A technique for speeding up heavily compute bound jobs using multiple, high speed, peripheral processors

A computer program is modelled as a series of K job steps, computing on blocks of N words which are available at intervals of time T. It is assumed that one of the job steps takes too long to complete. Several conventional solutions to this problem are examined: a faster processor, a single high speed peripheral processor and multiple high speed peripheral processors. It should be noted that using a different peripheral processor for each job step would normally require (2×N×K) data transfers between the host processor and the peripheral processor in each interval T. However, if the data is transferred directly between the peripheral processors only N peripheral processor data transfers in each interval T would be needed. Note that these transfers can be up to 10 times faster than host to peripheral processor transfers. This approach is expanded and modified to include other models of computer programs.