Previous | 
Home
Section: New Results
Variability of program execution times on multicore processors
Participants : Sid-Ahmed-Ali Touati, Matias Vara Larsen, Abdelhafid Mazouz.
The activity described here represents the finalization of previous efforts conducted by Sid Touati and members of his groups, initiated before he joined the AOSTE EPI, and which are progressively merged with our own objectives, for results to be reported hopefully next year).
With the massive introduction of multicore platforms on embedded systems, parallel applications gained in performance. However, we showed in previous studies that the performance gain comes with high instability: program execution times vary in important way. We investigated the reasons for this variations and tried to understand the factors that influence program performance variability, that we decompose intro multiple families: factors from the application itself (implemented algorithms, coding technique, synchronization barriers, etc.), factors from the execution environment (OS effects, thread scheduling, Input/Ouput operations) and factors from the underlying hardware (micro-architecture, memory hierarchy, speculative execution, hardware data prefectching, etc.). Now, we have better understanding to these factors thanks to the work of two students:
Mr. Abdelhafid Mazouz who defended his PhD under the direction of Sid Touati at the university of Versailles in 11th of December 2012. The title of his PhD is "An Empirical Study of Program Performance of OpenMP Applications on Multicore Platforms".
Mr. Matias Vara Larsen, intern under the supervision of Sid Touati from February to June 2012, inside the Aoste EPI in Sophia-Antipolis, co-funded under a grant from Inria international internship program). The topic of his internship was to study the influence of he Linux kernels (multiple versions) on the stability of parallel applications.
Last, we published a rigorous statistical protocol in [21] called the Speedup-Test. It is used to analyze valid speedups (performance gain) in presence of performance instability: The Speedup-Test protocol is implemented and distributed as an open source tool based on R software. Our statistical methodology defines a consistent improvement compared with the usual performance analysis method in high-performance computing.