Running Boinc 6.10.43 on Mac SnowLeopard with Einstein@Home it recognises the GPU on an nVidia 8800 GT card, is able to run one task but then starts downloading hundreds of GPU tasks. Often it then appears to stop running further GPU tasks but will continue running non-GPU tasks.

Some questions:
- is there any way to understand why it will only run one GPU task at a time?
- why does it continue to download lots of tasks when it has not completed the existing?
- any way to debug why it is not running further GPU tasks (guessing some issue with GPU such as memory left in use)?

Any help would be appreciated.

- is there any way to understand why it will only run one GPU task at a time?

Because you only have one GPU. All shader processors inside the GPU will all attack the task at the same time, hence why GPU crunching goes that fast, but you can't run as many tasks as you have shader processors. It's one task per GPU core, just as it's one task per CPU core.

- why does it continue to download lots of tasks when it has not completed the existing?

Because of a bug in the present BOINC client. Code is in place that should prevent the download of more work for this piece of hardware if something is stuck, or the GPU is out of memory, but the code doesn't work correctly. That's fixed in newer code, but Charlie hasn't added it to any alpha client yet. (It's not in 6.10.44, for those reading as well).

- any way to debug why it is not running further GPU tasks (guessing some issue with GPU such as memory left in use)?

I suspect your GPU ran out of memory. You can check that by making a core client configuration file in your BOINC Data directory (path to that can be found in the BOINC start-up messages).

Add into it these lines:

<cc_config>
<log_flags>
<coproc_debug>1</coproc_debug>
<cpu_sched_debug>1</cpu_sched_debug>
</log_flags>
</cc_config>

When saving the file make sure it's saved as ANSI format and that the extension of the file is .xml, not something extra (like .txt).

When done, open BOINC Manager, Advanced View, Advanced menu, Read config file.
You will now see in the Messages what's going on with your GPU.

To 'resurrect' the GPU, a reboot is needed. That will clear all GPU memory.

---

Thanks for the reply and debug details.

Issue is indeed GPU memory (366Mb < 400Mb). Where is the memory leak coming from in the GPU? CUDA itself or the BOINC processing?

Edward

Where is the memory leak coming from in the GPU? CUDA itself or the BOINC processing?

BOINC doesn't do any processing, it merely manages everything for the science programs. So it's something in the science program (CUDA) that's not releasing all memory after it's done, or in the drivers, or slight hiccup. Crashing (and restarting) drivers can cause this as well.

Now, I've asked the BOINC developer for the Mac when he thinks he can build a newer version that has the newer code. He told me that he can't do that until the present version 6.10.44, has been promoted to recommended level. And that can't happen until it's been tested by more people than it's been done at present. Either there aren't that many Mac testers anymore, or those that do test it haven't posted a report yet.

If you feel you can help out, check out BOINC Alpha test instructions.

---