LASR Colloquia - Chris Rossbach/Researcher, Microsoft Research-Silicon Valley Campus, "PTask: Operating System Abstractions To Manage GPUs as Compute Devices", ACES 2.402

Contact Name: 
Jenna Whitney
Date: 
Feb 9, 2012 10:30am - 11:30am

There is a sign-up schedule for this event that can be found at

http://apps.cs.utexas.edu/talkschedules/cgi/list_events.cgi

Type o

f Talk: LASR Colloquia

Speaker/Affiliation: Chris Rossbach/Researcher

, Microsoft Research-Silicon Valley Campus

Talk Audience: UTCS Faculty

, Graduate Students, Undergraduate Students, and Outside Interested Part

ies

Date/Time: February 9, 2012, 10:30 a.m.

Location: ACES 2.40

2

Host: Emmett Witchel

Talk Title: PTask: Operating System Abstra

ctions To Manage GPUs as Compute Devices

Abstract:
Improving perform

ance under increasingly strained power budgets demands specialization at th

e hardware level. Accelerators such as GPUs and FPGAs represent promising m

odes of specialization that help address this problem, but to be a viable

solution, developing code for these specialized platforms must become more
accessible. This talk focuses on PTask, a dataflow programming framework

for accelerators that addresses this need. PTask insulates the programmer f

rom low-level detailsĖœsuch as device-management, movement of data between
disjoint memories, and asynchronous communicationĖœwhile providing good p

erformance and freeing the programmer to focus on application-level algorit

hms and issues.

PTask relies on OS-level support for dataflow primitiv

es, as well as compiler support for simplifying the programmer''s interact

ion with those primitives. This talk will consider both aspects, providing
an overview of how PTask provides OS support for accelerators, and a tour
of recent work toward compiling LINQ queries to run in PTask. Preliminary

results show that PTask can improve performance of LINQ queries up to 20x o

ver CPU-based execution, while providing the programmer with a minimally a

nnotated C# interface that is simpler and more familiar than interfaces sup

ported by current GPU frameworks such as CUDA and OpenCL.

Bio:

Chris Rossbach earned a BS in computer systems engineering from Stanford Un

iversity and a PhD in computer science from the University of Texas at Aust

in. His research interests emphasize the development of better tools for ma

naging and exploiting concurrency. Chris joined Microsoft Research, Silico

n Valley as a researcher in 2010. He currently focuses on abstractions and

programming techniques for leveraging graphics processing units to accelera

te general-purpose computing tasks.