A few questions :) - Page 4

xtreme2k · 2002-08-23, 17:22

I dont think the cache bandwidth of the Celeron are half of what the equilvent. They cut the size, but not the bandwidth. This are proved since in Linpack, Celeron and P3 are pretty much 'neck to neck' until up to the matrix block size is larger than the Celeron's cache, then it drops off. If the Celeron's cache bandwidth are half of the P3, they wont be neck to neck at that point.

However, the P4 Celeron is SO BLOODY SLOW at Prime95 compared to the same clocked P4W. I think this is mostly because 128K L2 is really too small to cache any of the 'work' Prime95 is giving it, therefore, trashing the cache resulting in a significant higher FFT time...

garo · 2002-08-23, 19:29

Tasuke what is this P4 penalty on 33Ms? The benchmark pages show that a P4 2000 w/ RDRAM does 448K in 0.23 secs but a 33M in .120
An Athlon 2000+ does the same in .047 and .213 So a P4 does 33M in 5.2 times the time whereas an Athlon does it in 4.5 Keep in mind that the roundoffs mean there is a margin of error in my calculations.
My question is why???

Quote:

Today's report has 5837891 at 41.7 days to go. Rex, if you're out there, thanks for bumping up the priority!

Alas dswanson that was not to be and the exponent was poached by Tom Ehlert!!

Tasuke · 2002-08-23, 20:23

http://www.teamprimerib.com/xls/Prim...0Relations.xls

If you look, the P4 gets less credit/time as the FFt increases in size.
for a first time LL this is marginal, but at 33m, it exuates to near a 15% to 20% loss.

Quote:

So a P4 does 33M in 5.2 times the time whereas an Athlon does it in 4.5

An athlon may be slower, but it will get the same credit for all the different fft sizes. 4.5 is the break even point, with less being better, more meaning slower. Notice how the other cpu architectures are nearly @ parity in terms of the credit/time. Something on the P4 acts funny with the higher FFT's and the northwood didn't help it much.

P4's are still the fastest cpu's, no doubt, but they are not as efficient when doing the 33m testing.

dswanson · 2002-08-24, 00:43

Quote:

Originally Posted by garo

Alas dswanson that was not to be and the exponent was poached by Tom Ehlert!!

:( I certainly didn't expect my prediction to come true that rapidly!

Poaching exponents just before they expire, as shaneamy's been doing, is tolerable if annoying. Poaching exponents from someone who's been reporting back regularly, even if their progess is slow, is completely out of bounds

Quote:

Originally Posted by garo

Anyway several of them are sitting in our orphanage annie right now.... If you want some - drop me a line

If I take you up on this offer, who would get the credit? Are you willing to simply give me the exponent(s)? Or would I temporarily have to reassign one (or more) of my CPUs over to TPR?

Tasuke · 2002-08-24, 00:48

No, you can have them, but the more unscrupulous would have you join TPR. If I wasn't having such a bad I may have been one of them.

garo · 2002-08-24, 20:55

You can keep the credit but since they will be continued to be assigned to TPR you will have to be careful to set the Report completion dates figure to very large or to set communication to manual as if Prime95 connects to the server it will remove all the TPR exponents from your worktodo with the exception of the first.
Garo

dswanson · 2002-08-25, 04:29

Sounds like it would also work if I feed them into my worktodo file one at a time at the head of the file, and wait for each one to complete before I feed in the next, right?

Let's give it a shot. Pick a few and email them to me at the address in my profile.

Thanks much! :D

NickGlover · 2002-08-25, 23:19

Quote:

Originally Posted by dswanson

Sounds like it would also work if I feed them into my worktodo file one at a time at the head of the file, and wait for each one to complete before I feed in the next, right?

Let's give it a shot. Pick a few and email them to me at the address in my profile.

Thanks much! :D

Actually, the way it works is that the program only erases exponents from your worktodo.ini file that are not reserved to you if you have not already started working on them. So, one way to keep a bunch of exponents that aren't assigned to you in your worktodo.ini file is to do a few iterations on each exponent, and then put all of them in your worktodo.ini file in whatever order you like. After you have done this, you can contact that primenet server without any worries. Since you have technically started working on all of the exponents, the program will not delete any of them from your worktodo.ini file.

dswanson · 2002-08-26, 00:17

That would be simpler. Thanks for the tip!

garo · 2002-08-26, 06:00

Nick,
Unfortunately that doesn't work starting from 22.6 I think. Try it and see. I had all but the first guy removed. I'll send some stuff over from work tomorrow to you dswanson.
garo.

NickGlover · 2002-08-26, 06:18

Quote:

Originally Posted by garo

Nick,
Unfortunately that doesn't work starting from 22.6 I think. Try it and see. I had all but the first guy removed. I'll send some stuff over from work tomorrow to you dswanson.
garo.

Yeah, you're right! Sorry for the misinformation, dswanson. It works exactly as Garo said earlier.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Two questions:	Dubslow	GPU Computing	1	2011-08-05 18:22
Questions about the QS	Carmichael	Factoring	8	2007-04-10 11:30
5 questions	OmbooHankvald	Factoring	6	2005-08-28 19:31
Questions	OmbooHankvald	Prime Sierpinski Project	2	2005-08-01 20:18
LLR questions	OmbooHankvald	Math	6	2005-06-23 11:42

2002-08-23, 17:22	#34
xtreme2k Aug 2002 2×3×29 Posts	I dont think the cache bandwidth of the Celeron are half of what the equilvent. They cut the size, but not the bandwidth. This are proved since in Linpack, Celeron and P3 are pretty much 'neck to neck' until up to the matrix block size is larger than the Celeron's cache, then it drops off. If the Celeron's cache bandwidth are half of the P3, they wont be neck to neck at that point. However, the P4 Celeron is SO BLOODY SLOW at Prime95 compared to the same clocked P4W. I think this is mostly because 128K L2 is really too small to cache any of the 'work' Prime95 is giving it, therefore, trashing the cache resulting in a significant higher FFT time...

2002-08-24, 00:48	#38
Tasuke Aug 2002 2³×3² Posts	No, you can have them, but the more unscrupulous would have you join TPR. If I wasn't having such a bad I may have been one of them.

2002-08-24, 20:55	#39
garo Aug 2002 Termonfeckin, IE 2²·691 Posts	You can keep the credit but since they will be continued to be assigned to TPR you will have to be careful to set the Report completion dates figure to very large or to set communication to manual as if Prime95 connects to the server it will remove all the TPR exponents from your worktodo with the exception of the first. Garo

2002-08-25, 04:29	#40
dswanson Aug 2002 2³·5² Posts	Sounds like it would also work if I feed them into my worktodo file one at a time at the head of the file, and wait for each one to complete before I feed in the next, right? Let's give it a shot. Pick a few and email them to me at the address in my profile. Thanks much! :D

2002-08-26, 00:17	#42
dswanson Aug 2002 2³·5² Posts	That would be simpler. Thanks for the tip!

2002-08-26, 06:00	#43
garo Aug 2002 Termonfeckin, IE 2²×691 Posts	Nick, Unfortunately that doesn't work starting from 22.6 I think. Try it and see. I had all but the first guy removed. I'll send some stuff over from work tomorrow to you dswanson. garo.