Method for determining the structure of a hybrid computing system

摘要

A method comprises measuring the execution time T1 for a problem to be solved with a program being run by a single processor, measuring the execution time TM and TS of MIMD and SIMD program fragments being run by a single processor and a single accelerator correspondingly, determining the specific acceleration ρ of the execution time for an SIMD program fragment being run by a single accelerator in comparison with the execution time for the fragment being run by a single processor, determining a portion of the execution time for an MIMD fragment being run by a single processor and a portion of the execution time for an SIMD fragment being run by a single processor and adjusting the quantity of processors or accelerators comprised in a hybrid computing system structure according to the data obtained.

主权项

1. A method for determining the structure of a hybrid computing system including an Multiple Instruction, Multiple Data (MIMD) component containing at least a single processor and a Single Instruction, Multiple Data (SIMD) component containing at least a single arithmetic accelerator, wherein the method comprises measuring the execution time T1 for a problem to be solved with a program being run by a single processor, measuring the execution time TM and TS of MIMD and SIMD program fragments being run by a single processor and a single accelerator correspondingly, determining the specific acceleration ρ of the execution time for an SIMD program fragment being run by a single accelerator in comparison with the execution time for the fragment being run by a single processor and adjusting the quantity of processors or accelerators comprised into a hybrid computing system structure according to the data obtained, determining a portion φ of the execution time for an MIMD fragment being run by a single processor and a portion 1-φ of the execution time for an SIMD fragment being run by a single processor relative to the execution time of the program being run by a single processor; the ratio of the execution time portion for an SIMD fragment being run by a single processor and the execution time portion for an MIMD fragment being run by a single processor is compared with the specific acceleration value ρ of the execution time for an SIMD fragment being run by a single accelerator in comparison with the execution time for an SIMD fragment being run by a single processor, wherein forρ>1-φφthe quantity of MIMD component processors is being increased, and forρ<1-φφthe quantity of SIMD component accelerators is being increased.