-   Msieve (
-   -   Msieve announcements (

jasonp 2008-08-27 15:02

Msieve v1.37
Now available. Highlights include:

- A lot of cleanup in the NFS poly selection (still a work in progress)

- A lot of cleanup and optimization in the NFS filtering. The new version should also be faster because it removes several disk based passes. Also, this version now is a lot smarter about handling free relations, and your SNFS jobs will see a large number of extra free relations added.

- Added the '-p' switch to run the demo binary at idle priority

- Miscellaneous bug fixes everywhere

The prospect of describing the NFS filtering at the CADO workshop means that I'll be busy playing with this part of the code for the near future.

Happy factoring,

jasonp 2008-09-25 01:49

Msieve v1.38
Now available. Highlights include:

- A lot of ECM fixes, plus use of GMP-ECM v6.2.2; the demo binary does not have SSE2 support, and this really helps stage 2, so if you want the highest speed you should still compile your own

- Optimizations in the NFS square root that double the speed and halve the memory use

- Bug fixes everywhere, plus a few NFS filtering tweaks

Happy factoring,

PS: Forgot a minor tweak or two; v1.38 has been uploaded again

tmorrow 2008-10-03 12:02

Is GMP-ECM v6.2.2 available? ECMNET and Inria are only reporting v6.2.1.

jasonp 2008-10-03 12:34

[QUOTE=tmorrow;144383]Is GMP-ECM v6.2.2 available? ECMNET and Inria are only reporting v6.2.1.[/QUOTE]
No, that was a typo, sorry.

jasonp 2008-11-22 18:20

Msieve v1.39 beta
There are enough invasive changes in the next release that I wanted to post a beta first for people to test:


Highlights include:

- a completely rewritten NFS polynomial selector. The current version actually finds better polynomials than pol5 does, but in much longer time. In fact, when testing RSA120 it generated a polynomial with the best Murphy alpha I've ever seen (-7.92 !) The demo binary also allows you to specify leading coefficient ranges, and the library now listens to any time limits that are set on polynomial selection (sorry this has taken so long, I know I suck). The stage 2 optimization code, i.e. what pol51opt does, is also enormously better, and for large problems is hundreds of times faster while still finding the best polynomials.

- a lot of patches by Brian Gladman that make the inline asm compile correctly with Intel and Microsoft's compiler, on both windows and linux, both 32- and 64-bit (this was amazingly painful). I've also revamped some of the asm, so that QS is now actually (slightly) faster. We know that QS works, but were hoping someone could verify the NFS code still works in the expected time when compiled with MSVC.

- a set of fixes to the linear algebra that hopefully reduces the number of problem matrices that cause weeks of wasted work right now. Serge, can you verify that the problem factorization we went over now at least generates a matrix without empty and duplicate columns?

- changes to the NFS filtering so that the clique processing is only rerun after the merge phase proves that it would be advantageous to do so

If nobody finds stupid mistakes of mine, I'll make the full release next week.

Batalov 2008-11-23 02:00

my tests went fine. Good stuff!
My 2^743-3 (that had problems with 1.38 before, twice) went through fine. I did, from scratch, -nc1 10000000,10000000 (1.2 hrs) and started -nc2 and tested the resulting matrix for empty or duplicate columns (use the "dump.c" program that I PM'd). Cannot say if the BL will be successful (it takes 5 days*4 threads), but so far, so good. The dumper test has passed. I am pretty sure that this matrix looks good. (I don't know if dumper test is now simply built in. I am going to read the source and diffs. later)

So, I think 1.39-beta is ready for M857.

Bruce, Greg, please nc1, nc2 and after matrix is built -- the dumper test as per PM.

Thanks, Jason!


10metreh 2008-11-23 12:44

I did a quick QS test on R71. I found that v1.39b was 2% slower than v1.38! I hope this is only normal fluctuations, though.

jasonp 2008-11-26 13:55

Msieve v1.39 beta 2
Now available:

[url=""]win32 binary[/url]


- fixed 64-bit NFS poly selection; it was completely broken

- added poly skew to the factor base file format, logfile, and internal structures (not used internally, yet)

- lowered the size of most A5 coefficients that are searched, and increased the time per coefficient

- fixed almost all compile warnings

- added forgotten typo to the NFS filtering

- added more MSVC tweaks from Brian Gladman; the MSVC project will now link with GMP and GMP-ECM by default. See the readme in the build.vc9 directory if you don't want them.

The shoddy state of this beta has me peeved, but hopefully this will be the last one before a full release later in the week.

jasonp 2008-12-02 03:51

Msieve v1.39
Full release is now available (essentially no changes from beta2)

Don't get too attached to the current NFS polynomial selector, I have big plans for improvements in v1.40

Happy factoring,

Batalov 2008-12-10 04:27

Two simple patches to (for v.1.39)
Those who use may want to add somewhere around line 1317 the SKEW to .fb file (it started its internal life in msieve and is reported in ggnfs.log and may be used in versions >= 1.39):
open(OF, ">$FBNAME");
print OF "N $N\nSKEW $SKEW\n"; # patch #1
for my $i (reverse 0..$DEGREE) {
print OF "A$i $COEFVALS{$i}\n";
if(!$COEFHASH{Y1}) { # patch #2 for the Y1/Y0-less .poly files
print OF "R1 $COEFHASH{Y1}\n";
print OF "R0 $COEFHASH{Y0}\n";
(patch #2 is just a convenience thingy - sometimes one gets a poly from old-timer users (or a program-generated), who have "m:" but not "Y1:"/"Y0:")... my 2 cents...

jasonp 2009-03-02 04:03

Msieve v1.40 beta
I have not updated the web page, but a beta of the next version is now available (it's been a while, and bug fixes have piled up).


Highlights include:

- Major internal changes to NFS poly selection; the .p file now outputs polynomials in GGNFS format, and the code now optimizes and reports Murphy's E score, so polynomial scores are now directly comparable to that of the pol5 code in GGNFS

- A fix for a typo that made QS a lot slower

- Huge swaths of NFS code have been made into common code, for bsquared's impending attempt to add three large prime support to QS. This includes the batch factoring code, hashtable code and most of the NFS filtering code

- Miscellaneous bug fixes eveywhere (see the changelog for details). Optimizations to NFS with huge excess will have to wait until v1.41

Please make sure things compile on whatever platforms and compilers you have, and seem to work okay. Replies should go in other threads. I'll give it maybe a week before considering a full release.

All times are UTC. The time now is 07:54.

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