Providing a universal performance measurement framework for GPCPU
SESSION: Providing a universal performance measurement framework for GPCPU
EVENT TYPE: Birds of a Feather
TIME: 5:30PM - 7:00PM
SESSION LEADER(S):Bob Ciotti, Davin Chan
ABSTRACT: With the advent of GPGPU and Cloud Computing, current measures of large scale systems performance (e.g. Top500) are becoming less useful at providing insight into systems performance, performance trending, and ignore the emergence of Cloud Computing.
The science and engineering community that rely on General Purpose CPU (GPCPU) computing need a more reliable standard of measure. This has the potential to increase competition, leverage new technologies like Cloud Computing, reduce purchasing overheads and increase the use of scientific computing.
Problems to overcome are issues such as the lack of adoption, relevance to real-world performance, and completeness of the historical record. Further, the number and quality of submissions are often hindered by issues such as practical time constraint's, investment of staff resources to run them, physical access to a system, or willingness to participate.
This BOF will discuss alternative methods of ranking for large scale systems that might leverage existing measures including the Top500, SPEC, the NAS Parallel Benchmarks, etc. The ultimate goal to provide a relevant, timely and more universally adopted measure of systems performance for chip and system designers, system integrators, system providers, cloud providers and end users.
Panel members include representatives from Amazon, Argonne, IBM, Intel, NASA, SGI, and SPEC.
Participants will make a presentation that may discuss proposed measures, issues with current measures, or other relevant information. An open group discussion will follow.
Session Leader Details:
Bob Ciotti (Primary Session Leader) - NASA Ames Research Center