I have made some changes to the server to add some debugging code and some other smaller changes.

Let me know if you have been given a cuda3.1 workunit and you should not have receive that.

gdf

Thank you! When should we expect the change to be fully effective? Should we wait a day to make sure any older 3.1 tasks have cleared the queue?

This will make many crunchers very happy!

---

Thx - Paul

Note: Please don't use driver version 295 or 296! Recommended versions are 266 - 285.

it's in effect now for all new requests.

gdf

On a sample of one (http://www.gpugrid.net/results.php?hostid=93580), last week's 3.1 allocation has been replaced by 4.2

It will be interesting to see if this little 420M laptop can complete it within 24 hours.

Good for now.

gdf

This task is 3.1 but should be 4.2

http://www.gpugrid.net/result.php?resultid=5564707

The problem seems to be that your machine is marked as unreliable with the cuda4.2 application, so the server decides to give the cuda3.1 one which is reliable.

I'll contact Berkeley about it.

gdf

This host also gets 4.2 tasks.

All of my GTX5xx hosts get mixture of tasks. My one GTX680 host gets only 4.2 tasks.

Still getting a mix. ie http://www.gpugrid.net/results.php?hostid=124305

Is a project reset needed following this mornings update?

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

It should not be required, but you never know.
gdf

The problem seems to be that your machine is marked as unreliable with the cuda4.2 application, so the server decides to give the cuda3.1 one which is reliable.

Could this be the result of the high error count with

ERROR: file deven.cpp line 1106: # Energies have become nan

which some people got with the cuda4.2 app?

I had several myself with my GTX 470 (host 43404). That's not a good host to generalise from, because I run it under app_info.xml, but in case it helps, here are my observations.

For over 3 months, I was running the cuda3.1 app with a count of 0.5, and tasks from other projects running alongside GPUGrid on the same GPU (see thread 2897). A few tasks failed, but no more than usual.

Then I swapped to cuda4.2 in the same configuration. The failure rate soared - to over 50%, by eye - and all errors were of the type 'Energies have become nan'.

Finally, I set count=1 in app_info (so that GPUGrid has sole use of the GPU while running, although it is swapped out periodically so other projects can run). Since making that change, I haven't had a single error.

So, perhaps, other apps in GPU memory cause a problem? I see someone else was talking about memory being a possible suspect in the news threads.

All of which leads me to suspect a buffer overflow, or use of uninitialised memory, in the cuda4.2 app. I recently helped a developer on another project pin down an error which was causing invalid data to be processed: his comments after he'd found the bug were:

I recall I always got some junk at the end of arrays (array size can be any but processing is vectorized to float4) ....

Brook build took fold buffer from CPU memory. And there was guarding zeroes written at the end to allow vectorized fetch.
Now fetch done directly from GPU buffer. And (will check this) no guard elements at buffer end.

The test which let us track that one down was: "If the host is regularly producing errors, perform a complete cold restart (to zero GPU RAM), and then allow tasks to run while avoiding any application which might load large amounts of data into VRAM" - so no games, video playback, photo editing etc. If the errors go away when VRAM is kept 'clean', that might be a pointer.

I got this error a few times, i solved it by raising the voltage a bit. Or not overclocking as much would help I would think too.

This host is also getting an mix of cuda31 and cuda42 tasks.

Hasn't had an error since i installed the GTX560 in it and started running GPUGRID again.

Not had a 3.1 task since my last post, so looking promising.

We have from now implemented a correcting suggested by David A. in the scheduler which according to him should fix the problem.

Let me know.

gdf

Any comment? Is the problem solved?

gdf

Just checked. Looks good. No new mixed tasks for me.

3 Jul 2012 | 16:41:51 UTC Thats the date of my last 31 sent. Its after your 10 oclock. But i must wait for more wus the current one is 42 but this means nothing ^^ 285gtx is slowing barely down on 42 apps so i need more time to wait :/

---

DSKAG Austria Research Team: http://www.research.dskag.at

This computer has not received any cuda 4.2 work units since updating the driver on 6/30/2012. The last one just downloaded a few minutes ago, it was cuda 3.1 also. Any suggestions. http://www.gpugrid.net/show_host_detail.php?hostid=79921
Thanks Jim