View Single Post
Old 2015-10-08, 14:38   #374
If I May
chalsall's Avatar
"Chris Halsall"
Sep 2002

2×33×197 Posts

Originally Posted by Mark Rose View Post
For instance, recently GPU72 was holding on to a bunch of exponents at 73 bits, but PrimeNet started handing out exponents at 71 bits. It would have been better to release back some of those 73 bit TFed exponents.
Actually, what happened there is GPU72 /was/ releasing some at 73 bits when it "observed" that candidates were about to be assigned for P-1'ing at below that (up in the 80M range). But because it only "looks" at the situation every five minutes and Mr. P-1, and then later aurashift, were reserving candidates in such large batches, it wasn't keeping up with the releasing.

Originally Posted by Mark Rose View Post
If we know we need 300 exponents per day, Chris can program GPU72 to make sure an appropriate amount are left with PrimeNet at say 74 bits or even 73 bits, based on the available TF capacity. So having the deltas is the most important thing.
It's actually a bit more complicated than that, because of the non-linear nature of the assignment requests. Normally it's pretty steady, but sometimes large batches are requested. The ideal is to pull far enough ahead that we've got a comfortable buffer of candidates at 75 bits for both the P-1'ers, and separately, the LL'ers.

Having the deltas allow us to make an informed decision as to how many we can take to 75, and how many we should only take to 74. My goal is to not have assigned for LL'ing anything not at at least 75 bits above 65M. GPU72 will "recapture" anything released for P-1 at below 75 bits for final TF'ing, assuming a factor isn't found by the P-1 run of course.
chalsall is offline   Reply With Quote