Measuring and Analysing the Variations of Program Execution Times on Multicore Platforms: Case Study
Abdelhafid Mazouz, Sid Ahmed Ali Touati, Denis Barthou · INRIA a CCSD electronic archive server · 2010
The recent growth in the number of precessing units in today's multicore processor architectures enables multiple threads to execute simultanesiouly achieving better performances by exploiting thread level parallelism. With the architectural complexity of these new state of the art designs, comes a need to better understand the interactions between the operating system layers, the applications and the underlying hardware platforms. The ability to characterise and to quantify those interactions can be useful in the processes of performance evaluation and analysis, compiler optimisations and operating system job scheduling allowing to achieve better performance stability, reproducibility and predictability. We consider in our study performances instability as variations in program execution times. While these variations are statistically insignificant for large sequential applications, we observe that parallel native OpenMP programs have less performance stability. Understanding the performance instability in current multicore architectures is even more complicated by the variety of factors and sources influencing the applications performances.