I would be a little cautious about scaling processing linearly with CPU speed based on 300 MHz vs 1 GHz CPUs.
Cache size, speed and architecture can have and affect on performance. More modern CPUs have bigger, better designed caches which can enhance performance, even in excess of linear scaling, but perhaps limited to details of the code being executed. Benchmarking with real hardware is best. The highest performance applications are often action video games. CPU manufacturers actively design for that market.
Disk I/O performance hasn't kept pace with CPU performance. The fastest PCI SCSI controllers now operate at 160 Megabytes per second through the extended (32 bit) PCI bus. The fastest LVD-160 SCSI drives can deliver bursts of 160 MHz only. Some of the LVD-160 drives even have extra cache and are recommended for streaming video onto and off of the disks.
10K RPM disks have been available for several years. Manufacturers have recently introduced 15K RPM disks. I fear that these disks may be too fragile for the high altitude, low humidity environment of the pole. Icecube may have to limit itself to more robust lower performance disks. I recall hearing of significant disk failure rates at the pole. By configuring disks into certain RAID configurations one may be able to push the system performance back up near the PCI performance limit. A benchmarking program of just the right structure may be able to demonstrate sustained 100 Megabyte plus disk I/O performance, but these benchmarks don't resemble real-world computing.
1 GHz + CPUs are likely to spend time waiting for data from disk or waiting while data is being written.
One gigabit ethernet is 125 Megabyte per second. Ethernet I/O may be an important issue too.
My point is that disk I/O performance may be a better scale by which to measure system performance than CPU speed. It's harder to get good numbers on I/O performance too.
I suggest you ask someone closer to the field than for their opinion on scaling strategy.
Jerry Gerald Przybylski Phone: (510) 486 7165 FAX: (510) 486 6292 University of California - Lawrence Berkeley National Laboratory http://rust.lbl.gov/~gtp
I agree with much of what you say. My experiences are that the big uncertainty is the software and the science requirements. It is also likely that we will run much less than an optimized system.