Model-Driven Approach to In-Network Computing: A Case Study for Parallel Quicksort

Abstract

Abstract This article describes how Model-Driven Architecture (MDA) is applied in the development of parallel algorithms using In-Network Computing (INC). The parallel quicksort algorithm is used as a case study to demonstrate the incorporation of In-Network Computing into a specific parallel algorithm for ensuring load balance. The technique used for load balancing is intersecting prefix-sum and postfix-sum operations that determine the network split point for a good prospect of load balancing. Those two operations are done on network as an INC component of parallel quicksort algorithm. The network topology used in the algorithm is fat tree. Computations at compute nodes, INC computations at network switches, and fat tree topology are modeled together at a platform-independent level as part of MDA. Verification of the algorithm is done by translating the platform-independent model into a platform-specific model, which in this case is the simulator. The simulator can be used for both debugging and initial performance evaluation of the INC-enabled parallel quicksort algorithm.