Yes, this is late. Sorry. I'd like to discuss the following topic: -- Presently, NUMA layout is determined at boot time and never changes again. This setup works for real hardware, but virtual machines are more dynamic: they could be migrated between different hosts, and have to share the physical memory space with other VMs which are also being moved around or shut down while other new VMs are started up. As a result, the physical backing memory that a VM had when it started up changes at runtime. Problems to be overcome: - How should userspace be notified? Do we need new interfaces so applications can query memory to see if it was affected? - Can we make the NUMA layout initialization generic? This also implies that all initialization of struct zone/struct page/NODE_DATA() would be made (somewhat) generic. - Some one-time allocations now will know they are on a non-optimal node. - hotpluged per node data is (in general) not being allocated optimally) - NODE_DATA() for hotpluged nodes is allocated off-node (except for ia64). - SLUB's kmem_cache_node is always allocated off-node for hotpluged nodes. [Not a new problem, but one that needs solving]. Some more generic NUMA layout/mm init things: - boot-time and hotplug NUMA init don't share enough code. - architectures do not share mm init code - NUMA layout (from init) is kept (if it is kept at all) in only arch specific ways. Memblock _happens_ to contain this info, while also also tracking allocations, and every arch but powerpc discards it as __init/__initdata) A WIP patchset addressing initial reconfiguration of the page allocator: https://github.com/jmesmon/linux/tree/dnuma/v25 -- Cody P Schafer -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>