New baseline machine / figuring stats

[Prime95]

Originally Posted by ebx

In our new server code, make sure we define a dividor, or baseline if you would. If it is 1, we have P90 year. When it is 10, we have P900 year.

One other option. Let the user of the stats pages decide his preferred reference via a list box (classic P-90 days, modern GHz days, petaflops, whatever). You are right that it's all a simple matter of applying a different divisor.

The downside: Person A posts question about his stats in GHz-days and confuses the users that like to use P-90 years.

Anyway, we've probably beaten this horse to death.....

[QuintLeo]

1 Ghz machines aren't hard to find at all.

My company just bought a BUNCH of them earlier this year - from Dell.

I was *expecting* to get assigned a 1.1 Ghz Celeron box, from researching the Dell website at the time - instead I got a 1.0 Ghz Pentium-III box, which wasn't even listed as an available option at the time.


VIA C3 should be up to, or just hit, 1 Ghz as well - though I can NOT recommend them for any serious usage, their performance per clock is *horrible*. Tolerable for appliance servers, low-end home usage, and the power consumption makes them a good Blade Server choice (though the Crusoe blows them away in blade service in pretty much all ways).

Personally, I don't really *care* what the baseline is - I see valid arguments on all sides - but my K5 farm has a mild preference for staying with P90 years. 9-)

11 machines (14 CPUs) and growing....

[tha]

Originally Posted by Prime95

When contemplating the difficulty of a server rewrite a year ago, I pretty much convinced myself to change to GHz-days. That is the amount of work a 1 GHz machine (probably PIII) would accomplish in one day.

I think it would be good to use a benchmark that is based on the SSE2 LL-code, so maybe a 2.4 Ghz 'Pentium 4' divided by 2.4 to obtain 1 GHz.

YotN,

Henk.

[QuintLeo]

I don't like using SSE2 code as the baseline for "general" usage - the large majority of machines used are NOT SSE2 capable, though those machines do process a disproportionate amount of the LL work.

This opinion will likely change in 2-3 years, once P4s get more common and Hammers start displacing Athlons....

[cheesehead]

... Giddyup, dead horse! Beat, beat, beat! Giddyup, dead horse! Beat, beat, beat! Giddyup, dead horse! Beat, beat, beat! ...

- - -

Folks, this same thing happens every time someone proposes changing the GIMPS standard work unit reference -- there erupts a squabble over exactly which new unit to substitute. In particular, every new choice of specific CPU is objectionable to some faction.

Five and a half years ago, in September 1997, those wanting a change were proposing:

"... the CPU commonly sold today, the 200MHz P55c (aka Pentium w/MMX)"

and

"... it should be updated a little past a P90 to maybe a P133"

Did either of those suggestions stand the test of time? No.

_If_ we had adopted one of those then, would that have kept anyone from proposing another change now? No.

Will some object to adopting a new SSE2 P4 standard now? Yes.

Will some object to adopting a new non-SSE2 standard now? Yes.

If we adopt a new standard now, will that be satisfactory to all for the next five years? No. ... the next five months? No. ... the next five weeks? No.

There is, however, one advantageous reason to keep the current standard based on P90 performance that no other proposed CPU standard can ever match: The original system that George Woltman used when he founded GIMPS was a P90.

The original GIMPS work unit has historical reference value that cannot be equalled by any other unit.

- - -

Having said that, I agree it would be nice to have a feature that allows people to see their stats in their own favorite choice of unit, converted from the standard P90 CPU-year unit.

As George noted earlier, there's potential trouble if people discuss stats while unknowingly using different units. I don't think any serious misunderstanding would last too long ... but then there was that Boeing 767 that ran out of fuel in midflight over Canada because someone confused the conversion factor for pounds with the one for kilograms ... and the Mars Climate Orbiter spacecraft that was lost because one team used kilometers while another used miles ...

[tha]

Originally Posted by cheesehead

The original GIMPS work unit has historical reference value that cannot be equalled by any other unit.

This statement must have come from an american or british citizen. There is only one internationally recognized system, which is so because of its internal consistency. The metric system. We wouldn't have conversion factors if people would stop using obsolete, internally inconsistent and even non decimal systems.

P90*year is indeed the historic measure, but GHz*day as George considered is less trivial since it eliminates an arbitrary factor 90. Of course conversion to Mflops based on the conversionfactor that is used for quite some time now in this project would do as well.

[dswanson]

Originally Posted by cheesehead

and the Mars Climate Orbiter spacecraft that was lost because one team used kilometers while another used miles ...

Just for the record, the conversion error was between pounds and Newtons (the English versus metric units of force or weight). They differ by a factor of 4.448. In order to calculate the trajectory of the spacecraft, the navigation folks needed to correct for the force applied by occasional thruster firings. This force was given to them by the engineers who built the spacecraft. The spacecraft engineers handed over the numbers in pounds, but the navigation team interpreted them in Newtons. Because the thrusters were used only rarely and for short pulses, the difference wasn't noticed during the cruise to Mars. But by the time MCO reach Mars, the error amounted to tens of miles (that's sixteens of kilometers for you metric purists). That's not much of an error when you've just traveled a few hundred million miles, but it's fatal when you're trying to just skim the top of an atmosphere and miss on the low side.

The Yahoo News link is vague, so I don't blame cheesehead for guessing the wrong units.

[Tasuke]

Originally Posted by tha

This statement must have come from an american or british citizen. There is only one internationally recognized system, which is so because of its internal consistency. The metric system.

Most of the metric system is based on an arbitrary standard and because of this, P90 works just fine.

Some facts:

Kilogram is the only SI base unit to incorporate a prefix. (By the way, teachers have been known to test that fact.)

1 milliliter = 1 cubic centimeter
1 milliliter of water has a mass of approximately 1 gram
1 liter of water has a mass of approximately 1 kilogram
1 cubic meter of water has a mass of approximately 1 metric ton
0 Degrees celcius is the freezing point of water
100 Degrees celcius is the boiling point of water
absolute zero is actually 273.15 degrees below 0 <--why not use kelvins?

Here is the definition of the kilogram, from 1901:

"... it is equal to the mass of the international prototype of the kilogram."
There is a cylinder of 90% platnium/10% iridium in a vault in Paris, France and that lump of stuff has a mass of exactly one kilogram, by definition.

1792
The U.S. Mint was formed to produce the world's first decimal currency (the U.S. dollar consisting of 100 cents).

The definition of a meter
1793 1 / 10 000 000 of the distance from the pole to the equator.
1906 1 000 000 / 0.643 846 96 wavelengths in air of the red line of the cadmium spectrum. <-- not decimal
1960 1 650 763.73 wavelengths in vacuum of the radiation corresponding to the transition between levels 2p10 and 5d5 of the krypton 86 atom. <--noticing a patern
1983 Length traveled by light in vacuum during 1/299 792 458 of a second. <--you following

It does not matter what the metric is, it just has to be a measurable and equal 1. P90 years fit the bill as good as any of the metric definitions.

The metric system was a way to setup decimal relationships from within a unit of mesure, not to try to make everything fit.

Tasuke, a damn american that knows how to use the metric system, being an electrical engineer.

/Edit for errors

Originally Posted by dswanson

A nit on an otherwise excellent dissertation: The lump has a MASS of exactly one kilogram.

[dswanson]

Originally Posted by Tasuke

Here is the definition of the kilogram, from 1901:

"... it is equal to the mass of the international prototype of the kilogram."
There is a cylinder of 90% platnium/10% iridium in a vault in Paris, France and that lump of stuff weighs exactly one kilogram, by definition.

A nit on an otherwise excellent dissertation: The lump has a MASS of exactly one kilogram. What it weighs depends on where you weigh it. If you weigh it in Paris, France, you will get a reading of one kilogram-force (kgf), or what is conventionally thought of as a weight of one kilogram. But if you took it to the Moon it would only weigh 0.17 kgf. If you took it to Mars it would weigh about 0.33 kgf. If you weighed it while in orbit around Earth it would weigh nothing. But in all locations its mass is exactly the same, 1 kilogram.

Sorry, I'm horribly off-topic. I'll climb down off my soapbox now.

[trif]

Not only that, but I think that Cheesehead meant that the P90 has historical reference value that cannot be equalled by any other unit that might be used for measuring GIMPS work, and that is because the program was originally timed on a P90. No other such unit can make that claim, and I don't think it was intended to say anything about metric or English units.

Even if we were to move to a mythical GHz, the question arises of *which* GHz to measure. A P90 is pretty stock, but a GHz of Athlon is not equal to a GHz of PIII, and definitely not equal to a GHz of Northwood, and we have no idea what a GHz of Hammer is going to look like.

[ET_]

A P90 is pretty stock, but a GHz of Athlon is not equal to a GHz of PIII, and definitely not equal to a GHz of Northwood, and we have no idea what a GHz of Hammer is going to look like.

That's right.
We actually know the conversion factor between P90 years and MFLOPs, so we implicitly also know how "fast" a PIII or an Athlon is compared to a P90.

I should add that while MFLOP is based on seconds, a metric unit recognized by MKS, CGS and International System, comparing floating point operations on different processor families (like Intel, PowerPCs PARCs, Alphas and RISC) leads to incongruence, as they have different microcoded instructions; we could at most talk of MFLOPs relative to Intel architecture, that's another conversion/convention.

But again, Intel has had different processor families, anf George squeezed the most from each one, starting from 80486 on; he and Scott metered the power of a P90, and each new version of both software and hardware can be rated using that value.

That's why I think that P90 years can reasonably mantained as a baseline.

Luigi