Date: Thu, 21 Nov 1996 20:32:24 GMT
Server: NCSA/1.4.2
Content-type: text/html
Last-modified: Mon, 18 Mar 1996 22:50:51 GMT
Content-length: 5144
Memory Systems Research at the University of Washington
Memory Systems Research
Department of Computer Science & Engineering
University of Washington, FR-35
Seattle, WA 98195
Welcome to the home page for Memory Systems Research at UW CSE.
Description
Our research group is investigating techniques that use the operating
system to improve memory system performance. All of our work shares
the following features:
-
We rely a combination of simple hardware support and operating
system modifications to monitor the dynamic behavior of applications.
-
These monitoring mechanisms incur a small overhead at runtime, but the
information they collect can be used to identify sources of memory
system delays such as cache misses and TLB misses.
-
By identifying and resolving these bottlenecks, we not only pay for
the overhead of the monitoring mechanisms, but also significantly
improve overall system performance.
In our most recent project, we explored policies that monitor
application memory reference patterns in order to identify and resolve
TLB performance problems. Poor TLB performance results when the TLB
is too small to cover the current application's working set. Several
modern architectures support superpages: pages whose size is a
multiple of the system's base page size. On such systems TLB
performance can be improved by using larger pages, but at the cost of
wasted memory due to internal fragmentation.
We simulated several policies that adapt the page size dynamically to
different regions of an application's address space, constructing
superpages by copying the component pages to a contiguous region of
memory. We developed a policy that monitors TLB misses, and balances
the potential benefit of having a superpage (a reduction in future TLB
misses) against the cost of constructing the superpage (an in-memory
copy). By constructing superpages only when and where TLB miss
patterns warrant, this policy attains the TLB performance of large
pages without their internal fragmentation.
For more details on this project, see our paper
Reducing TLB and Memory Overhead Using Online Superpage Promotion
(ISCA '95, to appear).
We're looking for someone to implement these algorithms -- this would make
a good quals or masters project.
Project Description.
People
Faculty:
Current Students:
Papers
-
Reducing TLB and Memory Overhead Using Online Superpage Promotion .
Romer, Ohlrich, Karlin, and Bershad. ISCA '95, to appear.
-
Dynamic Page Mapping Policies for Cache Conflict Resolution on
Standard Hardware .
Romer, Lee, Bershad, and Chen. OSDI , pp. 255-266.
-
Avoiding Conflict Misses Dynamically in Large Direct-Mapped Caches .
Bershad, Lee, Romer, and Chen. ASPLOS VI, pp. 158-170.
-
A Comparison of the Memory Performance of the MIPS R3000 and DEC
Alpha 21064. Wong. Ph. D. Quals Project Report, University of
Washington.
-
Instruction Cache Effects of Different Code Reordering Algorithms.
Lee. Ph. D. Quals Project Report, University of Washington.
Ted Romer (romer@cs.washington.edu)