Date: December 5, 2019
Time: 2:30-3:30pm ET
Network transport performance is playing an important role in today’s and future cloud computing. As the increasingly diversified applications and the exponentially growing amount of data meet the end of Moore’s law, we have to rely on more servers and more special-purpose hardware devices to complete a task. Networking is unprecedentedly critical in blurring the lines between individual servers and devices, by providing ultra-low latency and high throughput at a large scale. However, traditional transport layer designs face challenges because they constrain their design spaces mostly within individual servers. In this talk, I will show that by breaking the unnecessary architectural boundaries in datacenters, we can leverage the strength of components beyond individual servers, to scalably and efficiently diagnose and optimize the transport performance. I will give two examples in our work. First, I will present DETER, a deterministic replay tool for detailed and efficient TCP diagnosis, enabled by co-designing servers and the network controller. Second, I will present HPCC, a high-precision congestion control scheme that enables raw hardware transport performance at a large scale, by leveraging the detailed link load information at switches and the fast rate adaptation at servers.
Yuliang Li is a PhD student in Computer Science at Harvard. He is advised by Prof. Minlan Yu. His research focuses on improving datacenter network performance and availability by leveraging the opportunity of co-designing diverse hardware or software components. Prior to Harvard, he received his Bachelor’s degree from Tsinghua University.