Re: eth_type_trans(): Re: [Bug #11308] tbench regression on eachkernel release from 2.6.22 -> 2.6.28

[Eric Dumazet]

Ingo Molnar a écrit :
> * David Miller <davem@xxxxxxxxxxxxx> wrote:
>
> > From: Ingo Molnar <mingo@xxxxxxx>
> > Date: Mon, 17 Nov 2008 22:26:57 +0100
> >
> > > eth->h_proto access.
> > Yes, this is the first time a packet is touched on receive.
> >
> > > Given that this workload does localhost networking, my guess would be 
> > > that eth->h_proto is bouncing around between 16 CPUs? At minimum this 
> > > read-mostly field should be separated from the bouncing bits.
> > It's the packet contents, there is no way to "seperate it".
> >
> > And it should be unlikely bouncing on your system under tbench, the 
> > senders and receivers should hang out on the same cpu unless the 
> > something completely stupid is happening.
> >
> > That's why I like running tbench with a num_threads command line 
> > argument equal to the number of cpus, every cpu gets the two thread 
> > talking to eachother over the TCP socket.
>
> yeah - and i posted the numbers for that too - it's the same 
> throughput, within ~1% of noise.

Thinking once again about loopback driver, I recall a previous attempt
to call netif_receive_skb() instead of netif_rx() and pay the price
of cache line ping-pongs between cpus.

http://kerneltrap.org/mailarchive/linux-netdev/2008/2/21/939644

Maybe we could do that, with a temporary percpu stack, like we do in softirq
when CONFIG_4KSTACKS=y

(arch/x86/kernel/irq_32.c  : call_on_stack(func, stack)

And do this only if the current cpu doesnt already use its softirq_stack
(think about loopback re-entering loopback xmit because of TCP ACK for example)

Oh well... black magic, you are going to kill me :)


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/