Title: Enhancing Data Center Networking with Multi-path Transport

时间:12月28日下午15:30

地点：东五楼210

Abstract:

Data Center has become one of the most important infrastructure in modern Internet era, hosting almost all the Internet services people use today (e.g., search, social network, E-commerce). All the data center computation and memory resources are connected through Data Center Networking (DCN), which consists of plenty of parallel networking paths. This talk focuses on how to utilize multi-path transport to enhance the throughput, latency and robustness of DCN. Specifically, I'll introduce two of our previous works:1. To achieve low TCP flow completion time in DCNs, it is critical and challenging to rapidly recover loss without adding extra congestion. Therefore, we propose a novel loss recovery approach FUSO that exploits multi-path diversity in DCN for transport loss recovery. In FUSO, when a multi-path transport sender suspects loss on one sub-flow, recovery packets are immediately sent over another sub-flow that is not or less lossy and has spare congestion window slots. FUSO is fast in that it does not need to wait for timeout on the lossy sub-flow, and it is cautious in that it does not violate congestion control algorithm. We have implemented FUSO in linux kernel and opensourced it on Github. This work has been published on UNSENIX ATC'16 and accepted by IEEE/ACM Transactions on Networking.2. RDMA over Converged Ethernet (RoCE) is becoming prevalent in DCN as it provides low latency and high throughput with low CPU overhead. However, current RDMA transmits data along a single path which is not robust against failures and is inferior to utilize the rich parallel paths in datacenters. To bring the benefit of multi-path into the RDMA world, we present a multi-path transport for RDMA, i.e. MP-RDMA. The core design challenge for MP-RDMA is to efficiently utilize the rich network paths in datacenters with minimal memory footprint. We have implemented MP-RDMA on a 40Gbps FPGA testbed. This work is going to appear on NSDI'18.

个人简介:

陈果博士现任微软亚研院网络研究组副研究员。本科毕业于武汉大学电子信息学院，博士就读于清华大学计算机系。2016年以优秀毕业生身份博士毕业，同年加入微软亚洲研究院网络组工作。陈果博士的研究方向为网络系统，近年注重于数据中心网络及异构高性能网络的研究。陈果博士迄今为止在NSDI、USENIX ATC、INFOCOM、IEEE/ACM Transactions on Networking等国际顶尖会议、期刊上发表多篇论文，同时担任INFOCOM，TON等会议、期刊的审稿人。