Section: New Results

Scheduling contexts for StarPU

Scheduling context is an extension of StarPU that allows multiple parallel codes to run concurrently with minimal interference. A scheduling context encapsulates an instance of the runtime system, and runs on top of a subset of the available processing units (i.e. regular cores or GPU accelerators). In order to maximize the overall efficiency of applications, contexts can be dynamically shrunk or expanded by a hypervisor that periodically gathers performance statistics inside each context (e.g. resource utilization, computation progress) and tries to determine how resources should be assigned to contexts so as to minimize the overall execution time. We have demonstrated the relevance of this approach using benchmarks invoking multiple high performance linear algebra kernels simultaneously on top of heterogeneous multicore machines. We have shown that our mechanism can dramatically improve the overall application run time (-34%), most notably by reducing the average cache miss ratio (-50%).