Open MPI logo

Hardware Locality Development Mailing List Archives

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

Subject: Re: [hwloc-devel] hwloc support added in Grid Engine / Grid Scheduler
From: Brice Goglin (Brice.Goglin_at_[hidden])
Date: 2011-04-13 16:53:16

Le 13/04/2011 22:47, Jeff Squyres a écrit :
> On Apr 13, 2011, at 1:32 PM, Brice Goglin wrote:
>> Don't worry about the depth. Even if you have less levels of caches in
>> some parts of the machine, or no hyperthreading or whatever different,
>> we still build levels of identical objects. So you still get a
>> consistent depth for all cores, threads, sockets, and different levels
>> of caches. But obviously, you will see less objects when iterating
>> through levels within the "smaller" part of the machine.
> This would be good to add to the documentation; I've been doing manual traversals of hwloc trees just to future-proof/heterogeneous-proof my code (because I didn't know that the depth would always be the same for a given object type even if some branches don't have all object types).
> So let's say I have a (fictitious) machine with one socket with an L1 and L2 cache for its cores, and another socket that only has an L1 cache for its cores:
> numa_node
> / \
> socket 0 socket 1
> L2 L1
> L1 core
> core
> If I traverse down through children in the tree, are you saying that there will be some kind of fake/empty object in the right sub tree corresponding to where the L2 is in the left sub tree?

There will no fake object. socket1 will be at depth 1 and its child will
be the second L1 at depth 3. There's nothing at depth 2 in the right
part of the machine.

We'll try to make this more clear in the doc.