Flexible hardware/software support for message passing on a distributed shared memory architecture

With the advent of multiprocessor systems on a chip, the interest for message passing libraries has revived. Message passing helps in mastering the design complexity of parallel systems. However to satisfy the stringent energy-budget of embedded applications, the message passing overhead should be limited. Recently, several hardware extensions have been proposed for reducing the transfer cost on a distributed memory architecture. Unfortunately, they ignore the synchronization cost between sender/receiver and/or require many dedicated hardware blocks. To overcome the above limitations, we present in this paper lightweight support for message passing. Moreover, we have made our library as flexible as possible such that we can optimally match the application with the target architecture. We demonstrate the benefits of our approach by means of representative benchmarks from the multimedia domain.