Multithreaded Linear Algebra threshold

[chris2be8]

By default msieve only uses multiple threads for matrices with more than 200000 columns. I've found this causes a step in the run time to solve a matrix:

grep -e '  lanczos halted' -e BLanczosTime -e 'commencing L' f*/ggnfs.log | less
...
f600_51+3/ggnfs.log:Tue Oct  9 21:48:53 2018  commencing Lanczos iteration (4 threads)
f600_51+3/ggnfs.log:Tue Oct  9 21:49:40 2018  lanczos halted after 3203 iterations (dim = 202523)
f600_51+3/ggnfs.log:Tue Oct  9 21:49:41 2018  BLanczosTime: 60
...
f61_80+7/ggnfs.log:Wed Oct 10 01:50:19 2018  commencing Lanczos iteration
f61_80+7/ggnfs.log:Wed Oct 10 01:52:15 2018  lanczos halted after 3134 iterations (dim = 197959)
f61_80+7/ggnfs.log:Wed Oct 10 01:52:15 2018  BLanczosTime: 128

You can adjust the threshold by updating common/lanczos/lanczos.h as follows:

chris@vega:~/msieve-svn/trunk> diff common/lanczos/lanczos.h ~/msieve-svn.old/trunk/common/lanczos/lanczos.h
98c98
< #define MIN_NROWS_TO_THREAD 100000
---
> #define MIN_NROWS_TO_THREAD 200000

Then run make all with the same parameters used to compile msieve previously.

On the same system after the change:

f791_47+6/ggnfs.log:Thu Oct 11 15:24:58 2018  commencing Lanczos iteration (4 threads)
f791_47+6/ggnfs.log:Thu Oct 11 15:25:39 2018  lanczos halted after 2973 iterations (dim = 187904)
f791_47+6/ggnfs.log:Thu Oct 11 15:25:39 2018  BLanczosTime: 51

So I'm saving 1-2 minutes per number with a matrix in this range. The crossover is smaller than 100000 but I've not had time to find out exactly where.

That system has 4 real cores, no hyperthreading. On another system where msieve is running on 1 core and it's hyprethreaded partner there's no step in run times.

Note it's only worth doing if you are factoring a lot of numbers with a matrix in this size range. And you do need to check if it will help on the system you are using.

Chris