[lkml]   [2020]   [Mar]   [29]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [LKP] Re: [mm] fd4d9c7d0c: stress-ng.switch.ops_per_sec -30.5% regression
Hi Linus,

On Fri, Mar 27, 2020 at 12:57:45AM +0800, Linus Torvalds wrote:
> On Wed, Mar 25, 2020 at 10:57 PM kernel test robot
> <> wrote:
> >
> > FYI, we noticed a -30.5% regression of stress-ng.switch.ops_per_sec due to commit:
> >
> > commit: fd4d9c7d0c71866ec0c2825189ebd2ce35bd95b8 ("mm: slub: add missing TID bump in kmem_cache_alloc_bulk()")
> This looks odd.
> I would not expect the update of c->tid to have that noticeable an
> impact, even on a big machine that might be close to some scaling
> limit.

The test machin is a Cascade Lake platform, 2 sockets, 48C/96T.

> It doesn't add any expensive atomic ops, and while it _could_ make a
> percpu cacheline dirty, I think that cacheline should already be dirty
> anyway under any load where this is noticeable. Plus this should be a
> relatively cold path anyway.
> So mind humoring me, and double-check that regression?
> Of course, it might be another "just magic cache placement" detail
> where code moved enough to make a difference.
> Or maybe it really ends up causing new tid mismatches and we end up
> failing the fast path in slub as a result. But looking at the stats
> that changed in your message doesn't make me go "yeah, that looks like
> a slub difference".

Per our check, the code movement does exist.

From the system map:

old map:
ffffffff812a1880 T kmem_cache_alloc_bulk
ffffffff812a1a80 t kmalloc_large_node
ffffffff812a1b10 t calculate_sizes
ffffffff812a1eb0 t store_user_store
ffffffff812a1f20 t poison_store
ffffffff812a1f90 t red_zone_store
ffffffff812a2000 t order_store

new map:
ffffffff812a1880 T kmem_cache_alloc_bulk
ffffffff812a1a90 t kmalloc_large_node
ffffffff812a1b20 T __kmalloc_node ---> relocated
ffffffff812a1e40 t calculate_sizes
ffffffff812a21e0 t store_user_store
ffffffff812a2250 t poison_store
ffffffff812a22c0 t red_zone_store
ffffffff812a2330 t order_store

In old map the 'kmem_cache_alloc_bulk' is cache aligned, and occupies
0x200 bytes, and the next function 'kmalloc_large_node' starts at
an alinged address. In new map 'kmem_cache_alloc_bulk' occupies
0x210 bytes, and offset of the alignment of many functions following
it. (please let us know if you need the full system map for the
2 vmlinuxs)

From the objdump, the direct chagne of "c->tid = next_tid(c->tid);"
is one line added "49 83 40 08 01 addq $0x1,0x8(%r8)"

We did experiments to make the kernel functions 32 bytes aligned,
diff --git a/Makefile b/Makefile
index 171f2b004c8a..63f28aaf78c9 100644
--- a/Makefile
+++ b/Makefile

-KBUILD_CFLAGS := -Wall -Wundef -Werror=strict-prototypes -Wno-trigraphs \
+KBUILD_CFLAGS := -Wall -Wundef -falign-functions=32 -Werror=strict-prototypes -Wno-trigraphs \

the regression is reduced to about 3%:

2060457 ± 4% -3.2% 1993685 ± 2% stress-ng.switch.ops_per_sec

which is pretty small for a micro-benchmark


 \ /
  Last update: 2020-03-30 03:11    [W:0.087 / U:10.332 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site