On Fri, Aug 30, 2019 at 08:39:43AM +0200, Greg KH wrote:
> On Fri, Aug 30, 2019 at 02:59:01AM +0800, Peikan Tsai wrote:
> > On Thu, Aug 29, 2019 at 05:27:22PM +0200, Christian Brauner wrote:
> > > On Thu, Aug 29, 2019 at 09:53:59AM -0400, Joel Fernandes wrote:
> > > > On Thu, Aug 29, 2019 at 08:42:29AM +0200, Greg KH wrote:
> > > > > On Thu, Aug 29, 2019 at 01:49:53PM +0800, Peikan Tsai wrote:
> > > > [snip]
> > > > > > The allocated size for each binder_thread is 512 bytes by kzalloc.
> > > > > > Because the size of binder_thread is fixed and it's only 304 bytes.
> > > > > > It will save 208 bytes per binder_thread when use create a kmem_cache
> > > > > > for the binder_thread.
> > > > >
> > > > > Are you _sure_ it really will save that much memory? You want to do
> > > > > allocations based on a nice alignment for lots of good reasons,
> > > > > especially for something that needs quick accesses.
> > > >
> > > > Alignment can be done for slab allocations, kmem_cache_create() takes an
> > > > align argument. I am not sure what the default alignment of objects is
> > > > though (probably no default alignment). What is an optimal alignment in your
> > > > view?
> > >
> > > Probably SLAB_HWCACHE_ALIGN would make most sense.
> > >
> >
> > Agree. Thanks for yours comments and suggestions.
> > I'll put SLAB_HWCACHE_ALIGN it in patch v2.
> >
> > > >
> > > > > Did you test your change on a system that relies on binder and find any
> > > > > speed improvement or decrease, and any actual memory savings?
> > > > >
> > > > > If so, can you post your results?
> > > >
> > > > That's certainly worth it and I thought of asking for the same, but spoke too
> > > > soon!
> > >
> > > Yeah, it'd be interesting to see what difference this actually makes.
> > >
> > > Christian
> >
> > I tested this change on an Android device(arm) with AOSP kernel 4.19 and
> > observed
> > memory usage of binder_thread. But I didn't do binder benchmark yet.
> >
> > On my platform the memory usage of binder_thread reduce about 90 KB as
> > the
> > following result.
> > nr obj obj size total
> > before: 624 512 319488 bytes
> > after: 728 312 227136 bytes
>
> You have more objects???
>
Sorry, it's total objects which include some inactive objects ...
And because I tested it on an Android platform so there may be some noise.
So I try 'adb stop' and 'echo 3 > /proc/sys/vm/drop_caches' before starting
test to reduce the noise, and the result are as following.
objs
kzalloc 220 (kmalloc-512 alloc by binder_get_thread)
active_objs total objs objperslab slabdata
kmem_cache 194 403 13 31
Seems there are more objects when use kmemcache for binder_thread...
But as I understand it, those inactive objects can be free by kmemcahe shrink?
Also, I tested the throughput by using performace test of Android VTS.
size(bytes) kzalloc(byte/ns) kmemcache(byte/ns)
4 0.17 0.17
8 0.33 0.32
16 0.66 0.66
32 1.36 1.42
64 2.66 2.61
128 5.4 5.26
256 10.29 10.77
512 21.51 21.36
1k 41 40.26
2k 82.12 80.28
4k 149.24 146.95
8k 262.34 256
16k 417.96 422.2
32k 596.66 590.23
64k 600.84 601.25
_______________________________________________
devel mailing list
devel@xxxxxxxxxxxxxxxxxxxxxx
http://driverdev.linuxdriverproject.org/mailman/listinfo/driverdev-devel
