Abstract:
High performance parallel computing is accomplished by dividing a large and complex problem into smaller tasks and computing them simultaneously on multiple processors. Problems that would require hopelessly long times by using serial programming techniques are solved in a speck of time by parallel programming techniques. There are three kinds of parallel computer architectures. First is shared memory architecture which includes Symmetric Multi-Processor (SMP) and multi-core processors. Second is distributed memory architecture which includes Massively Parallel Processors (MPP). Third is hybrid architecture which includes SMP or multi-core clusters. Prevalent programming models for parallel hardware include shared memory programming and message passing. Message Passing Interface (MPI) is the defacto standard for writing high performance computational software in C, C++ and Fortran. MPJ Express is a Java implementation of MPI standard. It provides thread safe communication for Java NIO and Myrinet.
In high performance compute clusters there are two types of communications. One is the inter-node communication (between processors residing on different nodes) and other is intra-node communication (between processors residing on the same node). MPJ Express has NIO device and MX device for inter-node communication using Ethernet and Myrinet as interconnects respectively. It doesn’t have any shared memory device for intra-node communication rather it involves a loop back from network card for intra-node communication. This project focuses on developing efficient shared memory communication device for MPJ Express. A goal is to achieve better communication performance than already existing devices based on Transmission Control Protocol (TCP) and Myrinet on SMPs and multi-core compute clusters. Another design goal is that existing parallel applications must not be modified for this purpose, thus relieving application developers from extra efforts of porting their applications to such modern clusters.
