SIQS Across Multiple Machines

[EdH]

Thank you! This helps a bunch. I had started out using the values from the tune.c info, but I strayed into doubling values and such as I moved up to larger composites, instead of going back to the listed values. I will probably want to address 105+ and, therefore 3LP eventually, but for now I need to let this settle. I still need to add even more machines to the mix and see what happens.

The rels/sec is definitely going to be important. The two 4 thread machines are widely different and the processing shows it when they are both given the same -siqsR. Would it be far off to consider the ratio between CPUs at one size to be similar to another size composite?

[bsquared]

Originally Posted by EdH

Would it be far off to consider the ratio between CPUs at one size to be similar to another size composite?

Seems reasonable.

[EdH]

I was able to calculate a pretty good balance for the original 4 machines and brought the execution time from the above 11:58 down to 8:18. I then was able to do the calculations to add five more machines and brought the time down to 6:52.

But, I'm definitely going to have to find a better way to do these balance calculations if I decide to make this a bigger set of machines and work with different sized composites.

Thanks for all the help!

[EdH]

Just to add a comparison, I ran the 90 and 100 digit composites with CADO-NFS on the exact same machines and the composites were factored in 2:25 and 6:20, respectively.

[EdH]

Originally Posted by bsquared

. . .
If you are thinking of doing this on C105+ numbers for some reason then 3LP is faster, but only with parameters that are not auto-selected. Also I haven't tested 3LP on many machines; it might get tricky with the batch factoring that is used in 3LP. We can talk over strategies if this is something you're interested in.

I suppose my next step "might" be to try this, but these questions come to mind:

How hard would it be for me to learn what it takes to do this?
Would it always take manual intervention?
Would there be a gain over CADO-NFS/Msieve?

A totally different question:

My understanding is that sieving for S/GNFS is too memory intensive for efficient GPU use? Would SIQS be any more capable of being performed by GPU?

[bsquared]

Originally Posted by EdH

I suppose my next step "might" be to try this, but these questions come to mind:

How hard would it be for me to learn what it takes to do this?
Would it always take manual intervention?
Would there be a gain over CADO-NFS/Msieve?

A totally different question:

My understanding is that sieving for S/GNFS is too memory intensive for efficient GPU use? Would SIQS be any more capable of being performed by GPU?

It would probably be non-trivial, and I don't have any plans to fully automate it anytime soon. It could be close but it's unlikely to be faster than NFS at C105.

SIQS is also pretty memory intensive at sizes that are interesting. Maybe it would be easier than NFS but that's not saying much.