Msieve 1.43 feedback

[Jeff Gilchrist]

Originally Posted by Batalov

for most computers the checkpoint should be written every 1-3 hours.

There will be no chk, if the matrix is less than 1M².

Thanks for the info. A checkpoint did arrive, it was coming about 1.5 - 2 hours now I think so thanks for the clarification.

What about when you hit CTRL-C, is msieve supposed to write a checkpoint on exit? If not, that would be a nice feature to have in case you need to kill it somewhere in between checkpoints or even before the first checkpoint so you don't have to rebuild the .mat file again.

[jasonp]

Originally Posted by Jeff Gilchrist

What about when you hit CTRL-C, is msieve supposed to write a checkpoint on exit? If not, that would be a nice feature to have in case you need to kill it somewhere in between checkpoints or even before the first checkpoint so you don't have to rebuild the .mat file again.

Yes, it's supposed to write a checkpoint on interrupt. Ctrl-C in windows is a little strange; a C signal handler will catch it the first time but the process will be killed immediately if you hit Ctrl-C a second time.

[Jeff Gilchrist]

Originally Posted by jasonp

Yes, it's supposed to write a checkpoint on interrupt. Ctrl-C in windows is a little strange; a C signal handler will catch it the first time but the process will be killed immediately if you hit Ctrl-C a second time.

Hmmm, I'm pretty sure I just hit CTRL-C once, but I was running the VS2008 binary in a cygwin shell so maybe the shell had some strange interaction with it.

I will try it again some time to make sure.

Jeff.

[Batalov]

((Capt.Obvious reports that as you are under CygWin, you can do ps -f and then kill <the_PID>. This should obviate the Windows key bindings and idiosyncrasies.))

[jasonp]

Not to say you did anything wrong, I've had other reports of interrupts not cleaning up like they were supposed to. Not sure what to do about them.

[Jeff Gilchrist]

Originally Posted by jasonp

Not to say you did anything wrong, I've had other reports of interrupts not cleaning up like they were supposed to. Not sure what to do about them.

I just tried it again to make sure I only hit CTRL-C once and it did generate a checkpoint file. So seems to be working fine.

Jeff.

[Jeff Gilchrist]

I'm now having a problem with -nc3 on a number with both 32bit and 64bit VS2008 binaries. Not sure if Linux has a problem, I'm uploading the files to a Linux box but it will take several hours because of their size. Here is the crash message from Windows:

Unhandled exception at 0x00486e33 in msieve32.exe: 0xC0000005: Access violation writing location 0x00000000.

Essentially the same thing for the 64bit version as well. It crashes in sqrt.c during this function call:

		/* read in only the relations for dependency i */
		nfs_read_cycles(obj, &fb, NULL, NULL,
				&num_relations, &rlist, 0, i);

The full log file is attached but it always happens when it is reading relations for depedency 36, it gets 0 cycles:

Mon Nov 16 08:11:32 2009  reading relations for dependency 35
Mon Nov 16 08:11:34 2009  read 2745153 cycles
Mon Nov 16 08:11:39 2009  cycles contain 8604690 unique relations
Mon Nov 16 08:13:21 2009  read 8604690 relations
Mon Nov 16 08:13:41 2009  algebraic side is not a square!
Mon Nov 16 08:13:41 2009  reading relations for dependency 36
Mon Nov 16 08:13:43 2009  read 0 cycles

Anyone have an idea what is going on here?

Jeff.

[jrk]

Unhandled exception at 0x00486e33 in msieve32.exe: 0xC0000005: Access violation writing location 0x00000000.

Looks like a null pointer dereference. Can you get a stack trace?

[frmky]

Originally Posted by Jeff Gilchrist

The full log file is attached but it always happens when it is reading relations for depedency 36, it gets 0 cycles:

Note that the matrix solution "recovered 35 nontrivial dependencies" so dependency 36 (and 35?) shouldn't exist. Though I suspect this is a symptom of a more fundamental problem...

[jasonp]

Looks like the last round of memory optimizations to the relation reading code cannot handle an empty dependency. I'll take a look when my schedule calms down a bit.

[frmky]

The real problem, though, seems to be all "the number of relations is not even" and "algebraic side is not a square!" All of the non-trivial dependencies failed.