Running GGNFS

[10metreh]

For the C104 that Luigi is doing, I would use factLat.pl and nfs2ms.pl rather than going through all the complicated stuff. nfs2ms.pl creates all the files for msieve, too.

[Andi47]

For GNFS: What composit length is the crossover between using gnfs-lasieve4I12e and 13e?

[henryzz]

Originally Posted by Andi47

For GNFS: What composit length is the crossover between using gnfs-lasieve4I12e and 13e?

110 digits according to factmsieve.pl

[fivemack]

I've been doing a bit of my own tuning for this.

One very obvious observation: starting at Q=alim and going backwards (rather than starting at Q=alim/2 and going forwards as factLat.pl does) seems to get you to a factorisation more quickly.

For GNFS

12/13 crossover does seem to be around 110 digits.

At 144 digits, lp=28 using 13e is worse than lp=29 using 14e, which is itself worse than lp=28 using 14e.

28/13 beats 28/14 at 138 digits, so the crossover there is around 140

29/14 beats 28/14 quite soundly by 150 digits.

For SNFS

On quartics it makes sense to go to /15 very early on (2^860+1 S208 is helped by using /15)

30/14 appears to be better than 29/14 by about S210, though rather worse at S200.

For 10^263-1, initial small experiments suggests 32/15 and 31/15 are very comparable, and since the linalg for lp=32 takes four times as long I'd use the latter.

[10metreh]

Andi47, you put "4.7 million?" for the number of relations needed for a C100 GNFS. According to schickel, it is around 3.8 million. When I did a C100 and oversieved to 5.2 million, msieve had trouble building a matrix.

[Andi47]

Originally Posted by 10metreh

Andi47, you put "4.7 million?" for the number of relations needed for a C100 GNFS. According to schickel, it is around 3.8 million. When I did a C100 and oversieved to 5.2 million, msieve had trouble building a matrix.

Thanks. These 4.7M were more or less a wild guess. I corrected it and attached the new file to this posting.

BTW: Data marked with a "?" are guesses based on interpolations, where I don't have enough / any data yet.

[10metreh]

Originally Posted by Andi47

Thanks. These 4.7M were more or less a wild guess. I corrected it and attached the new file to this posting.

BTW: Data marked with a "?" are guesses based on interpolations, where I don't have enough / any data yet.

In the end I knocked the number of relations down to 4 million and it worked. A tip: never oversieve too much.

[ET_]

I managed to download, compile and run GGNFS on my Linux64 box.

Thank you all for the help!

Luigi

P.S. is a yield of 0.0077 sec/rel a "good" parameter?

[Jeff Gilchrist]

Originally Posted by ET_

I managed to download, compile and run GGNFS on my Linux64 box.

Where did you download the source, what version? I'm having trouble compiling the ggnfs-0.77.1.tar.gz from sourceforge on my Linux64 box getting these errors:

echo "#define GGNFS_VERSION \"0.77.1\"" > include/version.h
make -C src
make[1]: Entering directory `/home/jeffg/ggnfs/src'
gcc -I../include -O3 -fomit-frame-pointer -Wall -march=core2 -c
blanczos64.c
blanczos64.c: In function ‘MultB64’:
blanczos64.c:316: warning: pointer targets in initialization differ in
signedness
blanczos64.c: In function ‘MultB_T64’:
blanczos64.c:469: warning: pointer targets in initialization differ in
signedness
/tmp/ccyfRMZS.s: Assembler messages:
/tmp/ccyfRMZS.s:164: Error: `(%rdx,%ecx,2)' is not a valid base/index
expression
/tmp/ccyfRMZS.s:167: Error: `8*1(%rdx,%ecx,2)' is not a valid
base/index expression
/tmp/ccyfRMZS.s:167: Error: `8*2(%rdx,%ecx,2)' is not a valid
base/index expression
/tmp/ccyfRMZS.s:167: Error: `8*3(%rdx,%ecx,2)' is not a valid
base/index expression
/tmp/ccyfRMZS.s:167: Error: `8*4(%rdx,%ecx,2)' is not a valid
base/index expression
/tmp/ccyfRMZS.s:167: Error: `8*5(%rdx,%ecx,2)' is not a valid
base/index expression
/tmp/ccyfRMZS.s:167: Error: `8*6(%rdx,%ecx,2)' is not a valid
base/index expression
/tmp/ccyfRMZS.s:169: Error: `(%rsi,%eax,8)' is not a valid base/index
expression
...
/tmp/ccgdcQdx.s:55976: Error: `8*n(%rbx,%eax,8)' is not a valid
base/index expression
/tmp/ccgdcQdx.s:55976: Error: `8*n(%rbx,%eax,8)' is not a valid
base/index expression
make[1]: *** [blanczos64.o] Error 1
make[1]: Leaving directory `/home/jeffg/ggnfs/src'

So the errors seem to come from the blancsoz64 source file which as
inline assembler. Do I need to so something special with GCC to
assemble this?

Jeff.

[Andi47]

Batalov posted in another thread, that he has added a -R (resume) option and an input sanity check to the siever (version 328). Are there win32 binaries (for P4 and/or C2D) available of gnfs-lasieve4I12e .... 4I15e, version 328?

edit: I just got an "issue" myself, which caused a windows error message, and turned out to be a typo in my input:

n 171715153617185465051393022673714527931465270135551225634357302947311904186640760284113383777747947192749639705628979
skew: 261187.54
Y0 -34809139703073288415331
Y1  2359110901579
c0  3824914116375045584368825974840
c1  15238950334340823938379382
c2 -87184131588771861153
c3 -433560750171960
c4  330570356
c5  3360
rlim: 4000000
lim: 4000000
lpbr: 27
lpba: 27
rlambda: 2.6
alambda: 2.6
mfbr: 50
mfba: 50

The "a" of alim was missing, thus causing a fatal crash of the siever (this was a version which is (much?) older than 328, so it would be nice if someone could post a (link to a) windows binary of the sievers.)

Edit2: and the colon between n and the number was missing. *grrr* (*wants an option for msieve to output ggnfs-formatted polys*)

[Jeff Gilchrist]

Ok so I found the latest trunk code:
http://ggnfs.svn.sourceforge.net/viewvc/ggnfs/

I was able to compile this code now. Thanks.