David Miller a écrit :
> From: Sven-Thorsten Dietrich <sven@xxxxxxxxxxxxxxxxxxxxx>
> Date: Wed, 18 Mar 2009 18:43:27 -0700
>
>> Do we have to rule-out per-CPU queues, that aggregate into a master
>> queue in a batch-wise manner?
>
> That would violate the properties and characteristics expected by
> the packet scheduler, wrt. to fair based fairness, rate limiting,
> etc.
>
> The only legal situation where we can parallelize to single device
> is where only the most trivial packet scheduler is attached to
> the device and the device is multiqueue, and that is exactly what
> we do right now.
I agree with you David.
Still, there is room for improvements, since :
1) default qdisc is pfifo_fast. This beast uses three sk_buff_head (96 bytes)
where it could use 3 smaller list_head (3 * 16 = 48 bytes on x86_64)
(assuming sizeof(spinlock_t) is only 4 bytes, but it's more than that
on various situations (LOCKDEP, ...)
2) struct Qdisc layout could be better, letting read mostly fields
at beginning of structure. (ie move 'dev_queue', 'next_sched', reshape_fail,
u32_node, __parent, ...)
'struct gnet_stats_basic' has a 32 bits hole
'gnet_stats_queue' could be split, at least in Qdisc, so that three
seldom use fields (drops, requeues, overlimits) go in a different cache line.
gnet_stats_rate_est might be also moved in a 'not very used' cache line, if
I am not mistaken ?
3) In stress situation a CPU A queues a skb to a sk_buff_head, but a CPU B
dequeues it to feed device, involving an expensive cache line miss
on the skb.{next|prev} (to set them to NULL)
We could:
Use a special dequeue op that doesnt touch skb.{next|prev}
Eventually set next/prev to NULL after q.lock is released
--
To unsubscribe from this list: send the line "unsubscribe linux-rt-users" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
