Skip site navigation (1)Skip section navigation (2)
Date:      Fri, 7 Nov 2014 03:20:39 +0300
From:      "Alexander V. Chernikov" <melifaro@ipfw.ru>
To:        Luigi Rizzo <rizzo@iet.unipi.it>
Cc:        svn-src-head@freebsd.org, Adrian Chadd <adrian@freebsd.org>, src-committers@freebsd.org, "Alexander V. Chernikov" <melifaro@freebsd.org>, svn-src-all@freebsd.org
Subject:   Re: svn commit: r273112 - head/sys/dev/ixgbe
Message-ID:  <67A66ECA-2550-4ABE-8405-8EC332FB84DA@ipfw.ru>
In-Reply-To: <20141107001507.GD17502@onelab2.iet.unipi.it>
References:  <201410150122.s9F1Mudu083306@svn.freebsd.org> <545B55FA.6070408@FreeBSD.org> <20141107001507.GD17502@onelab2.iet.unipi.it>

next in thread | previous in thread | raw e-mail | index | archive | help

> On 07 Nov 2014, at 03:15, Luigi Rizzo <rizzo@iet.unipi.it> wrote:
>=20
> On Thu, Nov 06, 2014 at 03:05:30PM +0400, Alexander V. Chernikov =
wrote:
>> On 15.10.2014 05:22, Adrian Chadd wrote:
>>> Author: adrian
>>> Date: Wed Oct 15 01:22:56 2014
>>> New Revision: 273112
>>> URL: https://svnweb.freebsd.org/changeset/base/273112
>>>=20
>>> Log:
>>>   Set the DROP_EN bit before the RX queue is brought up and active.
>>>=20
>>>   He noticed issues setting this bit in SRRCTL after the queue was =
up,
>>>   so doing it from the sysctl handler isn't enough and may not =
actually
>>>   work correctly.
>> While investigating/measuring ixgbe performance w/ and without flow =
control
>> I noticed that disabling fc entirely (e.g. disabling 512K skid buffer=20=

>> and turning drop_en on)
>> not only makes things a bit worse, but also makes NIC stop accounting=20=

>> tail-drops as errors
>> at all.
>> It is a bit unhandy with HW counters, since you, for example, see 2M=20=

>> packets being received
>> in netstat, but in reality system receives only 1M (And there is no =
easy=20
>> way to determine exact
>> count).
>=20
> i'd like to point out that the default does not change (flow control
> is still enabled by default) and the previous handling of DROP_EN
The default did change: now drop_en is turned on by default due to a =
small mistake (see my comment to commit).
> was ineffective.
>=20
> One may argue that FC and DROP_EN should be set independently,
> and i am all for it. In fact that would be the right thing to do
> in terms of POLA.
>=20
> This said, i do not understand what kind of performance degradation
> you experienced, can you be more specific ?
Well, I haven=E2=80=99t tested it in great details, but for my =
traditional shooting (IPv4 200 dsts to single rte) I observed some =
notable change,
like 9.2->8.6 MPPS or similar. I=E2=80=99m not insisting that something =
should be done about it, since there is no silver bullet here.
My point was that we=E2=80=99ve lost an ability to count ingress errors =
in stock HEAD.

>=20
> cheers
> luigi




Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?67A66ECA-2550-4ABE-8405-8EC332FB84DA>