Test Method and Hardware

In order to test the efficiency of dual nodes we can employ the Netperf and Netpipe benchmarks. These two benchmarks are in the Beowulf Performance Suite (BPS) http://www.plogic.com/bps. We used three systems:

• Tyan 2762 with dual Athlon MP 1600+ (1.4Ghz) processors, 2 GByte of RAM, dual 3Com 3c59x interfaces, running kernel 2.4.20
• Supermicro P3TDL3 with dual Tualatin 1.26 Ghz processors, 2 GByte of RAM, dual Intel EEPro100 interfaces, running kernel 2.4.18
• Intel SE7500WV2 with dual Xeon 2.2Ghz processors, 1 GByte of RAM, dual Intel EEPro100 interfaces, running kernel 2.4.20

The systems were arranged so that each node under test connected was two receiver nodes. The receiver nodes were dual PIII-800 systems with Intel EEpro100 interfaces.

A simple script has been written to run the tests as follow.

1. Run Netperf and Netpipe on one interface. This configuration is intended to simulate one processes on the same node using a single interface.
2. Run two copies of Netperf and Netpipe using the same interface sending to two different receiver nodes. This configuration is intended to simulate two processes on the same node using the same interface.
3. Run two copies of Netperf and Netpipe using two interfaces sending to two different receiver nodes.

Finally, the effect of using two interfaces for the NAS test suite was studied. A small 4 node (dual PIII-800) cluster was used for these tests.