View Single Post
Old 2007-10-19, 05:25   #27
gd_barnes
 
gd_barnes's Avatar
 
May 2007
Kansas; USA

100111101011002 Posts
Default More exact CPU specs and speeds

Quote:
Originally Posted by VBCurtis View Post
Gary- thanks for the timing report. I extrapolated from my 600k timing sample (550 sec on a P4-2750), which explains why I suddenly thought we were not under-sieving. If there really is a 30-40% jump in execution times at 633, the original sieve-depth estimates are still accurate (I wondered how I was so far off!). Nonetheless, removing the lower candidates is still likely wise.

If you were at 500 sec before the jump, I'm guessing you have a P4-3.2. Core2's LLR at roughly a P4 of 50-60% higher clock speed, so Carlos' 2.4 would get timings equivalent to P4-3.6 to 3.8, or 15-20% faster than Gary. That leaves some error in my estimates, since Carlos is under 700 sec at 775k on 57.... but I'm in the ballpark.

Gary and I are looking to sieve to 58T in the next 10 days, then release files from 700 to 840 or so. 700 to 760 are spoken for... get 'em while they're hot! One reservation at a time, from each batch, please. When we get sieving ahead of LLR reservations, we'll revert to the RPS-drive style of reserving the next range when the current range is about to finish.

Finally, there will be a large break at n=1.2M. Sheep's sieving farm is on hiatus at the moment, and we'll hit a wall at that point until he comes back online. In a perfect world, we'll hit 1M by 1/1/2008, 1.2M by 1 Mar, and Sheep will be back sieving before then. I'll hope.
-Curtis

Thanks for that explanation Curtis. I'm really a little confused about computer types and speeds. (It used to be so easy to go by Mhz/Ghz CPU speed but things have changed a lot in the last 2-3 years.) You were close on the machine type. I'm using one of my three Dell core-2 duos and it is only slightly slower than my single-core 3.2-Ghz P4 desktop that I use as my main personal machine. So I assumed it was a 3.0 Ghz or the equivalent. But it's actually 1.66 Ghz, which I could never quite understand until your explanation here. Even though half the Ghz speed, each core is only about 10% slower than my main desktop running one core. Carlos, your 2.4-Ghz quad machine must be like if I had a 2.4-Ghz Dell core-2 duo (vs. my 1.6 Ghz), i.e. 50% Ghz-speed faster than what I'm running. Although I'm not sure if that translates to 50% faster LLR time, i.e. 1/3rd less or 600 vs. 400 secs as an example.

Curtis, the 500 to 700-second jump was a slight exaggeration but not much. It actually went from about 520 to 690. About a 33% jump. How does that make the sieving look now? So you'll know exactly what I'm running and will have a better estimate for sieving, here is a cut-and-paste right out of LLR for my Dell core-2 duo as well as exact timings before and after the fftlen increase:

(Dell core-duo CPU specs):
Intel(R) Core(TM) 2 CPU T5500 @ 1.666Ghz
CPU speed: 1662.38 MHz
CPU features: RDTSC, CMOV, PREFETCH, MMX, SSE, SSE2
L1 cache size: 32 KB
L2 cache size: 2048 KB

Timings:
5*2^633108-1 519.754 secs.
5*2^633140-1 519.544 secs.
*****
5*2^634922-1 691.318 secs.
5*2^635020-1 692.568 secs.


Here's something confusing...My main sieving machine is an Athlon dual-core laptop that also runs at 1.66 Ghz. But it LLR's at half of the speed of the Dell core-duos although it sieves at > 75% of the speed of them. So I always only use it for sieving. By the way, I'm only 'assuming' it is an Athlon based on the CPU specs for it from LLR as follows:

(Assumed Athlon CPU specs):
AMD Turion(tm) 64 X2 Mobile Technology TL-50
CPU speed: 1607.28 Mhz
CPU features: RDTSC, CMOV, PREFETCH, MMX, SSE, SSE2
L1 cache size: 64 KB
L2 cache size: 256 KB


Curtis, I will probably stop my LLR effort somewhere in the n=800's. There will be plenty of people to continue from there. This is mostly why I don't pick up ranges from the team efforts much. I have nothing against them at all but there's almost always plenty of people on them and the tests take too long for my tastes and I prefer to LLR large n-ranges at once. As you know, I try to find dormant k's that need to be worked and k's with gaps that need to be filled and k's that are 'behind the curve' in how far they should have been tested. k=5 certainly filled the 'behing the curve' case.

It will be a stretch for us to get LLR to n=1M by year-end with or without me after n=800's, but it can be done. If Carlos can keep throwing 2 or 3 cores of that new very speedy quad-core of his at the effort, we may just get there. Kosmaj or Karsten may even want to throw in a CPU or two at it. Mostly, I'm confident that all ranges up to n=1M will have LLRing 'in the works' by year end. It would be a very significant accomplishment if all LLR tests were done to n=1M because of the amount of time each candidate will take to LLR at that point.


Gary

Last fiddled with by gd_barnes on 2007-10-19 at 05:40
gd_barnes is offline   Reply With Quote