Messages in this thread | | | Date | Fri, 12 Nov 2021 10:58:42 +0800 | Subject | Re: [PATCH v3] mm: migrate: Support multiple target nodes demotion | From | Baolin Wang <> |
| |
On 2021/11/12 10:44, Huang, Ying wrote: > Baolin Wang <baolin.wang@linux.alibaba.com> writes: > >> We have some machines with multiple memory types like below, which >> have one fast (DRAM) memory node and two slow (persistent memory) memory >> nodes. According to current node demotion policy, if node 0 fills up, >> its memory should be migrated to node 1, when node 1 fills up, its >> memory will be migrated to node 2: node 0 -> node 1 -> node 2 ->stop. >> >> But this is not efficient and suitbale memory migration route >> for our machine with multiple slow memory nodes. Since the distance >> between node 0 to node 1 and node 0 to node 2 is equal, and memory >> migration between slow memory nodes will increase persistent memory >> bandwidth greatly, which will hurt the whole system's performance. >> >> Thus for this case, we can treat the slow memory node 1 and node 2 >> as a whole slow memory region, and we should migrate memory from >> node 0 to node 1 and node 2 if node 0 fills up. >> >> This patch changes the node_demotion data structure to support multiple >> target nodes, and establishes the migration path to support multiple >> target nodes with validating if the node distance is the best or not. >> >> available: 3 nodes (0-2) >> node 0 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 >> node 0 size: 62153 MB >> node 0 free: 55135 MB >> node 1 cpus: >> node 1 size: 127007 MB >> node 1 free: 126930 MB >> node 2 cpus: >> node 2 size: 126968 MB >> node 2 free: 126878 MB >> node distances: >> node 0 1 2 >> 0: 10 20 20 >> 1: 20 10 20 >> 2: 20 20 10 >> >> Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
snip
>> /* >> * 'next_pass' contains nodes which became migration >> @@ -3192,6 +3281,14 @@ static int __init migrate_on_reclaim_init(void) >> { >> int ret; >> >> + /* >> + * Ignore allocation failure, if this kmalloc fails >> + * at boot time, we are likely in bigger trouble. >> + */ >> + node_demotion = kmalloc_array(nr_node_ids, >> + sizeof(struct demotion_nodes), >> + GFP_KERNEL); >> + > > I think we should WARN_ON() here.
In this unlikey case, I think the mm core will print more information, IMHO WARN_ON() will help little. Anyway no strong opinion on this. Other than that, can I get your reviewed-by tag with this nit fixed? Thanks.
| |