Mysterious slowdown on Ubuntu 9.04

[mdettweiler]

Hi all,

A friend of mine on this forum has a Ubuntu 9.04-based crunching box which has been exhibiting some mysterious slowdowns. He's got all four cores (it's a Core 2 Quad, I don't recall the exact model number but it may be Q9550) running LLRnet on the NPLB project right now, and normally would expect each core to do a test in about 350 seconds or so. However, a quick look through the logs shows that many tests will take as long as 450-500 seconds--usually the slowdown will affect one core for a while, but then will switch to another core.

(BTW, the slowdown affects other crunching applications as well, though LLRnet was the easiest one from which to get hard data on this.)

My first guess was that there's some background application siphoning CPU time. I checked System Monitor and didn't see anything with particularly high enough accumulated CPU time in order to have been the culprit (except for vino-server, which I use to access the system and only seems to siphon off CPU time when a remote user is connected, so that couldn't have been it since this has happened during long stretches where nobody got on remotely). As such, I'm guessing it's some sort of background app that does its thing for a while then goes away, hence removing its accumulated CPU time from System Monitor.

I first thought of search indexers like beagle or tracker, but neither of those are installed. I can't think of any others like them that would possibly be on a Ubuntu installation that should be completely default in that respect.

What's really strange is that for a while my friend had just three cores running LLRnet (with the fourth idle), but the slowdown still occurred. If it's due to a phantom background application, then I would have expected the background application to automatically gravitate to the empty core and not impact the others...unless it's multithreaded or something.

Any ideas as to what might be going on here?

Max

[axn]

Have you ruled out the use of a different FFT size due to differing k values?

[mdettweiler]

Originally Posted by axn

Have you ruled out the use of a different FFT size due to differing k values?

Yes. The problem occurs regardless of specific k-value, and in fact happens with other tasks such as sieving as well.

[joblack]

Originally Posted by mdettweiler

Yes. The problem occurs regardless of specific k-value, and in fact happens with other tasks such as sieving as well.

You might have activated the 'energy saving crap' in the BIOS and/or Linux/KDE ...

[mdettweiler]

Originally Posted by joblack

You might have activated the 'energy saving crap' in the BIOS and/or Linux/KDE ...

Hmm...well, I do believe the owner of the machine deactivated CPU frequency scaling in the BIOS (which is confirmed by the GNOME frequency scaling applet saying "CPU Frequency Scaling Unsupported"), which is his standard practice for all his crunching boxes. That would seem to rule out frequency scaling as the problem.

Do you know of any other means by which the computer might be slowed down by some sort of "energy saving" application? The problem appears to still exist, so it didn't go away on its own; you're right, it is looking a lot like some sort of energy saving thing, though I can't imagine what.