Want to trade work?

geoff · 2004-09-23, 16:49

Would anyone like to trade their P3, Athlon, etc. with my Celeron for a few weeks? It is sse2-capable, 2.88GHz, and good for GIMPS trial factoring. (GIMPS name is 97030014/cel). It takes about 19 hours to TF one of the current exponents in the 2710xxxx range from 2^62 to 2^67. If you are doing this sort of work on a non-sse2 machine, try adding one of these lines to your worktodo.ini and note the time taken for each curve:

ECM=737,11000000,1,1,0,0,1,0
ECM=749,11000000,1,1,0,0,1,0

(Don't report it to the primenet server, and you may have to temporarily increase the OutputIterations line in prime.ini to OutputIterations=999999999 to avoid masses of unwanted output affecting the timing). My celeron takes 178 seconds (for the equivalent work by a slightly different method) for each of these, which means I can do about 385 curves in the time it takes to TF from 2^62 to 2^67.

If your machine can do significantly more than this (in the time it takes to do the same TF on the same machine) then it might be worthwhile trading work for a while. I want up to 5000 each of these curves. You can send me worktodo.ini entries for whatever you want done on my Celeron.

(If you haven't done ECM before then the work will just involve downloading this file http://www.mersenne.org/gimps/lowp.txt into your Prime95 directory, adding the line "GmpEcmHook=1" to prime.ini and an entry ECM=737,11000000,1,N,0,0,1,0 to worktodo.ini, where N is the number of curves you want to do, and then sending me a copy of the resulting results.txt file, which may grow to a few MB. It will use no more memory than trial factoring does.)

edit: If you have a sse2-capable machine then it will not be worthwhile trading, but if you are interested in the times anyway then the equivalent lines to put in the worktodo.ini are:

ECM=1474,11000000,1,1,0,0,0,0
ECM=1498,11000000,1,1,0,0,0,0

geoff · 2004-09-24, 12:10

I timed my old 166MHz Pentium-mmx for comparison. It takes 738 hours to complete TF for a 2710xxxx exponent from 2^62 to 2^67 on this machine, and 2075 seconds per ECM curve on P737. This means it can do 1280 curves in the time it takes to do one TF, compared to the Celeron's 385.

If you had a machine like this doing GIMPS TF assignments and wanted to trade, then you would do 705 ECM curves for me and I would do one TF for you. It would take you 406 hours work and you would get 738 hours in return. I would do 19 hours work and get 35 in return.

We would each get over 80% more work done!

I realise there are probably not many 166MHz machines doing TF any more, but perhaps the numbers will work out like this for a P3 or Duron too?

akruppa · 2004-09-24, 12:39

I'm currently generating F12 residues on my home P3-500, but P737 and P749 would make better use of the length 32 FFT than F12 makes of length 256. Would you like to trade P737/P749 B1=11M residues against M32768 B1=44M residues? The exchange rate should be ~1:178. A curve takes me ~10 minutes, so I should be able to produce ~70-100 residues a day.

Alex

geoff · 2004-09-24, 15:01

Quote:

Originally Posted by akruppa

Would you like to trade P737/P749 B1=11M residues against M32768 B1=44M residues? The exchange rate should be ~1:178.

yes that would be good. I haven't timed M32768 yet, but I think they will take 5.5-6 hours each on my Celeron. At 1:178 I could provide you 60 B1=44M M32768 residues in exchange for 5300 each of P737, P749 at B1=11M. I will PM you my email address.

geoff · 2004-09-24, 22:38

It turns out M32768 is not so good on the Celeron, it would take 8hr 45min per curve. I will do them on a 2.9GHz P4/HT instead, it does M32768 B1=44M in 4hr 40min and P737/749 B1=11M in 152sec. How does that affect the exchange rate? At 1:178 I would be 61% better off :-)

akruppa · 2004-09-25, 08:58

I wouldn't have a problem with keeping the exchange rate at 1:178. That's the ratio derived from exponent and B1, and for the purpose of optimizing total throughput, I think it makes sense to rate machines only be how much work they get done rather than penalizing extremely efficient machines. Btw, I factored the exponent in linearly instead of n*log(n), but lets not be overly picky, especially if the error is in my favour!

Doing 2x5300 curves will probably take more than 3 months on my P3 which may be longer than you'd like to wait, so if you would like to barter residues with someone else as well, please feel free to!

Alex

geoff · 2004-09-25, 18:23

OK, lets just trade 30 of M32768 for 5300 of P737 for now and see how that goes.

If anyone else wants to trade work then post here.

Mystwalker · 2004-10-08, 19:14

I can throw in a Duron @ 800 (or was it 850?) MHz running Linux. Interested?

geoff · 2004-10-09, 07:44

Quote:

Originally Posted by Mystwalker

I can throw in a Duron @ 800 (or was it 850?) MHz running Linux. Interested?

Yes please. What sort of work do you want done in exchange? And how long does it take your Duron to generate a P749 residue? (Put GmpEcmHook=1 in your prime.ini and ECM=749,11000000,1,1,0,0,1,0 in your worktodo.ini to find out).

If anyone has an Athlon handy I would also be interested in finding out whether an Athlon is any faster dollar for dollar than a Duron at this sort of work. I am thinking of putting together a budget AMD system and the Duron 1800MHz chip is very nicely priced.

ET_ · 2004-10-09, 10:27

Quote:

Originally Posted by geoff

If anyone has an Athlon handy I would also be interested in finding out whether an Athlon is any faster dollar for dollar than a Duron at this sort of work. I am thinking of putting together a budget AMD system and the Duron 1800MHz chip is very nicely priced.

I'll have an Athlon 2100 XP (1736 MHz) available for tests next monday night.
Tell me what I have to check

Luigi

Mystwalker · 2004-10-09, 16:58

Quote:

Originally Posted by geoff

What sort of work do you want done in exchange?

If you don't mind, I'd like to have PRP tests of Prime Sierpinski numbers.
To do this, you have to

1) download PRP3 from http://www.mersenne.org/gimps (either Windows (.zip) or Linux (.tgz) version)
2) make a file with the following contents:

Code:

1000000000:P:0:2:257
222113 1226837
222113 1226901

3) take this file as "input file" under "Test --> Input Data..."

Quote:

And how long does it take your Duron to generate a P749 residue? (Put GmpEcmHook=1 in your prime.ini and ECM=749,11000000,1,1,0,0,1,0 in your worktodo.ini to find out).

I've tested it with gmp-ecm:

Code:

GMP-ECM 5.1-beta [powered by GMP 4.1.3] [ECM]
Input number is 2831308076778128190043701447911173778911219952694768521472311451
0480121 (189 digits)
Using B1=11000000, B2=25577181640, polynomial Dickson(12), sigma=3011671549
Step 1 took 652909ms
Step 2 took 440633ms
talis@Suse:~/bin/gmp-ecm> ./ecm 11e6 < p749
GMP-ECM 5.1-beta [powered by GMP 4.1.3] [ECM]
Input number is 283130807677812819004370144791117377891121995269476852147231145147056120358331224118442174975812354857658437068455078687726853215164019399440647197130552010023632611590549058806284940480121 (189 digits)
Using B1=11000000, B2=25577181640, polynomial Dickson(12), sigma=4161096674
Step 1 took 652649ms
Step 2 took 440083ms

So I need a little more than 18 minutes. I already took away the known factors.
Is the prime95 ecm'ing faster?

2004-09-23, 16:49	#1
geoff Mar 2003 New Zealand 13×89 Posts	Want to trade work? Would anyone like to trade their P3, Athlon, etc. with my Celeron for a few weeks? It is sse2-capable, 2.88GHz, and good for GIMPS trial factoring. (GIMPS name is 97030014/cel). It takes about 19 hours to TF one of the current exponents in the 2710xxxx range from 2^62 to 2^67. If you are doing this sort of work on a non-sse2 machine, try adding one of these lines to your worktodo.ini and note the time taken for each curve: ECM=737,11000000,1,1,0,0,1,0 ECM=749,11000000,1,1,0,0,1,0 (Don't report it to the primenet server, and you may have to temporarily increase the OutputIterations line in prime.ini to OutputIterations=999999999 to avoid masses of unwanted output affecting the timing). My celeron takes 178 seconds (for the equivalent work by a slightly different method) for each of these, which means I can do about 385 curves in the time it takes to TF from 2^62 to 2^67. If your machine can do significantly more than this (in the time it takes to do the same TF on the same machine) then it might be worthwhile trading work for a while. I want up to 5000 each of these curves. You can send me worktodo.ini entries for whatever you want done on my Celeron. (If you haven't done ECM before then the work will just involve downloading this file http://www.mersenne.org/gimps/lowp.txt into your Prime95 directory, adding the line "GmpEcmHook=1" to prime.ini and an entry ECM=737,11000000,1,N,0,0,1,0 to worktodo.ini, where N is the number of curves you want to do, and then sending me a copy of the resulting results.txt file, which may grow to a few MB. It will use no more memory than trial factoring does.) edit: If you have a sse2-capable machine then it will not be worthwhile trading, but if you are interested in the times anyway then the equivalent lines to put in the worktodo.ini are: ECM=1474,11000000,1,1,0,0,0,0 ECM=1498,11000000,1,1,0,0,0,0 Last fiddled with by geoff on 2004-09-23 at 16:57

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Looking for possible trade	bcp19	Hardware	1	2013-05-17 12:10
How to calculate work/effort for PRP work?	James Heinrich	PrimeNet	0	2011-06-28 19:29
Waxman-Markey AKA Cap-and-Trade Bill	AES	Soap Box	114	2009-07-06 03:11
No Work	Pilgrim	Information & Answers	1	2008-01-31 18:53
It seems to work, but why ?	T.Rex	Math	15	2005-10-15 10:38

2004-09-24, 12:10	#2
geoff Mar 2003 New Zealand 13×89 Posts	I timed my old 166MHz Pentium-mmx for comparison. It takes 738 hours to complete TF for a 2710xxxx exponent from 2^62 to 2^67 on this machine, and 2075 seconds per ECM curve on P737. This means it can do 1280 curves in the time it takes to do one TF, compared to the Celeron's 385. If you had a machine like this doing GIMPS TF assignments and wanted to trade, then you would do 705 ECM curves for me and I would do one TF for you. It would take you 406 hours work and you would get 738 hours in return. I would do 19 hours work and get 35 in return. We would each get over 80% more work done! I realise there are probably not many 166MHz machines doing TF any more, but perhaps the numbers will work out like this for a P3 or Duron too?

2004-09-24, 12:39	#3
akruppa "Nancy" Aug 2002 Alexandria 9A3₁₆ Posts	I'm currently generating F12 residues on my home P3-500, but P737 and P749 would make better use of the length 32 FFT than F12 makes of length 256. Would you like to trade P737/P749 B1=11M residues against M32768 B1=44M residues? The exchange rate should be ~1:178. A curve takes me ~10 minutes, so I should be able to produce ~70-100 residues a day. Alex

2004-09-24, 22:38	#5
geoff Mar 2003 New Zealand 10010000101₂ Posts	It turns out M32768 is not so good on the Celeron, it would take 8hr 45min per curve. I will do them on a 2.9GHz P4/HT instead, it does M32768 B1=44M in 4hr 40min and P737/749 B1=11M in 152sec. How does that affect the exchange rate? At 1:178 I would be 61% better off :-)

2004-09-25, 08:58	#6
akruppa "Nancy" Aug 2002 Alexandria 2,467 Posts	I wouldn't have a problem with keeping the exchange rate at 1:178. That's the ratio derived from exponent and B1, and for the purpose of optimizing total throughput, I think it makes sense to rate machines only be how much work they get done rather than penalizing extremely efficient machines. Btw, I factored the exponent in linearly instead of n*log(n), but lets not be overly picky, especially if the error is in my favour! Doing 2x5300 curves will probably take more than 3 months on my P3 which may be longer than you'd like to wait, so if you would like to barter residues with someone else as well, please feel free to! Alex

2004-09-25, 18:23	#7
geoff Mar 2003 New Zealand 13·89 Posts	OK, lets just trade 30 of M32768 for 5300 of P737 for now and see how that goes. If anyone else wants to trade work then post here.

2004-10-08, 19:14	#8
Mystwalker Jul 2004 Potsdam, Germany 831₁₀ Posts	I can throw in a Duron @ 800 (or was it 850?) MHz running Linux. Interested?