Improving the performance of scientific parallel applications in a cluster of workstations

Recent improvements in LANs make network of workstations a good alternative to traditional parallel computers in some applications. However, in this platform the communication performance is over two orders of magnitude inferior to state-of-art multiprocessors. Currently networking technologies have put the pressure in the software overhead. Because of this, applications could not take advantage of this communication performance potential. In this paper, we present an implementation of Virtual Circuit Caching that reduces the software overhead by allocating communication resource once and re-using them for multiple messages to the same destinations. With this approach, the communication overhead is reduced by approximately a 35% far long messages; this reduction should enable the extensive use of networks of workstations for scientific parallel applications.