Section: New Results

Switcheable scheduling

Parallel applications used to be executed alone until their termination on partitions of supercomputers. The recent shift to multicore architectures for desktop and embedded systems is raising the problem of the coexistence of several parallel programs. Operating systems already take into account the affinity mechanism to ensure a thread will run only onto a subset of available processors (e.g., to reuse data remaining in the cache since its previous execution). But this is not enough, as demonstrated by the large performance gaps between executions of a given parallel program on desktop computers running several processes. To support many parallel applications, advances must be made on the system side (scheduling policies, runtimes, memory management...). However, automatic optimization and parallelization can play a significant role by generating programs with dynamic-auto-tuning capabilities to adapt themselves to the complete execution context, including the system load.

Our approach is to design at compile-time programs that can adapt at run-time to the execution context. The originality of our solution is to rely on switcheable scheduling, a selected set of program restructuring which allows to swap between program versions at some meeting points without backtracking. A first step selects pertinent versions according to their performance behavior on some execution contexts. The second step builds the auto-adaptive program with the various versions. Then at runtime the program selects the best version by a low overhead sampling and profiling of the versions, ensuring every computation is useful.

This work is an addition to the research directions of CAMUS related to dynamic optimization. It has been started at Paris-Sud University by Cédric Bastoul before he joined CAMUS during this year. This is an ongoing work with the PhD student Lénaïc Bagnères (GRAND-LARGE Team at Inria Saclay-Île-de-France, co-advised by Christine Eisenbeis and Cédric Bastoul). The first results have been presented in 2013 at the HiPEAC Computing System Week (http://www.hipeac.net/thematic-session/let-us-push-thread-level-speculation ) and at the Rencontres Françaises de Compilation (http://compil13.cri.mines-paristech.fr ).