CS Talk - Rohan Gandhi
Title: Improving cloud load-balancer infrastructure for online services.
Host: Minlan Yu
Load-balancer is a foundational building block for the online services to support billions of users and connections. Traditionally, load balancing is done either through specialized hardware or software. In this talk, I will first show that these traditional approaches suffer from glaring limitations including poor performance, high cost and poor availability at cloud scale. I will present Duet that addresses these limitations to build a high performance and ultra-low cost load balancer using a hybrid architecture consisting of commodity hardware and software. I will also present Yoda, which provides high availability using a novel principle of decoupling the TCP state so that any server can take over any TCP connection at any time. Together, these systems meet all the key requirements for load balancing at cloud scale.
Rohan Gandhi is a software engineer in the Microsoft Azure networking group. Previously, he completed his Ph.D. at Purdue University advised by Prof. Y. Charlie Hu. His research work focuses on improving middlebox designs, big data analytics systems and network management systems. He has published multiple papers at ACM SIGCOMM, ACM Eurosys and USENIX ATC.