The alpha farm is a general purpose machine, used by many people. For most cases, this is a good thing, but for getting exact I/O measurements it is not. We have to try to second guess the disk cache to see true disk performance, and use huge files (300MB). In the case of the micro-benchmarks, there simply is not enough disk space to allow every reader to process 300MB. More disk would help aleviate this problem, and some mechanism to turning off the disk cache would help as well.
Another problem is that large 8 node jobs must be run at night, while at the same time system administration processes and backups may be running. All disks are crossmounted, and the honor system is used for controlling access to reserved nodes. While very flexible, this allows for many events to skew the timings. The machine should come off the network and used exclusively for timing runs.