Re: [PATCH 2/6] mm/page_alloc: Disassociate the pcp->high from pcp->batch

From: Dave Hansen
Date: Fri May 21 2021 - 17:52:43 EST


On 5/21/21 3:28 AM, Mel Gorman wrote:
> Note that in this patch the pcp->high values are adjusted after memory
> hotplug events, min_free_kbytes adjustments and watermark scale factor
> adjustments but not CPU hotplug events.

Not that it was a long wait to figure it out, but I'd probably say:

"CPU hotplug events are handled later in the series".

instead of just saying they're not handled.

> Before grep -E "high:|batch" /proc/zoneinfo | tail -2
> high: 378
> batch: 63
>
> After grep -E "high:|batch" /proc/zoneinfo | tail -2
> high: 649
> batch: 63

You noted the relationship between pcp->high and zone lock contention.
Larger ->high values mean less contention. It's probably also worth
noting the trend of having more logical CPUs per NUMA node.

I have the feeling when this was put in place it wasn't uncommon to have
somewhere between 1 and 8 CPUs in a node pounding on a zone.

Today, having ~60 is common. I've occasionally resorted to recommending
that folks enable hardware features like Sub-NUMA-Clustering [1] since
it increases the number of zones and decreases the number of CPUs
pounding on each zone lock.

1.
https://software.intel.com/content/www/us/en/develop/articles/intel-xeon-processor-scalable-family-technical-overview.html

> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index a48f305f0381..bf5cdc466e6c 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -2163,14 +2163,6 @@ void __init page_alloc_init_late(void)
> /* Block until all are initialised */
> wait_for_completion(&pgdat_init_all_done_comp);
>
> - /*
> - * The number of managed pages has changed due to the initialisation
> - * so the pcpu batch and high limits needs to be updated or the limits
> - * will be artificially small.
> - */
> - for_each_populated_zone(zone)
> - zone_pcp_update(zone);
> -
> /*
> * We initialized the rest of the deferred pages. Permanently disable
> * on-demand struct page initialization.
> @@ -6594,13 +6586,12 @@ static int zone_batchsize(struct zone *zone)
> int batch;
>
> /*
> - * The per-cpu-pages pools are set to around 1000th of the
> - * size of the zone.
> + * The number of pages to batch allocate is either 0.1%

Probably worth making that "~0.1%" just in case someone goes looking for
the /1000 and can't find it.

> + * of the zone or 1MB, whichever is smaller. The batch
> + * size is striking a balance between allocation latency
> + * and zone lock contention.
> */
> - batch = zone_managed_pages(zone) / 1024;
> - /* But no more than a meg. */
> - if (batch * PAGE_SIZE > 1024 * 1024)
> - batch = (1024 * 1024) / PAGE_SIZE;
> + batch = min(zone_managed_pages(zone) >> 10, (1024 * 1024) / PAGE_SIZE);
> batch /= 4; /* We effectively *= 4 below */
> if (batch < 1)
> batch = 1;
> @@ -6637,6 +6628,27 @@ static int zone_batchsize(struct zone *zone)
> #endif
> }
>
> +static int zone_highsize(struct zone *zone)
> +{
> +#ifdef CONFIG_MMU
> + int high;
> + int nr_local_cpus;
> +
> + /*
> + * The high value of the pcp is based on the zone low watermark
> + * when reclaim is potentially active spread across the online
> + * CPUs local to a zone. Note that early in boot that CPUs may
> + * not be online yet.
> + */

FWIW, I like the way the changelog talked about this a bit better, with
the goal of avoiding background reclaim even in the face of a bunch of
full pcp's.

> + nr_local_cpus = max(1U, cpumask_weight(cpumask_of_node(zone_to_nid(zone))));
> + high = low_wmark_pages(zone) / nr_local_cpus;

I'm a little concerned that this might get out of hand on really big
nodes with no CPUs. For persistent memory (which we *do* toss into the
page allocator for volatile use), we can have multi-terabyte zones with
no CPUs in the node.

Also, while the CPUs which are on the node are the ones *most* likely to
be hitting the ->high limit, we do *keep* a pcp for each possible CPU.
So, the amount of memory which can actually be sequestered is
num_online_cpus()*high. Right?

*That* might really get out of hand if we have nr_local_cpus=1.

We might want some overall cap on 'high', or even to scale it
differently for the zone-local cpus' pcps versus remote.