1 Performance Evaluation of Gigabit Ethernet & Myrinet- 2000
2 Performance Evaluation Methodology Measurement Based 3 classes of benchmarks used 1. Protocol Independent micro-benchmarks (NetPIPE) 2. Message Passing Interface (MPI) micro-benchmarks (SKaMPI) 3. Parallel applications using MPI communication (NAS PB) Each class respectively targets 1. Raw latency & bandwidth 2. Latency & bandwidth with MPI added overhead 3. Overall effect on parallel applications performance
3 Test Setup 9 nodes cluster with 1 NFS node Dell 2.2 GHz Intel Xeon 3Com 3C996B Gigabit Ethernet NIC (copper) Cisco Catalyst T Switch Myrinet 1.2 Gbps LANai9 adaptor Myrinet port crossbar switch
4 Raw Latency & Bandwidth Results Myrinet saturates at 1100 Mbps Ethernet saturates at 930 Mbps (900 Mbps with IC) Ethernet NIC Interrupt Coalescing feature reduces latency by 30 s
5 MPI Latency & Bandwidth Results TCP-MPI 50 s slower than Myrinet TCP-ED-MPI 65 s slower than Myrinet
6 NAS Parallel Benchmark Results Name format: benchmark.data size.processors Runtimes normalized to Myrinet Runtimes
7 Analysis of Results Simple micro-benchmarks show Myrinet consistently enables lower latency and higher bandwidth MPI library using TCP-ED messaging outperform or match Myrinet library version on 6/15 benchmark configurations TCP-ED outperforms Myrinet by effectively overlapping communication and computation
8 Conclusions Optimized TCP/IP can outperform raw performance of Myrinet in some cases Optimized MPI libraries with OS support can achieve better performance than MPI over user-level libraries such as Myrinet GM Gigabit Ethernet can serve as a cost-effective cluster computing solution IF aggressive TCP/IP optimizations are implemented for cluster computing