David S. Miller a écrit :
> From: Andrew Morton <akpm@osdl.org>
> Date: Fri, 23 Sep 2005 00:10:13 -0700
> 
> 
>>Ravikiran G Thirumalai <kiran@scalex86.org> wrote:
>>
>>>Hi Andrew,
>>>This patchset contains alloc_percpu + bigrefs + bigrefs for netdevice
>>>refcount.  This patchset improves tbench lo performance by 6% on a 8 way IBM
>>>x445.
>>
>>I think we'd need more comprehensive benchmarks than this before adding
>>large amounts of complex core code.
> 
> 
> I agree.  tbench over loopback is about as far from real life
> as it gets.
> 
> I'm still against expanding these networking datastructures with
> bigrefs just for this stuff.  Some people have per-cpu and per-node on
> the brain, and it's starting to bloat things up a little bit too much.

Hi David

I agree with you about bigref might be a bit of over-engineering for netdevice 
refcount.

I sent one month ago a patch to reorder 'struct net_device' to at least make 
it SMP friendly, with no per-cpu or per-node additional memory but got no comment.

It's important to place mostly read parts together, so that a cache lines can 
be shared by CPUS. It's important to separate hot fields in three different 
cache lines to that a cpu can work on the transmit while another is receiving 
frames without paying cache line ping pongs.

Please (re)consider this patch since it really reduce CPU load and/or memory 
bus trafic between CPUS.

I ask this question :
  - Is the following comment found in netdevice.h still true ?

/*
  * This is the first field of the "visible" part of this structure
  * (i.e. as seen by users in the "Space.c" file).  It is the name
  * the interface.
  */

If it is still true, then I suspect name_hlist cannot be moved like I did.


Thank you


- [NET] : Reorder some hot fields of struct net_device and place them on 
separate cache lines in SMP to lower memory bouncing between multiple CPU 
accessing the device.

     - One part is mostly used on receive path (including eth_type_trans())
      (poll_list, poll, quota, weight, last_rx, dev_addr, broadcast)

     - One part is mostly used on queue transmit path (qdisc)
      (queue_lock, qdisc, qdisc_sleeping, qdisc_list, tx_queue_len)

     - One part is mostly used on xmit path (device)
      (xmit_lock, xmit_lock_owner, priv, hard_start_xmit, trans_start)

'features' is placed outside of these hot points, in a location that may be 
shared by all cpus (because mostly read)

Minor but maybe not possible because of ABI :
	name_hlist is moved close to name[IFNAMSIZ] to speedup __dev_get_by_name()

Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>