I've run a couple of benchmarks using different numruns on the same hardware, calculated the variations and deviations, and found that numruns should be 4 or more on my current setup [P3, linux-2.2.14-Raid1, mem=16M, runlevel 1, single AHA2940UW1, two Quantum Atlas IV] to keep variance down enough. From numruns from 1 to 4 I get relatively high variances which usually means the numbers aren't solid enough. numruns about 4 get very low variances ( all in all, about 0.005 ) and to me seem solid enough. Anyway, I'm posting the benchmarks I've done over the last week [with numruns=1, so I'm not following my advice], in the next few posts. I'll post a summary of what I think I've obtained - which isn't very much, because of the low numruns - but you can look at the numbers and judge for yourselves. Cheers, -- _/\ Christian Reis is sometimes [EMAIL PROTECTED] \/~ suicide architect | free software advocate | mountain biker