High performance and grid computing with quality of service control
Sadiq M. Sait, Raed A. Al-Shaikh · 2014
Up to writing this paper, existing High Performance Computing (HPC) systems do not provide proper quality of service (QoS) controls and reliability features because of two limitations: first, standard middleware libraries such as Message Passing Interface (MPI) and Parallel Virtual Machine (PVM) do not provide means for applications to specify service quality for computation and communication. Second, modern high-speed interconnects such as Infiniband, Myrinet and Quadrics are optimized for performance rather than fault-tolerance and QoS control. The Data-Centric Publish-Subscribe (DCPS) model - the core of Data Distribution Service (DDS) systems - defines standards that enable applications running on heterogeneous platforms to control various QoS policies in a net-centric system. In this paper, we present our novel model of incorporating DDS QoS and reliability controls into HPC systems. Our results show that DDS integration into HPC adds considerable overheard in terms of performance and network utilization, when the application is mainly communication.