Jasonp, I can live with msieve crash just before square root starts but I would like to help out to fix this issue. Because I use a batch file to start msieve it won't write nothing to the log if it crashes. So what do I have to do, what do you need to know what happened? Should I save Win7 error? Please advice. Just want to prepare myself...

Carlos

What's wrong with ETA? lol

[code]linear algebra completed 2498967 of 11770836 dimensions (21.2%, ETA 185h 3m)
linear algebra completed 2499222 of 11770836 dimensions (21.2%, ETA 185h 3m)
linear algebra completed 2499476 of 11770836 dimensions (21.2%, ETA 185h 3m)
linear algebra completed 2499730 of 11770836 dimensions (21.2%, ETA 185h 2m)
linear algebra completed 2499982 of 11770836 dimensions (21.2%, ETA 185h 2m)
linear algebra completed 2500235 of 11770836 dimensions (21.2%, ETA 185h 4m)
linear algebra completed 2500489 of 11770836 dimensions (21.2%, ETA 185h 5m)
linear algebra completed 2500739 of 11770836 dimensions (21.2%, ETA 185h 7m)
linear algebra completed 2500991 of 11770836 dimensions (21.2%, ETA 185h 8m)
linear algebra completed 2501246 of 11770836 dimensions (21.2%, ETA 185h 8m)
linear algebra completed 2501496 of 11770836 dimensions (21.3%, ETA 185h 9m)
linear algebra completed 2501748 of 11770836 dimensions (21.3%, ETA 185h10m)
linear algebra completed 2502000 of 11770836 dimensions (21.3%, ETA 185h10m)
linear algebra completed 2502254 of 11770836 dimensions (21.3%, ETA 185h11m)
linear algebra completed 2502508 of 11770836 dimensions (21.3%, ETA 185h11m)
linear algebra completed 2502762 of 11770836 dimensions (21.3%, ETA 185h11m)
linear algebra completed 2503015 of 11770836 dimensions (21.3%, ETA 185h12m)
linear algebra completed 2503270 of 11770836 dimensions (21.3%, ETA 185h12m)
linear algebra completed 2503521 of 11770836 dimensions (21.3%, ETA 185h12m)
linear algebra completed 2503774 of 11770836 dimensions (21.3%, ETA 185h12m)
linear algebra completed 2504030 of 11770836 dimensions (21.3%, ETA 185h12m)
linear algebra completed 2504286 of 11770836 dimensions (21.3%, ETA 185h13m)
linear algebra completed 2504541 of 11770836 dimensions (21.3%, ETA 185h13m)
linear algebra completed 2504793 of 11770836 dimensions (21.3%, ETA 185h13m)
linear algebra completed 2505047 of 11770836 dimensions (21.3%, ETA 185h14m)
linear algebra completed 2505300 of 11770836 dimensions (21.3%, ETA 185h14m)
linear algebra completed 2505553 of 11770836 dimensions (21.3%, ETA 185h15m)
linear algebra completed 2505805 of 11770836 dimensions (21.3%, ETA 185h15m)
linear algebra completed 2506058 of 11770836 dimensions (21.3%, ETA 185h16m)
[/code]

jasonp,

When I start a post-processing job Under Msieve v. 1.51 (SVN Official Release - LARGEBLOCKS) with "-nc", without stopping it, it stalls on square root phase.
If during LA I stop and then resume from last checkpoint with "-nc3 -ncr", square root doesn't stall, it resumes LA and starts/finishes square root. I don't know why this happens. My OS is Win7 Enterprise SP1.

Carlos

Carlos,

This was a silly problem caused by recent changes in the NFS square root. It should be fixed as of SVN 889. COuld you confirm that latest SVN works for you?

[QUOTE=jasonp;343558]Carlos,

This was a silly problem caused by recent changes in the NFS square root. It should be fixed as of SVN 889. COuld you confirm that latest SVN works for you?[/QUOTE]

Jasonp, I am now in Brazil with bandwidth issues so I can't run at the moment NFS@Home post-processing jobs. I am very sorry.

Carlos

No worries, I'm pretty sure the above is what was wrong. v1.52 is really accumulating a lot of bug fixes.

The latest SVN now includes a major overhaul of the linear algebra, the first since about 2008. The new code uses a thread pool and coordinates much more tightly between threads. Major changes include

- no more LARGEBLOCKS; everybody gets large blocks without having to compile them in
- way less memory use than LARGEBLOCKS needed; the memory use also does not go up as more threads are added
- added the ability to override the automatically chosen block size from the command line

On my older machines at home, the result is slightly faster, but with a more modern CPU with lots of memory bandwidth and large caches, the new code is much faster. On an Ivy Bridge system with 8MB of cache, the new code runs 30% faster single-threaded and more than 50% faster with 4 threads. It's an interesting question whether we can now achieve higher performance using MPI+threads compared to MPI alone.

Anyway, if you're crunching through large problems give it a try. Hopefully this will make a nice dent in NFS@Home's backlog. Windows users will now need a pthreads DLL (the one from MinGW will do). A win32 build plus pthreads DLL is available [url="www.boo.net/~jasonp/msieve_svn900.zip"]here[/url].

I'm looking forward to trying out this newest version once I move to the linear algebra sections of some of my current numbers!

Edit: I tried the -ns step with the svn_899 and svn_900, and it starts fine the 1st time. If I stop it, though, I get this error the next time I try using -ns again.

[CODE]error: NFS input does not match polynomials
check that input doesn't have small factors[/CODE]

Even if I delete everything in the factor base file except for the polynomial parameters (essentially resetting it), I get the error. Any idea what I'm doing wrong?

Disregard that last post--it was just something fubar with the factor base file.

Another neat data point: on the quad-core IVB systems I have access to, no combination of MPI+threads across two machines runs faster for solving a 7.5M matrix than using 4 threads on one of those machines. In fact the ETA with 4 threads on one machine is over 30% less! That definitely was not the case before.

Dumb question(s) amnesty time:

1) I'm on an AMD Phenom II X4 955--do you expect a performance bump with it?

2) Is the performance increase (in general) at the -ns stage, or at the -nc stage?