Work Scheduling on GPU and Phi
On MIMD parallel architectures general efficient scheduling methods have been devised. Among these are static threads, thread pools, and work stealing. For more specialized applications, such as combinatorial optimization problems such as puzzles and games, specialized scheduling methods have been found to work even better. One such method is Transposition Driven Scheduling, TDS in short. In TDS the data is not brought to the jobs, but the jobs are executed where the data is. This results in a decoupling of the algorithm, less data transport, with very good speedups for certain applications.
Parallel architectures, however, have become more exotic. Some of the hottest hardware of the last few years, GPUs, are more SIMD in nature, and Intel’s latest, the Phi, is even more heterogeneous. TDS was devised in 1999. The topic of this thesis project is to see if TDS, or a TDS-like scheduling algorithm, can be devised for GPU and Phi.
It is an open problem if one of the fastest specialized scheduling methods, TDS, can be adapted for GPU and Phi. Both success and failure may result in exciting, publishable, results.
Information: aske.plaat at gmail.com