LL tests more credit-efficient than P-1?

[ixfd64]

I've assigned three desktops (all with i5-650 processors) to do P-1 factoring exclusively. From what I've noticed, each P-1 assignments takes about 1.8 days to complete gives exactly 4.6787 GHz-days of credit, provided that no factor is found in stage 1.

4.6787 / 1.8 = 2.6 GHz-days / (core * day)

However, another i5-650 desktop is LL-testing a 52M exponent. Each iteration takes 0.058 seconds; at this rate, it would take about 35 days for the assignment to finish.

An exponent of this size gives about 110.5 GHz-days of credit, or about 3.15 GHz-days / (core * day).

Is there an explanation for this discrepancy? Are LL tests more computationally intensive? Or does the code for calculating P-1 credit need to be adjusted?

[Uncwilly]

You may want to see this thread:
http://www.mersenneforum.org/showthread.php?t=11411

[nucleon]

How much memory did you allocate to P-1? I find allocating 500MB+ gave quicker results, than say 200MB.

Sorry I don't have exact figures. But I have 4x cores doing P-1 stage at the time, the 500MB core was sitting at 40% compelte, while the other was sitting at 25% ish. Note this figures are from memory. It's more qualative comparison than quantitive.

-- Craig

[ixfd64]

I'm using 384 MB per core.

[Parshan]

Originally Posted by nucleon

How much memory did you allocate to P-1? I find allocating 500MB+ gave quicker results, than say 200MB.

Sorry I don't have exact figures. But I have 4x cores doing P-1 stage at the time, the 500MB core was sitting at 40% compelte, while the other was sitting at 25% ish. Note this figures are from memory. It's more qualative comparison than quantitive.

-- Craig

How much will it affect the processing if i allocate 700MB+?

[kriesel]

What software are you running P-1 upon? If prime95 v30.8 or later, which have a superior algorithm, see https://www.mersenneforum.org/showth...130#post614130 and its attachments, and https://mersenneforum.org/showthread.php?t=28038. Mlucas and mprime both benefit from lots of GB of ram available for P-1 stage 2. More ram allowed = more buffers can be used = more efficiency, until the point thrashing or sluggishness of other applications sets in.

[petrw1]

Originally Posted by ixfd64

I'm using 384 MB per core.

That seems low; might be all you have though.
You need to be at version 30.8+.
It's been out for almost 2 years
It is very fast in Stage 2 with high RAM; high being about 16Gig. And, assignments with larger B2 are more credit generous.
There's a whole thread on this here:
https://www.mersenneforum.org/showthread.php?t=27366

[kriesel]

Originally Posted by petrw1

That seems low; might be all you have though

Possibly because the first four posts in this thread are from over 12 years ago. https://www.gsmarena.com/countercloc...news-30756.php
Then, ~1 GiB system ram was common. Now, 16 GiB and up are. I have an old dual-xeon doing production first-test wavefront P-1 with 24 GiB installed, and that's on the low side for such work in V30.8 mprime/prime95 or later. Other systems with 64 GiB, 128 GiB, & 1 maxed out at 1/2 TiB for P-1 on large exponents (~250. GiB allowed / stage 2 instance). However, the new polynomial multiplication & high pairing efficiency of v30.8 delivers additional performance down to around 8 GiB allowed or somewhat less.

I like how George adapts his software to changed circumstances over time. (Both hardware advances, and number theory.)

[petrw1]

Originally Posted by kriesel

Possibly because the first four posts in this thread are from over 12 years ago.