mersenneforum.org

mersenneforum.org (https://www.mersenneforum.org/index.php)
-   Software (https://www.mersenneforum.org/forumdisplay.php?f=10)
-   -   P3 TF time (https://www.mersenneforum.org/showthread.php?t=1633)

PrimeCruncher 2003-12-10 23:40

P3 TF time
 
I have a potential opportunity to install Prime95 on 50+ PCs. They are all P3s in the vicinity of 800-1000 MHz. They'll be on about 8 hours per day Monday through Friday, so I was wondering if anyone had such a P3 running trial factoring so I would know about how long a TF would take on these PCs.

Xyzzy 2003-12-10 23:56

My 450MHz P3 takes around 10 days to do a current factoring exponent...

nfortino 2003-12-10 23:58

My 533MHz P3 takes 8.5-9 days to do a current TF

Complex33 2003-12-10 23:59

Oh goody, more computers. That will definately get you to the first few TF positions. As for hard numbers, I don't have any although, I got a hold of about 20 PIII's once ranging from 500-1000 Mhz and the account I put it under rose quickly on the TPR stats, sadly lost them after only a few weeks. Good luck!

axn 2003-12-11 00:20

My P3 M 1.2G can do a TF in about 3.5 days running 24x7

PrimeCruncher 2003-12-11 00:42

Thanks for all your responses. The provided data indicates a P3 900 should take about 108 hours to do a TF (assuming I've got the math right...). These computers belong to my middle school, so they won't be on all year, but I've estimated that for the 180 days they are active they will produce 210 CPU years worth of work in total. :shock: Note to self: install Prime95 ASAP!! :grin:

dsouza123 2003-12-11 02:26

On a Celeron 1 Ghz it takes 4.25 days (102 hours) to do a current (23 million ) TF.

QuintLeo 2003-12-12 12:41

Frak. Someone else trying to knock me out of 100'th (again, I just got back there) on the TF rankings.

9-)

nfortino 2003-12-12 15:26

[QUOTE][i]Originally posted by QuintLeo [/i]
[B]Frak. Someone else trying to knock me out of 100'th (again, I just got back there) on the TF rankings.

9-) [/B][/QUOTE]

You’re in luck. PrimeCruncher is already 55th in TF. Your top one hundred spot is safe for now...

PrimeCruncher 2003-12-12 19:54

For a while I was stuck in the mid 60s. Then I got the PCs at my dad's office and my friend's house involved. Since then we've had a pretty steady climb. +11 in the last 3 months according to the TPR reports. QuintLeo, if you want to have ANY chance of staying in the Top 100 factorers, you're going to have to do better than 4.135 CPU years per 90 days. If you really want to stay on the list you should get that up to at least 6, preferably more.

nitro 2003-12-12 22:19

[QUOTE][i]Originally posted by PrimeCruncher [/i]
[B]For a while I was stuck in the mid 60s. Then I got the PCs at my dad's office and my friend's house involved. Since then we've had a pretty steady climb. +11 in the last 3 months according to the TPR reports. QuintLeo, if you want to have ANY chance of staying in the Top 100 factorers, you're going to have to do better than 4.135 CPU years per 90 days. If you really want to stay on the list you should get that up to at least 6, preferably more. [/B][/QUOTE]

My few 2.4 P4's spit that out per WEEK, if they didn't suck at factoring I might switch them over....

nfortino 2003-12-12 22:46

[QUOTE][i]Originally posted by nitro [/i]
[B]My few 2.4 P4's spit that out per WEEK, if they didn't suck at factoring I might switch them over.... [/B][/QUOTE]

Except TF is not credited the same as LL testing is. I don't know the exact conversion, but it is discussed at [url]http://mersenne.org/ips/faq.html[/url] under Why is factoring CPU time not ranked with equal weight as primality (Lucas-Lehmer) test CPU time?

PrimeCruncher 2003-12-12 23:51

To get people to do LL tests. George figures if you get more credit for LL, then more people will do it.

And P4s don't suck at TF. My Celeron 2.0GHz does one about every 32 hours and is my best machine.

dsouza123 2003-12-13 00:20

With trial factoring, for 64 bits and below, P4s or Celerons with SSE2 are very slow for their clock speed but at 65+ they are superb.

A few of them could really boost TF ranking, maybe with specific optimizations the Athlon64s/Opterons will be able to be superb at both, and become super trial factoring CPUs. :grin:

PrimeCruncher 2003-12-13 00:36

Pehaps. But for purely TF purposes I think Celeron will be the best buy for some time. It has the best SSE2 optimizations (obviously, since Intel made it), it has the speed, and it has a low price tag. A Celeron node can be had for $150 where a similar P4 node would be at least $250 and I don't want to think about Athlon64/Opteron. But you are correct that they add a lot of power. Given 0.105 CPU years per current TF (is that still the figure, dsouza123?) my Celeron will push over 28 CPU years per year :shock: Note to self: complete farm and buy more nodes ASAP!!

Just goes to show that the answer is ALWAYS D: [B]MORE BOXEN!![/B] :grin:

dsouza123 2003-12-13 00:48

The .105 figure varies somewhat depending on the TF but it is very close.

I agree with you, until things change alot, right now the Celeron with SSE2 gives an excellent (perhaps best) bang for the buck at trial factoring.

nitro 2003-12-13 01:56

[QUOTE][i]Originally posted by PrimeCruncher [/i]
[B]To get people to do LL tests. George figures if you get more credit for LL, then more people will do it.

And P4s don't suck at TF. My Celeron 2.0GHz does one about every 32 hours and is my best machine. [/B][/QUOTE]

P4's suck at anything less than ~64 bits, above that they wipe the floor with everything.

I just sent a P4 to get a couple of factoring exponents, the 2.4GHz reckons it will go from 60 to 67 bits in just over 28.5 hours...

PrimeCruncher 2003-12-13 03:06

[QUOTE][i]Originally posted by nitro [/i]
[B]P4's suck at anything less than ~64 bits, above that they wipe the floor with everything.[/B][/QUOTE]

Yes, but 57-64 is an extremely small portion of the TF. The stuff that takes a long time on current PCs is 65 and above, which P4s are optimal for. It's a small sacrifice for a huge gain.

Complex33 2003-12-13 04:49

Too bad the server can't split the TF's up between the instruction sets... below 64 to non SSE2 machines and above to SSE2 machines. I tried to convince a friend who has a couple of dual processor Athlons and a dual 3 gig Xeon that he should physically transer factoring work between the two along these lines, too much work, oh well :) Then again if there were just a P-1 factoring assignment I'd have him doing that.... 12 gigs of ram between these boxen is good :grin:

PrimeCruncher 2003-12-13 14:07

[QUOTE][i]Originally posted by Complex33 [/i]
[B]Too bad the server can't split the TF's up between the instruction sets... below 64 to non SSE2 machines and above to SSE2 machines.[/B][/QUOTE]

Maybe we should ask Old man PrimeNet for that feature in v5. :wink:

nfortino 2003-12-13 14:40

[QUOTE][i]Originally posted by Complex33 [/i]
[B]Too bad the server can't split the TF's up between the instruction sets... below 64 to non SSE2 machines and above to SSE2 machines. [/B][/QUOTE]
A variation of the first part of that can, and in my opinion should, be done with an edit to prime95. There are some computers doing TF that take forever to complete it. In fact, some take so much time, that the LL front wave overtakes them. When this happens, it defeats the entire purpose of TF. The v4 server works with the FactorOverride=n command, so all that needs to be done is tell prime95 to automatically put this command in for really slow machines at a reasonable bound. That way, the slow machines can take the easy parts of TF, and the fast ones can finish the job. The only complexity I see is how to make sure the slow machine does not continuously take the same exponent which has already been factored past its bound, and then release it, and take it again.

QuintLeo 2003-12-14 22:58

The V4 server does NOT work well with the Factoroverride option - there have been a lot of "stuck exponent" problems that appear to be attributed to that. Scott is aware of it already, though - I suspect v5 will be more ... robust ,,, when dealing with that option.

I think it would take a Prime program change to say "well, if this is an Athlon do the TF up to 2^64 then hand it back in for a P4 or P4-Celeron to work on the rest of the way" automagically.

nfortino 2003-12-14 23:00

I was just about to apologize after re-reading the undoc entry. Sorry. :redface:

Complex33 2003-12-20 02:18

Ok I ran an experiment using the FactorOverride=n switch with the V4 server. The procedure I used allows a machine to factor to a specific bit depth, return the results and "un-stick" the exponents with a little manual work. I also verified that the server has an updated bit depth for the said exponents when assigning them to other factoring accounts. I'd post the procedure but I'm not sure how it would be received by the powers that be. Anyone want to condone this?

PrimeCruncher 2003-12-20 02:52

Well, the point of doing that would be to get non-SSE2 CPUs to factor only through 2^64, would it not? This procedure would be a lot of work for miniscule amounts of credit. You could post the procedure but I doubt many people would use it. I, for one, will not.

sdbardwick 2003-12-20 05:40

Complex33:
I'd be interested, depending on the amount of intevention required to make sure the exponents are 'un-stuck'. If it can be done in a semi-automated batch process I'd love to do the initial step on my various Athlons so P4s can be used in the most efficient manner.

Obviously, I would only take this course of action if it does not cause problems with the server, and George and Scott don't object. I seem to recall there being some discussion that the next version of the server would allow for this degree of granularity in assigning TF work.

GP2 2003-12-20 15:31

If you look at [url]http://www.mersenne.org/primenet[/url], you don't see stuck exponents any more. I think Scott made a recent change to automatically handle them.

Of course you do see the daily expiries, which sometimes take hours to clear, but that's not the same as permanently stuck exponents.

garo 2003-12-20 20:00

The problem with this procedure - which was verified by trif and myself about a year ago - is that these exponents will become stuck again when the time comes to release them for LL testing. And if any of the exponents expire they are likely to become stuck as well. Primenet had huge gobs of stuck exponents until recently that rendered the summary page a not so pretty sight.

Therefore, we decided it is not worth the hassle and the screw-up on Primenet.

garo 2003-12-20 20:12

Oops! i did not read GP2's message before posting and forgot about the change Scott made. Hopefully, that takes care of our past concerns. I'd be interested to see if someone can verify that Scott's new mechanisms "unsticks" all stuck exponents successfully.

Complex33 2003-12-21 04:46

When running the expierment TF exponents were stuck after sending in lower than normal factoring bit depths. I had to manually un-stick them and then release them.

garo 2003-12-21 05:26

Yeah but Scott's unstick routine runs at 0600 UTC everyday. So leave a few stuck and come back the next day to see if they got unstuck.


All times are UTC. The time now is 11:29.

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2021, Jelsoft Enterprises Ltd.