Optimizing the concurrency level of network artificial intelligence computations

Abstract

Recent achievements in development of artificial intelligence systems would be impossible without the use of high performance distributed computing platforms. This paper presents a graph model of concurrent network schedules and a technique that estimates the execution time and implementation cost over maximum weight cliques in task graphs. It proposes an algorithm for recalculating clique sets after changing the concurrency level of a schedule by adding an edge to the concurrency graph and removing the edge from the complement graph. Since the set of pairs of concurrent tasks has been found, it treats a schedule existence problem as solving a combined logical equation. The proposed model and technique are a basis for the development and implementation of network algorithms.