On 01/03/2018 01:55 PM, Michal Hocko wrote: > From: Michal Hocko <mhocko@xxxxxxxx> > > do_pages_move is supposed to move user defined memory (an array of > addresses) to the user defined numa nodes (an array of nodes one for > each address). The user provided status array then contains resulting > numa node for each address or an error. The semantic of this function is > little bit confusing because only some errors are reported back. Notably > migrate_pages error is only reported via the return value. This patch > doesn't try to address these semantic nuances but rather change the > underlying implementation. > > Currently we are processing user input (which can be really large) > in batches which are stored to a temporarily allocated page. Each > address is resolved to its struct page and stored to page_to_node > structure along with the requested target numa node. The array of these > structures is then conveyed down the page migration path via private > argument. new_page_node then finds the corresponding structure and > allocates the proper target page. > > What is the problem with the current implementation and why to change > it? Apart from being quite ugly it also doesn't cope with unexpected > pages showing up on the migration list inside migrate_pages path. > That doesn't happen currently but the follow up patch would like to > make the thp migration code more clear and that would need to split a > THP into the list for some cases. > > How does the new implementation work? Well, instead of batching into a > fixed size array we simply batch all pages that should be migrated to > the same node and isolate all of them into a linked list which doesn't > require any additional storage. This should work reasonably well because > page migration usually migrates larger ranges of memory to a specific > node. So the common case should work equally well as the current > implementation. Even if somebody constructs an input where the target > numa nodes would be interleaved we shouldn't see a large performance > impact because page migration alone doesn't really benefit from > batching. mmap_sem batching for the lookup is quite questionable and > isolate_lru_page which would benefit from batching is not using it even > in the current implementation. Hi Michal, After slightly modifying your test case (like fixing the page size for powerpc and just doing simple migration from node 0 to 8 instead of the interleaving), I tried to measure the migration speed with and without the patches on mainline. Its interesting.... 10000 pages | 100000 pages -------------------------- Mainline 165 ms 1674 ms Mainline + first patch (move_pages) 191 ms 1952 ms Mainline + all three patches 146 ms 1469 ms Though overall it gives performance improvement, some how it slows down migration after the first patch. Will look into this further. -- 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>
