Can anyone explain this? I ran CPU benchmarks with BOINC v5.8.16 on Windows 2000 and got:
1095 Floating-point MIPS
1849 Integer MIPS

Using the exact same hardware, I ran the same test with BOINC v5.8.15 on Fedora kernel v2.6.17 and got:
1033 floating-point MIPS
1512 Integer MIPS

Why is the floating point performance 22% lower in Linux?

The only possibility I could find was on the difference of BOINC versions.

I found this on the changelog from version 15 to 16:
http://boinc.berkeley.edu/rev_history.php

\"Set processor affinity mask for benchmark threads (Windows)\"

It\'s something to do with how the BOINC benchmarks work, and has been discussed on the BOINC/dev boards.

The last time this was asked about was about 2-4 weeks back.
Search option at the bottom of the page.

I guess my question is something of an academic one. As long as the projects are running fast, the benchmark doesn\'t <i>have</i> to be accurate. But I know Boinc uses these to estimate time to completion and scheduling priority.

Here\'s an interesting test I did. I booted OpenSUSE 10.2 and ran both Boinc versions 5.8.15 and 5.8.17. Version 17 included some \"benchmark tweaks\" and also uses glibc2.4. (I don\'t have glibc2.4 on my regular Fedora install.) Version 5.8.15 under OpenSUSE ran faster! Here\'s a table of my tests:

Boinc Windows Fedora 2.6.14 SUSE 2.6.18
---------------------------------------------------
5.8.15 n/a 1033f; 1512i 1021f; 1606i
5.8.16 1095f; 1849i n/a n/a
5.8.17 n/a n/a 1186f; 2623i

The integer calculation is vastly different between Boinc versions and statistically significant between these two Linux distros on the same Boinc version. I wonder if updating my glibc on Fedora will help integer performance of the models?

New posts at BOINC/dev on the problem
here.