mersenneforum.org

mersenneforum.org (https://www.mersenneforum.org/index.php)
-   Factoring (https://www.mersenneforum.org/forumdisplay.php?f=19)
-   -   32/33 and 15e/16e crossover point (https://www.mersenneforum.org/showthread.php?t=11435)

fivemack 2009-01-31 19:50

32/33 and 15e/16e crossover point
 
These are with N=2^941-1, rlim=alim=200M, sieving 200M .. 200M+10^4

[code]
side lp siever yield time/rel
alg 32 15 6237 2.29344
alg 33 15 11693 1.21814
alg 32 16 13320 2.79733
alg 33 16 25553 1.45897
rat 32 15 8559 1.99026
rat 33 15 16440 1.03673
rat 32 16 17536 2.57357
rat 33 16 33922 1.30827
[/code]

which looks as if 32-bit large primes and 16e is the right way to go for numbers of this size (changing the siever doubles the yield at a fairly small cost in runtime; lp=33 doubles the yield and the number required at the same time so is no net benefit). Probably rational side 0-300M. rlim=alim=200M was a guess, I'll do some more runs to optimise that.

This would be a Big Calculation with capital Big; 2.6 seconds per relation and we need half a billion, so 40 CPU-years. 16e is a prodigious user of memory (about 4G virtual of which just over 1G used), so this may be more a project for people with clusters than for random home user - indeed, that might be a bit more of a strain on clusters than their administrators really want.

FactorEyes 2009-02-01 03:16

News I can use?
 
I may write this down somewhere.

Then again...:ermm:

Andi47 2009-02-02 10:06

[QUOTE=fivemack;161210]This would be a Big Calculation with capital Big; 2.6 seconds per relation and we need half a billion, so 40 CPU-years. 16e is a prodigious user of memory (about 4G virtual of which just over 1G used), so this may be more a project for people with clusters than for random home user - indeed, that might be a bit more of a strain on clusters than their administrators really want.[/QUOTE]

Can you please post your polynomial file (including mfbr/a, a/rlambda) - I want to do some benchmarks on my PC (running on 32-bit Windows XP) to see if it would fit in 1 GB or if it needs more.

(BTW: my office box where I used to run my ranges in your factoring projects has got 1 GB RAM, so running 16e will be almost certainly too much if I am present and want to use word, excel, etc., but I will be out of office for ~2 weeks in the beginning of march and therfore can possibly run a small range with 16e in the meantime - iff it fits into 1 GB.)

10metreh 2009-02-02 10:24

[quote=fivemack;161210]These are with N=2^941-1, rlim=alim=200M, sieving 200M .. 200M+10^4

[code]
side lp siever yield time/rel
alg 32 15 6237 2.29344
alg 33 15 11693 1.21814
alg 32 16 13320 2.79733
alg 33 16 25553 1.45897
rat 32 15 8559 1.99026
rat 33 15 16440 1.03673
rat 32 16 17536 2.57357
rat 33 16 33922 1.30827
[/code]

which looks as if 32-bit large primes and 16e is the right way to go for numbers of this size (changing the siever doubles the yield at a fairly small cost in runtime; lp=33 doubles the yield and the number required at the same time so is no net benefit). Probably rational side 0-300M. rlim=alim=200M was a guess, I'll do some more runs to optimise that.

This would be a Big Calculation with capital Big; 2.6 seconds per relation and we need half a billion, so 40 CPU-years. 16e is a prodigious user of memory (about 4G virtual of which just over 1G used), so this may be more a project for people with clusters than for random home user - indeed, that might be a bit more of a strain on clusters than their administrators really want.[/quote]

I see you're testing M941. What is your plan after 10^263-1 is finished?

fivemack 2009-02-02 11:49

andi47: No feasible -16e run will fit on a 32-bit machine. Sorry.

10metreh: next project will be GNFS on 109!+1, starting with a month of polynomial selection; I'm currently trying to find good parameters for the polynomial selection run.

Andi47 2009-02-02 18:20

[QUOTE=fivemack;161365]andi47: No feasible -16e run will fit on a 32-bit machine. Sorry.
[/QUOTE]

Is it way to slow or way to memory-hungry on a 32 bit machine?

Batalov 2009-02-02 19:18

[quote=fivemack;161210]...16e is a prodigious user of memory (about 4G virtual of which just over 1G used), so this may be more a project for people with clusters than for random home user - indeed, that might be a bit more of a strain on clusters than their administrators really want.[/quote]
See the very beginning. Memory hungry. And slow, too.

When Tom says 4Gb - he means and 4Gb and a bit more, too. (and Win usually needs a big chunk of memory for itself.)
This is a project for the 21st century and we'll all have to tweak the tools first. It is possible (and ultimately desirable) that a modified version of the 16e siever will later fit (Tom, I think that for many ranges SCHED_TOL can be actually lowered from 2). But now just yet.

Batalov 2009-04-21 07:59

Here are the 32-bit re-runs with [URL="http://mersenneforum.org/showthread.php?p=169381#post169381"]the experimental siever[/URL] (on a Phenom940):
[code]side lp siever yield1 sec/rel yield2 sec/rel
alg 32 15e 6237 2.29344 6238 (1.01389 sec/rel)
rat 32 15e 8559 1.99026 8558 (0.87265 sec/rel)

alg 32 16e 13320 2.79733 13322 (1.29829 sec/rel)
rat 32 16e 17536 2.57357 17537 (1.16108 sec/rel
[/code]

The time/rel are simply proportional (the CPU and the code is different).
The number of relns is +- a few, too; this is a known effect.
The memory footprint, though, for 15e and 16e was
586m (1031m virt.) for 15e
682m (1446m virt.) for 16e

Not as bad as with old memory allocations. Not 4g/1g!

[SIZE="1"]
Note, that I have not quite figured out all new Kleinjung code (lasieve5), indeed there are many changes there ...and there are 15e, 15f, 15g siever variants![/SIZE]


All times are UTC. The time now is 06:41.

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2020, Jelsoft Enterprises Ltd.