Negotiated Coscheduling in a P2P Based HPC Platform

To promote peer-to-peer (P2P) based high performance computing (HPC), new scheduling strategies are necessarily designed to break the limitation in current systems, which only support applications with independent tasks. Therefore, we introduce a negotiated coscheduling policy, which supports applications with interactive tasks in a structured P2P based HPC platform, P2HP-2. In this policy, with the help of a project description file, a negotiated dispatching metric with two priority- entitling algorithms is proposed, which distributes tasks across computing resources. The coscheduling is further processed with two steps: spinning for dependent data and performing tasks' computation. Furthermore, a reliable message-event transmission method and a load balancer are developed to improve the performance. All our preliminary results show that this scheduling policy is efficient and fault-tolerant to support applications with task-dependent relationships.