An Adaptive Software-Based Deadlock Recovery Technique

Deadlock management has a direct effect on making a reliable connection between processing nodes in parallel computers. Networks using wormhole switching are the most vulnerable networks to deadlock occurrence due to chained blocking nature of this switching method. Different hardware based techniques for deadlock recovery were proposed in the literature which have considerable design complexity, while deadlock occurrence in a network is rare. A software based technique can reduce this cost while preserving performance. The only software based technique proposed in the literature is static and independent of network workload and working conditions. In this paper we present an adaptive software based technique for deadlock recovery, and validate its performance in the presence of different traffic patterns including uniform, hot spot, local and first matrix transpose (FMT) patterns in 8-ary 3-cube network (torus). Simulation results exhibit about 21% 20% 98% and 20% performance improvement under local, FMT, hotspot and uniform traffic patterns, respectively.