Optimal additional data layers amount determining for interconnect latency hiding scheme

Abstract

The key component of parallel computing efficiency is the structure of data exchange between computing nodes. It is necessary to reduce the latency of data exchange to improve the efficiency of parallel computing. A B+2R algorithm for overlapping delays in the data exchange was offered for this purpose. Existing works do not offer a method for selecting the additionally transmitted data layer count R. We introduce the models of all systems affecting the parallel calculation time. It makes possible to apply the analytical optimization. We introduce a data transmission network and a parallel computing application models. We consider the cell calculation time is a constant value. The cell calculation time depends on the specific calculation parameters. We introduce an estimate of the computation time. Computation time depends on the additionally transmitted data layers count. Further we introduce the derivative of computation time equation. We use lowest positive real roots of the cubic equation. It’s possible that the final cubic equation hasn’t real positive roots. It’s mean that local structured grid calculation time is much bigger than network latency. In that case, it’s not recommended to use latency hiding schemes. Otherwise we recommend to use R equal to 1. Purpose of out research is to find optimal R. Optimal R value should lead to a calculation time equation minimum. The method proposed in the paper correspond to experimental result. Designed analytical model for B+2R algorithm makes possible to select optimal R value, which leads to the best calculation speedup.