PSI, mpg, GHz P4-days and P90 years

[cheesehead]

Originally Posted by davieddy

BTW what was your PC-independent standard?

On a previous occasion (IIRC, about September of whichever year it was) when someone complained that the unit of work needed to be "updated" from the P90 CPU-year to a base related to a then-current CPU model, I posted a sincere, though heavily larded with humor, retort. I started by pointing out (just as others have here) that "updating" to some unit based on a then-current CPU would just find itself similarly "obsoleted" a few years later. I pointed out that the P90 CPU-year had a historical relevance that no other standard could have -- that it was benchmarked on the very CPU on which George Woltman had started GIMPS!

Then I presented a humorous, but sincerely intended, example of a PC-independent standard unit of work: the mass of turkey that could be cooked with the waste heat generated by a PC running an L-L test on a particular exponent. (The Canadian and U.S. Thanksgiving holidays were in the near future.) Also to be included was a waveguide steering the radio-frequency emissions from the CPU to the combination microwave-convection oven to assist in the cooking along with the heat.

There are, of course, real difficulties with such a standard if attempted in reality, such as differing amounts of heat and RF generated during such a trial by different models of CPUs, varying rates of cooking of differing types of turkey tissue, and so on ...

My intent was to show that any unit of measure of GIMPS work would necessarily be arbitrary in some sense. Thus, there was no fundamental reason to abandon the P-90 CPU-year in favor of some other CPU-time unit.

However, I find that that thread seems to have disappeared from this forum! (Else I'd have posted a link.)

I think that thread should be restored.

I am distressed to learn that the C2GHD is to be the new standard, and ask for a reversal of that decision.

[S485122]

Originally Posted by cheesehead

My intent was to show that any unit of measure of GIMPS work would necessarily be arbitrary in some sense. Thus, there was no fundamental reason to abandon the P-90 CPU-year in favor of some other CPU-time unit.
[...]
I am distressed to learn that the C2GHD is to be the new standard, and ask for a reversal of that decision.

I agree that any CPU based unit would be arbitrary, and if the units are CPU based I say "P90 years for ever !"

It is possible to create a CPU independant measure though : to LL test exponent N, you have to do N-2 squarings followed by a substraction modulo 2^N-1, almost all of the numbers will have about the size of 2^N-1 so the weight of the test is N-2 *squarings of 2^N-1. Such a unit would penalise exponents near the low limit for a FFT size. So it could be refined by including the FFT size, but it could be argued that the FFT size is dependant on the instruction set supported by the CPU... Still it could be argued that this unit would reflect the work done by the algoritm on the processor that did do the work and this is different from crediting the work that would have been done by a particular version of the program on an particular processor. But is this worth the effort ?

What should be changed urgently (in my opinion) is the way trial and P-1 factroring is measured : it should earn a user as much credit per time unit spent on a processor as LL testing on the same processor. This is especially urgent as P-1 work will be an independant type of workunit on the v5 server, but I think the credit awarding system has already been changed for factoring jobs on the v5 server.


Jacob

[petrw1]

I was part of an earlier thread too where we discussed P90 years ... I can't remember my line of commentary but it earned me the phrase "1976 Toyota Corona years forever!" over my Avatar.

Anyway for the record as I mentioned earlier in this post I thought we were going to use GFLOPS ... works for me.

[Xyzzy]

http://www.mersenneforum.org/showthread.php?t=8129

[jinydu]

Sorry to resurrect an old thread; but I thought I would weigh in with one more reason to continue using P90 years that isn't mentioned in this thread:

The GHz-day is a smaller unit than the P90 year. At least in spirit, this is contrary to what normally happens in the computer world. The relentless march from kilobyte to megabyte to gigabyte to terabyte gives a feeling of progress, that we should raise our expectations over time. Switching to a smaller unit could give some people a feeling of "going backwards". The change from 'year' to 'day' could worsen this feeling that standards are going down.

(Ok, the Kb/Mb/Gb/Tb analogy isn't perfect, as all those units are part of the same system; but I think the point still holds)

EDIT: By this argument, one could make a case for the GHz-year. But I don't think that would be a good idea either. 1 P90 year would be worth about 0.0139 GHz-years. Seeing a first test earn barely more than 0.1 GHz-years could discourage a new user. In this regard, I think the P90 year is actually quite a good unit for this stage of the project. Almost all tests earn between 1-10 P90 years, numbers large enough to not discourage new users but small enough to not make the unit look "cheap", a feeling that is reinforced by the word 'year', which conveys an 'impressively long time'.

EDIT 2: Come to think of it, if we continue to follow the reasoning of the previous paragraph, the best progression of units for the future would be: P90 year, P90 decade, P90 century, P90 millennium... It would be easy to convert between the units, we could continue to make most tests worth between 1 and 10 units, and the names would sound increasingly impressive.

[retina]

I find nothing wrong with the much simpler FLOPS. kFLOPS, MFLOPS, GFLOPS, TFLOPS etc. I don't know why we have the strange P90 years, GHz-Days, atto-parsecs per micro-fortnight, nano-century-clocks per peta-hurts-days and other seemingly unneeded specially made up units.

[davieddy]

Originally Posted by retina

I find nothing wrong with the much simpler FLOPS. kFLOPS, MFLOPS, GFLOPS, TFLOPS etc. I don't know why we have the strange P90 years, GHz-Days, atto-parsecs per micro-fortnight, nano-century-clocks per peta-hurts-days and other seemingly unneeded specially made up units.

Since you mention parsecs, I might suggest that the problem of
quantifying computing work is comparable to that of measuring
cosmological distance, or temperature before the thermodynamic
scale was established.

David

[cheesehead]

Originally Posted by retina

I find nothing wrong with the much simpler FLOPS. kFLOPS, MFLOPS, GFLOPS, TFLOPS etc.

Fine, but those are all speeds of computation (floating-point operations per second), not amounts of computation (total number of operations) accomplished.

If we said that an L-L test was credited with, say, 900 MFLOPS, an appropriate response would be, "Yes, but for how long did that speed have to be maintained in order to finish the L-L test?" So then we'd have to multiply that by time needed to complete the test (e.g., 900 MFLOPS * 24 days = 21600 MFLOPS-day or 21.6 GFLOPS-day), to quantify the amount of computation accomplished.

If what you want is just a speed, without regard to how long that speed is maintained to accomplish something, FLOPS are fine.

I don't know why we have the strange P90 years, GHz-Days,

Those are speed of computation multiplied by time over which that speed was maintained, to give total computation accomplished.

P90-year = how much computation a P90 (Pentium at 90 MHz) could accomplish in a year

GHz-Day = how much computation a 1-GHz CPU could accomplish in a day.
(It's also how much a 2-GHz CPU could accomplish in half a day, or how much a 333-MHz CPU could accomplish in 3 days)

atto-parsecs per micro-fortnight,

Now, you're back to speed (distance per time) again -- though with a silly combination of prefixes and probably-impractical choice of units.

nano-century-clocks per peta-hurts-days

Okay, that's whacko.

[retina]

Originally Posted by cheesehead

Fine, but those are all speeds of computation (floating-point operations per second), not amounts of computation (total number of operations) accomplished.

Okay, I just looked it up, you're correct. What I mistakenly though it meant was FLoating-point OPerationS.

Originally Posted by cheesehead

If we said that an L-L test was credited with, say, 900 MFLOPS, an appropriate response would be, "Yes, but for how long did that speed have to be maintained in order to finish the L-L test?" So then we'd have to multiply that by time needed to complete the test (e.g., 900 MFLOPS * 24 days = 21600 MFLOPS-day or 21.6 GFLOPS-day), to quantify the amount of computation accomplished.

Okay, my use of FLOPS was bad. But substitute for a unit that means floating-point operations and that is what I would like to see. Maybe FLO? (makes me think of Andy Capp)

Originally Posted by cheesehead

If what you want is just a speed, without regard to how long that speed is maintained to accomplish something, FLOPS are fine.

Those are speed of computation multiplied by time over which that speed was maintained, to give total computation accomplished.

P90-year = how much computation a P90 (Pentium at 90 MHz) could accomplish in a year

Sorry I didn't mean to mislead, I knew what p90-years was intended to be, my post was meant kind of rhetorically.

Originally Posted by cheesehead

GHz-Day = how much computation a 1-GHz CPU could accomplish in a day.
(It's also how much a 2-GHz CPU could accomplish in half a day, or how much a 333-MHz CPU could accomplish in 3 days)

Now this one is not a good choice. What is a CPU? Really! I don't mean what does CPU stand for, but is a CPU a Pentium core or a PII core, or Core2 ... umm ... core? They are clearly not the same thing. And once AVX starts becoming mainstream the lines become even more blurred.

Originally Posted by cheesehead

Now, you're back to speed (distance per time) again -- though with a silly combination of prefixes and probably-impractical choice of units.

Okay, that's whacko.

[cheesehead]

Originally Posted by retina

Now this one is not a good choice. What is a CPU? Really! I don't mean what does CPU stand for, but is a CPU a Pentium core or a PII core, or Core2 ... umm ... core? They are clearly not the same thing.

See? That's one reason I continue to defend using the P90-year as an appropriate GIMPS unit of work measurement -- it specifies a specific CPU (original Pentium) at a specific clock rate (90MHz) maintained for a specific length of time (year), thus together denoting a specific amount of computing accomplished.

And once AVX starts becoming mainstream the lines become even more blurred.

But we can always refer to the historical benchmark George established for how long his P90 (where GIMPS originated) took to perform an FFT. Then one can multiply by appropriate factors for other FFT lengths and clock rates.

Then one can compare the total P90 CPU time needed to perform an L-L test on a specific exponent to the times required by other CPU types to perform that same test, and derive conversion factors to account for the differing efficiencies with which other CPU types perform L-L tests compared to a Pentium (I).

Then we can use all these numbers to calculate how much credit to give, measured in units of P90-years, for an L-L test performed on any of those CPU types.

Sure, that's a number of conversion factors -- but there will be just as many conversion factors needed for _any other_ choice of basic measurement unit!

Sure, there are still other variable factors -- but those factors will be there _regardless_ of the basic unit of measurement.

No other unit of measurement has any real technical advantage over the P90-year, some of them might be slightly simpler to understand, on the surface, by newcomers than the P90-year, but no other unit can _ever_ match the P90-year in regard to historical significance for GIMPS.

[Uncwilly]

Originally Posted by cheesehead

Sure, that's a number of conversion factors -- but there will be just as many conversion factors needed for _any other_ choice of basic measurement unit!

I thought that P90 years was the SI base unit of computing power.