Assignment Strategy

[diamonddave]

It occurs to me that we are TFing "No P-1" Exponent faster then we can effectively P-1 them.

Right Now we have released 2,436 in the LL Range, yet we are sitting on more then 730 exponent TF to 72 that just need a P-1 test (Not counting the 1200 assigned witch I don't know the bit-level).

Perhaps we should change the default of the "Only take Exponents with no P-1 work done." option in the assignment page so that we produce more work that can be returned faster to PrimeNet?

By the same token I think we are currently dedicating too much fire-power to DC-TF

in the last week PrimeNet did:

DC : 912
LL : 1301

So clearly we need to release LL exponent faster then DC.

Yet this is what we released so far:

DC : 4200
LL : 2436 (+730 available exponent waiting P-1 at the 72 bit-level)

Personally I changed to TF in the LL range until levels get better.

[KyleAskine]

Originally Posted by diamonddave

It occurs to me that we are TFing "No P-1" Exponent faster then we can effectively P-1 them.

Right Now we have released 2,436 in the LL Range, yet we are sitting on more then 730 exponent TF to 72 that just need a P-1 test (Not counting the 1200 assigned witch I don't know the bit-level).

Perhaps we should change the default of the "Only take Exponents with no P-1 work done." option in the assignment page so that we produce more work that can be returned faster to PrimeNet?

By the same token I think we are currently dedicating too much fire-power to DC-TF

in the last week PrimeNet did:

DC : 912
LL : 1301

So clearly we need to release LL exponent faster then DC.

Yet this is what we released so far:

DC : 4200
LL : 2436 (+730 available exponent waiting P-1 at the 72 bit-level)

Personally I changed to TF in the LL range until levels get better.

I always do LL as well - it is unquestionable it is more helpful that TF DC.

Remember though, that to release a DC, we generally only have to TF one level. It is usually three for LL.

[chalsall]

Agree completely diamonddave.

I had planned to start releasing LL candidates which had been TFed to at least 72 bits, but not P-1ed after we have 1000 of them pending assignment.

The good news is my last call for P-1 work had a reasonable response, and the number awaiting assignment to P-1 workers dropped by several hundred.

With regards to DC vs. LL, again I think your suggestion is a good one. The good news there, also, is there are a finite number left to do -- the ranges above where we're currently working have already mostly been taken to 69.

But, as always with GIMPS, individual workers can decide what they want to work on with their hardware and their electrical bill.

[petrw1]

Originally Posted by chalsall

The good news is my last call for P-1 work had a reasonable response, and the number awaiting assignment to P-1 workers dropped by several hundred.

Guilty.
I decided to fill up my 13 cores to year end and took 210 myself....just over 2 weeks work.

[petrw1]

Originally Posted by diamonddave

By the same token I think we are currently dedicating too much fire-power to DC-TF

Some might disagree....in fact at one time George said we could use more P-1 and DC workers. That is partly based on the fact that DC is about 28M behind LL.

If theoretically we continued to complete 912 per week these 28 ranges will take about a dozen years.

And if we continue to complete more LL than DC it will only get farther and farther behind.

TF'ing the DC ranges reduces the number of DC tests and directly helps the catch-up / keep-up effort.

[chalsall]

Originally Posted by diamonddave

Perhaps we should change the default of the "Only take Exponents with no P-1 work done." option in the assignment page so that we produce more work that can be returned faster to PrimeNet?

OK -- I've removed the "No P-1 work done" option entirely since, as you pointed out, it's holding back returning assignments to PrimeNet; it doesn't significantly change your chances of finding a factor (only two out of 48 P-1 tests coordinated by the system found a factor below 72 "bits"); and there were only about 700 left to be assigned anyway.

[Dubslow]

Originally Posted by petrw1

Some might disagree....in fact at one time George said we could use more P-1 and DC workers. That is partly based on the fact that DC is about 28M behind LL.

If theoretically we continued to complete 912 per week these 28 ranges will take about a dozen years.

And if we continue to complete more LL than DC it will only get farther and farther behind.

TF'ing the DC ranges reduces the number of DC tests and directly helps the catch-up / keep-up effort.

But we're already doing significantly more work to find a DC factor than to find an LL factor, at far less work saved. I think more exponents would actually be cleared if the CPU core running mfaktc and the GPU were both put towards DC, rather than DC-TF.

I would back this up with numbers, but @chalsall, gpu.mersenne.info appears to be down. (And also there's a / in the <title> for mersenne.info)

[diamonddave]

Originally Posted by chalsall

OK -- I've removed the "No P-1 work done" option entirely since, as you pointed out, it's holding back returning assignments to PrimeNet; it doesn't significantly change your chances of finding a factor (only two out of 48 P-1 tests coordinated by the system found a factor below 72 "bits"); and there were only about 700 left to be assigned anyway.

[chalsall]

Originally Posted by Dubslow

I would back this up with numbers, but @chalsall, gpu.mersenne.info appears to be down.

No it's not. I'm logged into it right now via SSH, and several people are successfully browsing it as I "speak".

Originally Posted by Dubslow

(And also there's a / in the <title> for mersenne.info)

Could you please be a little more specific?

[chalsall]

Originally Posted by Dubslow

But we're already doing significantly more work to find a DC factor than to find an LL factor, at far less work saved. I think more exponents would actually be cleared if the CPU core running mfaktc and the GPU were both put towards DC, rather than DC-TF.

Actually, that's only half true...

DC TF: 6.401 GHz Days Work Done, 24.797 GHz Days Work Saved.
LL TF: 12.675 GHz Days Work Done, 182.163 GHz Days Work Saved.

(Average per factor found. GHz Days Work Done is "Normalized".)

[diamonddave]

Originally Posted by petrw1

Some might disagree....in fact at one time George said we could use more P-1 and DC workers. That is partly based on the fact that DC is about 28M behind LL.

If theoretically we continued to complete 912 per week these 28 ranges will take about a dozen years.

And if we continue to complete more LL than DC it will only get farther and farther behind.

TF'ing the DC ranges reduces the number of DC tests and directly helps the catch-up / keep-up effort.

I agree with you

But the point is to keep the Beast feed (PrimeNet). Right now the DC part is overflowing a bit (~600 exponent before the wavefront are available to DC) but the LL part isn't (hard to give a number with all the "eyesore" in the PrimeNet Summary Report) clearly the wavefront at 56M is still advancing with now only 1090 Exponent left.

A little test asking for 24 LL exponent gave me:

45M - 1
56M - 23

So we have a 1 exponent buffer...