Latency Optimization In RDMA-enabled Network Through Erasure Coding

Modern data centers employ large-scale multi-tier network topology and offer higher aggregate bandwidth by creating multiple paths.To utilize this bandwidth re-quires different flows take different paths,traditional single-path transport is ill-suited.Low latency is critical for interactive networked applications.But it can be much more difficult to reduce latency—especially the tail of the latency distribution—compared to increasing capacity.Both theory and implementation have shown that redundancy is a powerful technique to reduce latency in network regardless of the bandwidth con-sumption.Based on two key insights that redundancy can reduce latency in networked sys-tem and the commercial RDMA network interface card supports EC(Erasure Coding)offloading which dramatically reduce the CPU consumption.This paper presents the design and implementation of ECoff an user-level RDMA middleware which providing portable APIs.Here are the two key contributions of this paper:(1)Designed and developed the user-level RDMA middleware ECoff which lever-aging the hardware EC offloading and the multipath in data centers to reduce latency and the tail of the latency distribution.Ecoff hides the underlying RDMA interface for upper applications to provide a socket-like portable API.(2)Designed a general method of optimizing latency through erasure coding of-floading and multipath transmission.This method can be used in any system tightly-coupled with RDMA.This method registers data memory directly and avoids local memory copying and improves the performance of transferring large blocks of data.