Re: [Ggnfs-devel] compressed relations

[Sten <st...@ma...>]

The idea behind these latest modifications is to reduce
size of the spairs.out files in large scale projects. Say,
c155 and higher.

I've already implemented and commited changes to the both
sievers making them able to produce what I call
"short relation format". When siever is invoked with -s command
line argument it outputs only a,b values.

procrels now recognizes such "short relation" automatically and
calls factRel() routine for it instead of completePartialRelFact().

I've tested yesterday "short relations" on snfs_small and found
critical performance degradation on procrels step. On my test
machine the relation processing speed dropped from 13000 rels/sec
to 360 rels/sec.

It seems the factRel() -> factor() routine seems to be a bit unoptimal
(it tried to factor relations using Trial Division and Pollard Rho).
And after aSL! commited his changes to factor(), a call to squfof(),
the speed increased to 550 rels/sec. But it still too low..

So, I think we can analyze how sievers factor relations - they
should have more optimal approach - and, maybe, mimic CWI-like
behaviour saving only large factors.

 Sten

JP> With the recent checkins to enable compressed relation 
JP> handling, I was wondering...

JP> Could the lattice siever be modified to not dump the full
JP> multiplicity of each factor to the output? Procrels figures
JP> out the multiplicity on its own, and duplicate factors
JP> take about 25% of the volume in a relation file.

JP> Another idea would mimic what the CWI suite does, and
JP> only print factors larger than a specified lower bound.

JP> jasonp