The International Conference for High Performance Computing, Networking, Storage and Analysis
Hierarchical Scheduling Frameworks for Heterogeneous Clusters with GPUs.
Student: Kittisak Sajjapongse (University of Missouri)
Advisor: Michela Becchi (University of Missouri)
Abstract: Graphic Processing Units have increasingly been adopted for a wide-range of scientific applications, and have become part of HPC clusters. Distributed GPU applications typically offload computation to GPUs using CUDA and OpenCL and distribute tasks through MPI and SHMEM. Despite the availability of these frameworks, coding such applications is still non-trivial. In addition, the use of batch schedulers to handle these applications on shared clusters often leads to performance and underutilization issues.
In our research, we propose the design of hierarchical scheduling frameworks consisting of node- and cluster-level runtime to support concurrent distributed GPU applications. Our proposed node-level runtime enables GPU virtualization, sharing and flexible scheduling mechanisms. The cluster-level scheduler allows administrator to define scheduling policies and configure node-level sharing. Results show that our framework outperforms existing batch schedulers while improving GPU utilization. Current work focuses on increasing the programmability of hybrid nodes while enabling effective sharing and load-balancing mechanisms.