Making Linux TCP Fast

We describe several major changes that have recently been made and are still underway within the Linux TCP implementation. The Linux send engine has been refactored to decouple loss recovery from congestion control, allowing the sending rate to be independent of packet loss. RACK loss recovery reworks loss detection to rest on a more general foundation of time-based analysis, instead of counting duplicate ACKs or sequence numbers. BBR congestion control sets the sending rate based on the network’s delivery rate and limits the data in-flight based on the estimated BDP, instead of reacting only to packet losses. Both algorithms are founded on a high-performance packet scheduler. The scheduler sizes the TSO bursts and releases the packets based on the rate specified by BBR congestion control, and mixes packets from different flows. These tightly-coupled but modularly-structured mechanisms together produce a low-latency and high-throughput