Architectural support for efficient utilization of interconnection network resources.

摘要

Efficient and reliable communication is essential for achieving high performance in networked computing environments. Limited network resources bring about unavoidable competition among in-flight packets, resulting in network congestion and, possibly, deadlock. The effectiveness of deadlock handling mechanisms in a network often determines the upper bound in the throughput and utilization that the network can deliver.; Much research has been conducted to improve network performance by efficiently handling deadlock anomalies. However, the existence of multiple message types and associated inter-message dependencies at network end nodes may cause message-dependent deadlocks in networks that are designed to be free of routing deadlock. Most methods currently used for dealing with message-dependent deadlock require more system resources than are necessary and/or do not use system resources efficiently. This may have an adverse effect on system performance if resources are scarce.; To provide useful insight into the severity of message-dependent deadlocks, this work characterizes the frequency of such deadlocks in multiprocessor/multicomputer systems. It then proposes a handling technique based on progressive deadlock recovery and evaluates its performance. Results show that message-dependent deadlocks occur very infrequently under typical circumstances. The proposed technique relaxes restrictions on the use of network resources considerably, allowing the routing of packets and the handling of message-dependent deadlocks to be much more efficient.; One of the drawbacks of recovery-based deadlock handling techniques is that networks might suffer from significant performance degradation when they are deeply saturated. There has been a body of research done on solving this problem by preventing networks from being overloaded with packets. However, none of these provide an efficient way of accelerating the movement of packets involved in congestion onward to their destinations. This dissertation proposes a new mechanism for the detection and resolution of network congestion and deadlocks. The proposed mechanism is based on increasing the scheduling priority of packets involved in congestion and providing necessary resources for those packets to make forward progress.