Impact of virtual channels and adaptive routing on application performance

Research on multiprocessor interconnection networks has primarily focused o n wormhole switching, virtual channel flow control, and routing algorithms to enhance their performance. The rationale behind this research is that by alleviating the network latency for high network loads, the overall system performance would improve. Many studies have used synthetic workloads to s upport this claim. However, such workloads may not necessarily capture the behavior of real applications. In this paper, we have used parallel applica tions for a closer examination of the network behavior. In particular, the performance benefit from enhancing a 2D mesh with virtual channels (VCs) an d a fully adaptive routing algorithm is examined with a set of shared-memor y and message passing applications. Execution time and average message late ncy of shared memory applications are measured using execution-driven simul ation and by varying many architectural attributes that affect the network workload. The communication traces of message passing applications, collect ed on an IBM-SP2, are used to run a trace-driven simulation of the mesh arc hitecture to obtain message latency. Simulation results show that VCs and a daptive routing can reduce the network latency to varying degrees depending on the application. However, these modest benefits do not translate to sig nificant improvements in the overall execution time because the load on the network is not high enough to exploit the advantages of the network enhanc ements. Moreover, this benefit may be negated if the architectural enhancem ents increase the network cycle time. Rather, emphasis should be placed on improving the raw network bandwidth and faster network interfaces.