Poor Elapsed Time for IO-bound Process

[Shaul Wallach]

We have obtained longer than expected elapsed times for
a benchmark test run of the ab initio molecular orbital
program GAMESS on our RS/6000 Model 530 server. The test
run makes intensive use of scratch files totaling about
500 Mb of disk space. When IBM ran the benchmark for us,
on an empty 530, CPU utilization for the job was very
good (ca. 90%). However, on our 530, while the actual CPU
time for the job was about the same, CPU utilization was
only about 50%, meaning that elapsed time was almost twice
as long. We have two 320 Mb and two 670 Mb disks, while
IBM also used a 670 Mb disk. We have 64 Mb of memory versus
32 Mb on IBM's machine. While our 530 is connected to other
work stations by NFS, all the scratch files for the job are
local. Does anybody have any suggestions?

Thanks,

Shaul Wallach