HPCSA Seminar Series

Dr. Daniel Wong

UC Riverside

Faculty Host: Hyeran Jeon

Date: 05/11/2021, Tuesday

Time: 11:00 AM - 12:00 PM

Zoom: https://ucmerced.zoom.us/j/86287104866?pwd=VnV3Z20xaWtRenhCbjdUdU8rc1M3QT09

Title: Emerging communication- and power-centric architectural challenges of cloud microservices

Abstract: The complex, distributed nature of data centers have spawned the adoption of distributed, multi-tiered software architectures, consisting of many loosely-coupled fine-grain microservices. These microservices present many unique efficiency challenges in modern data centers. In this talk, we will discuss two challenges that emerging cloud microservice workloads present on energy efficiency and communication efficiency.

First, due to the fine-grain nature of microservices, service times are typically in the order of microseconds. These “killer microsecond” service times can cause state-of-the-art dynamic power management techniques to break down and become ineffective. We propose μDPM, a dynamic power management scheme for the microsecond era that carefully coordinates request delaying, per-core sleep states, and voltage frequency scaling. μDPM reduces processor energy consumption by up to 32% and consistently outperforms state-of-the-art techniques by 2x.

Second, many emerging microservice workloads are increasingly accelerated by GPUs, or even multiple GPUs, and can exhibit significant inter-accelerator communication. However, existing resource allocation policies are agnostic to the communication requirements of jobs, resulting in fragmented resource allocations that cause degraded bandwidth allocations. We propose MGAP, a multi-GPU allocation policy that defragments resource allocation and improves inter-accelerator communication.

Bio: Daniel Wong is an Assistant Professor in the Department of Electrical and Computer Engineering at the University of California, Riverside. His research spans Computer Architecture, High-Performance Computing, and Datacenter-scale Computing. His current research work focuses on improving the power, performance, and programmability efficiencies of accelerator-based computer systems. His work on data center energy efficiency was recognized as an IEEE Micro Top Picks from the Computer Architecture Conferences in 2012 and a recipient of the NSF CAREER award in 2021.

Daniel completed his PhD in Electrical Engineering, MS in Electrical Engineering, and BS in Computer Engineering and Computer Science, all at the University of Southern California. In the past, he also worked on cyber-security and non-volatile memory management at Lawrence Livermore National Labs and emerging non-volatile memory architecture at Samsung Semiconductor, Inc. R&D. Additional information about Daniel can be found on his webpages: danielwong.org.

For more information, please contact Hyeran Jeon: hjeon7@ucmerced.edu