Originally posted by: mrmag

#CPLEXOptimizers
#DecisionOptimization

There is no direct way to do that. One option would be to do your own timing in the user cut callback (this gives you the time spent for creating user cuts) and also use a solve callback to time the LP (re)solves.

#CPLEXOptimizers
#DecisionOptimization

Originally posted by: mrmag

Dear Daniel, thank you for the response. I realized the timing in the user callback. But if I do the same in the solve callback it will compute the time for all LP resolutions also with those after branches without generation of user cuts. Do you have any idea how to compute the time only for LP resolutions after a user cut addition?

Marat

#CPLEXOptimizers
#DecisionOptimization

It should be possible to do that using a node's user data. Assume you have two markers, TIME and NOTIME.

In the solve callback you only measure the time if the current node's user data is TIME. Before the solve callback returns you always set the current node's user data to NOTIME.

After adding cuts in the cut callback add/change the current node's user data to TIME.

 

#CPLEXOptimizers
#DecisionOptimization

Originally posted by: mrmag

That is a good idea. I will report the results after implementation.

The problem could of course occur if i solve the program in parallel. Then the markers will be mixed.

#CPLEXOptimizers
#DecisionOptimization

Why would the markers be mixed? Multiple threads may process different nodes simultaneously but multiple threads will never process the same node simultaneously. I don't see how the markers could get mixed up.

#CPLEXOptimizers
#DecisionOptimization

Originally posted by: mrmag

Sorry, I thought about a global variable... Can you give me a hint about a node's user data? Where can I find the info about them?

#CPLEXOptimizers
#DecisionOptimization

I would do it like this:

static int mark_time = 0; /* We don't use that, we only use its address. */
#define TIME &mark_time
#define NOTIME NULL

Then use CPXcallbacksetuserhandle() to get and/or change the current node's user data. As user data just use TIME and NOTIME. You can also use CPXgetcallbacknodeinfo() with CPX_CALLBACK_INFO_NODE_USERHANDLE to get a node's data.

#CPLEXOptimizers
#DecisionOptimization

Originally posted by: mrmag

Dear Daniel,

#CPLEXOptimizers
#DecisionOptimization

In C++ something like this should do the trick:

struct TimeDummy : public IloCplex::MIPCallbackI::NodeData {
  // Just an empty class.
};
ILOSOLVECALLBACK1(UserSolve, outStructure &, nfo) {
   // If TIME then calculate the time
   IloCplex::MIPCallbackI::NodeData *mark = getNodeData();
   if ( mark ) {
     // update nfo
     setNodeData(0);
     delete mark;
   }
   // else do nothing
}
ILOUSERCUTCALLBACK2(UserCut, two_cont&, M, outStructure &, nfo) {
  // generate cut
  // set user data to TIME
  setNodeData(new TimeDummy());
}

 

 

#CPLEXOptimizers
#DecisionOptimization

Originally posted by: mrmag

Marat

#CPLEXOptimizers
#DecisionOptimization

Originally posted by: mrmag

I solved the problem by the follwing code.

But I observe a very strange effect. For example, for the root node 10000 cuts are generated. The total time for the root node relaxation with cuts is 30 sec (in real), but the value of nfo.cutTime is only 4 sec. What else does CPLEX in the root node reoptimization?

 

ILOSOLVECALLBACK1(UserSolve,

#CPLEXOptimizers
#DecisionOptimization

CPLEX may invoke heuristics during the root node. Do the timings look better if you disable heuristics (IloCplex::HeurFreq = -1)?

#CPLEXOptimizers
#DecisionOptimization

Originally posted by: mrmag

I initiated a new topic on this: https://www.ibm.com/developerworks/community/forums/html/topic?id=2ccc9f85-ba4a-4e01-bfb3-5119293dd797&ps=25

 

#CPLEXOptimizers
#DecisionOptimization