Analyzing Effect of Network Processor´s Cache Dependent Parameter on MPI Broadcast Performance

For cluster based parallel computing, optimization at every layer of Network Protocol to improve the performance is an ongoing process. Use of dedicated Network Processor has become good choice to cope up with continuous and rapid growth of network link speed. GPP or RISC can be used for packet processing, because of their higher processing speed and flexibility. To improve overall performance, it is necessary to program Network Processor by making efficient use of its resources. In our experiment, we have formed Hot Connection pool that is heavily dependent on Network Processor cache. Our experimental result shows performance improvement when pool value is chosen proportionate to cache size of Network Processor.