Quote:
Originally Posted by cgy606
Thank you Carlos, the one with CUDA7 works (at least is seems to generate a ton of number triplets on the screen which I think is stage 1). Now, without asking a ton of questions, is there something I can read that will tell me how to configure the parameters of stage 1 poly search given a particular input size (the readme isn't too informative in this regard). What I mean is, what should the bounds be for the stage 1/2 norms, min E-Value, leading coeffs, poly search ranges and time limits per coeff and for the entire process. Also, can gpu do anything other than a degree 5 poly? Thanks...
|
Within this forum, you'll find a one-post thread on stage1norm selection, as well as quite a lot of discussion about using the various poly-select flags throughout the poly request thread. The poly request thread includes actual invocations of msieve flags, which may help you figure out how to generate a full polynomial.
A super-abridged version:
Run -np1 once for a few seconds to see what the log file says are msieve's choices for stage1norm and stage2 norm.
Then, run msieve -np1 -nps -stage1_norm={default norm divided by 8 to 10, depending on how fast your GPU is; experiment) -stage2_norm={default norm divided by 25 to 30, depending on how big the project is} -s output
This will create an output file "output.ms" with a few dozen size-optimized hits, ready for the root-opt phase. Setting the norms as I suggest should produce a couple hits per hour, though of course that varies with GPU and CPU speed. I try to collect 80 to 150 before bothering to run -npr.
When you have a batch you wish to run, msieve -npr -s output tells msieve to look in output.ms for the polys to run root-opt on. This step takes something like an hour per 100 hits, give or take a factor of two. If you want less output, you can use -min_evalue={something bigger than default but less than the log's hoped-for poly score}.
Let it finish, and the best-scoring poly will be written to (I think) msieve.fb (or maybe msieve.poly, or output.fb).
I allocate GPU time based on a guess of 3% of total project run time. A GNFS-155 will take about 35 core-days, so a full GPU-day of -np1 -nps sounds about right.