TAS was originally primarily geared towards handling intensive RPC workloads with a large number of connections and small packet sizes. The current version of TAS lacks several optimizations that cater to larger flow workloads such as bandwidth-intensive data transfers, features that are notably present in Linux and other networking stacks (e.g., TCP Segmentation Offload, Large Receive Offload, etc.). The objective of this project is to investigate TAS performance with large flows, recognize its limitations, and iteratively update the design to rectify these weaknesses. Our ideal outcome is to gain insights that would contribute towards a refined design of TAS that can adeptly manage a wider variety of workloads efficiently, without compromising on performance for RPC-centric workloads.
Improving TAS Large-Flow Performance
Optimize bulk connection TCP performance while minimizing penalties for small RPC latency and throughput.