(2014/09/02 22:51), Johannes Weiner wrote:
On Mon, Sep 01, 2014 at 01:55:51PM +0100, Mel Gorman wrote:
Zones are allocated by the page allocator in either node or zone order.
Node ordering is preferred in terms of locality and is applied automatically
in one of three cases.
1. If a node has only low memory
2. If DMA/DMA32 is a high percentage of memory
3. If low memory on a single node is greater than 70% of the node size
Otherwise zone ordering is used to preserve low memory. Unfortunately
a consequence of this is that a machine with balanced NUMA nodes will
experience different performance characteristics depending on which node
they happen to start from.
The point of zone ordering is to protect lower nodes for devices that require
DMA/DMA32 memory. When NUMA was first introduced, this was critical as 32-bit
NUMA machines commonly suffered from low memory exhaustion problems. On
64-bit machines the primary concern is devices that are 32-bit only which
is less severe than the low memory exhaustion problem on 32-bit NUMA. It
seems there are really few devices that depends on it.
AGP -- I assume this is getting more rare but even then I think the allocations
happen early in boot time where lowmem pressure is less of a problem
DRM -- If the device is 32-bit only then there may be low pressure. I didn't
evaluate these in detail but it looks like some of these are mobile
graphics card. Not many NUMA laptops out there. DRM folk should know
better though.
Some TV cards -- Much demand for 32-bit capable TV cards on NUMA machines?
B43 wireless card -- again not really a NUMA thing.
I cannot find a good reason to incur a performance penalty on all 64-bit NUMA
machines in case someone throws a brain damanged TV or graphics card in there.
This patch defaults to node-ordering on 64-bit NUMA machines. I was tempted
to make it default everywhere but I understand that some embedded arches may
be using 32-bit NUMA where I cannot predict the consequences.
This patch is a step in the right direction, but I'm not too fond of
further fragmenting this code and where it applies, while leaving all
the complexity from the heuristics and the zonelist building in, just
on spec. Could we at least remove the heuristics too? If anybody is
affected by this, they can always override the default on the cmdline.
I'm okay with removing heuristics. There were a request to add "automatic detection"
at the time this feature was developped. But I'm not sure whether the logic is
still required. i.e. at that age, node-0 memory was small and default node order
can cause OOM easily.
Thanks,
-Kame
--
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>