[Coin-bcpdiscuss] Performance

[acw at ascent.com]

We are in the process of migrating from OSL to COIN/Bcp.  Mostly the 
process has gone quite well, but recently I received quite a shock 
concerning performance.

I have a MIP problem which OSL completes in about 25 seconds; the same 
problem takes COIN/Bcp about 290 seconds.  This is more than an order of 
magnitude discrepancy, which would be unacceptable on some of our larger 
problems.

I can think of several possible explanations.

We are using the BranchAndCut example almost unchanged; the only 
modifications we have made were to interface the code to our LP server. 
Perhaps BranchAndCut uses cut generation and variable selection that is 
only appropriate to small problems, and we have to do more research to 
figure out how to guide the branching process.  Maybe it's as simple as 
selecting the right values for the branching control parameters.

Perhaps the main time-sink is communication between the TM and LP modules. 
 BranchAndCut uses the "BCP_message_single" communications protocol, which 
has to pack and unpack every LP problem and solution.  Doubtless this 
could be streamlined, but we don't want to open that can of worms unless 
the potential gain is great (like a factor of 3 or more).

And of course there is the possibility that we are doing something else 
completely stupid that we're not even aware of.

So I'm appealing to the community.  What experiences have you had 
regarding performance?  What approaches have worked?  What stupid mistakes 
should we be careful to avoid?

If there is interest, I can post an MPS file containing the test problem 
in question.

Thanks in advance.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://list.coin-or.org/pipermail/bcp/attachments/20051212/bc5b06bf/attachment.html