Considering current hardware on the status page

petrw1 · 2007-05-09, 22:15

http://www.mersenne.org/status.htm has 2 columns that reference OLD hardware.

1. P90 CPU Years: I can see a valid case made to continue to report points in P90 years as that has always been the scale. Mind you, if you converted this to current hardware (something like P4 - 3Ghz) all the numbers would reduce proportionally without changing the relative placings. As well the almost overwhelming total of almost 23 million years at the bottom would change to a more apparent 20 years. Obviously these are just relative numbers and changing them in NO way changes the project duration. In any case I won't lost sleep over this one.

2. PII-400 Speec (sec): I am supposing this column is meant as a benchmark to give members an idea of how long a particular LL/DC assignment will take. However, per http://www.mersenne.org/primenet/status.shtml, only a small fraction still have PCs of that approximate size. Changing this column to report in more recent/popular sizes (P4 - 3 Ghz?) and occassionally updating it would be of more use as a benchmark.

TimSorbet · 2007-05-10, 01:34

I think that the P90 years should stay the same forever, and that the PII-400 years should be updated every now and then to stay relevant.

axn · 2007-05-10, 04:09

Quote:

Originally Posted by petrw1

2. PII-400 Speec (sec): I am supposing this column is meant as a benchmark to give members an idea of how long a particular LL/DC assignment will take.

Nope. That is what the benchmark page is for. The PII-400 speed is how you calculate P90 cpu years!! The formula is:

(PII-400 time*5.5*n)/(86400*365) = P90 cpu years.

Here 5.5 is the conversion factor from PII-400 iteration time to P90 iteration time (as explained in the status page) and n is your exponent.

Xyzzy · 2007-05-10, 04:28

petrw1 · 2007-05-10, 15:31

Quote:

Originally Posted by axn1

The PII-400 speed is how you calculate P90 cpu years!! The formula is:

(PII-400 time*5.5*n)/(86400*365) = P90 cpu years.

Here 5.5 is the conversion factor from PII-400 iteration time to P90 iteration time (as explained in the status page) and n is your exponent.

But since close to 90% of us now have P-4 type PCs, a conversion factor for P-4s would be more relevant.

S485122 · 2007-05-10, 18:25

I think the units should be hardware independent : number of elemental operations done times the amount of affected data. That way all type of PrimeNet work is on equal footing, trial factoring LL tests AND P-1 factoring. Of course this would benefit new hardware because it can do more elemental operations in fewer steps. One could define the "elemental operations" by the mathematical operations done : multiply, subtract, add and compute a reminder or a modulo... It does not need to be precise to the bit, for instance a LL test could be computed as if all steps where done with the same amount of data, even if the first steps are lighter.

For instance the LL test of M7 would get a weight of 5 squarings over 7 bits, 5 subtractions over 7 bits and 5 modulos over 7 bits. Counting the square as the number of bits additions it gives 7*5*7+5*7+5*7=45 operations over 7 bits or 315 obs (operation bits).

For an exponent N it would be N-2 multiplications over N bits, N-2 subtractions over n bits and n-2 modulos over n bits. N*(N-2)*N+(N-2)*N+(n-2)*N=N*(N-2)*(N+2)=N³-4*N. And since the exponents that PrimeNet has worked on have always been very big the -4*N term can be ignored.

I ignored the FFT multiplication method in this crude sketch. And it gives weird results since in my formula the computed credit is proportional to the cube of the exponent, with PrimeNet the credit is roughly proportional to the square of the exponent. But the idea is to have a simple formula for LLR testing, trial factoring and P-1 factoring.

The advantage is simplicity, hardware independence and a kind of mathematical justification.

Jacob

cheesehead · 2007-05-12, 08:11

As has been pointed out before (http://mersenneforum.org/showpost.ph...04&postcount=3):

One P90 CPU-year is 1.0407 * 10¹⁵ floating-point operations, only 4% more than a petaflop.

Since a P90 performed so close to a petaflop (of the exact operation mixture used in L-L, no less!) in a year, just substitute "petaflop" for "P90 CPU-year" if the latter bothers you. (IMO the 4% discrepancy will rarely be a significant problem, but if it is, just apply a 1.0407 correction factor.)

Quote:

Originally Posted by S485122

The advantage is simplicity, hardware independence and a kind of mathematical justification.

Jacob

... and the petaflop is already a widely-recognized (after a brief review of "kilo-, mega-, giga-, tera-, peta-" ... or a quick lookup via Google, or on page 600 of The World Almanac and Book of Facts 1997 in some cases) standard term.

S485122 · 2007-05-14, 19:43

The problem is just that the PrimeNet credits are not really related to Petaflops except for LLR work and that some work is over credited and other under credited.

All numbers following are P90 CPU years per day.
On a PIII an unsuccessful factorisation is worth 0,020, a successful one 0,080

On a Q6700 unsuccessful factorisation is worth 0,200, successful factorisation is worth 0,380, an unsuccessful LL test 0,300, successful P-1 0,480 and an unsuccessful P-1 is worth only 0,001. Since it is the same hardware, that the P-1, finding a factor or not, used 140 times more memory (3600 MB instead of 25 MB) the credit for all type of works should be more or less equivalent.

This is why I think that a more objective unit should be used. Of course there would still be differences between different types of work units depending on exponent size and type of work.

If one is after PrimeNet or GIMPS credit it is evident that under the current system one should not invest in memory to do P-1 tests. On the contrary, one should change the 0 to a 1 in the worktodo.ini so as to avoid that time consuming step.

I agree that the credit system is not a goal as such, but making it a bit more consistent would not be bad.

Jacob

cheesehead · 2007-05-15, 07:04

Quote:

Originally Posted by S485122

The problem is just that the PrimeNet credits are not really related to Petaflops except for LLR work and that some work is over credited and other under credited.

... which is irrelevant to replacing "P90 CPU year" by "petaflop".

Quote:

All numbers following are P90 CPU years per day.
On a PIII an unsuccessful factorisation is worth 0,020, a successful one 0,080

That inequity might be clearer to some folks if it were expressed in petaflops.

Quote:

This is why I think that a more objective unit should be used.

... which is exactly why I proposed using the petaflop.

Quote:

On the contrary, one should change the 0 to a 1 in the worktodo.ini so as to avoid that time consuming step.

I prefer not to advocate actions that reduce GIMPS's efficiency.

Quote:

I agree that the credit system is not a goal as such, but making it a bit more consistent would not be bad.

I think the credit system is a useful incentive that would be more useful if it were more accurate and did not encourage folks to take actions that reduce GIMPS's efficiency.

S485122 · 2007-05-15, 10:20

Quote:

Originally Posted by cheesehead

I think the credit system is a useful incentive that would be more useful if it were more accurate and did not encourage folks to take actions that reduce GIMPS's efficiency.

But that is exactly my point !

petrw1 · 2007-05-15, 15:50

I agree that PetaFlops would be more CPU Speed and Era Agnostic ... that is you could always relate to it, now and in the future, whereas at the present it is difficult to relate to P90 speeds since most of us have long since converted our P90 (and similar) to boat anchors.

That being said, I think it would be a major effort to convert the current standings and a major mind shift to think in terms of new units ... kind of like many countries went through a generation ago when we converted from Imperial measurements to Metric. I'm not sure the perceived (or real) benefits are worth this effort ... but like the Metric system, if it did change we would all eventually get used to it.

On the other hand, converting the Status chart from P400 (boat anchors in waiting) to something like P3000 would be much easier to implement, to wrap our heads around and to keep current in the future assuming it continues to be updated every year or so as hardware changes.

And all this being said I will close with: "Whether anything or nothing changes I will not lose sleep over it."

2007-05-09, 22:15	#1
petrw1 1976 Toyota Corona years forever! "Wayne" Nov 2006 Saskatchewan, Canada 2×7×13×29 Posts	Considering current hardware on the status page http://www.mersenne.org/status.htm has 2 columns that reference OLD hardware. 1. P90 CPU Years: I can see a valid case made to continue to report points in P90 years as that has always been the scale. Mind you, if you converted this to current hardware (something like P4 - 3Ghz) all the numbers would reduce proportionally without changing the relative placings. As well the almost overwhelming total of almost 23 million years at the bottom would change to a more apparent 20 years. Obviously these are just relative numbers and changing them in NO way changes the project duration. In any case I won't lost sleep over this one. 2. PII-400 Speec (sec): I am supposing this column is meant as a benchmark to give members an idea of how long a particular LL/DC assignment will take. However, per http://www.mersenne.org/primenet/status.shtml, only a small fraction still have PCs of that approximate size. Changing this column to report in more recent/popular sizes (P4 - 3 Ghz?) and occassionally updating it would be of more use as a benchmark.

2007-05-10, 04:28	#4
Xyzzy Aug 2002 2171₁₆ Posts	Attached Thumbnails

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Newbie question about current users hardware running GIMPS	JonRussell	Hardware	42	2017-09-13 17:10
Current status	fivemack	NFSNET Discussion	97	2009-04-17 22:50
Current status	fivemack	NFSNET Discussion	90	2006-11-13 13:37
Current Status	moo	LMH > 100M	0	2006-09-02 01:15
Current Status of the Cunningham Tables	rogue	Cunningham Tables	4	2005-06-10 18:28

2007-05-10, 01:34	#2
TimSorbet Account Deleted "Tim Sorbera" Aug 2006 San Antonio, TX USA 10267₈ Posts	I think that the P90 years should stay the same forever, and that the PII-400 years should be updated every now and then to stay relevant.

2007-05-10, 18:25	#6
S485122 "Jacob" Sep 2006 Brussels, Belgium 11101110011₂ Posts	I think the units should be hardware independent : number of elemental operations done times the amount of affected data. That way all type of PrimeNet work is on equal footing, trial factoring LL tests AND P-1 factoring. Of course this would benefit new hardware because it can do more elemental operations in fewer steps. One could define the "elemental operations" by the mathematical operations done : multiply, subtract, add and compute a reminder or a modulo... It does not need to be precise to the bit, for instance a LL test could be computed as if all steps where done with the same amount of data, even if the first steps are lighter. For instance the LL test of M7 would get a weight of 5 squarings over 7 bits, 5 subtractions over 7 bits and 5 modulos over 7 bits. Counting the square as the number of bits additions it gives 757+57+57=45 operations over 7 bits or 315 obs (operation bits). For an exponent N it would be N-2 multiplications over N bits, N-2 subtractions over n bits and n-2 modulos over n bits. N(N-2)N+(N-2)N+(n-2)N=N(N-2)(N+2)=N³-4N. And since the exponents that PrimeNet has worked on have always been very big the -4N term can be ignored. I ignored the FFT multiplication method in this crude sketch. And it gives weird results since in my formula the computed credit is proportional to the cube of the exponent, with PrimeNet the credit is roughly proportional to the square of the exponent. But the idea is to have a simple formula for LLR testing, trial factoring and P-1 factoring. The advantage is simplicity, hardware independence and a kind of mathematical justification. Jacob

2007-05-14, 19:43	#8
S485122 "Jacob" Sep 2006 Brussels, Belgium 1,907 Posts	The problem is just that the PrimeNet credits are not really related to Petaflops except for LLR work and that some work is over credited and other under credited. All numbers following are P90 CPU years per day. On a PIII an unsuccessful factorisation is worth 0,020, a successful one 0,080 On a Q6700 unsuccessful factorisation is worth 0,200, successful factorisation is worth 0,380, an unsuccessful LL test 0,300, successful P-1 0,480 and an unsuccessful P-1 is worth only 0,001. Since it is the same hardware, that the P-1, finding a factor or not, used 140 times more memory (3600 MB instead of 25 MB) the credit for all type of works should be more or less equivalent. This is why I think that a more objective unit should be used. Of course there would still be differences between different types of work units depending on exponent size and type of work. If one is after PrimeNet or GIMPS credit it is evident that under the current system one should not invest in memory to do P-1 tests. On the contrary, one should change the 0 to a 1 in the worktodo.ini so as to avoid that time consuming step. I agree that the credit system is not a goal as such, but making it a bit more consistent would not be bad. Jacob

2007-05-15, 15:50	#11
petrw1 1976 Toyota Corona years forever! "Wayne" Nov 2006 Saskatchewan, Canada 2×7×13×29 Posts	I agree that PetaFlops would be more CPU Speed and Era Agnostic ... that is you could always relate to it, now and in the future, whereas at the present it is difficult to relate to P90 speeds since most of us have long since converted our P90 (and similar) to boat anchors. That being said, I think it would be a major effort to convert the current standings and a major mind shift to think in terms of new units ... kind of like many countries went through a generation ago when we converted from Imperial measurements to Metric. I'm not sure the perceived (or real) benefits are worth this effort ... but like the Metric system, if it did change we would all eventually get used to it. On the other hand, converting the Status chart from P400 (boat anchors in waiting) to something like P3000 would be much easier to implement, to wrap our heads around and to keep current in the future assuming it continues to be updated every year or so as hardware changes. And all this being said I will close with: "Whether anything or nothing changes I will not lose sleep over it."