Safe and Effective Fine-grained TCP Retransmissions for Datacenter Communication

Summary

This paper explores the Incast Problem which is generally seen in TCP workloads in datacenter networks. The incast problem occurs when a client simultaneously requests data from many servers and can't continue until it receives all the data. All the servers respond at the same time, filling the client's inbound link, reducing overall throughput drastically.

This paper focuses on using high-resolution timers to enable more granular TCP timeouts. The authors test their solution in simulations and real-world experiments. Finally, they show that this solution is safe for wide area networks.

While explaining the background of the incast problem, the authors outline 3 main features that generally lead to TCP collapse: (1) high bandwidth, low latency networks with small switch buffers (2) clients that issue barrier-synchronized requests in parallel and doesn't issue next request until all come back from previous request and (3) servers that return small amount of data each request.

For current-day datacenters, the authors find that reducing the RTO to at least 1 ms will avoid degradation when having up to 16 servers. For next-generation datacenters, they suggest adding a randomized component to the RTO.

Finally, they investigate how well their solution would work in the wide area network. They conclude that reducing RTO to 200 microseconds doesn't affect TCP performance in the wide area network.

Criticism & Questions

I think this paper did a good job of explaining the incast problem and especially the preconditions required to go into TCP incast collapse. I think they also did a good job exploring how changes in RTO affects TCP performance. I would have liked to see more information on what else could be changed to fix the incast problem. This paper makes it seem like the timer was the biggest problem, but I'm unsure if that's the case, especially after reading the second paper.