mersenneforum.org  

Go Back   mersenneforum.org > Great Internet Mersenne Prime Search > Data

Reply
 
Thread Tools
Old 2009-01-04, 04:52   #12
stars10250
 
stars10250's Avatar
 
Jul 2008
San Francisco, CA

C916 Posts
Default

Quote:
Originally Posted by tha View Post
I believe (literally, not based on figures) the extra throughput is entirely wiped out by errors in the results. Of course a little overclocking does less harm than more aggressive overclocking etc. Still I don't see a reason to overclock for GIMPS machines.
Is it possible to get these figures? I know when I run benchmarks, prime95 knows my system is OC'd because it reports two speeds, for ex:

Intel(R) Core(TM)2 Quad CPU Q6600 @ 2.40GHz
CPU speed: 3005.39 MHz, 4 cores

I know a benchmark run isn't the same as when a LL completes and reports back, but maybe there could be a way for it to keep track so that eventually we can examine the statistics on error rates. Maybe there could be a checkbox on prime95 or on primenet where people could indicate if their computer is OC'd (and maybe even by how much). I think this is very important for GIMPS in the long term. Even those that just want to win the next prize might be doing themselves a disservice by OCing, and I bet they would like to know.

Last fiddled with by stars10250 on 2009-01-04 at 04:57
stars10250 is offline   Reply With Quote
Old 2009-01-04, 05:13   #13
Xyzzy
 
Xyzzy's Avatar
 
"Mike"
Aug 2002

164548 Posts
Default

We sure wish there was an option with PrimeNet to alternate regular LL tests with double-check tests. (Or some user-selectable ratio.) We think we've asked for this option several times over the years. At least with this option you would know if you were turning in bad results a lot sooner.

How many hours of testing an overclock for stability can you spare before you lose time against a regular machine that started work right from the beginning?
Xyzzy is offline   Reply With Quote
Old 2009-01-04, 15:18   #14
stars10250
 
stars10250's Avatar
 
Jul 2008
San Francisco, CA

C916 Posts
Default

Quote:
Originally Posted by Xyzzy View Post
How many hours of testing an overclock for stability can you spare before you lose time against a regular machine that started work right from the beginning?
It would depend on how long you tinker with an unstable system and lose valuable time, and how much you OC. But if overclocking is about as reliable (equivalent error rate to non-OC system) then I believe it is worth it. In my case, it was pretty clear when my system was stable vs unstable using the torture test on prime95. If it passed 5 minutes running stable, then it also passed two weeks running stable. If my system was going to fail it would do so in the first 5 minutes. I've heard others noting instability after several days of tests, but I haven't experienced that. That would cause one to lose more time tinkering and be rather frustrating. Another user also recommended I run a double-check (~2 week run) as a way to further verify my stable overclock. I learned I was stable and got credit while doing it.

In my case, I'm overclocking a Q6600. The non-OC'd iteration times for my particular exponents were 96 ms (for all 4 cores) and are now 61 ms (for all 4 cores) when OC'd. That's nearly a 45% performance difference for a cost difference of $14 for a better heatsink (newegg rebates made my motherboard and memory cost the same as components for a stock, non-OC'd, system). This is why I'm eager to know the statistics about error rates for OC'd systems. The performance gain vs cost for OC is simply too large to ignore.

I won't answer your question with a specific number, but my 45% performance improvement would probably let me tinker with my system for a few years (assuming a computer life of 7 years?) before I'd lose out to the stock speed computer.

This brings up another question. If one of your results get double-checked and found to differ, does primenet let you know? Meaning, if one has a bad computer that is constantly turning in results that are later proved to be wrong, does primenet let you know you have a problem computer on your hands? I haven't been around long enough to know, but I suspect not.
stars10250 is offline   Reply With Quote
Old 2009-01-04, 21:18   #15
Xyzzy
 
Xyzzy's Avatar
 
"Mike"
Aug 2002

22·1,867 Posts
Default

Quote:
If one of your results get double-checked and found to differ, does primenet let you know?
With the old PrimeNet, not directly. You can find out but you have to work at it a bit.

The new PrimeNet may be able to but we haven't looked into it.

This is our old Team Prime Rib stats page. At the bottom it shows the bad results.

http://www.teamprimerib.com/rr1/bin/...er.php?u=Xyzzy
Xyzzy is offline   Reply With Quote
Old 2009-01-04, 22:55   #16
garo
 
garo's Avatar
 
Aug 2002
Termonfeckin, IE

267510 Posts
Default

Yes you can with the new Primenet. You just need to customize your report. For example:
http://v5www.mersenne.org/report_LL/...B1=Get+LL+data

Replace xxxxx with your userid.
garo is offline   Reply With Quote
Old 2009-01-05, 01:38   #17
stars10250
 
stars10250's Avatar
 
Jul 2008
San Francisco, CA

3·67 Posts
Default

Ok, I think I understand. Based on that url, I have two verified test results and then a whole bunch that are unverified. The unverified have an empty column for "Error code (if any)," so I guess if I see anything show up in there I'll know my computer messed that one up. I can try to monitor this and compare OC'd to non-OC'd computers, as I have a few of each. It will take some time for those to get verified, but at least I can keep alert for any problems. Thanks much.

Last fiddled with by stars10250 on 2009-01-05 at 01:40
stars10250 is offline   Reply With Quote
Old 2009-01-05, 11:48   #18
garo
 
garo's Avatar
 
Aug 2002
Termonfeckin, IE

52×107 Posts
Default

No the error code tells you if there is an error code generated during the run. But not all error codes are harmful. If you have an incorrect test there is a separate heading under unverified LLs called bad LLs or something like that.
garo is offline   Reply With Quote
Old 2009-01-06, 05:00   #19
joblack
 
joblack's Avatar
 
Oct 2008
n00bville

10110101012 Posts
Default

Quote:
Originally Posted by tha View Post
Still I don't see a reason to overclock for GIMPS machines.
You overclock mostly because you're games (applications, ...) have to be faster. That prime95 will work faster is a side effect.

I´m a fan of overclocking, BUT I think if a newbie wants to do that he/she has to be informed and extra cautious about it. Imho you have to be experienced in overclocking to avoid getting hardware errors.
joblack is offline   Reply With Quote
Old 2009-01-06, 10:55   #20
garo
 
garo's Avatar
 
Aug 2002
Termonfeckin, IE

52×107 Posts
Default

Perhaps there was no reason to overclock in the pre-Core2 era when you usually got 10-15% if you were lucky. With Core2 however you can easily get 40-50% without tinkering with any other settings. For a 40% speedup, a bad LL test or two a year is quite acceptable IMHO.
garo is offline   Reply With Quote
Old 2009-01-06, 18:35   #21
cheesehead
 
cheesehead's Avatar
 
"Richard B. Woods"
Aug 2002
Wisconsin USA

22·3·641 Posts
Default

Quote:
Originally Posted by garo View Post
For a 40% speedup, a bad LL test or two a year is quite acceptable IMHO.
Hmmm... Let's do some numbers.

System A is standard-clocked and does 10 LL tests in a year, all reliable. System B is 40% OCed and does 14 LL tests in a year but 2 of those have incorrect residues. System C, doing doublechecks, and system D, doing triplechecks, are standard-clocked and do 10 reliable LL tests per year.

Assume that only one exponent is assigned (first-time or DC or TC, but not overlapping any other) at a time.

With only system A performing first-time tests, GIMPS gets 14 first-time LLs in 1.4 years, then gets 14 DCs on those from system C the next 1.4 years. Result: 14 double-checked exponents in 2.8 years.

With only system B performing first-time tests, GIMPS gets 14 first-time LLs in one year, then gets 12 DCs with matching residues plus 2 DCs with nonmatching residues from system C in the next 1.4 years. Then it gets 2 TCs with matching (to DCs) residues from system D in the next 0.2 year. Result: 14 double-checked exponents in 2.6 years.

Another way of looking at system B's output is that it produces 12 correct results per year (which are DCed in 1.2 years) and provokes 2 more correct results by the combination of systems C and D in 0.2+0.2 year. Same result: 14 double-checked exponents in 1.0+1.2+0.2+0.2 = 2.6 years.

These numbers seem to show that system B is a net plus for GIMPS throughput. Any oversights?

In order to slow GIMPS throughput to 14 exponents in 2.8 years in the second case, system B would have to err at the rate of 4 erroneous residues in 14 LL tests.
cheesehead is offline   Reply With Quote
Old 2009-01-06, 19:34   #22
garo
 
garo's Avatar
 
Aug 2002
Termonfeckin, IE

51638 Posts
Default

I must add here that I only do doublechecks and last year I did about 120 tests. Two of those were bad. And two still unverified.
garo is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
error rate and mitigation ixfd64 Hardware 4 2011-04-12 02:14
EFF prize and error rate S485122 PrimeNet 15 2009-01-16 11:27
A plot of Log2 (P) vs N for the Mersenne primes GP2 Data 3 2003-12-01 20:24
What ( if tracked ) is the error rate for Trial Factoring dsouza123 Data 6 2003-10-23 22:26
Error rate for LL tests GP2 Data 5 2003-09-15 23:34

All times are UTC. The time now is 22:50.

Sat Apr 4 22:50:00 UTC 2020 up 10 days, 20:23, 0 users, load averages: 1.02, 1.15, 1.32

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2020, Jelsoft Enterprises Ltd.

This forum has received and complied with 0 (zero) government requests for information.

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.
A copy of the license is included in the FAQ.