Open MPI logo

Hardware Locality Development Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Hardware Locality Development mailing list

Subject: [hwloc-devel] HWLOC_OBJ_GROUP & hwloc_topology_support
From: Fawzi Mohamed (fawzi_at_[hidden])
Date: 2010-04-14 08:05:45


Always a bit sow, but here are my comments after the latest changes (I
wake up when I get time to update my bindings...)

HWLOC_OBJ_GROUP

nice!, I like it, it makes my job easier, as one of the main things I
use hwloc for is exactly having groups of uniform latency.
I suppose that those groups still form a partition, if not it should
be noted in the documentation.
One should note that a priori nodes of uniform latency do *not* build
a partition, for example in a ring topology with 4 elements (4,1,2)
(1,2,3) (2,3,4) (3,4,1) are the groups of uniform latency coming from
1,2,3,4.
But I guess that breaking the simple partitioning is not a good idea
in the "normal" hwloc view, and an non partitioning hierarchy is
better kept separated.

hwloc_topology_support

I know that was changed also due to my comments, but I am not sure the
change really better: the structure is not really hidden, so adding a
flags it breaks binary compatibility, well it is not as bad as it
sounds, because the normally the structure should be only read, and
one receives only pointers, not directly the structure, so even using
old bindings should not really lead to catastrophic failures, but I
find it still non clean.
With a user visible structure I find enum flags bla=(1<<x) still the
best solution.
Anyway for a readonly structure also the current solution is certainly
viable...

ciao
Fawzi