[RFC,net-next] tcp: reduce cpu usage under tcp memory pressure when SO_SNDBUF is set

On 08/10/2015 10:47 AM, Eric Dumazet wrote:
> On Fri, 2015-08-07 at 18:31 +0000, Jason Baron wrote:
>> From: Jason Baron <jbaron@akamai.com>
>>
>> When SO_SNDBUF is set and we are under tcp memory pressure, the effective write
>> buffer space can be much lower than what was set using SO_SNDBUF. For example,
>> we may have set the buffer to 100kb, but we may only be able to write 10kb. In
>> this scenario poll()/select()/epoll(), are going to continuously return POLLOUT,
>> followed by -EAGAIN from write() in a very tight loop.
>>
>> Introduce sk->sk_effective_sndbuf, such that we can track the 'effective' size
>> of the sndbuf, when we have a short write due to memory pressure. By using the
>> sk->sk_effective_sndbuf instead of the sk->sk_sndbuf when we are under memory
>> pressure, we can delay the POLLOUT until 1/3 of the buffer clears as we normally
>> do. There is no issue here when SO_SNDBUF is not set, since the tcp layer will
>> auto tune the sk->sndbuf.
>>
>> In my testing, this brought a single threaad's cpu usage down from 100% to 1%
>> while maintaining the same level of throughput when under memory pressure.
>>
> 
> I am not sure we need to grow socket for something that looks like a
> flag ?
>


So I added a new field because I needed to store the new 'effective'
sndbuf somewhere and then restore the original value that was set via
SO_SNDBUF. So its really b/c of SO_SNDBUF. We could perhaps use the fact
that we are in memory pressure to signal wakeups differently, but I'm
not sure exactly how.


> Also you add a race in sk_stream_wspace() as sk_effective_sndbuf value
> can change under us.
> 
> +       if (sk->sk_effective_sndbuf)
> +               return sk->sk_effective_sndbuf - sk->sk_wmem_queued;
> +
> 
> 
> 
> 

thanks. better?

 }


Thanks,

-Jason
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

On Mon, 2015-08-10 at 13:29 -0400, Jason Baron wrote:
> > +
> > 
> > 
> > 
> > 
> 
> thanks. better?
> 
> --- a/include/net/sock.h
> +++ b/include/net/sock.h
> @@ -798,8 +798,10 @@ static inline int sk_stream_min_wspace(const struct
> sock *sk)
> 
>  static inline int sk_stream_wspace(const struct sock *sk)
>  {
> -       if (sk->sk_effective_sndbuf)
> -               return sk->sk_effective_sndbuf - sk->sk_wmem_queued;
> +       int effective_sndbuf = sk->sk_effective_sndbuf;
> +
> +       if (effective_sndbuf)
> +               return effective_sndbuf - sk->sk_wmem_queued;
> 
>         return sk->sk_sndbuf - sk->sk_wmem_queued;
>  }
> 
> 

You need to use instead :

int effective_sndbuf = READ_ONCE(sk->sk_effective_sndbuf);

--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html