Friday, February 23, 2024
10:20 a.m. – 11:10 a.m. (CST)
Dr. Anand Sivasubramaniam
Distinguished Professor, Department of Computer Science and Engineering
Pennsylvania State University
ETB 1020
Title: “Revisiting Memory Hierarchy Management in the GPU Context”
Abstract
As GPUs become mainstream to take on the computational challenges of Big Data applications, it is increasingly critical that we continue to maintain a steady stream of data flowing into their plentiful computational engines, to sustain their throughput capabilities. Despite on-chip caches and high bandwidth memories, these storage capacities are just not sufficient to hold the large working sets exhibited by Big Data applications close to the computational engines. There are several reasons why this problem is not just more acute in the GPU context, but also requires a different set of solutions than those employed for this decades-old problem in the CPU context. In this talk, apart from highlighting these issues, I will present 2 recent management solutions we have implemented on today’s GPUs. The first explicitly differentiates spatial and temporal locality in pages, and places data selectively in GPU memory and/or its caches based on this differentiation. In the second work, I will show how we have extended the GPU memory hierarchy to spill beyond its memory into host memory, and even into SSDs, and develop a bypass/placement strategy for selectively placing pages.
Biography
Dr. Anand Sivasubramaniam is a Distinguished Professor at Penn State, where he has been on the faculty since 1995, immediately after completing his PhD from Georgia Tech. His research interests are broadly in computer systems, covering both hardware and software aspects towards improving performance, energy efficiencies and system reliability/availability. He has published around 300 articles in highly competitive venues, and his work has been funded by NSF, DARPA, and several industries. He is an ACM and IEEE Fellow, recognized for his work and contributions to power management of computer systems.
More on CESG Seminars: HERE
Please join on Friday, 02/23/24 at 10:20 a.m. in ETB 1020.