Many-core graphics processing units have been used to accelerate a wide variety of applications and are increasingly a part of high performance computing clusters. Despite the rich body of literature on job scheduling and on virtualization of CPU clusters, resource management frameworks and cloud computing services for heterogeneous clusters that include many-core devices are still in their infancy. The project targets this problem and proposes the design of a set of virtualization and scheduling technologies to allow for the efficient use of heterogeneous clusters that include general purpose processors and many-core coprocessor devices. In particular, the proposed research focuses on the following aspects and on their interactions. First, many-core sharing is explored as a means to improve system utilization while keeping infrastructure costs low. Second, scheduling mechanisms targeting both single- and multi-node applications accelerated using coprocessor devices are proposed. Third, memory unification technologies are proposed in order to hide from the users the complexity of the heterogeneous memory system. Finally, diverse software stacks are integrated into a coherent runtime system.
The results of this research will be disseminated through publications and conference presentation. In addition, open-source software modules will be released and made available on the PI?s Lab website. The knowledge generated by this research will allow creating instructional material to teach cluster and cloud computing at both the undergraduate and graduate levels. Undergraduate student involvement will be promoted leveraging the University of Missouri Undergraduate Research Program and a partnership with Truman State University and Lincoln University.
