Re: tbench regression in 2.6.25-rc1

["Zhang, Yanmin" <yanmin_zhang@linux.intel.com>]

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: text/plain; charset=utf-8, Size: 5384 bytes --]

On Tue, 2008-02-19 at 08:35 +0100, Eric Dumazet wrote:
> Zhang, Yanmin a écrit :
> > On Mon, 2008-02-18 at 11:11 +0100, Eric Dumazet wrote:
> >> On Mon, 18 Feb 2008 16:12:38 +0800
> >> "Zhang, Yanmin" <yanmin_zhang@linux.intel.com> wrote:
> >>
> >>> On Fri, 2008-02-15 at 15:22 -0800, David Miller wrote:
> >>>> From: Eric Dumazet <dada1@cosmosbay.com>
> >>>> Date: Fri, 15 Feb 2008 15:21:48 +0100
> >>>>
> >>>>> On linux-2.6.25-rc1 x86_64 :
> >>>>>
> >>>>> offsetof(struct dst_entry, lastuse)=0xb0
> >>>>> offsetof(struct dst_entry, __refcnt)=0xb8
> >>>>> offsetof(struct dst_entry, __use)=0xbc
> >>>>> offsetof(struct dst_entry, next)=0xc0
> >>>>>
> >>>>> So it should be optimal... I dont know why tbench prefers __refcnt being 
> >>>>> on 0xc0, since in this case lastuse will be on a different cache line...
> >>>>>
> >>>>> Each incoming IP packet will need to change lastuse, __refcnt and __use, 
> >>>>> so keeping them in the same cache line is a win.
> >>>>>
> >>>>> I suspect then that even this patch could help tbench, since it avoids 
> >>>>> writing lastuse...
> >>>> I think your suspicions are right, and even moreso
> >>>> it helps to keep __refcnt out of the same cache line
> >>>> as input/output/ops which are read-almost-entirely :-
> >>> I think you are right. The issue is these three variables sharing the same cache line
> >>> with input/output/ops.
> >>>
> >>>> )
> >>>>
> >>>> I haven't done an exhaustive analysis, but it seems that
> >>>> the write traffic to lastuse and __refcnt are about the
> >>>> same.  However if we find that __refcnt gets hit more
> >>>> than lastuse in this workload, it explains the regression.
> >>> I also think __refcnt is the key. I did a new testing by adding 2 unsigned long
> >>> pading before lastuse, so the 3 members are moved to next cache line. The performance is
> >>> recovered.
> >>>
> >>> How about below patch? Almost all performance is recovered with the new patch.
> >>>
> >>> Signed-off-by: Zhang Yanmin <yanmin.zhang@intel.com>
> >>>
> >>> ---
> >>>
> >>> --- linux-2.6.25-rc1/include/net/dst.h	2008-02-21 14:33:43.000000000 +0800
> >>> +++ linux-2.6.25-rc1_work/include/net/dst.h	2008-02-21 14:36:22.000000000 +0800
> >>> @@ -52,11 +52,10 @@ struct dst_entry
> >>>  	unsigned short		header_len;	/* more space at head required */
> >>>  	unsigned short		trailer_len;	/* space to reserve at tail */
> >>>  
> >>> -	u32			metrics[RTAX_MAX];
> >>> -	struct dst_entry	*path;
> >>> -
> >>> -	unsigned long		rate_last;	/* rate limiting for ICMP */
> >>>  	unsigned int		rate_tokens;
> >>> +	unsigned long		rate_last;	/* rate limiting for ICMP */
> >>> +
> >>> +	struct dst_entry	*path;
> >>>  
> >>>  #ifdef CONFIG_NET_CLS_ROUTE
> >>>  	__u32			tclassid;
> >>> @@ -70,10 +69,12 @@ struct dst_entry
> >>>  	int			(*output)(struct sk_buff*);
> >>>  
> >>>  	struct  dst_ops	        *ops;
> >>> -		
> >>> -	unsigned long		lastuse;
> >>> +
> >>> +	u32			metrics[RTAX_MAX];
> >>> +
> >>>  	atomic_t		__refcnt;	/* client references	*/
> >>>  	int			__use;
> >>> +	unsigned long		lastuse;
> >>>  	union {
> >>>  		struct dst_entry *next;
> >>>  		struct rtable    *rt_next;
> >>>
> >>>
> >> Well, after this patch, we grow dst_entry by 8 bytes :
> > With my .config, it doesn't grow. Perhaps because of CONFIG_NET_CLS_ROUTE, I don't
> > enable it. I will move tclassid under ops.
> > 
> >> sizeof(struct dst_entry)=0xd0
> >> offsetof(struct dst_entry, input)=0x68
> >> offsetof(struct dst_entry, output)=0x70
> >> offsetof(struct dst_entry, __refcnt)=0xb4
> >> offsetof(struct dst_entry, lastuse)=0xc0
> >> offsetof(struct dst_entry, __use)=0xb8
> >> sizeof(struct rtable)=0x140
> >>
> >>
> >> So we dirty two cache lines instead of one, unless your cpu have 128 bytes cache lines ?
> >>
> >> I am quite suprised that my patch to not change lastuse if already set to jiffies changes nothing...
> >>
> >> If you have some time, could you also test this (unrelated) patch ?
> >>
> >> We can avoid dirty all the time a cache line of loopback device.
> >>
> >> diff --git a/drivers/net/loopback.c b/drivers/net/loopback.c
> >> index f2a6e71..0a4186a 100644
> >> --- a/drivers/net/loopback.c
> >> +++ b/drivers/net/loopback.c
> >> @@ -150,7 +150,10 @@ static int loopback_xmit(struct sk_buff *skb, struct net_device *dev)
> >>                 return 0;
> >>         }
> >>  #endif
> >> -       dev->last_rx = jiffies;
> >> +#ifdef CONFIG_SMP
> >> +       if (dev->last_rx != jiffies)
> >> +#endif
> >> +               dev->last_rx = jiffies;
> >>  
> >>         /* it's OK to use per_cpu_ptr() because BHs are off */
> >>         pcpu_lstats = netdev_priv(dev);
> >>
> > Although I didn't test it, I don't think it's ok. The key is __refcnt shares the same
> > cache line with ops/input/output.
> > 
> 
> Note it was unrelated to struct dst, but dirtying of one cache line of 
> 'loopback netdevice'
> 
> I tested it, and tbench result was better with this patch : 890 MB/s instead 
> of 870 MB/s on a bi dual core machine.
I tested your new patch and it doesn't help tbench.

On my 8-core stoakley machine, the regression is only 5%, but it's 30% on 16-core tigerton.
It looks like the scalability is poor.

> 
> 
> I was curious of the potential gain on your 16 cores (4x4) machine.