Cambridge Cluster
NAS Parallel Benchmarks

Jonathan Hardwick, October 1998

This page shows the performance of the Cambridge cluster on the NAS parallel benchmarks (version 2.3), which are designed to represent the kind of problems that NASA is interested in solving on parallel computers. The Cambridge cluster consists of 16 Windows NT nodes connected by both a Myrinet interconnect and switched Fast Ethernet. Each node contains two PII-300 CPUs (although only one was used for these benchmarks), 128 MB of SDRAM, and a 4 GB EIDE disk.

Summary

The table below summarizes the performance, in aggregate Mops/s, of 16 processors of various machines on the NAS parallel benchmarks (version 2.3, class A problems). The first four entries are for the Cambridge cluster using different interconnects (Myrinet versus Fast Ethernet), MPI implementations (HPVM 2.1 versus MPI/Pro 1.2.0), and compilers (Microsoft Fortran PowerStation 4.0 versus Digital Visual Fortran 5.0a). The cluster was swapping on the FT and SP benchmarks due to lack of memory. The Loki and RWC machines are Beowulf-class clusters, while the Hitachi, Cray, SGI and IBM machines are current-generation supercomputers.

To further illustrate the effect of compilers, I also show results on a single PII-300 node for the serial NAS benchmarks (version 2.3, class W problems). For all these benchmarks I used the optimization flags "/Ox /G5" for Microsoft Fortran Powerstation, and "/Ox /fast" for Digital Visual Fortran.

While the performance of the two compilers on the majority of benchmarks is similar, on the FT benchmark the DEC compiler is 1.8 times faster, whereas on the BT and LU benchmarks the Microsoft compiler is 1.3 and 1.4 times faster, respectively. Seen another way: while the DEC compiler is consistently faster on the five kernels, the Microsoft compiler is consistently faster on the three application benchmarks (BT, LU, and SP).

Full Results

For each of the NAS parallel benchmarks, and for the two classes of problem size (A and B), I give the number of nodes on which I've actually run it on a particular combination of network, MPI implementation, and Fortran compiler. If there was an unexpected failure I've listed the symptoms. Otherwise, I include a link to the results file for that run.

Note the improvement in reliability from MPI/Pro 1.1.3 to 1.2.0. However, there is a corresponding decrease in performance.