|From:||Domingo Alvarez Duarte|
|Subject:||Re: GLPK and OPENMP - GLP_SIMPLEX running concurrently|
|Date:||Wed, 13 May 2020 15:32:23 +0200|
|User-agent:||Mozilla/5.0 (X11; Linux x86_64; rv:68.0) Gecko/20100101 Thunderbird/68.7.0|
Hello Sierra !
My 2 cents, have you tried to compile with -fsanitize=thread to see if it catch the problem ?
La información contenida en este mensaje y en cualquier archivo adjunto, es confidencial y está dirigido únicamente al destinatario del mensaje. Si Ud. no es el destinatario correcto por favor notifique al remitente respondiendo este mensaje y elimine inmediatamente de su sistema el e-mail y los posibles archivos adjuntos. Está prohibida cualquier utilización, difusión o copia de este e-mail por cualquier persona o entidad que no sean las especificas destinatarias del mensaje. UTE no acepta ninguna responsabilidad con respecto a cualquier comunicación que haya sido emitida incumpliendo nuestra Política de Seguridad de la Información, así como lo previsto en la Ley 18.331 de Protección de Datos Personales y Ley 18381 de Acceso a la Información PúblicaHello, and sorry for insisting about this issue.
The program is returning correct values as long as the call to glp_simplex is enclosed inside an OMP CRITICAL region. This is the only part requiring a mutex. The rest of the glp_* functions are called concurrently by several threads with no issue at all.
When the OMP CRITICAL surrounding glp_simplex is removed, the calls to glp_get_row_dual and glp_get_col_prim return different values.
Each thread uses a different glp_prob, which is built from scratch by the same thread and used by the same thread.
Each thread also changes some values on the RHS and calls glp_simplex several times. After that, each thread deletes its own GLPK problem.
What could be affecting the way glp_simplex works? Is there anything you would suggest looking at?
I'm very grateful for the help I've received so far.
Subject: Re: [Fwd: Possible bug in GLPK - Error detected in file env/alloc.c at line 72]
I have been working on the project that I have been assigned and applied the changes you suggested with successful results.
- glp_prob objects are created, used and deleted by the same thread- GLPK lib compiled with reentrancy support (checking whether to enable reentrancy support... yes)
The program is no longer crashing. Thank you.
I have looked at the multiseed.c example. The main difference between that example and the project that I am working on is that I build a glp_prob from scratch, then using glp_copy_prob make as many copies as processors are available on the machine. Each thread changes a couple values on the right hand side and runs the simplex again, without building the glp_prob again or starting the search from scratch.
However, the problem that I am now facing is that when several threads are calling glp_simplex at the same time, some values appear to be mixed up such as those returned by glp_get_row_dual. Adding a mutex (OMP CRITICAL) that contains only the glp_simplex call makes the problem go away and no issues can be seen, the program generates the expected output data and ends without errors.
I have looked at the error and output handling contained in the multiseed.c example. I ended up using glp_open_tee and redirecting the output of each thread to a separate file. No warnings were displayed.
Thank you very much for your help.
On 3/25/20 8:23 PM, Andrew Makhorin wrote:
> -------- Forwarded Message --------
> Date: Wed, 25 Mar 2020 18:34:56 +0000
>> Dear GLPK mantainer(s),
>> I am having problems when trying to run the GLPK solver in a multi-
>> threaded fashion using its C API and OpenMP. While the solver does
>> appear to work, an error comes up when trying to free the memory with
>> glp_free: memory allocation error
>> Error detected in file env/alloc.c at line 72
>> I am no GLPK expert but I would be very grateful if you could check
>> the provided source code and determine whether this is being caused
>> by a bug or a simple beginner mistake.
>> Running on an eight core processor, 8GiB ram. Stack size limit: 2GB
>> (ulimit -s 2000000), Stack size limit per thread: 200MB (export
>> OMP_STACKSIZE=200000). Provided source makes little sense but is
>> quite short and reproduces the problem every time. Compile command is
>> as follows:
>> gcc test_c.c -o test_c -fopenmp -I/home/user/libglpk/glpk-
>> 4.65/include /home/user/libglpk/glpk-4.65/src/.libs/libglpk.so.40.3.0
>> Juan Pablo Sierra
Did you compile GLPK with '--enable-reentrant=yes'?
You are creating all problems in one thread and then call the solver in
separate threads. This is not supported by GLPK.
Please, have a look at:
Here all GLPK library calls needed for each individual model are called
within the same separate thread.
Please, have a look at the error and output handling in the example too.
|[Prev in Thread]||Current Thread||[Next in Thread]|