View Single Post
Old 2007-10-21, 07:33   #40
VBCurtis's Avatar
Feb 2005
Riverside, CA

3,989 Posts

Originally Posted by gd_barnes View Post
Holy cow. Both Sheep and Curtis, IMHO dedicating any more than 12 cores to one k for sieving is overkill. If you look back at previous team drives, I don't recall seeing one where we sieved so deeply for n<1M.
There is no other k sieved so deeply below 1M, because no other k had a single sieve done to 4M. Efficiency increases greatly with a larger n-range; combined with empirical evidence that removing the small n's provide very little speed benefit to the overall sieve, it makes sense to sieve extremely deeply for large sieves. Sheep's farm(s) have sieved k=15, my group of k's (11,13,31,45,99,127, and now 5), rieselsieve, and some other stuff I'm forgetting. PowerPCs have no LLR client, alas.

The 14 k's drive has quite a lot of life left for LLR before we need a new team drive; quite a few users prefer to do everything themselves (me included). I think building a new team drive with a dozen or so k's is a great idea, using Sheep to sieve; however, moving 100+ k's to a group effort puts global efficiency above individualism. A possible compromise is to build such a global sieve, but continue to allow users to reserve entire k's out of the sieve, rather than team LLR everything. Those who like to sieve (or have athlons or laptops that aren't well suited to LLR) can reserve ranges from this massive sieve. If we ever do this, we should still leave some k's untouched for those interested to process entirely on their own. Perhaps put 30 k's of various weights into a mass sieve... eventually.

Another way to look at this is max efficiency comes from the largest n-range, but we do not usually sieve very-large n-ranges. Why not? The same long-run "greater good" is achieved by sieving a range twice as high as sieving twice as many k-values. One chooses limits on both for practical, project-completion reasons. Only for very large n-ranges does one not need to finish sieving for max efficiency before starting LLR; if sieving 100 k's, that sieving step would be many CPU-years. And what about file size? The sieve needs to be practically emailed, which puts a 10MB or 20MB limit.

I tried 33 k's from 260k to 300k, in the 900-1000 range (see primesearch), as an experiment. It was NOT efficient. Many k's with a narrow n range is bad bad bad for sr1/sr2. Big fat n ranges are incredibly efficient. Thus, my decision to run this k=5 sieve to 4M, and the resulting ~100T sieve depth for n<1M.
VBCurtis is offline   Reply With Quote