On Tue, Nov 22, 2022, at 17:59, Vlastimil Babka wrote:
> On 11/22/22 17:33, Arnd Bergmann wrote:
>> On Mon, Nov 21, 2022, at 18:11, Vlastimil Babka wrote:
>> I can imagine those machines wanting to use sysfs in general
>> but not for the slab caches, so having a separate knob to
>> configure out the sysfs stuff could be useful without having
>> to go all the way to SLUB_TINY.
>
> Right, but AFAIK that wouldn't save much except some text size and kobjects,
> so probably negligible for >32MB?
Makes sense, I assume you have a better idea of how much this
could save. I'm not at all worried about the .text size, but
my initial guess was that the metadata for sysfs would be
noticeable.
>> For the options that trade off performance against lower
>> fragmentation (MIN/MAX_PARTIAL, KMALLOC_RECLAIM, percpu
>> slabs), I wonder if it's possible to have a boot time
>> default based on the amount of RAM per CPU to have a better
>> tuned system on most cases, rather than having to go
>> to one extreme or the other at compile time.
>
> Possible for some of these things, but for others that brings us back to the
> question what are the actual observed issues. If it's low memory in absolute
> number of pages, these can help, but if it's fragmentation (and the kind if
> RAM sizes should have page grouping by mobility enabled), ditching e.g. the
> KMALLOC_RECLAIM could make it worse. Unfortunately some of these tradeoffs
> can be rather unpredictable.
Are there any obvious wins on memory uage? I would guess that it
would be safe to e.g. ditch percpu slabs when running with less
128MB per CPU, and the MIN/MAX_PARTIAL values could easily
be a function of the number of pages in total or per cpu,
whichever makes most sense. As a side-effect, those could also
grow slightly larger on huge systems by scaling them with
log2(totalpages).
Arnd
