TCP handles packet loss in the network by retransmitting lost packets, which in turn increases latency. Many connections in data centers are short-lived and consist only of a few packets (e.g., RPCs). Such connections suffer disproportionately from packet retransmissions. We address this issue by introducing a new transport layer protocol called ATP: ATP uses ample forward error correction at the beginning of a connection, allowing short-lived flows to recover from packet loss without retransmissions – but at the same time not congesting long-lived flows. Our experiments show that in an environment with background traffic, the latency’s 99th percentile can be reduced by a fa

Top