Extending TCP for Low Round Trip Delay

Abstract

Transmission Control Protocol (TCP) is a widespread protocol and has evolved since the very beginning of the Internet. TCP implements congestion control to hinder congestion collapse and to shareout the capacity evenly across the participating flows. Recent studies have shown that the acknowledgement clock TCP uses to control its transmission rate has a remarkable weakness which makes TCP perform not so well under shallow Round Trip Time (RTT). The current deployment of DataCenreTCP (DCTCP) in data centres and Low Latency Low Loss Scalable throughput (L4S) over broadband has helped in cutting the generally deep queue of the Internet. However, Linux TCP has a weakness which causes it to override the AQM whenever the base RTT is also shallow because of its minimum transmission rate of two segments per RTT. Shallow base RTTs is the typical environment within data centres so by overriding the AQM TCP makes the job of the AQM undoubtedly harder. TCP by going too fast brings back the once removed queue and diminishes any benefits of the low queue AQM. We propose Logarithmically Scaled Additive Increase Multiplicative Decrease (LS-AIMD) in place for AIMD to scale over shallow base RTTs. We evaluate LS-AIMD against AIMD under both Reno and DCTCP. Our initial screening shows that LS-AIMD performs under shallow RTT more than ten times better than AIMD. The gain of LS-AIMD continues to grow as base RTT continues to descend to even lower values.

Transmission Control Protocol (TCP) is a widespread protocol and has evolved since the very beginning of the Internet. TCP implements congestion control to hinder congestion collapse and to shareout the capacity evenly across the participating flows. Recent studies have shown that the acknowledgement clock TCP uses to control its transmission rate has a remarkable weakness which makes TCP perform not so well under shallow Round Trip Time (RTT). The current deployment of DataCenreTCP (DCTCP) in data centres and Low Latency Low Loss Scalable throughput (L4S) over broadband has helped in cutting the generally deep queue of the Internet. However, Linux TCP has a weakness which causes it to override the AQM whenever the base RTT is also shallow because of its minimum transmission rate of two segments per RTT. Shallow base RTTs is the typical environment within data centres so by overriding the AQM TCP makes the job of the AQM undoubtedly harder. TCP by going too fast brings back the once removed queue and diminishes any benefits of the low queue AQM. We propose Logarithmically Scaled Additive Increase Multiplicative Decrease (LS-AIMD) in place for AIMD to scale over shallow base RTTs. We evaluate LS-AIMD against AIMD under both Reno and DCTCP. Our initial screening shows that LS-AIMD performs under shallow RTT more than ten times better than AIMD. The gain of LS-AIMD continues to grow as base RTT continues to descend to even lower values.