jclu52 wrote:
Like you said, you don't have to kill the completed WU. Just shutdown then start the GPU client / SMP client. The WU will be kept in the queue to be submitted.
However, comma, if you are running v6, there is a round-robin or circular queue 10 WUs long where results are kept. In my own case, I would process a p6801 WU from 171.64.65.64 every 2.2 hours, giving ~22 hours before the circular queue got back to an item.
I'm observing that since WUs aren't available from 171.64.65.64 for the fermi, the assignment server is sending my GPU client to 171.67.108.32, where it is serving up a series of 109xx and 112xx WUs. My GPU processes these in about 1.2 hours, giving the v6 circular buffer about 12 hours for the circular queue to wrap around.
When 171.64.65.64 went down, I was processing a p6801 in queue slot 9 that couldn't be uploaded. Every time the client has a completed new WU, it also tries to get rid of the old p6801. I'm now processing p109xx or p112xx from queue slot 2, so there are still 7 hours or so until the circular queue will wrap around and this WU overwritten (I think). Bad news if 171.64.65.64 isn't accepting uploads by then is that Stanford may lose the results from my computations, and I may lose the 1348 points for the P6801 WU. Good news for my particular GTX560 Ti card is that the smaller WUs seem to be more productive on my hardware (~19k ppd vs ~14,378 PPD).
v7 does better than v6 in that it will keep a pending WU upload indefinitely until it can upload. The reports that I and many others have made about stuck v7 uploads were related to uploading partial results from EUEs, not uploading full results from successful runs.