Software Knows Best: Portable Parallelism Requires Standardized Measurements of Transparent Hardware

摘要

The hardware trend of the last 15 years of dynamically tryingrnto improve performance with little software visibility isrnnot only irrelevant today, its counterproductive; adaptivityrnmust be at the software level if parallel software is going to bernportable, fast, and energy-efficient. A portable parallel programrnis an oxymoron today; there is no reason to be parallelrnif it’s slow, and parallel can’t be fast if it’s portable. Hence,rnportable parallel programs of the future must be able to understandrnand measure /any/ computer on which it runs sornthat it can adapt effectively, which suggests that hardwarernmeasurement should be standardized and processor performancernand energy consumption should become transparent.rnIn addition to software-controlled adaptivity for executionrnefficiency by using techniques like autotuning and dynamicrnscheduling, modern software environments adapt to improvern/programmer/ efficiency [1]. Classic examples include dynamicrnlinking, dynamic memory allocation, garbage collection,rninterpreters, just-in-time compilers, and debugger support.rnExamples that are more recent are selective embeddedrnjust in time specialization (SEJITS) [2] for highly productivernlanguages like Python and Ruby. Thus, the future of programmingrnis likely to involve program generators at manyrnlevels of the hierarchy tailoring the application to the machine.rnThese productivity advances via adaptivity should bernreflected in modern benchmarks: virtually no one writes thernstatically linked, highest-level-optimized C programs thatrnare the foundation of most benchmark suites.rnThe dream is to improve productivity without sacrificing toornmuch performance. Indeed, how often have you heard thernclaim that a new productive environment is now ”almost asrnfast as C” or ”almost as fast as Java?” The implication ofrnthe necessary tie between productivity and performance inrnthe manycore era is that these modern environments mustrnbe able to utilize manycore well, or the gap between highlyrnefficient code and highly productive code will grow with thernnumber of cores.rnFor industry’s bet on manycore to win, therefore, both veryrnhigh level and very low level programming environments willrnneed to be able to understand and measure their underlyingrnhardware and adapt their execution so as to be portable,rnrelatively fast, and energy-efficient.rnHence, we argue that a standard of accurate hardware operationrntrackers (SHOT) would have a huge positive impactrnon making parallel software portable with good performancernand energy efficiency, similar to the impact of the IEEE-rn754 standard had on portability of numerical software. Inrnparticular, we believe SHOT will lead to much larger improvementsrnin portability, performance, energy efficiency ofrnparallel codes than recent architectural fads like opportunisticrn”turbo modes,” transactional memory, or reconfigurablerncomputing.