> Yes, if there's an interface to get the information from Xen into (for
> exampl) Dom0, then we'd be able to use that for the userspace daemon,
> and then use a kernel module for Dom0 (or some other Dom, but it makes
> sense that it's a privilige domain handling this).
Yep. It'd be nice if we could present the stats to the standard cpufreqd so
that it does the right thing despite the multiple domains on the machine.
Otherwise, we'd need to have a Xen-aware cpufreqd, although that shouldn't be
too hard.
> Yes, I'm pretty sure that there's the Cool'n'Quiet option on AMD
> Opterons too (but I'm a bit removed from the products side, so I
> wouldn't know exactly).
OK, I thought there might be.
> problems with PowerNow/Cool'n'Quiet (and probably in Intels solution
> too, but I wouldn't know for sure, as I haven't worked with that
> technology) is that it's somewhat costly to change the frequency, so you
I suspect it's still an issue. They got the resync times down quite a bit in
the Dothan Pentium M but that's a whole lot removed from their server chips.
> However, if you have a server sitting doing some work that is based on,
> for instance, the working day of your employees, then it would make
> sense to lower the CPU-speed overnight. A multinational mail-server that
> receives mail spread over the whole 24 hour period probably don't get
> much gain from this, as it would probably be roughly at the same
> processing level all throughout the 24 hour period [and if it's only
> running 10% performance for 90% of the time, you'd be shuffling some
> other load onto that machine, right?]
Yep, indeed. CPU freq scaling will need doing at some stage but as you
rightly point out, you shouldn't have so many idle machines if you can
migrate workloads around the cluster.
Cheers,
Mark
> --
> Mats
>
> [snip]
_______________________________________________
Xen-users mailing list
Xen-users@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-users
|
Home