Departmental Papers (CIS)

Date of this Version

June 2005

Document Type

Conference Paper

Comments

Copyright 2005 IEEE. Reprinted from Proceedings of the 32nd International Symposium on Computer Architecture 2005 (ISCA 2005), pages 322-333.

This material is posted here with permission of the IEEE. Such permission of the IEEE does not in any way imply IEEE endorsement of any of the University of Pennsylvania's products or services. Internal or personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution must be obtained from the IEEE by writing to pubs-permissions@ieee.org. By choosing to view this document, you agree to all provisions of the copyright laws protecting it.

Abstract

Pre-execution removes the microarchitectural latency of "problem" loads from a program’s critical path by redundantly executing copies of their computations in parallel with the main program. There have been several proposed pre-execution systems, a quantitative framework (PTHSEL) for analytical pre-execution thread (p-thread) selection, and even a research prototype. To date, however, the energy aspects of pre-execution have not been studied.

Cycle-level performance and energy simulations on SPEC2000 integer benchmarks that suffer from L2 misses show that energy-blind pre-execution naturally has a linear latency/energy trade-off, improving performance by 13.8% while increasing energy consumption by 11.9%.

To improve this trade-off, we propose two extensions to PTHSEL. First, we replace the flat cycle-for-cycle load cost model with a model based on a critical-path estimation. This extension increases p-thread efficiency in an energy-independent way. Second, we add a parameterized energy model to PTHSEL (forming PTHSEL+E) that allows it to actively select p-threads that reduce energy rather than (or in combination with) execution latency.

Experiments show that PTHSEL+E manipulates preexecution’s latency/energy more effectively. Latency targeted selection benefits from the improved load cost model: its performance improvements grow to an average of 16.4% while energy costs drop to 8.7%. ED targeted selection produces p-threads that improve performance by only 12.9%, but ED by 8.8%. Targeting p-thread selection for energy reduction, results in "energy-free" pre-execution, with average speedup of 5.4%, and a small decrease in total energy consumption (0.7%).

Share

COinS
 

Date Posted: 25 February 2006