Re: [pooma-dev] [PATCH] parallel particle bctest3 crash

[Arno Candel]

okay, this patch fixes the UniformLayout particle swapping, and 
particularly the bctest3 particle test.

There was a floating point exception triggered as soon as less particles 
were created than patches existed -> sizePerPatch=0.
A there was an assertion error when the particle was being sent to a 
patch that didn't even exist (npid >= patchesGlobal()).


Arno.



Index: UniformLayout.h
===================================================================
RCS file: /home/pooma/Repository/r2/src/Particles/UniformLayout.h,v
retrieving revision 1.23
diff -u -r1.23 UniformLayout.h
--- UniformLayout.h     14 Jul 2004 15:44:59 -0000      1.23
+++ UniformLayout.h     10 Oct 2004 13:51:36 -0000
@@ -311,11 +311,11 @@

    for (int i = 0; i < size; ++i)
    {
-      int npid = (i+offset) / sizePerPatch;
+      int npid = (i+offset) / (sizePerPatch>0?sizePerPatch:1);

      // check for a leftover particle

-      if (npid == patchesGlobal())
+      if (npid >= patchesGlobal())
       npid = (i+offset) - (sizePerPatch*patchesGlobal());

      // Make sure this is kosher


Richard Guenther wrote:

> Arno Candel wrote:
>
> > > This is no wonder:
> > >
> > >   if (Pooma::context() == 0)
> > >     P.create(10,0,false);
> > >   P.sync(P.a1);
> > >
> > > i.e. we create 10 particles - distributing over 11 contexts isn't 
> > > going to work.  We don't handle contexts with zero particles 
> > > gracefully.  I think there are similar problems with #patches < 
> > > #contexts.  But I'd qualify these cases as user error.
> > >
> > > Hope this helps,
> > > Richard.
> >
> >
> > Unfortunately, that's not the problem. It still doesn't work after 
> > creating 11 or more particles.
>
>
> Hm ok, that was just a guess.
>
> > I've been using position-dependent distribution of particles for 
> > months now in my own code, often with many contexts not containing 
> > any particles. The problem only arises when the number of contexts is 
> > getting too high (mostly 8-10 being the limit). However, I have no 
> > idea what triggers the error...
>
>
> I don't have time at the moment to dig into this myself, maybe you can 
> try to debug this yourself - usually doing printf-debugging is the 
> only way to debug these parallel failures though (unless you have 
> expensive tools around, which I don't have).
>
> Richard.