Davieddy's World

[kracker]

Originally Posted by davieddy

Who deleted my reply to this post?

(PM me for it, assuming you are over 18.)

...
Go to another forum for that.


This thread has REALLY...

[davieddy]

Originally Posted by kracker

...
Go to another forum for that.

Where and for what exactly?
Surely you can't be talking about "it", can you?

[swl551]

[NBtarheel_33]

I concur. This thread is about the closest that this forum has ever come to being "non-family-friendly". I dearly hope that this is not to become a trend.

[Uncwilly]

Originally Posted by NBtarheel_33

I concur. This thread is about the closest that this forum has ever come to being "non-family-friendly". I dearly hope that this is not to become a trend.

I think that it has crossed the line. I think the Davie needs to be sat in the corner for a couple of weeks.

[davieddy]

Originally Posted by Aramis Wyler

Anything can happen here, it's a forum thread of joy and anguish, free love and bitterness, sarcasm and shame. Maybe we'll even do fan fiction!

Originally Posted by davieddy

Yep.
Most of us seem to be entering into the spirit of the thread.
I hope it stays in this subforum, because although we seem to veer towards the Soapbox genre from time to time, we also get heated about Primenet and GPUto72 occasionally.

To restore a more highbrow tone...

I spoke too soon

Originally Posted by Uncwilly

I think that it has crossed the line. I think the Davie needs to be sat in the corner for a couple of weeks.

Remedial courses are available for the culturally deprived, bad boy

[kracker]

Originally Posted by davieddy

I spoke too soon

Remedial courses are available for the culturally deprived, bad boy

"This video contains content from EMI, who has blocked it in your country on copyright grounds."

[davieddy]

Originally Posted by davieddy

Indeed.
A week is a good time to choose as our sampling period, since the submission rate fluctuates during it.

Allow me to cast my trained eye over last week's figures:

2300 LLs completed.
2300 TFed to 73+, of which ~1000 were taken to 73 bits exactly.

A good week's work all round, and as I have been saying all week:
IF IT AIN'T BROKE, DON'T FIX IT.

Cheers,
David

Originally Posted by Uncwilly

How many of those that were taken to 73 were done so outside of GPUto72?
Remember that Chris in his day job manages networks, etc. Watching for trends is part of what he does.

If 1000 were taken to 73, what about the other 1300, were they taken to 74?
If so, that would be ~31 LL tests saved, beyond those saved upto 73.
Taking them to 73 saved ~32 LL's, beyond those saved taking to 72.
Taking them to 72 saved ~33 LL's, beyond those saved taking to 71.
Taking them to 71 saved ~33 LL's, beyond those saved taking to 70 (the default PrimeNet level.)
This represent a speed up of the LL testing of ~5% (extra tests not needing being done). This also benefits the DC's that will happen in that range. This also frees up the P-1'ers do look for factors even higher (if I understand P-1 correctly, the 'floor' can get pushed up by selecting the right B1 value.)

Seems like GPUto 72 is a great thing. Chris is keeping it tuned for maximum effect, based upon available power.

To revert to the serious side of Davieddy's mad, bad and dangerous to know world:

It is desirable and feasible for each LL completion to be replenished with a new LL test TFed by GPU to at least 73 bits, and Chris has assured me (ad nauseam) that this is and will continue to be the case. This minimum bit level can be upped to 74 as soon as GPU firepower allows.

Let us suppose that a GPU is going to take an exponent from 71 to 73+.
If the time for 71 to 72 is T, then it will take 3T to go to 73, 7T to 74 and 15T to 75.
So TFing two exponents to 74 will save 50% more work in less than the time it takes to TF one exponent to 75.

So in answer to "Were the remaining 1300 TFed to exactly 74?", I am pleased to say "YES".

I think Chris and I just like a fight, even when we agree with each other 99%.

D

[Aramis Wyler]

Originally Posted by davieddy

So TFing two exponents to 74 will save 50% more work in less than the time it takes to TF one exponent to 75.

So in answer to "Were the remaining 1300 TFed to exactly 74?", I am pleased to say "YES".

Technically, it would be 44% more work in 93% of the time. But you could say the same for factoring two of them from 69 to 70 instead of taking them from 69 to 71. There are always diminishing returns going up a bitlevel, but we try to get as close as we can to optimal with the firepower we have.

[davieddy]

I have come in for a lot of nasty stick because people deliberately suggest that:
1) I advocate what I do because it's what I want.
2) They intentionally misinterpret what I am advocating for effect.

I advocate what I do because
A) It is possible
B) Minimizes the expected time to the next prime discovery.

If there were a choice of options fulfilling these two criteria, I would say so, and let people do whatever they want. There isn't.

Primenet has TFed everything up to 90M to ~70 bits. Useless, but at least you might think comparatively harmless. But there is a backlog between 60M and 62.5M which is a hangover from this strategy, interfering with both GPUto73 and the allocation of new LLs. Also most GPUers would rather take an exponent from 69 to 74 than 72 to 73 or 74 (from studying Chris's tables), for obvious reasons.

ATM we could TF to 73 and exceed the LL completion rate, thereby increasing the the lead TF has over LL allocation (which IMO (note) is desirable). Then we can take as many to 74 as we can while keeping up with (or preferably slightly exceeding) the LL completion rate, which is increasing nicely, THX mostly to the uptake of AVX.

Glancing through the range of a billion, the only significant source of extra GPU fire power is from the grossly excessive DC-TF of which the self-styled "king" BCP19 is so proud (3000/week to 70 bits I ASK YOU). But enlisting this still wouldn't make taking everything to 74 sustainable, anymore than the generous offer to club together and buy me a GPU would. I do have several ideas about how we could recruit more firepower (or at least not deter it) however. More anon.

David

[Aramis Wyler]

I used to TF from 69 to 74. It is my favorite way to TF as far as satisfaction goes, because you get the low bitlevels that feel like basically free factors, and yet every number you release is either ready for release to primenet or only missing the p-1. I'd get tired of grinding them from 73 to 74 all day long, with other people skimming along the 69-70 range and racking up millions of hours saved.

Thing is though, I only thought that was a good idea because I thought primenet had a basic fitness function; that having a number (say 63,000,000) tf'd to 73 and another one (63,000,001) tf'd to 74 and p-1'd it would release the second number and any other similar numbers before it released the first one so that the first one would have more time to get factored. It doesn't - it releases from the bottom up. At that point I switched back to just factoring the lowest to 74, personal satisfaction be damned, because that was the only way to get the best factored numbers out to the LL'ers.

Personally, I'd love to see a fitness function alter the primenet distribution, but I don't see that happening.

What sounds like the worst case though, even worse than the rather poor case of doing what I used to do as described in the first paragraph, is to simply not do what factoring we can for the sake of increasing our lead time. Honestly I can't fathom any gain from increasing the lead time unless we are not currently covering the range of numbers being assigned. I'm not closed to that, I just don't see any upside.

I'm not really arguing anything there (except maybe to promote a fitness test to primenet, heh!) just my

I have good news though! We'll have more gpu power working on factoring in just a few weeks when the DCTF runs out. Currently My and Chuck's daily output on DCTF is higher even than Pete's (though we'll never catch up to the work he did) and then that should all go to LLTF.

EDIT: You may be surprised to discover, Davieddy, that bcp19 does twice as much LLTF work as he does DCTF work, despite still doing more DCTF than anyone else.