Task Scheduling for NUMA Architectures
In certain scheduling policies, parallel tasks are allocated to different processing elements at the beginning of the execution. Faster processing elements, at the end of their own task queue, can steal work from slower ones. Conventional wisdom has it that stealing from a randomly chosen processing element gives the best results, but this does not take into account non uniform memory access architectures, where memory access have different costs depending on the involved processing elements.
We offer a thesis (for a single student) that aims at exploring the best work-stealing policies for NUMA architectures, depending on the memory hierarchy and its geometry.
The work requires programming in C/C++. Familiarity with OS concepts and simulators is a plus, but can be developed during the thesis.
