I like to live on "the bleeding edge" so I used to compile BOINC from SVN. Recently I though I upgrade, so I did- and everything worked smoothly...

..until today: I was intrigued by a rather heavy load on my machine, and it turned out it was boinc_client deciding to run SIX s@h apps, instead of regular two (Pentium D). As of now, it was running and swapping all of the sa@h WUs from daily stockpile crunching six at the time and swapping with another each 2 secs. Any thoughts how this might have happened? There's nothing out of the oridinary in the logs, except it "restarted" the six WU's and constantly swapping.

UPDATE: The hint for this unusual behaviour is CUDA, which suddenly discovered a compatible device (never did that before)

16-Oct-2008 20:27:21 [---] CUDA devices found
16-Oct-2008 20:27:21 [---] Coprocessor: Device Emulation (CPU) (1)

Upgrade to 6.3.14

---

lol, yup, CUDA stopped working, crunching 2 WUs again...

16-Oct-2008 20:33:09 [---] No CUDA devices found
16-Oct-2008 20:33:09 [---] No coprocessors
16-Oct-2008 20:33:09 [---] Version change (6.3.10 -> 6.3.14)

That depends on:
A) The actual GPU you have, and
B) which CUDA drivers you have installed.

If I am not mistaken, CUDA 2.0 is required. But the best place to report these 'anomalies' is on the GPUGrid forums, who do the active development and testing.

---

well, that the thing, I was aware that I had no CUDA compatible environment. The catch is that suddenly boinc_client discovered a CUDA device out of nowhere.

Still investigating how that could happen.

In 6.3.10, you mean? As now it sure didn't find a CUDA device.
On the link I gave you (to the front page of GPUGrid), you'll find links to the latest CUDA drivers as well as a list of nVidia GPUs that are supported.

---

In 6.3.10, you mean? As now it sure didn't find a CUDA device.

Well, to elaborate, with 7100 GS I can only dream of running CUDA, and was not surprised that 6.3.10 I compiled from SVN did not find any CUDA devices.

Until today that is...

But I think I found a clue why: libcudart.so that is bundled with BOINC was compiled with glibc version 2.4, while on my system- a remnant of debian etch had version 2.3.6. Yesterday, frustrated that a binary of something completely different wasn't firing up because of such reason, I upgraded libc to 2.7, and thus, during next run, libcudart.so fired up. As of why it did find a device is beyond me though.

Last week I had a similar chat with Eric Myers of Pirates@Home. He was trying to work around it with libcudart.so as well. I'll see if I can get hold of him and ask him to look in here, as I don't know whether or not he managed to fix it in the end.

---

16-Oct-2008 20:27:21 [---] CUDA devices found
16-Oct-2008 20:27:21 [---] Coprocessor: Device Emulation (CPU) (1)

I think the "device emulation" note is the hint here. The version of BOINC you had likely was for testing. It didnt' really detect a CUDA device, it just decided to simulate it. And hence your load went up dramatically. The later software didn't have the emulation turned on.

But this is just a theory.

The cutting edge is also the bleeding edge. ;-)

I too have a 7100 GS, but it wasn't detected, and I did not pursue it further.
It was a cheap card that I bought because it has composite video out, not for CUDA. If I learn anything new I'll share it here, and I'' check back to see if anyone else has any insights.

---

-- Eric Myers

"Education is not the filling of a pail, but the lighting of a fire." -- William Butler Yeats