2022年5月20日学术报告——田臣

报告题目：数据中心网络拥塞控制与流控机制研究

报告人：田臣教授，南京大学

时间：2022年5月20日15:30 – 16:30

地点：计算机学科楼327

报告人简介：

田臣，南京大学登峰计划引进人才，教授，博士生导师。田臣老师在计算机网络和分布式系统领域多个顶级学术会议（含6篇SIGCOMM，2篇NSDI和1篇FAST）和知名国际期刊上录用和发表论文70多篇，包括南京大学第1篇SIGCOMM和第1篇FAST。他的工作受到了国内外研究者的广泛引用和关注，根据Google Scholar 最新学术搜索数据，迄今为止论文累计被引用约4000次。研究成果被ACM SIGCOMM，USENIX NSDI等顶级学术会议多次引用，引用者包括多位美国工程院院士，引用单位包括伯克利大学，卡耐基梅隆大学，普林斯顿大学等国际顶级名校。研究工作获得工业界广泛应用，是华为中央研究院2020年度“最佳合作教授”唯一获奖者。

报告内容：

Congestion control protocols play an important role in improving the performance of datacenter applications. Reactive congestion control protocols (RCC) have been successfully deployed in commercial datacenters for decades. Recently a new branch, i.e., proactive congestion control protocols (PCC), seems promising. A natural question is: is it really worth switching to PCC? To answer the question, we first thoroughly analyze different design choices of PCC, and discuss their pros and cons. By assembling a group of (our) preferred choices, we get OPTIMUS as a delegate for PCC, so that large-scale evaluations can be conducted to compare PCC with RCC (with DCQCN as the delegate). Results show that PCC provides stable throughput and low buffer occupancy to benefit small flows. We share our one year experience implementing different PCC design options in our private programmable NICs. Testbed results are consistent with simulations. Further, incast cannot be handled effectively by CC. We demonstrate our recent progress in flow control published in CoNEXT'21.