Parallel computation model logPQ

Abstract

We propose a parallel computation model LogPQ by taking account of the communication queue into the LogP model. We assume that the message length is fixed and the three parameters SQ, RQ and TQ indicate the capacities of the sending queue, the receiving queue, and the transport channel queue, respectively. Since real parallel machines avoid the overflow of communication queues by using hardware, the message queue management takes up a lot of time in every communication. A parallel computation model should take into account the flow of processor communications. The LogPQ model considers the message length, and the communication queue management depended on its physical restriction in contrast with the LogP model. First, the LogP model is analyzed to show the requirements of the processor communication. We propose the LogPQ model by taking account of the communication requirement in parallel computers. To evaluate the accuracy of the LogPQ model, we compare with the LogP model for the behavior of parallel matrix multiplication algorithm on CM5.

This research is partly supported by Grant-in-Aide for Scientific Research (B) 09480051.