Envelope: A New Approach to Estimating the Delivered Performance of High Performance Processors
Abstract
Simulating a computer run can be an excellent method for identifying performance bottlenecks and is especially valuable when discussing systems that do not yet exist. Traditional simulations collect a program trace and then have a simulator execute some subset of the trace one instruction at a time. Unfortunately, all of the standard variants of this technique are far too slow to use on jobs for high-end High Performance Computers and Supercomputers. We have developed an approach based primarily on an analysis of the memory access patterns and the number of floating point operations being executed that will estimate the performance of any run in a small fixed amount of time (e.g., a few seconds or less). Experience has shown that the results are nearly always within a factor of 2 of the measured results and frequently are within 15% or better of the measured results.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 01, 2002
- Accession Number
- ADA399723
Entities
People
- Daniel M. Pressel
Organizations
- United States Army Research Laboratory