Gamma-ray pulsar search CREDIT
10 Aug 2014 3:06:25 UTC
Topic 197674
(moderation:
Howdy folks,
This GPU task takes much longer than others yet gives the fewest points.
Can we adjust the points up a bit please to sort of match the GPU time involved?
Completed and validated 3,088.53 880.95 9.85 1,000.00 Binary Radio Pulsar Search (Arecibo, GPU) v1.39
Completed and validated 6,104.52 1,935.13 21.63 3,333.00 Binary Radio Pulsar Search (Perseus Arm Survey) v1.39
Completed and validated 7,558.83 7,389.71 82.58 660.00 Gamma-ray pulsar search #3 v1.11
These all on the same GTX-580 setup...
8-)
Gamma-ray pulsar search CREDIT
)
You mention "GPU time" but none of the figures you quote measure GPU time - it's probably quite minimal for this initial version of the FGRP3 app.
Take a look at this post which announced the start of the FGRP3 run and the provision of a new app for partial use of a suitable GPU. In particular, note the following comment :-
If you look at both elapsed time and CPU time for the example you give, you can appreciate how relatively large the CPU involvement still is - less than 200 secs below the elapsed time. It would probably be fair to call these essentially "CPU tasks with a bit of GPU involvement.
As mentioned, the app will be refined over time to shift as much as possible of the remaining CPU calculations to the GPU. When that is done, I would expect both CPU time and elapsed time to decrease dramatically. As was done with BRP apps when they matured, you may even see multiple tasks being combined into one large task with appropriate increase in the awarded credit. At that point you could make a much better assessment of the fairness or otherwise of the credit allocation.
Cheers,
Gary.
Well, I didn't pay much
)
Well, I didn't pay much attention to the CPU time as the GPU time still looked like a lot..
But, what you say makes sense.
Thanks!
8-)
RE: Thanks! You're
)
You're welcome!
I should have made it clearer that the first mentioned time (elapsed time) is not a good measure of GPU involvement as the GPU is idle for those periods when waiting for CPU stuff to be done. You can see this by looking at GPU load which is quite low a lot of the time and spikes upward when the GPU is actually doing some work.
For the more efficient BRP apps, GPU load is more uniformly high so the elapsed time is a better measure of the GPU involvement.
Cheers,
Gary.