Section: New Results

Dynamic version selector

Adaptive version selection between different parallel versions of code is necessary when the execution context of a program is not known. The execution contexts includes all or some of these possibly variable parameters: the target architecture, the load of the computer at execution time, and the input data.

We have developed a framework handling loops in the polyhedral model, that is able to take a runtime decision about which version to execute. It is based on :

• the generation of different code versions of a loop nest;

• an install-time profiling to take into account the architecture parameters, that builds a parametric ranking table between the versions;

• a runtime selection, predicting the load balance and the execution time of each code version, before executing the best one.

We showed that different versions of a code are required on several polyhedral loop nest benchmarks, depending on both the target architecture and the input data. And we showed speedups compared to any statically chosen version in all execution contexts. More details are available in publication [19] .