A Cost Effective and Performance Enhanced Deadlock Recovery Scheme for Wormhole Routed Networks

Abstract

Directly connected multicomputer systems provide massively high computational capacity when connected with an interconnection network, having their own components as local memory, processor and other peripheral devices and connected with some topology as mesh, hypercube, and tree [1]. These nodes communicate with each other by exchanging messages. Each node equipped with a router which deals with communication-related implications [2]. Various switching techniques are used to route the messages from source to destination node. The wormhole routing technique has been proved as most suitable switching technique in these systems [3], but faces challenges like flow control and deadlocks (Mohapatra in Wormhole routing techniques for directly connected multicomputer systems [4]). There are two common approaches to deadlock handling: deadlock detection and deadlock recovery. The proposed work is based on deadlock recovery. This is based on the existing scheme ‘Disha’ (Anjan and Pinkston in An efficient fully adaptive deadlock recovery scheme: DISHA, pp. 201–210 (1995) [5]). Full adaptive routing is allowed without any restriction, and a deadlock detection mechanism is applied to monitor the deadlocks, and if deadlock is detected, then a deadlock recovery mechanism is in place to resolve the deadlock cycles. To optimize the cost the proposed scheme suggests, not equipping every node with additional buffer but alternate nodes in the columns of a mesh can be equipped with additional flit buffer, and blocked packets can switch to the recovery path, and eventually reach to the destination. So enhanced performance is achieved on reduced complexity and optimized cost.