Conventional Wisdom vs. Why Different? Quantifying the Effects Our - PowerPoint PPT Presentation

Conventional Wisdom vs. Why Different? Quantifying the Effects Our Findings • complex traffic model of Recent Protocol • focus on web performance • large range of RTTs Improvements to TCP: • two-way congestion not Web Performance • TCP SACK is better than TCP Reno ^ not clearly Michele Weigle, Kevin Jeffay, and Don Smith • RED is better than Drop Tail ^ MASCOTS 2003 clearly • ECN is better than dropping ^ 2 The UNIVERSITY of NORTH CAROLINA at CHAPEL HILL The UNIVERSITY of NORTH CAROLINA at CHAPEL HILL Evaluation Evaluation Queuing TCP data • Drop Tail • TCP Reno 4 3 2 1 X X _ high loss with bursts of X _ cumulative ACKs packets 1 1 • Adaptive RED ACKs X _ Random Early Detection _ lowers queue size data • TCP SACK 4 3 2 1 X X _ selective ACKs • Adaptive RED with ECN _ lets sender infer which _ Explicit Congestion packets were lost 1 4 Notification _ helps avoid timeouts ACKs _ marks instead of drops 3 4 The UNIVERSITY of NORTH CAROLINA at CHAPEL HILL The UNIVERSITY of NORTH CAROLINA at CHAPEL HILL

Offered Load (%) Network Setup Simulation Setup • ns-2 web web forward congested path New Requests per Second _ two-way web traffic clients servers • Bell Labs HTTP model _ 250,000 request-response pairs _ offered loads of 50-105% of 10 Mbps link 10 Mbps _ each TCP paired with each queuing mechanism web web • Main Performance Metric clients reverse congested path servers _ HTTP response time - time between sending HTTP request and receiving HTTP response • no major differences below 80% load RTTs vary from 1 ms to 3.5 seconds 5 6 The UNIVERSITY of NORTH CAROLINA at CHAPEL HILL The UNIVERSITY of NORTH CAROLINA at CHAPEL HILL Drop Tail: Reno vs. SACK Drop Tail vs. Adaptive RED 80% and 105% load 80% and 105% load 80% load crossover point No difference Tradeoffs between Reno between Drop and SACK Tail and ARED 105% load crossover point 7 8 The UNIVERSITY of NORTH CAROLINA at CHAPEL HILL The UNIVERSITY of NORTH CAROLINA at CHAPEL HILL

ARED vs. ARED+ECN Drop Tail vs. ARED+ECN 80% and 105% load 80% and 105% load 80% load crossover point Tradeoffs between ECN beats Drop Tail and dropping ARED+ECN 105% load crossover point 9 10 The UNIVERSITY of NORTH CAROLINA at CHAPEL HILL The UNIVERSITY of NORTH CAROLINA at CHAPEL HILL Our Findings Quantifying the Effects For HTTP Traffic of Recent Protocol • No benefit to using SACK over Reno Improvements to TCP: _ not enough flows can take advantage of SACK Web Performance • Complex tradeoffs exist when comparing Drop Tail and ARED (even Michele Weigle, Kevin Jeffay, and Don Smith with ECN) MASCOTS 2003 • ARED with ECN performs better than http://www.cs.unc.edu/Research/dirt/ ARED with dropping http://www.cs.unc.edu/~mcweigle/ _ drops cause retransmissions, which only increases response times 11 The UNIVERSITY of NORTH CAROLINA at CHAPEL HILL The UNIVERSITY of NORTH CAROLINA at CHAPEL HILL