From mboxrd@z Thu Jan  1 00:00:00 1970
From: Eric Dumazet <eric.dumazet@gmail.com>
Subject: Re: [Pv-drivers] rps and pvdrivers
Date: Sat, 04 Sep 2010 16:31:04 +0200
Message-ID: <1283610664.3402.34.camel@edumazet-laptop>
References: <AANLkTimrHh1ZasO=GWUOT9YJbu2EqeAx6OkRXPUSSNJ6@mail.gmail.com>
	 <8B1F619C9F5F454E81D90D3C161698D70337A49A5A@EXCH-MBX-3.vmware.com>
	 <AANLkTi=o8GYdLyY8EzfRno39O2kUSh3wSp9raZCUp_Zz@mail.gmail.com>
	 <1283583095.2320.11.camel@edumazet-laptop>
	 <AANLkTimC4z7mCKwyK8=LaacjTtZGEGN_BYsFMydQiwBk@mail.gmail.com>
	 <1283608311.3402.30.camel@edumazet-laptop>
	 <AANLkTik+T8Sv1tEmy4PU0O-dOeymJLEkEd9QOYy7ZJzC@mail.gmail.com>
Mime-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: QUOTED-PRINTABLE
Cc: Bhavesh Davda <bhavesh@vmware.com>,
	"netdev@vger.kernel.org" <netdev@vger.kernel.org>,
	"pv-drivers@vmware.com" <pv-drivers@vmware.com>,
	"therbert@google.com" <therbert@google.com>
To: Chetan Loke <chetanloke@gmail.com>
Return-path: <netdev-owner@vger.kernel.org>
Received: from mail-fx0-f46.google.com ([209.85.161.46]:54431 "EHLO
	mail-fx0-f46.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org
	with ESMTP id S1752647Ab0IDObL (ORCPT
	<rfc822;netdev@vger.kernel.org>); Sat, 4 Sep 2010 10:31:11 -0400
Received: by fxm13 with SMTP id 13so1806566fxm.19
        for <netdev@vger.kernel.org>; Sat, 04 Sep 2010 07:31:09 -0700 (PDT)
In-Reply-To: <AANLkTik+T8Sv1tEmy4PU0O-dOeymJLEkEd9QOYy7ZJzC@mail.gmail.com>
Sender: netdev-owner@vger.kernel.org
List-ID: <netdev.vger.kernel.org>

Le samedi 04 septembre 2010 =C3=A0 10:02 -0400, Chetan Loke a =C3=A9cri=
t :
> On Sat, Sep 4, 2010 at 9:51 AM, Eric Dumazet <eric.dumazet@gmail.com>=
 wrote:
> >
> > If packets have same rxhash, (same src IP, dst IP, src port, dst po=
rt),
> > they are directed on a single CPU, and this might explain you canno=
t
> > handle the load, RPS or not.
>=20
> I tried incrementing the src-macs from the load-generator. Or only th=
e
> above 4-tuple entries are used?
>=20

MAC addresses are not used to compute rxhash

>=20
> >
> > cat /proc/net/softnet_stat
> >
>=20
> cat /proc/net/softnet_stat
> 0655cb87 006dd31b 00096a78 00000000 00000000 00000000 00000000
> 00000000 00000000 00000dbe
> 0042feec 00000000 0008ae33 00000000 00000000 00000000 00000000
> 00000000 00000000 00000000
>=20
> cat /proc/net/softnet_stat
> 065d444e 006e669d 00097565 00000000 00000000 00000000 00000000
> 00000000 00000000 00000dbe
> 0042ff8c 00000000 0008b83f 00000000 00000000 00000000 00000000
> 00000000 00000000 00000000
>=20
> cat /proc/net/softnet_stat
> 066a6fcf 006fefd9 000988c9 00000000 00000000 00000000 00000000
> 00000000 00000000 00000dbe
> 004300bd 00000000 0008c945 00000000 00000000 00000000 00000000
> 00000000 00000000 00000000
>=20
>=20
>=20
> > RPS is good to handle multiple flows, because it can distribute loa=
d on
> > several cpus. But with a single flow, I guess it might be not that
> > useful.
>=20
> For single flow if it doesn't improve performance that's fine. But it
> should atleast work as good as N-2(2.6.33).
>=20

Nope. Because if each packet goes through two cpus instead of one befor=
e
being queued to socket queue, you pay overhead and memory trafic betwee=
n
these cpus.

This is the reason why RPS is not automatically switched on. It might b=
e
slower on some workloads.


=46or UDP trafic, single flow, RPS is not a win, because UDP stack is
quite small.