Datacenter switching hardware trends are delivering higher bisection bandwidth but not proportionally larger SRAM packet buffer capacity. The lack of buffer space challenges congestion control protocols to combine high link utilization and minimal queuing. Additionally, protocols must ensure low latency delivery, ease of deployment, and robust handling of congestion across all possible bottlenecks. We present SIRD, a datacenter transport that fulfills all these requirements by combining receiver-driven (RD) scheduling with a sender-informed (SI) control loop that optimizes credit allocation. The combined approach, informed overcommitment, achieves both high throughput and minimal buffering because it allows receivers to efficiently allocate a very limited amount of credit. Since little credit is used, receivers face little inbound traffic which minimizes in-network buffering. As such, SIRD can deliver messages with low latency without having to reserve Ethernet priorities. In addition to sender information, SIRD integrates network feedback into scheduling decisions and explicitly deals with core congestion, a typical limitation of end-to-end RD protocols. We compare SIRD to two state-of-the-art receiver-driven protocols (Homa and dcPIM) and two production-grade reactive protocols (Swift and DCTCP) and show that SIRD is the only one that can consistently maximize link utilization, minimize queuing, obtain near-optimal latency across a wide set of workloads and traffic patterns, and never suffer from congestion collapse under high load. SIRD causes 13% less peak buffering than Homa and achieves competitive latency and utilization without requiring Ethernet priorities. Unlike dcPIM, SIRD operates without latency-inducing message exchange rounds and outperforms it in utilization, buffering, and tail latency by 9%, 44%, and 46% respectively.
翻译:暂无翻译