[CoinMp] New Patch for 1.4.0 CoinMP
Bjarni Kristjansson
bjarni at maximalsoftware.com
Thu Oct 29 18:20:53 EDT 2009
Hi Stuart,
Thanks for sending me the patch file, it was very helpful. I have
already updated CoinMP stable 1.4 with fixes for following items (in
some cases I had to make different changes than suggested in the patch
file, for example the write MPS):
1) corrects the setting of some parameters that are real not ints
4) fixes a bug with the msgcallback
6) puts in an option for FractionGap
8) is able to write MPS files for integer problems
I have some questions on the other 4 items or need to review them further:
2) changes some solver options for better performance
COIN_INT_SCALING set to 5 was 3
COIN_INT_PERTURBATION set to 50 was 100
COIN_INT_PRIMALPIVOTALG set to 4 was 1
COIN_INT_DUALPIVOTALG set to 3 was 1
Were these option changes based on different defaults in the CLP code
or rather based on testing different options that are giving better
performance on average? Typically default values in CoinMP should
match the default values that John Forrest uses in his testing code
and only changed from those if there is a really good reason to. If
the new suggested option values are from the CLP code, please let me
know where exactly, so I can look them up. I was not able to
determine if scaling=5 is actually a valid option.
3) adds branch priorities
Need to review this bit further, but can probably go in soon.
5) puts in an option for MaxSeconds
The existing COIN_REAL_MAXSECONDS option seems to be missing a
corresponding call in the CoinMP. Could not see any new code for
MaxSeconds in the patch file. Should be relatively easy to figure
out.
7) outputs better status messages from cbc
Need also to review this bit further.
There were also number of additions in the patch file to unitTest.cpp,
which I have not looked at yet.
Best regards and thanks for keeping track of all these requests.
Bjarni
On Wed, Oct 28, 2009 at 9:50 PM, Stuart Mitchell
<s.mitchell at auckland.ac.nz> wrote:
> It has taken me all morning but I have the patch file for this now
>
> This Patch takes contributions from Larry Taylor, Ron Frowd and Myself and
> 1) corrects the setting of some parameters that are real not ints
> 2) changes some solver options for better performance
> 3) adds branch priorities
> 4) fixes a bug with the msgcallback
> 5) puts in an option for MaxSeconds
> 6) puts in an option for FractionGap
> 7) outputs better status messages from cbc
> 8) is able to write MPS files for integer problems
>
> Can we PLEASE have this integrated with the code as I need it for pulp-or
> which is dependent on this solver.
>
> Stu
>
>
> Stuart Mitchell wrote:
>>
>> hey there is a version 1.4.0 that fixes this problem
>>
>> Finally!!!
>>
>> There are still the other issues that we identified.
>>
>> So I am building a patch for this as we speak
>>
>> Stu
>>
>>
>> Rod Frowd wrote:
>>>
>>> Ming:
>>>
>>> I posted this several years ago:
>>>
>>> "I have been running the CBC MIP repetitively from CoinMP.dll and it
>>> seems to be leaking memory and growing in size.
>>>
>>> I have added a delete of the cbc object into CoinMP.cpp as follows. It
>>> looks like this will also delete the cut generator objects created as its
>>> destructor includes deletion of the cut generators.
>>>
>>> SOLVAPI int CoinUnloadProblem(HPROB hProb)
>>> {
>>> PCOIN pCoin = (PCOIN)hProb;
>>> if (pCoin) {
>>> delete pCoin->clp;
>>> if (pCoin->RowLower) free(pCoin->RowLower);
>>> if (pCoin->RowUpper) free(pCoin->RowUpper);
>>> if (pCoin->IsInt) free(pCoin->IsInt);
>>> delete pCoin->cbc;
>>> pCoin->cbc = NULL;
>>> }
>>> free(pCoin);
>>> pCoin = NULL;
>>> return SOLV_SUCCESS;
>>> }
>>>
>>> "
>>>
>>> And I think Stuart Mitchell from NZ also raised a ticket about this.
>>>
>>> Rod
>>>
>>> ming lu wrote:
>>>>>
>>>>> Hi All,
>>>>>
>>>>> It looks that CoinUnloadProblem doesn't release resource correctly.
>>>>> The cbc was not released, this causes a huge memory leak problem. Also
>>>>> following members are not explicitly released:
>>>>>
>>>>> ClpSolve *clp_presolve;
>>>>> OsiClpSolverInterface *osi;
>>>>> CBMessageHandler *msghandler;
>>>>> CBIterHandler *iterhandler;
>>>>> CBNodeHandler *nodehandler;
>>>>>
>>>>> CglProbing *probing;
>>>>> CglGomory *gomory;
>>>>> CglKnapsackCover *knapsack;
>>>>> CglOddHole *oddhole;
>>>>> CglClique *clique;
>>>>> CglLiftAndProject *liftpro;
>>>>> CglSimpleRounding *rounding;
>>>>>
>>>>> Could someone take a look?
>>>>>
>>>>> Thanks,
>>>>> Ming
>>>>>
>>>>>
>>>>
>>>> _______________________________________________
>>>> CoinMp mailing list
>>>> CoinMp at list.coin-or.org
>>>> http://list.coin-or.org/mailman/listinfo/coinmp
>>>>
>>>>
>>>
>>> _______________________________________________
>>> CoinMp mailing list
>>> CoinMp at list.coin-or.org
>>> http://list.coin-or.org/mailman/listinfo/coinmp
>>
>
> --
> ___________________________________
> Dr Stuart Mitchell
> Research Fellow
> Light Metals Research Centre (LMRC)
> University of Auckland
> Private Bag 92019
> Auckland
> New Zealand
>
> Ph (wk) +64 9 3737599 ext 84867
> (ddi) +64 9 9234867
> (fax) +64 9 3737925
> (mb) +64 21 441331
> ___________________________________
>
>
>
> ---------- Forwarded message ----------
> From: "Larry A. Taylor" <ltaylor at seas.ucla.edu>
> To: Rod Frowd <rod at frowd.net>, coinmp at list.coin-or.org
> Date: Mon, 18 Feb 2008 16:39:46 -0800
> Subject: Re: [CoinMp] CoinMP.dll errors
> See Ticket #1 for this project:
> https://projects.coin-or.org/CoinMP/ticket/1
>
> LAT
>
> At 01:33 AM 2/18/2008, Rod Frowd wrote:
>>I found that the CoinMP.dll did not work when built from source, though
>>the downloaded binary dll did. It was giving primal infeasible
>>errors, when the problem was feasible,
>>
>>Upon looking at the source code I found the problem below, some of the
>>real tolerances were being set to zero as the CoinGetIntOption was being
>>called instead of the CoinGetRealOption? The built from source dll works
>>after this change.
>>
>>Index: CoinMP.cpp
>>===================================================================
>>--- CoinMP.cpp (revision 53)
>>+++ CoinMP.cpp (working copy)
>>@@ -646,10 +646,10 @@
>>
>> pCoin->clp->setMaximumIterations(CoinGetIntOption(hProb,
>>COIN_INT_MAXITER));
>>
>>- pCoin->clp->setPrimalObjectiveLimit(CoinGetIntOption(hProb,
>>COIN_REAL_PRIMALOBJLIM));
>>- pCoin->clp->setDualObjectiveLimit(CoinGetIntOption(hProb,
>>COIN_REAL_DUALOBJLIM));
>>- pCoin->clp->setPrimalTolerance(CoinGetIntOption(hProb,
>>COIN_REAL_PRIMALOBJTOL));
>>- pCoin->clp->setDualTolerance(CoinGetIntOption(hProb,
>>COIN_REAL_DUALOBJTOL));
>>+ pCoin->clp->setPrimalObjectiveLimit(CoinGetRealOption(hProb,
>>COIN_REAL_PRIMALOBJLIM));
>>+ pCoin->clp->setDualObjectiveLimit(CoinGetRealOption(hProb,
>>COIN_REAL_DUALOBJLIM));
>>+ pCoin->clp->setPrimalTolerance(CoinGetRealOption(hProb,
>>COIN_REAL_PRIMALOBJTOL));
>>+ pCoin->clp->setDualTolerance(CoinGetRealOption(hProb,
>>COIN_REAL_DUALOBJTOL));
>>
>> /* check if it has been changed, leave alone otherwise */
>> ClpPrimalColumnSteepest primalSteepest(CoinGetIntOption(hProb,
>>COIN_INT_PRIMALPIVOTALG));
>>@@ -659,7 +659,7 @@
>> pCoin->clp->setDualRowPivotAlgorithm(dualSteepest);
>>
>> if (CoinGetIntOption(hProb, COIN_INT_CRASHIND)) {
>>- pCoin->clp->crash(CoinGetIntOption(hProb, COIN_REAL_CRASHGAP),
>>+ pCoin->clp->crash(CoinGetRealOption(hProb, COIN_REAL_CRASHGAP),
>> CoinGetIntOption(hProb,
>>COIN_INT_CRASHPIVOT));
>> }
>> switch (CoinGetIntOption(hProb,COIN_INT_SOLVEMETHOD)) {
>>
>>
>>
>>Another issue I found is that the default options were not set the same
>>as in the CLP executable, making CoinMP.dll a lot slower. I was able to
>>speed it up by a factor of 3 with the following parameter changes:
>>
>>COIN_INT_SCALING set to 5 was 3
>>COIN_INT_PERTURBATION set to 50 was 100
>>COIN_INT_PRIMALPIVOTALG set to 4 was 1
>>COIN_INT_DUALPIVOTALG set to 3 was 1
>>
>>
>>Rod Frowd
>>
>>
>>
>>
>>_______________________________________________
>>CoinMp mailing list
>>CoinMp at list.coin-or.org
>>http://list.coin-or.org/mailman/listinfo/coinmp
>
> _______________________________________________
> CoinMp mailing list
> CoinMp at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/coinmp
>
>
> ---------- Forwarded message ----------
> From: Rod Frowd <rod at frowd.net>
> To: coinmp at list.coin-or.org
> Date: Mon, 18 Feb 2008 19:33:51 +1000
> Subject: [CoinMp] CoinMP.dll errors
> I found that the CoinMP.dll did not work when built from source, though
> the downloaded binary dll did. It was giving primal infeasible errors, when
> the problem was feasible,
>
> Upon looking at the source code I found the problem below, some of the
> real tolerances were being set to zero as the CoinGetIntOption was being
> called instead of the CoinGetRealOption? The built from source dll works
> after this change.
>
> Index: CoinMP.cpp
> ===================================================================
> --- CoinMP.cpp (revision 53)
> +++ CoinMP.cpp (working copy)
> @@ -646,10 +646,10 @@
>
> pCoin->clp->setMaximumIterations(CoinGetIntOption(hProb,
> COIN_INT_MAXITER));
>
> - pCoin->clp->setPrimalObjectiveLimit(CoinGetIntOption(hProb,
> COIN_REAL_PRIMALOBJLIM));
> - pCoin->clp->setDualObjectiveLimit(CoinGetIntOption(hProb,
> COIN_REAL_DUALOBJLIM));
> - pCoin->clp->setPrimalTolerance(CoinGetIntOption(hProb,
> COIN_REAL_PRIMALOBJTOL));
> - pCoin->clp->setDualTolerance(CoinGetIntOption(hProb,
> COIN_REAL_DUALOBJTOL));
> + pCoin->clp->setPrimalObjectiveLimit(CoinGetRealOption(hProb,
> COIN_REAL_PRIMALOBJLIM));
> + pCoin->clp->setDualObjectiveLimit(CoinGetRealOption(hProb,
> COIN_REAL_DUALOBJLIM));
> + pCoin->clp->setPrimalTolerance(CoinGetRealOption(hProb,
> COIN_REAL_PRIMALOBJTOL));
> + pCoin->clp->setDualTolerance(CoinGetRealOption(hProb,
> COIN_REAL_DUALOBJTOL));
>
> /* check if it has been changed, leave alone otherwise */
> ClpPrimalColumnSteepest primalSteepest(CoinGetIntOption(hProb,
> COIN_INT_PRIMALPIVOTALG));
> @@ -659,7 +659,7 @@
> pCoin->clp->setDualRowPivotAlgorithm(dualSteepest);
>
> if (CoinGetIntOption(hProb, COIN_INT_CRASHIND)) {
> - pCoin->clp->crash(CoinGetIntOption(hProb, COIN_REAL_CRASHGAP),
> + pCoin->clp->crash(CoinGetRealOption(hProb, COIN_REAL_CRASHGAP),
> CoinGetIntOption(hProb,
> COIN_INT_CRASHPIVOT));
> }
> switch (CoinGetIntOption(hProb,COIN_INT_SOLVEMETHOD)) {
>
>
>
> Another issue I found is that the default options were not set the same
> as in the CLP executable, making CoinMP.dll a lot slower. I was able to
> speed it up by a factor of 3 with the following parameter changes:
>
> COIN_INT_SCALING set to 5 was 3
> COIN_INT_PERTURBATION set to 50 was 100
> COIN_INT_PRIMALPIVOTALG set to 4 was 1
> COIN_INT_DUALPIVOTALG set to 3 was 1
>
>
> Rod Frowd
>
>
> _______________________________________________
> CoinMp mailing list
> CoinMp at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/coinmp
>
>
More information about the CoinMp
mailing list