CSC 252 Lecture Notes - Lecture 7: Standard Performance Evaluation Corporation, Polskie Radio Program Iii, Polskie Radio Program I
Document Summary
Today"s agenda: amdhal"s law, instruction set architecture. Benchmark suites e. g. spec benchmarks (standard performance evaluation corporation) [consortium by several big companies] (~20 real world applications) Measure execution time on your hardware using your compiler, and report the results (20-30 benchmarks). Normalized: execution time: 0. 9 * as long as previous generation hardware. At this point, we realize, arithmetic mean is bogus for normalized execution times. Instead, we seek to use the geometric mean. Geometric mean = n ( n i=1 (normalized execution timei)) The fastest cpu under the geometric mean is always the same regardless of the reference machine to which execution times are normalized. Gives an upper bound on the performance improvement one can achieve from any optimization. S = 1/((1-fe)+fe/se) where s = speedup fe = fraction fo the execution time in the original (i. e. unoptimized) system when the enhancement is applicable.