Le 13/12/2011 19:30, Samuel Thibault a écrit : > Ok, I thought you were referring to an architectural detail. Well, > actually NUMA nodes and embedded memory should just both use a MEM > object, instead of duplicating all kinds of objects. We won't duplicate > such things for the MIC, will we?
My feeling about MIC is that it would something very similar to the current contents of a socket object inside a PCI device. And if MIC evers bypasses PCI and comes onto a real socket, it would make sense too. However I am far from understanding the implication on the core code and API of having such a hierarchy inside a PCI device, so I don't know for sure if that's a good idea. > IIRC the problems I had with using NODE was things like: make sure > total_memory doesn't count it, having a "name" to distinguish "shared" > and "global", avoid letting it enter the "levels". I think I agree that embedded mem and NUMA node should be MEM. New attributes could then be used distinguish between them to solve your problems above. This reminds me of the discussion before merging the PCI branch. It took way too long before we released a PCI-enabled hwloc, but at least we had time to mature the interface. Merging this new branch is probably even more important because it may have implications on future support for MIC and friends. > But all of these > could just be fixed by not browsing inside PCI devices, which we might > already be doing, actually. Yeah not done yet, but easy to do. Brice
