[RFC] net: further seperate dst_entry.__refcnt from cache contention

[RFC] net: further seperate dst_entry.__refcnt from cache contention

[Nathan Zimmer]

After some investigation on large machines I found that
dst_entry.__refcnt particpates in false cache sharing issues that show
when scaling past 12 threads who communicate via tcp with loopback addresses.
I adjusted refcnt to be on its own cache line and that helped quite a bit.
But perhaps a bit of a waste of space?  Is there some better way?

Here is some preliminary data I had gathered.  It shows nicely improved scaling.

Threads baseline   afterchange
2       1328.03    1340.67
4       2430.31    2282.09
6       3087.65    3258.12
8       3560.34    4165.88
10      3900.34    4962.28
12      3933.38    5613.76
14      3876.98    6113.85
16      3706.01    6338.00
18      3742.48    6634.77
20      3670.15    6641.25
22      3660.98    6799.55
24      3503.13    6613.45
26      3525.73    6702.67
28      3440.16    6801.27
30      3497.59    6911.52
32      3498.25    6540.06

I should say something about this test.  It is a dead simple test in which a
pair of threads simply pass data to each other.  They were placed in the same
socket to avoid cross node overhead.

CC: "David S. Miller" <davem@davemloft.net> 
Signed-off-by: Nathan Zimmer <nzimmer@sgi.com>

---
 include/net/dst.h |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

diff --git a/include/net/dst.h b/include/net/dst.h
index 8197ead..3898643 100644
--- a/include/net/dst.h
+++ b/include/net/dst.h
@@ -84,7 +84,7 @@ struct dst_entry {
 	 * input/output/ops or performance tanks badly
 	 */
 	atomic_t		__refcnt;	/* client references	*/
-	int			__use;
+	int			__use	____cacheline_aligned;
 	unsigned long		lastuse;
 	union {
 		struct dst_entry	*next;
-- 
1.6.0.2

Re: [RFC] net: further seperate dst_entry.__refcnt from cache contention

[Eric Dumazet]

On Fri, 2012-07-20 at 14:46 -0500, Nathan Zimmer wrote:
> After some investigation on large machines I found that
> dst_entry.__refcnt particpates in false cache sharing issues that show
> when scaling past 12 threads who communicate via tcp with loopback addresses.
> I adjusted refcnt to be on its own cache line and that helped quite a bit.
> But perhaps a bit of a waste of space?  Is there some better way?
> 
> Here is some preliminary data I had gathered.  It shows nicely improved scaling.
> 
> Threads baseline   afterchange
> 2       1328.03    1340.67
> 4       2430.31    2282.09
> 6       3087.65    3258.12
> 8       3560.34    4165.88
> 10      3900.34    4962.28
> 12      3933.38    5613.76
> 14      3876.98    6113.85
> 16      3706.01    6338.00
> 18      3742.48    6634.77
> 20      3670.15    6641.25
> 22      3660.98    6799.55
> 24      3503.13    6613.45
> 26      3525.73    6702.67
> 28      3440.16    6801.27
> 30      3497.59    6911.52
> 32      3498.25    6540.06
> 
> I should say something about this test.  It is a dead simple test in which a
> pair of threads simply pass data to each other.  They were placed in the same
> socket to avoid cross node overhead.
> 
> CC: "David S. Miller" <davem@davemloft.net> 
> Signed-off-by: Nathan Zimmer <nzimmer@sgi.com>
> 
> ---
>  include/net/dst.h |    2 +-
>  1 files changed, 1 insertions(+), 1 deletions(-)
> 
> diff --git a/include/net/dst.h b/include/net/dst.h
> index 8197ead..3898643 100644
> --- a/include/net/dst.h
> +++ b/include/net/dst.h
> @@ -84,7 +84,7 @@ struct dst_entry {
>  	 * input/output/ops or performance tanks badly
>  	 */
>  	atomic_t		__refcnt;	/* client references	*/
> -	int			__use;
> +	int			__use	____cacheline_aligned;
>  	unsigned long		lastuse;
>  	union {
>  		struct dst_entry	*next;

Its a known problem, and we are waiting IP cache removal to address it.

Before the cache removal, a machine can have million of dst

Another idea concerning very hot dst would be to clone them on demand.

Re: [RFC] net: further seperate dst_entry.__refcnt from cache contention

[David Miller]

From: Eric Dumazet <eric.dumazet@gmail.com>
Date: Fri, 20 Jul 2012 22:16:51 +0200

> Another idea concerning very hot dst would be to clone them on demand.

The FIB info cached dsts could also be made per-cpu.

And yes, the only reason I'm entertaining that idea now is the fact
that we'll have the number of dsts in the system under control.