Benchmarking di reti di comunicazione ad alte prestazioni: la "in-memory" network di A3Cube

This document describes the features of a new non-IP communication technology which is an element of A3Cube's Massively Parallel Data Access Platform and the benchmarks we run on it to assess its latency and throughput under the Message Passing Interface (MPI) programming model. This new architecture involves both hardware and software and is addressed to fast cluster networks and intensive communication operations. The features A3Cube has been developing for this technology allow for very low latency and first-class scalability besides high bandwidth, fault tolerance and self-healing. The architecture relies on a fully parallel design where storage, CPUs, network and co-processors are strictly linked in a fabric that allows for maximum performance in every aspect. All the project is based upon a PCI-Express card A3Cube developed and built: RONNIEE. The platform supports an experimental version of MPI-2, engineered by A3Cube. We run tests on A3Cube's network and a gigabit Ethernet network to make a first comparison. Benchmarking involved a quite traditional suite of micro-benchmarks (OSU Micro-Benchmarks) and a real-world application, namely the Lattice Boltzmann Code (LBC). The tested network was found to deliver truly high performance (low latency and high bandwidth) and exhibits very little jitter with regard to both latency and throughput (or at least it is far more stable than the Ethernet network).