[gecode-users] Cloning problems

[Christian Schulte]

Hi,

that sounds ugly ;-) Before I start with rambling about cloning, a tip:
Gecode supports unary branchings (with a single alternatiuve only). So you
could have a unary branching and don't have to have a mock-up empty branch.
The search engines even optimize that case. If you are looking for an
example, check the assign function in gecode/int.hh for assigning variables.

Okay, cloning it is. So, there are two stages: copying and updating. During
copying you first create a new space and copy all actors. While copying all
actors you copy each variable and its domain used by an actor and collect a
list of all these variables (actually a list for each variable type). The
only thing that you do not do is copying the dependency array of a variable
(the array of pointers that point to propagators that depend on a variable).
Then in updating, you update the dependency arrays (and also deal with
advisors).

Updating a dependency array just copies for each old entry the new entry
(stored via a forwarding pointer in each actor) and restore some information
for variables that has been stored in the copy of the dependency array.
After that you cleanup the space, the actors, and you are done.

In VarImp<VIC>::enter, the --idx[0] makes room for an additional entry. Then
you keep moving actors to a free entry and by this create a new free entry
for the propagation condition i. When the loop is done, you know that there
is a free entry for propagation condition pc and that's where the propagator
is entered.

Your guess about x->idx[0] in VarImp<VIC>::update is correct: this actually
stores the forwarding pointer for a variable implementation. Also your
suspicion about ActorLink::prev is correct: there the forwarding pointer for
actors is stored.

Let me come up with some idea where the trouble could be:
 - you forget to update a variable or view at some place. Either in your
branching, propagator, or model. 
   Then you can see an invalid pointer in the dependency array.
   You might want to change the default constructor of VarBase in
gecode/kernel/var.icc and the default constructor
   of VarViewBase in gecode/kernel/view.icc so that they set varimp to NULL
(actually, I just did that yesterday in the trunk as well). Then you will be
sure that when you forget to update, you'll see a NULL pointer.
 - you commit a heinous crime in that you change the dependencies of a
variable with cancel/subscribe during 
   cloning. That is illegal and will for sure break everything. Note that
creating a propagator during
   cloning will commit that crime, too.

It is very very unlikely (but not impossible, of course) that the cloning in
Gecode has a snatch but, as said, very unlikely given that its used by
anything else.

Hope that helps. Please do not hesitate to ask some more
Christian

--
Christian Schulte, www.ict.kth.se/~cschulte/


-----Original Message-----
From: users-bounces at gecode.org [mailto:users-bounces at gecode.org] On Behalf
Of Filip Konvicka
Sent: Thursday, February 12, 2009 2:21 PM
To: users at gecode.org
Cc: Luboš Moric
Subject: [gecode-users] Cloning problems

Hi,

[Sorry, this is a looong message...]

we're hunting a serious bug that occurs during space cloning in 2.2.0. 
The bug occurs very rarely, but we have a testcase that triggers this 
behavior.

We have many constraints in the problem instance and the solver should 
post as many propagators as possible. We have a custom branching for 
this, which posts one propagator at a time in commit(), while the 
alternative is not to post the propagator (i.e. a no-op). Because we're 
only looking for the first solution, in the case of a failure we no 
longer need the path back to the root in the recomputation tree, so we 
decided to use our own simple search engine for this. The standard DFS 
search engine exhibits exactly the same behavior (both with 
recomputation on and off), and we don't see any problems with our search 
engine.

Everything seems to work for the vast majority of the test cases, but 
there are a few instances that cause problems (probably) during cloning 
(can be probably also be caused by some earlier bad subscibe or 
unsubscribe). From our point of view, there is nothing wrong or special 
about the instances. The crashes occur at the same location both on 
Linux and Windows, in both release and debug builds. Changing memory 
management (e.g. never deleting Spaces in the search engine) can cause 
the crash to occur at slightly different places (e.g. some propagation 
during status() after clone() finishes).

One particular case we're looking at now crashes at core.icc:2270, where 
f[0] is a bad pointer (0xfeeefeee at Windows). We're not sure how this 
can happen - we know that in this case n==2 at core.icc:2255, so idx[0] 
is bad pointer at core.icc:2252. This is also what Valgrind says on 
Linux (bad read of size 4).

When we were trying to debug the other cases, we found out that the 
subscription list in a variable in the cloned space contained an actor 
link that was probably copied incorrectly as it seemed as a pure 
ActorLink like Space::a_actors, having a totally different address than 
the rest of the actors (probably belonging to the original space 
object). When we tried to find out when this actor link entered the 
list, we ended up in VarImp<VIC>::update again.

We're (of course:-)) using FloatVars in the model, and we eliminated all 
other kinds of variables and propagators. In our case, pc_max==1 and 
free_bits==0.

We find it difficult to understand what is happening during cloning. We 
would appreciate if someone explaned the basic idea. We only have 
floatvars, propagators and one branching (no advisors or other types of 
actors/branchings/advisors).

We know how VarImp<VIC>::resize works, that's easy. In 
VarImp<VIC>::enter, we can't see why you do "--idx[0];" as the first 
iteration of the for cycle overwrites it (as long as pc>0, of course). 
May be just optimization of course. As for VarImp<VIC>::update, we only 
guess...we suspect that a) the original x->idx[0] is destroyed somewhere 
so it needs to get restored from a memcpy backup at idx[0], b) 
ActorLink::_prev is probably used to map old actors to new ones (thus 
the "->prev()". We did not dig deep enough to be sure though, so we'd 
welcome some guidance here.

Cheers,
Filip


_______________________________________________
Gecode users mailing list
users at gecode.org
https://www.gecode.org/mailman/listinfo/gecode-users