Higher slub memory consumption on 64K page-size systems?

From: Bharata B Rao
Date: Wed Oct 28 2020 - 17:46:31 EST


Hi,

On POWER systems, where 64K PAGE_SIZE is default, I see that slub
consumes higher amount of memory compared to any 4K page-size system.
While slub is obviously going to consume more memory on 64K page-size
systems compared to 4K as slabs are allocated in page-size granularity,
I want to check if there are any obvious tuning (via existing tunables
or via some code change) that we can do to reduce the amount of memory
consumed by slub.

Here is a comparision of the slab memory consumption between 4K and
64K page-size pseries hash KVM guest with 16 cores and 16G memory
configuration immediately after boot:

64K 209280 kB
4K 67636 kB

64K configuration may never be able to consume as less as a 4K configuration,
but it certainly shows that the slub can be optimized for 64K page-size better.

slub_max_order
--------------
The most promising tunable that shows consistent reduction in slab memory
is slub_max_order. Here is a table that shows the number of slabs that
end up with different orders and the total slab consumption at boot
for different values of slub_max_order:
-------------------------------------------
slub_max_order Order NrSlabs Slab memory
-------------------------------------------
0 276
3 1 16 207488 kB
(default) 2 4
3 11
-------------------------------------------
0 276
2 1 16 166656 kB
2 4
-------------------------------------------
0 276 144128 kB
1 1 31
-------------------------------------------

Though only a few bigger sized caches fall into order-2 or order-3, they
seem to make a considerable difference to the overall slab consumption.
If we take task_struct cache as an example, this is how it ends up when
slub_max_order is varied:

task_struct, objsize=9856
--------------------------------------------
slub_max_order objperslab pagesperslab
--------------------------------------------
3 53 8
2 26 4
1 13 2
--------------------------------------------

The slab page-order and hence the number of objects in a slab has a
bearing on the performance, but I wonder if some caches like task_struct
above can be auto-tuned to fall into a conservative order and do good
both wrt both memory and performance?

mm/slub.c:calulate_order() has the logic which determines the the
page-order for the slab. It starts with min_objects and attempts
to arrive at the best configuration for the slab. The min_objects
is starts like this:

min_objects = 4 * (fls(nr_cpu_ids) + 1);

Here nr_cpu_ids depends on the maxcpus and hence this can have a
significant effect on those systems which define maxcpus. Slab numbers
post-boot for a KVM pseries guest that has 16 boottime CPUs and varying
number of maxcpus look like this:
-------------------------------
maxcpus Slab memory(kB)
-------------------------------
64 209280
256 253824
512 293824
-------------------------------

Page-order is a one time setting and obviously can't be tweaked dynamically
on CPU hotplug, but just wanted to bring out the effect of the same.

And that constant multiplicative factor of 4 was infact added by the commit
9b2cd506e5f2 - "slub: Calculate min_objects based on number of processors."

Reducing that to say 2, does give some reduction in the slab memory
and also same hackbench performance with reduced slab memory, but I am not
sure if that could be assumed to be beneficial for all scenarios.

MIN_PARTIAL
-----------
This determines the number of slabs left on the partial list even if they
are empty. My initial thought was that the default MIN_PARTIAL value of 5
is on the higher side and we are accumulating MIN_PARTIAL number of
empty slabs in all caches without freeing them. However I hardly find
the case where an empty slab is retained during freeing on account of
partial slabs being lesser than MIN_PARTIAL.

However what I find in practice is that we are accumulating a lot of partial
slabs with just one in-use object in the whole slab. High number of such
partial slabs is indeed contributing to the increased slab memory consumption.

For example, after a hackbench run, I find the distribution of objects
like this for kmalloc-2k cache:

total_objects 3168
objects 1611
Nr partial slabs 54
Nr parital slabs with
just 1 inuse object 38

With 64K page-size, so many partial slabs with just 1 inuse object can
result in high memory usage. Is there any workaround possible prevent this
kind of situation?

cpu_partial
-----------
Here is how the slab consumption post-boot varies when all the slab
caches are forced with the fixed cpu_partial value:
---------------------------
cpu_partial Slab Memory
---------------------------
0 175872 kB
2 187136 kB
4 191616 kB
default 204864 kB
---------------------------

It has been suggested earlier that reducing cpu_partial and/or making
cpu_partial 64K page-size aware will benefit. In set_cpu_partial(),
for bigger sized slabs (size > PAGE_SIZE), cpu_partial is already set
to 2. A bit of tweaking there to introduce cpu_partial=1 for certain
slabs does give some benefit.

diff --git a/mm/slub.c b/mm/slub.c
index a28ed9b8fc61..e09eff1199bf 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -3626,7 +3626,9 @@ static void set_cpu_partial(struct kmem_cache *s)
*/
if (!kmem_cache_has_cpu_partial(s))
slub_set_cpu_partial(s, 0);
- else if (s->size >= PAGE_SIZE)
+ else if (s->size >= 8192)
+ slub_set_cpu_partial(s, 1);
+ else if (s->size >= 4096)
slub_set_cpu_partial(s, 2);
else if (s->size >= 1024)
slub_set_cpu_partial(s, 6);

With the above change, the slab consumption post-boot reduces to 186048 kB.
Also, here are the hackbench numbers with and w/o the above change:

Average of 10 runs of 'hackbench -s 1024 -l 200 -g 200 -f 25 -P'
Slab consumption captured at the end of each run
--------------------------------------------------------------
Time Slab memory
--------------------------------------------------------------
Default 11.124s 645580 kB
Patched 11.032s 584352 kB
--------------------------------------------------------------

I have mostly looked at reducing the slab memory consumption here.
But I do understand that default tunable values have been arrived
at based on some benchmark numbers. Are there ways or possibilities
to reduce the slub memory consumption with the existing level of
performance is what I would like to understand and explore.

Regards,
Bharata.