msieve help

[jasonp]

It doesn't quite work that way; the data structures used in the filtering will also use as much memory as they require, but when intermediate files are written the number of passes through those files depends on whether they will fit in half the memory you specify. So singleton removal will happen in one pass if the intermediate singleton (.lp) file is at most half the size specified. If it is larger than that, singleton removal happens on disk and the comparison is performed again. If it's still too big, the filtering bound is raised and the process repeats. Clique removal and merging always happen in memory, with no regard to the specified bound.

I really should add on-disk clique removal, it will make the filtering much faster for large datasets because lots of cliques can be removed at once; but that isn't a high priority, because jobs large enough to benefit from that are often run on huge machines.

[fivemack]

I was very pleased with the performance of msieve-1.43 RelProc on 2^877-1; 9000 seconds to handle relations for a 31-bit gnfs, where similar-sized jobs had taken twenty to thirty hours with earlier versions. Even for 12^256+1, the RelProc took about 25 hours compared with 1100 for the matrix; I don't think more improvement there is particularly essential.