Towards Efficient Many-Task Computing on Accelerators in High-End Computing Systems

Current software and hardware limitations prevent Many-Task Computing (MTC) workloads from leveraging hardware accelerators boasting Many Core Computing architectures. This work aims to address the programmability gap between MTC and accelerators, through the innovative CUDA middleware GeMTC. By working at the warp level, GeMTC enables heterogeneous task scheduling and 10x number of workers compared to CUDA. In order to span multiple accelerators across nodes, we have adopted the Swift parallel programming system, which can both support fine grained millisecond tasks and extreme scale supercomputers at 100K-cores. Keywords-Many-Task Computing, Swift, GPGPU, CUDA