CONCLUSIONS

The computational speed and the I/O speed achieved by the Linux server (SGI 1400L) when running our benchmarks give us confidence that we can run efficiently ``production'' codes on Intel-based multi-processors running Linux as OS. The price-performances of these systems are very attractive compared with the MIPS-based SGI systems. One drawback of the Linux-Intel solution is that only four-processor systems are available at the moment, and eight-processor systems in the near future. SGI's Origin 2000 can scale up to 128 processors. However, the price-performances of the Origin 2000 is not attractive when compared with clustering several Intel based multi-processors.

We measured a slight difference in computational performances between the dual-processor Pentium and the four-processor Xeon; but probably these differences are not enough to justify the large cost differential between Pentium III processors with small caches (512 KB) and large caches (2MB). The substantial difference in I/O performance between the two systems is safely attributable to the fact that the dual-processor system was packaged as a workstation instead of as a server.

The Portland Group's F90 compiler proved to be a workable solution to compile our F90 codes. It generates efficient code, though it does not support all F90 and OpenMP features as robustly as we wished. Being a relatively new product, we are confident that the problems we observed will rapidly disappear as new versions are released.

We are grateful to SGI for loaning SEP the 1400L system that we used for our tests. We are also grateful to PGI for granting us several extensions to our evaluation licenses for the pgf90 compiler, and for giving us access to a pre-release version of their 3.3-1 release.