A Cloud-Optimized Transport Protocol for Elastic and Scalable HPC
Leah Shalev, Hani Ayoub, Nafea Bshara, Erez Sabbag
Abstract
Amazon Web Services (AWS) took a fresh look at the network to provide consistently low latency required for supercomputing applications, while keeping the benefits of public cloud: scalability, elastic on-demand capacity, cost effectiveness, and fast adoption of newer CPUs and GPUs. We built a new network transport protocol, scalable reliable datagram (SRD), designed to utilize modern commodity multitenant datacenter networks (with a large number of network paths) while overcoming their limitations (load imbalance and inconsistent latency when unrelated flows collide). Instead of preserving packets order, SRD sends the packets over as many network paths as possible, while avoiding overloaded paths. To minimize jitter and to ensure the fastest response to network congestion fluctuations, SRD is implemented in the AWS custom Nitro networking card. SRD is used by HPC/ML frameworks on EC2 hosts via AWS elastic fabric adapter kernel-bypass interface.