Sangeetha Jyothi
University of Illinois, Urbana-Champaign
abdujyo2@illinois.edu
Bio
Sangeetha Abdu Jyothi is a PhD candidate at the University of Illinois at Urbana-Champaign, where she is advised by Brighten Godfrey. Her research interests lie in the areas of computer networking and systems with a focus on automated resource management in large-scale networked systems. She is a winner of the Facebook Graduate Fellowship (2017-2019) and the Mavis Future Faculty Fellowship (2017-2018). She was a research intern at Microsoft Research during the summer of 2015 and a software engineering intern at Google during the summer of 2014. Prior to UIUC, she completed her Masters at the University of Pennsylvania in 2012. She received her Bachelors at the National Institute of Technology, Calicut, India, in 2010, where she was awarded the gold medal for being the top graduating student.
Automated Resource Management in Large-Scale Networked Systems
Automated Resource Management in Large-Scale Networked Systems
Data centers (DCs) constitute a critical component in today’s Internet infrastructure, with most applications relying on DCs partly or wholly. DCs are typically multi-tenanted — the servers and the interconnecting network are shared across multiple users. In this environment, the goals of the various stakeholders are diverse. The objective of the provider is to increase revenue by utilizing the resources effectively. The applications, on the other hand, have a variety of performance requirements and time-varying demands. My research is centered around automated, closed-loop control in large-scale networked systems to simultaneously satisfy the utilization requirements of the provider and varied performance requirements of applications with dynamic loads. Towards this goal, I have worked on (a) estimating the bounds of performance achievable in a given environment and (b) designing and building automated systems to utilize the resources efficiently. We built Morpheus, a system for cloud resource management, which is currently deployed in Microsoft and open-sourced in Hadoop 2.9. Morpheus achieves high resource utilization in enterprise clusters using three main components: a learning module which accurately learns the requirements of users, a reservation module which schedules jobs based on the estimated requirements, and a dynamic reprovisioning module which adapts the allocation in real time. I also extend these ideas to other large-scale systems with distinct challenges: (i) geo-distributed micro data centers deployed by cellular providers supporting applications with stringent latency constraints and (ii) Deep Neural Network (DNN) frameworks with iterative network-heavy workloads.