CUDA MB and ATI OpenCL AP on one host

[Frizz]

I have my host configured to accept GPU work for CUDA MB and ATI OpenCL AP. Basically this works fine.

Problem is the BOINC scheduler. Since there a way more MB tasks available, I usually and up getting lots of work for CUDA - and almost none for ATI.

Is there a way to tell BOINC that this are actually two different GPUs - and as such should have different work queues?

I am using BOINC version 6.10.58 (with Jasons mod).

What to do?

[Richard Haselgrove]

Provided you have

<plan_class>ati13ati</plan_class>

and

<coproc>
<type>ATI</type>

in your app_info.xml, you should see entries like

Requesting new tasks for ATI GPU

in your message log. Provided you ask for ATI GPU work only (the client is capable of asking for work for multiple resources at the same time), you should get work back that is flagged for the ATI card. If you are still requesting work for NVidia GPU as well, it may be that the scheduler fills those requests first, and finds it has run out of AP work in the current feeder cycle.

If you wait until the current feeding frenzy is over, and there is a steady supply of results "ready to send" of both types, we could check it out further. But I doubt we'll reach that state until the new year.

[Frizz]

Provided you have

<plan_class>ati13ati</plan_class>

and

<coproc>
<type>ATI</type>

in your app_info.xml, you should see entries like

Requesting new tasks for ATI GPU

My app_info.xml is configured as you said - still I'm seeing no lines like "Requesting new tasks for ATI GPU".

What I get is this here:


...
17/12/2010 10:48:45   SETI@home   Requesting new tasks for GPU
17/12/2010 10:48:52   SETI@home   Scheduler request completed: got 0 new tasks
17/12/2010 10:48:52   SETI@home   Message from server: No work sent
17/12/2010 10:48:52   SETI@home   Message from server: No work is available for Astropulse v5
...



Based on the status page there are always around 3,500 AP tasks ready to send.

[Richard Haselgrove]

Based on the status page there are always around 3,500 AP tasks ready to send.

They would be Astropulse v505 - the message is right, there are no AP v5, and haven't been for ? over a year.

I'll look into that "requesting" thing.

Ah - they've dumbed down the messages again. Try a debug flag, as Mike posted last night:

17.12.2010 00:47:25 SETI@home Sending scheduler request: To fetch work.
17.12.2010 00:47:25 SETI@home Reporting 7 completed tasks, requesting new tasks for GPU
17.12.2010 00:47:25 SETI@home [sched_op_debug] CPU work request: 0.00 seconds; 0.00 CPUs
17.12.2010 00:47:25 SETI@home [sched_op_debug] ATI GPU work request: 950400.00 seconds; 1.00 GPUs
17.12.2010 00:48:33 SETI@home Scheduler request completed: got 3 new tasks

[Frizz]

I posted my question in the SAH forum:
http://setiathome.berkeley.edu/forum_thread.php?id=62388