Scheduling Irregular Workloads on GPUs

This doctoral research aims at understanding the nature of the overhead for data irregular GPU workloads, proposing a solution, and examining the consequences of the result. We propose a novel, retry-free GPU workload scheduler for irregular workloads. When used in a Breadth First Search (BFS) algorithm, the proposed simple, monolithic concurrent queue scales to within 10% of ideal scalability on AMD’s Fiji GPU with 14,336 active threads. The dissertation presents an important finding that the retry overhead associated with Compare and Swap (CAS) operations is the principle reason why concurrent queues do not scale well as the number of clients increases in a massively multi-threaded environment