Final Report: Performance Modeling Activities in PERC2

Snavely, Allan

University of California, San Diego

关键词: Forecasting;    Performance Modeling Memory Tracing High Performance Computing;    Communications;    42 Engineering;    Simulation Performance Modeling Memory Tracing High Performance Computing;   

DOI  :  10.2172/899886
RP-ID  :  FC02/01ER2549/1
RP-ID  :  FC02-01ER25491
RP-ID  :  899886

美国|英语

来源: UNT Digital Library

【 摘 要 】

Progress in Performance Modeling for PERC2 resulted in: • Automated modeling tools that are robust, able to characterize large applications running at scale while simultaneously simulating the memory hierarchies of mul-tiple machines in parallel. • Porting of the requisite tracer tools to multiple platforms. • Improved performance models by using higher resolution memory models that ever before. • Adding control-flow and data dependency analysis to the tracers used in perform-ance tools. • Exploring and developing several new modeling methodologies. • Using modeling tools to develop performance models for strategic codes. • Application of modeling methodology to make a large number of “blind” per-formance predictions on certain mission partner applications, targeting most cur-rently available system architectures. • Error analysis to correct some systematic biases encountered as part of the large-scale blind prediction exercises. • Addition of instrumentation capabilities for communication libraries other than MPI. • Dissemination the tools and modeling methods to several mission partners, in-cluding DoD HPCMO and two DARPA HPCS vendors (Cray and IBM), as well as to the wider HPC community via a series of tutorials.