Reducing the Cost of Measuring Memory Hierarchy Communication Parameters

Abstract

For parallelizing compilers that target multi-core machines, some hardware information is very useful in producing optimized code. One approach to obtain such information is extracting the characteristics of the hardware from a set of performance data collected by performing extensive measurements on the computer system. However, such measurements tend to be heavy-weight and time-consuming, which makes them unable to complete in a reasonable amount of time, especially on a largescale system. We propose an algorithm to minimize the required measurements on a computer system using the tree structure of the memory hierarchy, that divides resources into equivalence classes so that different members of an equivalence class do not require separate measurements. Correctness arguments for the algorithms are presented, along with results that show the effectiveness and performance of the algorithms.