Messages in this thread | | | Date | Thu, 13 Feb 2020 11:18:02 +0000 | From | Mel Gorman <> | Subject | Re: [PATCH 08/11] sched/numa: Bias swapping tasks based on their preferred node |
| |
On Thu, Feb 13, 2020 at 11:31:08AM +0100, Peter Zijlstra wrote: > On Wed, Feb 12, 2020 at 09:36:51AM +0000, Mel Gorman wrote: > > When swapping tasks for NUMA balancing, it is preferred that tasks move > > to or remain on their preferred node. When considering an imbalance, > > encourage tasks to move to their preferred node and discourage tasks from > > moving away from their preferred node. > > Wasn't there an issue for workloads that span multiple nodes? >
Sortof, yes -- specifically workloads that could not fit inside a node for whatever reason.
> Say a 4 node system with 2 warehouses? Then each JVM will want 2 nodes, > instead of a single node, and strong preferred node stuff makes it > difficult to achieve this. > > I forgot how we dealt with these cases, just something I worry about > when reading this.
We deal with it in task_numa_migrate() by considering nodes other than the preferred node for placement -- see "Look at other nodes in these cases" followed by a sched_setnuma if the preferred node doesn't match.
We do not do any special casing as such in task_numa_compare other than finding the best improvement so we can pick a task belonging to a group spanning multiple nodes with or without this patch. A workload spanning multiple nodes in itself does not justify a full search if it can be avoided.
-- Mel Gorman SUSE Labs
| |