From mboxrd@z Thu Jan  1 00:00:00 1970
From: Dmitry Golubev <dmitry@mikrotik.com>
Date: Wed, 16 Jun 2004 15:31:44 +0000
Subject: Re: [LARTC] HTB theory?
Message-Id: <200406161831.44991.dmitry@mikrotik.com>
List-Id: <lartc.vger.kernel.org>
References: <200406121446.08416.dmitry@mikrotik.com>
In-Reply-To: <200406121446.08416.dmitry@mikrotik.com>
MIME-Version: 1.0
Content-Type: text/plain; charset="windows-1252"
Content-Transfer-Encoding: quoted-printable
To: lartc@vger.kernel.org

Thanks very much, Devik and Andy, I had seminar today and I think it has so=
me=20
success, and (for now?) I do not feel like having unanswered questions.

Dmitry

On Sunday 13 June 2004 21:41, Martin Devera <devik@cdi.cz> wrote:
> > 1. In order of priority, we satisfy all leaf classes' rates (while the
> > class is =93green=94)
> > 2. When the leaf classes' rate is reached (all the leaf classes are
> > =93yellow=94), borrow the unused speed from parent classes if they have
> > something to give (if they are not =93red=94). In this case, each leaf =
class
> > is equal to others until their ceil is met or until they will have no
> > parent to borrow from... Or maybe their level matters?
>
> In both cases DRR is used to cycle classes to dequeue - this gives us
> fairness (weighted by "quantums").
>
> > I also think that:
> > 1. Until all leaf classes' rate is met we even do not check anything el=
se
> > (including ceil of parent classes)
>
> correct
>
> > 2. the root queue is unlimited feed for the classes which are attached =
to
> > it directly
>
> yes
>
> > One think I do not understand neither for SFQ nor for HTB (please expla=
in
> > for both) - how can we maintain fairness in case of differently-sizes
> > packets. As I understand, one packet is atomic unit, and interface is
> > requesting not more and not less than one packet. Now given a quantum f=
or
> > each leaf class in HTB or the same for SQF (are they different in this
> > process?) of 1500 bytes - do they wait for more packets from one subcla=
ss
> > (asking it once again in the next turn, of course, if they have nothing
> > to send, we switch to the next...)? And if one subclass is providing 15=
00
> > packets in each turn, and another subclass is providing 800-byte packets
> > - will they be equal on the amount of traffic they are allowed to send =
in
> > case the real channel is smaller than the sum of their effective ceil
> > values?
>
> both SFQ and HTB (and CBQ) uses DRR which uses BYTES as count not packets.
> Look into literature for "leaky bucket" description. To answer your last
> mail, cannonical HTB theory (v.3 =3D latest one) is:
> http://luxik.cdi.cz/~devik/qos/htb/manual/theory.htm. It is up to date -
> mainly because algorithm is the same since 2002.
>
> > (I am preparing for a seminar in which I will tell people how it works,
> > so I need to know this perfectly :))) )
>
> Hmm I have talk at SUCON (http://www.suug.ch/sucon/04/speakers.html#7) so
> that
> I'll need to refresh it too :)
>
> devik
>
> _______________________________________________
> LARTC mailing list / LARTC@mailman.ds9a.nl
> http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/
_______________________________________________
LARTC mailing list / LARTC@mailman.ds9a.nl
http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/