Towards architecture independent metrics for multicore performance analysis, in SIGMETRICS Performance Evaluation Review Dec 2010. The motivation of architecture independence in this paper is often used in many papers in the IISWC community, “most performance analyses are architecture dependent, and hence insights gleaned from an application’s behavior on one platform may not apply when the application is run on another”. However, insights gleaned from a universal platform or no platform are limited. From a, likely over-simplified, signal and system point of view, an application running on an architecture is like a signal being filtered in a system. We usually use frequency spectrum to characterize the input signal. In a similar way, to get some kind of spectrum properties of an application, we should observe the response of such application running on different architectures, comparable to different frequency filters for frequency analysis. Frequency analysis of a signal is independent of the system which the signal will be input, but frequency analysis of a signal also requires some specially designed filter to get its response to a specific frequency. Frequency analysis of a signal, obtained from a set of systems, is system independent. In a similar way, workload characterization, obtained from a set of architectures, can also be architecture independent.
Why Parallel Performance Results Don’t Matter…much, by Clay Breshears. He analogizes the recent trend, in which people port applications to multi-/many-core and then report X times speedup, to the old days of proving a problem to be NP-complete.
Disclaimer: opinions expressed in this blog do not express the views or opinions of my employer.