[Coin-bcpdiscuss] Re: BCP: node buffering?

[Joerg Herbers]

Laci,

thanks for the changes. I have finally been able to test the
enhancement with regard to my original implementation (all inner nodes
and user data is kept) and Francois' proposition of deleting only the
user data in the tree manager when the node is processed (as in the BAC
example).

However, when deactivating the explicit deletion of user data (i.e.
only using the RemoveExploredBranches option), I initially did not
obtain significant enhancements with regard to the original
implementation. Actually, it seems that there is a memory leak with
regard to user data. Adding "delete _user_data" to
BCP_tm_node::~BCP_tm_node(), memory requirements were reduced from 1.6GB
to 1.4GB. If I use the explicit deletion of user data from the BAC
example, memory requirements drop to 1.3GB, using both options yields
memory requirements in the same order of magnitude.

The RemoveExploredBranches seems to delete a node (along with those
parents that don't have additional children) when it is pruned. This may
explain the stronger effect of explicitly deleting user in the tree
manager. (In the latter case, we already delete the user data already
when the node has been processed.)

Could you check if I'm right with the  memory leak assumption (and
possibly correct the file in the repository)?

Thanks,
Joerg



>>> On 3/31/2006 at 4:44 AM, in message
<Pine.A41.4.21.0603302143180.23606-100000 at oslpp.watson.ibm.com>, Laszlo
Ladanyi
<ladanyi at us.ibm.com> wrote:
> Hello,
> 
> I have added code that removes the "deadwood" from the search tree,
i.e.,
> immediately removes subtrees that are completely explored. I have
done some
> quick testing and it seems to work. However, by default it is turned
off, 
> just
> to make sure that noone's code breaks. I wonder if people could give
it a 
> try
> and let me know whether it works fine or not. I am interested in both
serial
> and parallel settings. To enable the feature add
>     BCP_RemoveExploredBranches 1
> to your parameter file.
> 
> --Laci
> 
> 
>> Actually, BCP is very wasteful from this point of view :-(. The
whole tree
>> is kept until the end of computation, even if you do depth first
>> search. When I'm back from vacation I'll try to do something about
this.
>> 
>> --Laci
> 
> 
> 
> _______________________________________________
> Coin-bcpdiscuss mailing list
> Coin-bcpdiscuss at list.coin-or.org 
> http://list.coin-or.org/mailman/listinfo/coin-bcpdiscuss