Title: Informed Caching Environments
Speaker: Alvin R. Lebeck

Abstract:
As the importance of memory hierarchy performance increases,
techniques for improved cache management become ever more important.
In this talk, I present our ongoing efforts to develop an Informed
Caching Environment (ICE).  We are focusing on three key aspects
required to implement such an environment: 1.) What information can
be obtained from the program, 2.) How the information is passed to the
memory hierarchy, and 3.) How the memory hierarchy exploits the
information for improved performance.
  
In the first part of this talk, I show that
exposing cache management mechanisms to control cache block size or
replacement decisions can decrease the execution time of three media
benchmarks (epic, pegwit, ijpeg) between 11% and 17% on a 4-issue
dynamically scheduled processor.  These improvements are contingent on
the existence of a low-overhead mechanism for software to convey
information to the memory hierarchy.

The second part of this talk presents a technique to quantitatively
evaluate load latency tolerance in dynamically scheduled processors.
We define load latency tolerance as the amount of time a load can be
outstanding without degrading processor performance with respect to an
ideal memory system, where all loads complete in one cycle.
Our results show that a significant number of all loads miss in the level one
cache yet have latency demands lower than second level cache access
times. We also show that a similar percentage of loads hit in the
level one cache even though they possess enough latency tolerance to
be satisfied by lower levels of the memory hierarchy.


Biography:
Alvin Lebeck is an Assistant Professor of Computer Science and of
Electrical and Computer Engineering at Duke University, where he is
co-director of the NSF and DARPA funded TUNE project.  His research 
interests include
hardware and software techniques for improved memory hierarchy
performance, efficient messaging systems for distributed operating
system services, and energy efficient computing.

Lebeck received the BS in Electrical and Computer Engineering, and the
MS and PhD in Computer Science at the University of
Wisconsin---Madison.  He is the recipient of a 1997 NSF CAREER Award and
a member of ACM and IEEE.

-----------------------------------------------------------------------
Alvin R. Lebeck                         Office  D304 LSRC
Assistant Professor                     Phone   (919) 660-6551
Dept. of Computer Science               FAX     (919) 660-6519
Box 90129                               e-mail: alvy@cs.duke.edu
Duke University                         http://www.cs.duke.edu/~alvy
Durham, NC 27708-0129
-----------------------------------------------------------------------