heterogeneous topology support

[ompiteam]

what if we have a machine with different processors? for instance if one socket has a shared L3 and the other one doesn't?

• sthibaul: should work fine already

support levels that do not cover the whole machine? (no L3 above the cores of the second socket above)

• sthibaul: I do not understand

support object whose children are not in the exact next level? (socket pointing to cores instead of cache above)

• sthibaul: should work fine already

need to be decided if we want to put GPUs as hwloc_obj_t, see ticket:5

[ompiteam]

Imported from trac issue 6. Created by bgoglin on 2009-09-24T01:06:50, last modified: 2012-12-22T02:52:25

[ompiteam]

Trac comment by bgoglin on 2009-09-28 01:54:58:

"support levels that do not cover the whole machine? (no L3 above the cores of the second socket above)"

There's currently an implicit rule about every PROC being covered by one object of each level. If it's not the case, we need to rethink some functions. It's clear that some current functions in the helpers need to be their behavior clarified/fixed in this case. I haven't actually tested the code yet, will built a fake topology today.

[ompiteam]

Trac comment by sthibaul on 2009-09-28 13:07:32:

Yes, I mean the GPU level (i.e. just a list of the available GPUs) would be independent from core/cache levels for instance.

[ompiteam]

Trac comment by bgoglin on 2009-09-29 07:50:29:

After a deeper look at the code and some tests:

• There seems to be some breakage in core discovery, causing multiple PROC levels to be created in the heterogeneous linux test.
• The helper functions actually look ok. I didn't try them for real but reviewing the code makes me feel good about them. They don't seem to rely on having enough objects in a single level to cover any given cpuset (which was my first fear, 2nd point of initial report).
• I am not sure what hwloc_distribute() should do if children are not at the same level. Maybe it doesn't matter for real cases but we'll probably see some "missbalanced" distribution on heterogeneous topologies :)

Once the heterogeneous tests/linux case works as expected, this bug could probably be closed.

[ompiteam]

Trac comment by sthibaul on 2009-09-29 08:17:23:

linux testcase fixed by r1026 (stupid error while coding the algorithm).