Re: Solver performance solving examples/life_goe.mod

[Domingo Alvarez Duarte]

Hello Chris !

Thank you for reply !

Please commit your "probing" code to a new branch in your repository, this way maybe someone can give you some help polishing it.

Also about the problem with "avl_tree" do you plan to review and fix your repository ?

Cheers !

On 26/9/20 17:30, Chris Matrakidis wrote:

A brief description of the changes is in https://lists.gnu.org/archive/html/help-glpk/2018-02/msg00018.html - those considerably increased the number of miplib problems that can be solved, in particular using pseudocost branching. Do ask if you have any specific questions, but it was some time ago so it will take me a bit to dig up the details.

Note that for problems like the one you are using with lots of binary variables, a pre-processing technique called probing is very effective but not available in GLPK. I have a draft implementation but I'm not very happy with it, so it isn't committed to my tree.

Best Regards,

Chris Matrakids

On Sat, 26 Sep 2020 at 17:55, Domingo Alvarez Duarte <mingodad@gmail.com> wrote:

Hello Chris !

Thank you for reply !

Can you describe in a few lines what your improvements achieved ?

I've been looking at your changes and found that you've bitten by "src/env/avl.c" avl_tree do not reject duplicate keys but do not provide a way to recover duplicates, I fixed this here https://github.com/mingodad/GLPK/commit/502da3ae23ffb4c1731cc437fd6b420ac82443d5 and I found it while trying to update your code to work with splay_tree.

See this thread https://lists.gnu.org/archive/html/bug-glpk/2020-08/msg00018.html .

Comparing your glpsol with mine solving hashi.mod we get this:

=====

# your repository compiled with  -O3 -DNDEBUG -march=native -ffp-contract=off

...

INTEGER OPTIMAL SOLUTION FOUND
Time used:   91.6 secs
Memory used: 19.6 Mb (20560460 bytes)

...

Model has been successfully processed
91.72user 0.64system 1:32.37elapsed 100%CPU (0avgtext+0avgdata 23032maxresident)k
0inputs+0outputs (0major+764597minor)pagefaults 0swaps

=====

# your repository compiled with  -O2 -MT -MD -MP -MF

...

INTEGER OPTIMAL SOLUTION FOUND
Time used:   120.9 secs
Memory used: 19.6 Mb (20560460 bytes)

...

Model has been successfully processed
120.93user 1.23system 2:02.16elapsed 99%CPU (0avgtext+0avgdata 22900maxresident)k
0inputs+0outputs (0major+764575minor)pagefaults 0swaps

=====

#my repository compiled with -g -O3 -DNDEBUG -march=native -ffp-contract=off -DWITH_SPLAYTREE

INTEGER OPTIMAL SOLUTION FOUND
Time used:   59.2 secs
Memory used: 10.6 Mb (11130654 bytes)
...

Model has been successfully processed
58.99user 0.60system 0:59.59elapsed 99%CPU (0avgtext+0avgdata 13664maxresident)k
0inputs+0outputs (0major+397605minor)pagefaults 0swaps

=====

# ubuntu 18.04 glpsol package

INTEGER OPTIMAL SOLUTION FOUND
Time used:   70.5 secs
Memory used: 19.8 Mb (20725782 bytes)
...

Model has been successfully processed
71.49user 0.22system 1:11.71elapsed 99%CPU (0avgtext+0avgdata 23500maxresident)k
0inputs+0outputs (0major+135565minor)pagefaults 0swaps

=====

Cheers !

On 26/9/20 16:20, Chris Matrakidis wrote:

I made some performance improving patches a few years ago. You can find them in: https://github.com/cmatraki/GLPK

Best Regards,

Chris Matrakidis

On Sat, 26 Sep 2020 at 14:54, Domingo Alvarez Duarte <mingodad@gmail.com> wrote:

Hello !

Testing GLPK I left it solving examples/lie_goe.mod for more than 2
hours and it didn't found a solution (wasm and native) then I stopped
then and tried with cplex/gurobi/xpress/scip all of then gives a
solution instantly (except scip that takes 3s).

The difference is so big, have someone managed through command line
options or other means managed to get a solution quickly with glpsol ?

Any idea of how to improve GLPK to not be so behind ?

Cheers !