GCASR 2016‎ > ‎Presentations‎ > ‎

Cross-Layer Optimizations between Network and Compute in Online Services

Cross-Layer Optimizations between Network and Compute in Online Services

Balajee Vamanan (University of Illinois at Chicago)



Abstract:

Datacenters have emerged as the de facto computing platform for organizing and accessing most of world’s information. Modern datacenter applications that enable users to interactively query and access such large datasets (e.g., Web Search) introduce newer challenges to systems community. These applications have a partition-aggregate architecture in which each end-user query looks up data from hundreds-to-thousands of physical servers and and performs sophisticated, distributed processing at these servers. Thus, the overall latency of each individual query is highly sensitive to both the network and the compute parts of the application. Achieving high efficiency (i.e., high query throughput, low latency, and low energy) in datacenters that host these applications requires cross-layer optimization across network and architecture layers with a broader understanding of the entire system (i.e., software architecture, CPU, network, and I/O). 

In this talk, I will present two instances of such cross-layer optimization: (1) D2TCP: A transport layer protocol that leverages applications' time budgets to improve network flow completion times. (2) TimeTrader: A power governor that leverages global knowledge of the applications' time budgets to identify system-wide, critical parts of a datacenter-scale application and slows down the sub-critical parts to save energy. 


Biography: 
Balajee Vamanan is an assistant professor in the department of computer science at the University of Illinois at Chicago (UIC). He joined UIC after graduating from Purdue University in October 2015. His research straddles networking, systems, and architecture. Specifically, he is interested in the emerging networking-paradigms such as Software-Defined Networking (SDN) and Network Function Virtualization (NFV), and network optimizations for interactive, online services (e.g., Web search, social networks). He has published in top networking and architecture conferences. Before graduate school, he was a hardware designer at NVIDIA where he designed the memory controller for the first unibody MacBook Air.
 
Comments