An applications approach to benchmarking and performance modelling low latency interconnection networks

As the field of High Performance Computing (HPC) approaches the Exascale era we see larger systems coming online with a rich set of applications and programming paradigms given the diverse system architecture employed to deliver petascale levels of performance. Underpinning these distributed applications is the use of interconnected nodes; something which can contribute to significant performance degradation when a machine is highly utilised. This thesis examines the interactions between communication patterns commonly seen in distributed applications written on top of Message Passing Interface (MPI), with a benchmark framework (StressBench) designed to orchestrate concurrent communication patterns. Application replay through StressBench yields reproducible applications with in 15% difference in runtime, showing that it provides a useful; abstraction from commercially sensitive production applications. A congested workload is distributed across two supercomputers demonstrating a slow down for application and Input Output (I/O) traffic, and the effects of job placement on I/O and application traffic is investigated with the benchmark framework. This thesis documents a validation methodology for a layered simulator built on top of Structural Simulation Toolkit (SST). Using the validated hardware platforms accurate performance models are developed for four systems, and two applications on top of the SST which are then used to evaluate future network designs, both to support the development of next generation interconnection networks and design responses for an Request for Proposal (RFP)