Task mapping in rectangular twisted tori

Abstract

Twisted torus topologies have been proposed as an alternative to toroidal rectangular networks, improving distance parameters and providing network symmetry. However, twisting is apparently less amenable to task mapping algorithms of real life applications. In this paper we make an analytical study of different mapping and concentration techniques on 2D twisted tori that try to compensate for the twisted peripheral links. We introduce a performance model based on the network average distance and the detection of the set of links which receive the highest load. The model also considers the amount of local and global communications in the network. Our model shows that the twisted torus can improve latency and maximum throughput over rectangular torus, especially when global communications dominate over local ones and when some concentration is employed. Simulation results corroborate our synthetic model. For real applications from the NPB benchmark suite, the use of the twisted topologies with an appropriate mapping provides overall average application speedups of 2.9%, which increase to 4.9% when concentrated topologies (c = 2) are considered.