Performance Issues (1) Bandwidth Asymmetry TCP throughput over the forward link restricted by the normalized asymmetry factor k (defined as the ratio of the raw bandwidths to the ratio of the packet sizes in both directions) e.g. Wireless Cable Modem Network (up to 10 Mb/s downstream and 14.4 kb/s to 56 kb/s modem dial-up upstream) Relay Tower Wireless Transmitter Client Network PC PC Internet Server Internet Internet Modem Dial-up Line 4
Notions of Asymmetry in Wireless Networks and Related TCP Performance Issues (2) Delay Asymmetry Due to half-duplex links Switching between transmitting and receiving modes Finite turn-around time e.g. Metricom s Ricochet Packet Radio Network Due to hybrid (wireline-wireless) nature of networks e.g. Large delay over satellite links Impact on TCP slow start algorithm and bandwidth utilization T1 T1 T1 T1 Host 1 Host 2 Router 1 SES-1 SES-2 Router 2 5
Notions of Asymmetry in Wireless Networks and Related TCP Performance Issues (3) Media-access Asymmetry Use of polling scheme, similar to the Request-To-Send/Clear- To-Send (RTS-CTS) Variable round trip time (RTT) estimation e.g. Metricom s Ricochet Packet Radio Network Radio Interface Fixed Host ER PT Internet Internet GW ER ER PT PT MH MH: Mobile Host GW: Metricom Gateway PT: Poletop Radios ER: Ethernet Radio Stations 6
Notions of Asymmetry in Wireless Networks and Related TCP Performance Issues (4) Packet Loss Asymmetry Lossy nature of wireless channels Every packet loss causes TCP to invoke congestion control Greater packet loss ratio over the reverse link (radio unit to base station) as compared to the forward link (base station to radio unit) Attributed mainly to much lower transmission power (signal strength) at the radio unit e.g. Cellular data networks such as CDPD 7
Review of Certain Proposed Solutions (1) TCP/IP Header Compression Useful in bandwidth asymmetric networks with unidirectional data transfer Alleviates load over the slower upstream, improving TCP throughput over the forward link ACK Congestion Control Reduces number of TCP acknowledgements (ACKs) appearing over the reverse link by delayed response to received TCP data packets Useful in presence of unidirectional data transfer over bandwidth asymmetric networks and for networks with variable round-trip time (RTT) estimation (delay asymmetry) 8
Review of Certain Proposed Solutions (2) ACK Filtering Selective dropping of TCP ACKs Useful for both unidirectional and bi-directional data transfer over bandwidth asymmetric networks and networks with variable RTT estimation (delay asymmetry) Performs better than ACK congestion control scheme ACK Prioritization Useful in presence of bi-directional data transfer over bandwidth asymmetric networks May starve the reverse link data of bandwidth 9
Simulation Models and Results (1) Model A: Bandwidth Asymmetry (1) Degradation in forward link utilization for unidirectional data transfer Pronounced degradation in forward link utilization for bidirectional data transfer (Forward Link bandwidth = 400 kb/s, reverse link bandwidth = 14.4 kb/s) 10
Simulation Models and Results (2) Model A: Bandwidth Asymmetry (2) Effect of packet size variation on TCP throughput in forward direction in presence of unidirectional data transfer (Forward Link bandwidth = 400 kb/s, reverse link bandwidth = 14.4 kb/s) Forward Link Utilization (%) Forward Link Utilization vs. Packet Size (MTU) 95 90 85 80 75 88 82.67 90 90 70 65 60 70 62.67 55 50 0 500 1000 1500 2000 Packet Size (MTU) (bytes) Packet size (bytes) Normalized Asymmetry (k) Forward Link Utilization (%) 512 2.17 62.67 640 1.73 70 832 1.33 82.67 960 1.15 88 1540 0.72 90 2000 0.55 90 11
Simulation Models and Results (3) Model B: Delay Asymmetry The greater delay (aggregate delay over all TCP connections) over the satellite link (T1) results in much lesser data being transferred than that over terrestrial link (T1) 12
Simulation Models and Results (4) Model C: Packet Loss Asymmetry Greater packet loss over reverse link results in lower forward link utilization 13
Simulation Models and Results (5) Model D: Simulation of Two Proposed Solutions Improvement in forward link utilization using TCP/IP header compression technique Improvement in forward link utilization using ACK congestion control scheme 14
Conclusions (1) In bandwidth asymmetric networks, the slower reverse (upstream) channel is the primary bottleneck for the TCP throughput over the faster forward (downstream) channel. Bi-directional data transfer further exacerbates the degrading effect of bandwidth asymmetry on TCP throughput over the forward link. The large delay encountered over a satellite link results in much lesser amount of data being transferred over such a link as compared to a terrestrial link with the same bandwidth. Media-access asymmetry causes variability in RTT estimation and hence considerable variability in TCP performance. Greater packet loss over the reverse link results in inefficient bandwidth utilization in the forward direction. 15
Conclusions (2) TCP/IP header compression and ACK congestion control techniques are useful for unidirectional data transfer over bandwidth asymmetric networks. ACK congestion control scheme is also useful in improving TCP performance in networks with delay variability, however ACK filtering is proposed to perform better. ACK prioritization scheme improves forward link utilization for bidirectional data transfer over bandwidth asymmetric networks. However, this may in turn adversely affect data transfer over the reverse link. There is further scope for research in TCP performance in networks exhibiting media-access asymmetry Need to devise better techniques to improve TCP performance in presence of media-access and packet loss asymmetry 16
Summary Varied notions of asymmetry present in many popular wireless or hybrid (wireline-wireless) networks pose challenge to reliable transport protocol such as TCP Due to large-scale deployment of TCP, the study of its performance over asymmetric networks is crucial Research so far has provided considerable insight into this problem Certain techniques have been proposed to mitigate the adverse effects of network asymmetry on TCP performance Scope for further research in media-access asymmetric networks and need to devise better techniques for improving TCP performance in such networks 17