Am 09.10.2011 06:22, schrieb Tom Gundersen:

Could you point me to some more info about this (I couldn't find any bug reports)?

If I remember right, if you load p4-clockmod and try to set ondemand as governor, cpufreq will refuse the to change the governor. On i686, linux even refuses to build the performance governor as a module, even if it is not the default.

I'm interested in following up on it so we can one day move to ondemand by default.

There is absolutely no reason for that.

For what it's worth, fedora and opensuse have been using ondemand for some time. Is it known why we are seeing problems and they are not?

I'm not sure if we would actually see any problems, it just seemed like a bad idea to me.

My understanding was that ondemand is recommended by upstream (http://www.codon.org.uk/~mjg59/power/good_practices.html), and that it should automatically fall back to performance if the HW latencies are too high. Please correct me if I'm wrong.

I would hope so - but at least on x86_64, the performance governor might be unavailable, so there is nothing to fall back to.