Impact Of Thread Scheduling On Modern Gpus

The Graphics Processing Unit (GPU) has become a more important component in high-performance computing systems as it accelerates data and compute intensive applications significantly with less cost and power. The GPU achieves high performance by executing massive number of threads in parallel in a SPMD (Single Program Multiple Data) fashion. Threads are grouped into workgroups by programmer and workgroups are then assigned to each compute core on the GPU by hardware. Once assigned, a workgroup is further subgrouped into wavefronts of the fixed number of threads by hardware when executed in a SIMD (Single Instruction Multiple Data) fashion. In this thesis, we investigated the impact of thread (at workgroup and wavefront level) scheduling on overall hardware utilization and performance. We implement four different thread schedulers: Two-level wavefront scheduler, Lookahead wavefront scheduler and Two-level + Lookahead wavefront scheduler, and Block workgroup scheduler. We implement and test these schedulers on a cycle accurate detailed architectural simulator called Multi2Sim targeting AMD\u27s latest Graphics Core Next (GCN) architecture. Our extensive evaluation and analysis show that using some of these alternate mechanisms, cache hit rate is improved by an average of 30% compared to the baseline round-robin scheduler, thus drastically reducing the number of stalls caused by long latency memory operations. We also observe that some of these schedulers improve overall performance by an average of 17% compared to the baseline