I have had two 8800GT's crunching for about 3.5 months. That always downloads 4 WU's -- 2 that crunch immediately and 2 in waiting. Given our recent work issues, I figured I throw in a 9500GT (factory OC'd) that I had laying around, expecting to get 6 WU's. The 9500GT is slow, but can finish 1 WU with plenty of time before the deadline and the 2nd WU would get crunched by one of the 8800GT's. However, after getting the 9500GT in the system, it crunches 3 WU's but only has 1 in reserve (same # of WU's before the 9500GT).

It definitely identified the 3 CUDA devices:
12/31/08 00:37:06||CUDA devices: GeForce 8800 GT, GeForce 8800 GT, GeForce 9500 GT


It is also definitely trying to get more work every few minutes, but running into the message that I am already saturated with WU's. Any ideas?

12/31/08 07:47:14|GPUGRID|Sending scheduler request: To fetch work. Requesting 130899 seconds of work, reporting 0 completed tasks
12/31/08 07:47:19|GPUGRID|Scheduler request completed: got 0 new tasks
12/31/08 07:47:19|GPUGRID|Message from server: No work sent
12/31/08 07:47:19|GPUGRID|Message from server: (reached per-CPU limit of 1 tasks)

...

12/31/08 08:23:32|GPUGRID|Sending scheduler request: To fetch work. Requesting 130899 seconds of work, reporting 0 completed tasks
12/31/08 08:23:37|GPUGRID|Scheduler request completed: got 0 new tasks
12/31/08 08:23:37|GPUGRID|Message from server: No work sent
12/31/08 08:23:37|GPUGRID|Message from server: (reached per-CPU limit of 1 tasks)

Taking a closer look at the server message, I see it specifically says "per-CPU limit". I have a Q6600, so that must be where the 4 tasks for 4 CPU's comes from. However, given I can crunch 3 at a time now, can this limit be raised to 2? That will let a Q6600 download 8 WU's, which even with just 2 8800GT's I can crunch in a little under 3 days.

This might also help others with speedier cards like 2xx series that can crunch 4 WU's / day. I'd expected a Q6600 with a 2xx card to download a max of 4 WU's at a time, which is only 1 day's worth. With the per-cpu limit at 2, the 2xx's would have 2 days worth of WU's available at all times.

Ideally, this would be tied to the # of graphics cards, not the number of CPU's. But, I understand GPU-processing is in its infancy and I can't expect the world. However, the suggestion above for a 2 WU per cpu limit might relieve a lot of WU issues we are currently experiencing.

I think the idea was to go straight to a per-GPU limit. The main problem with increasing the per-CPU limit is that the BOINC work fetch (in the 6.4.x clients) and the estimated WU completion times are still screwed. BOINC downloads too much work on slower cards, which has to be aborted by the user.. this only gets worse if more work per CPU is allowed.

MrS

True. Ideally it will be tied to # of GPU's and be able to determine the estimated time for a particular GPU. That's a wish that isn't probably that close to being fulfilled.

I have been able to babysit the work -- sometimes I still get the "not available" message, but I generally have been able to keep all three GPU's crunching. If the fetch issues (and 24 hour connection delay) are fixed, I think I will been fine. I can keep an eye on it until that day comes.

True. Ideally it will be tied to # of GPU's and be able to determine the estimated time for a particular GPU. That's a wish that isn't probably that close to being fulfilled.

I have been able to babysit the work -- sometimes I still get the "not available" message, but I generally have been able to keep all three GPU's crunching. If the fetch issues (and 24 hour connection delay) are fixed, I think I will been fine. I can keep an eye on it until that day comes.