Linear algebra at 600%

[CRGreathouse]

This seems wrong to me:

linear algebra completed 764732 of 125729 dimensions (608.2%, ETA 1027h56m)

[R.D. Silverman]

Originally Posted by CRGreathouse

This seems wrong to me:

linear algebra completed 764732 of 125729 dimensions (608.2%, ETA 1027h56m)

Yep! This looks B-A-D. bad. Almost surely a purported basis for one of the
sub-spaces is not orthogonal. Indeed, it probably does not even span the
sub-space.

[jasonp]

You should restart the linear algebra from scratch. v1.42 has an orthogonality check that runs at regular intervals during the linear algebra, but the matrix has to be of size > 1M for checking to be turned on. I figured that was a good compromise because checkpointing is turned on at that size too, so that you can actually hope to restart from some point beyond the beginning when an error is detected.

Note that in the past, the occurrence of this error has brought the stability of the machine into question. Is your machine overclocked at all?

[joral]

If not, I would recommend getting a CD with memtest and letting it run for a while. I was having the same problems, and found it was a faulty DIMM. Of course, that has taken my system from 3GB to 1GB, which is probably limiting my matrix sizes, but..

[Batalov]

The new orthogonality checks + restarts (see below*) will work very well on a slightly compromised computer. (A severely, irresponsibly OC'd computer will not be helped, though.) If errors happen only once/twice a day - you will do fine.

In a way, I have a computer like that, not because of OC'ing but because with 8Gb of memory [which I need, the numbers I am dealing with are that big] doesn't work completely flawlessly with the memory controller and whatnot. For control purposes, 4Gb works excellent for days through: with mprime or with msieve. 8Gb, however, and an error or a couple per day -- at the hottest time of the day, of course. No errors at night.

I have loosened memory timings to 5-6-6 (even though it is rated 5-5-5, 1066) and upped the NB voltages just a tad. The tuning is ongoing, but the calculation simply goes forward. Finished a couple in this regime already. Summer!! It's going to get cooler soon.

___________
*a simple perl script, feel free to make it better:

 
#!/usr/bin/perl -w
$ret = 1;
while($ret && -e "t.dat.chk") {
$cmd = "./msieve -s t.dat -i t.ini -l ggnfs.log -v -nf t.fb -t 4 -ncr";
$ret = system($cmd);
}
1;

P.S. And no, the system without any changes does not fail neither memtest nor even mprime! Parallel msieve is harder on the memory controller than these tests. And yes, I have a decent cooler. :-)

[CRGreathouse]

Originally Posted by jasonp

You should restart the linear algebra from scratch. v1.42 has an orthogonality check that runs at regular intervals during the linear algebra, but the matrix has to be of size > 1M for checking to be turned on. I figured that was a good compromise because checkpointing is turned on at that size too, so that you can actually hope to restart from some point beyond the beginning when an error is detected.

I stopped the process, ran the perl script, and re-started msieve. It looks like it's failing the test repeatedly:

charles@zed:~/mth/msieve-1.41$ ./msieve

sieving complete, commencing postprocessing
linear algebra completed 2025 of 125729 dimensions (1.6%, ETA 0h 2m)    

linear algebra completed 2014 of 125729 dimensions (1.6%, ETA 0h 2m)    

linear algebra completed 10121 of 125729 dimensions (8.0%, ETA 0h 1m)    

linear algebra completed 4055 of 125729 dimensions (3.2%, ETA 0h 2m)    


linear algebra completed 4048 of 125729 dimensions (3.2%, ETA 0h 2m)    

linear algebra completed 8079 of 125729 dimensions (6.4%, ETA 0h 1m)    

linear algebra completed 14163 of 125729 dimensions (11.3%, ETA 0h 1m)    



linear algebra completed 6072 of 125729 dimensions (4.8%, ETA 0h 1m)    
linear algebra completed 20254 of 125729 dimensions (16.1%, ETA 0h 1m)    
linear algebra completed 6080 of 125729 dimensions (4.8%, ETA 0h 1m)    

linear algebra completed 6072 of 125729 dimensions (4.8%, ETA 0h 1m)    
linear algebra completed 2022 of 125729 dimensions (1.6%, ETA 0h 1m)    
linear algebra completed 4046 of 125729 dimensions (3.2%, ETA 0h 1m)    

linear algebra completed 2025 of 125729 dimensions (1.6%, ETA 0h 2m)    
linear algebra completed 4045 of 125729 dimensions (3.2%, ETA 0h 1m)    
linear algebra completed 2021 of 125729 dimensions (1.6%, ETA 0h 2m)    
linear algebra completed 18212 of 125729 dimensions (14.5%, ETA 0h 1m)    
linear algebra completed 16177 of 125729 dimensions (12.9%, ETA 0h 1m)    
linear algebra completed 6064 of 125729 dimensions (4.8%, ETA 0h 1m)    
linear algebra completed 18207 of 125729 dimensions (14.5%, ETA 0h 1m)    
linear algebra completed 14162 of 125729 dimensions (11.3%, ETA 0h 1m)    


linear algebra completed 10112 of 125729 dimensions (8.0%, ETA 0h 1m)    
linear algebra completed 10109 of 125729 dimensions (8.0%, ETA 0h 1m)    


linear algebra completed 2025 of 125729 dimensions (1.6%, ETA 0h 2m)    
linear algebra completed 4048 of 125729 dimensions (3.2%, ETA 0h 2m)    
linear algebra completed 8093 of 125729 dimensions (6.4%, ETA 0h 1m)    
linear algebra completed 6072 of 125729 dimensions (4.8%, ETA 0h 1m)    
linear algebra completed 6068 of 125729 dimensions (4.8%, ETA 0h 1m)    

linear algebra completed 6072 of 125729 dimensions (4.8%, ETA 0h 1m)    

linear algebra completed 2023 of 125729 dimensions (1.6%, ETA 0h 1m)

Originally Posted by jasonp

Note that in the past, the occurrence of this error has brought the stability of the machine into question. Is your machine overclocked at all?

Not overclocked at all (neither the memory nor the CPU). My browser has been crashing a lot recently though, not sure if that's related. I'm running 64-bit Ubuntu if it matters.

[Batalov]

The easiest thing to try without rebooting is the "mprime -m", and then invoke torture test option.

[CRGreathouse]

I rebooted into memtest, and it looks like my (not that new) RAM is bad. I pulled out the offending sticks and am retrying. Edit: It worked just fine -- it seems only the memory was an issue. I guess it makes sense that it took until the linear algebra phase to notice this.

Hmm, I guess it's time to look up Corsair's warranty information. I bought it 4-5 weeks ago, do you think they'll replace it?

[Batalov]

They should.
Patriot did -- for a 2x2Gb, when I got the same effect as you. Usually they will not cross-mail, of course. And do start the formal RMA process.
(My trick of getting a particular contact person's email off of the Patriot forum was effective. Some other people who write to the generic RMA email boxes usually complain that there's no answer for a while.)
____

Your avatar reminds me of my original avatar and a quote to go with it:
"The Eagle never lost so much time as when he submitted to learn from the Crow." /W.B./