linesiever RAM

[junky]

hi, i've remark the linesiever.exe is gobbing upto 89 mb of ram, i'd like to know if a version taking like 300 mb of ram would go faster for users.
if yes, that would be really interesting for users who can afford that amount of ram (which, i think, the most users).

Thanks.

[Wacky]

linesiever.exe is gobbing up to 89 mb of ram, i'd like to know if a version taking like 300 mb of ram would go faster

The short answer is "Yes, a little". But not really all that much.

Of more interest is a fundamentally different siever that REQUIRES significantly more memory. We are hoping to have it available for additional users soon. Until we do, we are unlikely to attempt another extremely large factorization.

[junky]

with "REQUIRES significantly more memory" is like what ? 1 G ?
are ya alone to work on that new version?


i can go upto 1G of RAM on my own computer, so i've no trouble for the moment. In 2 months, i think im gonna change it.

[Wacky]

Originally Posted by junky

with "REQUIRES significantly more memory" is like what ? 1 G ?

No, nothing that significant. Only our post processing needs G's of memory.

The 90MB line siever can usually be run on a machine with 256MB of RAM.
But this siever needs about that much itself with no room for the OS or shared libraries. But I do have in running on this machine with "only" 512MB and still have room to run everything else I normally run. (I do the compiles, etc. on a different machine)

At the moment, we are trying to sort out some licensing issues as well as improve our infrastructure to handle multiple sievers.

[junky]

multiple sievers would be really great.

or maybe at initialization, it could detect how much RAM the system has, and based on these infos, it could the more RAM it can (or it's allow to used).

btw, is it just paul doing the post-processing ? its take what ? 2 G of RAM or more ?

[R.D. Silverman]

Beyond a certain threshold extra memory does not help a 'basic' line siever.
One needs a certain amount to hold the primes, polynomial roots, and
sieve start points etc., and a certain amount to hold the sieve line itself, but
the latter gets partitioned into pieces that fit in cache anyway. More
memory allows a bigger factor base, but if the size of the factor base is
already optimal, then making it bigger will SLOW DOWN the computation.

However, in *some* circumstances it can help a lot.

Consider a number represented by a polynomial all of whose exponents
are EVEN. In this case norm(a + alpha) = norm(-a + b alpha), thus
once one sieves a > 0, the results can be saved so that a < 0 does not
have to be sieved. But it is a LOT of state information to save.
I actually added most of the required code to my siever then realized
the required memory would keep me from running on machines with
less than 256M. It needs a LOT of extra memory.

It only applies in certain cases. And it yields a 20-25% speedup.

One can get a similar speedup in some cases by doing two numbers at once.
Suppose one wants to do (say) 2,761+ and 2,761-. The linear polynomial
is the same for both. Thus one could sieve the linear polynomial and then
use it for BOTH numbers. This would save 20-25% of the total time to
do both numbers. BUT: it greatly increases the required memory.

To take advantage one would need machines with at least 256M of memory.
Many potential users do not have that much.

[junky]

bob: on one machine here, i've no trouble with memory. i really appreciate if ya can send me the client for i can test it or run it?
it's using what? 256 mb of RAM ?

and don't forget to send me ur papers for NFS.

Thanks.

[R.D. Silverman]

Originally Posted by junky

bob: on one machine here, i've no trouble with memory. i really appreciate if ya can send me the client for i can test it or run it?
it's using what? 256 mb of RAM ?

and don't forget to send me ur papers for NFS.

Thanks.

If you are "grepmoo@hotmail" then I already tried sending you the paper.
The mail bounced. Noone else has sent me mail requesting the paper.

When you refer to "the client", to which piece of software do you refer?
I have my own siever. It is not part of NFSNET. I never completed the
code that takes advantage of symmetry because it required too much
memory.

[nfortino]

Originally Posted by Bob Silverman

Noone else has sent me mail requesting the paper.

I sent you a PM yesterday at 6:30pm (GMT). If you ment an e-mail, I am about to try that now.

[junky]

Originally Posted by Bob Silverman

If you are "grepmoo@hotmail" then I already tried sending you the paper.
The mail bounced. Noone else has sent me mail requesting the paper.
memory.

ya, its me gremoo AT hotmail DOT com, how big is ur file ? my box is currently loaded at 22%, gonna delete everything to be sure to have enough space.

and i thought is was a custom client for NFSnet. :)
but if ya wanna send me ur code, i'll can take a look on it.

Thanks.

[junky]

now account droped at 1%, so go ahead for files :)