Message-ID: <1323258799.2312.18.camel@edumazet-HP-Compaq-6005-Pro-SFF-PC> (sfid-20111207_125333_452276_CF2D2B73)
Subject: Re: [Bloat] Time in Queue, bufferbloat, and... our accidentally
 interplanetary network
From: Eric Dumazet <eric.dumazet@gmail.com>
To: Hagen Paul Pfeifer <hagen@jauu.net>
Cc: Dave Taht <dave.taht@gmail.com>,
	linux-wireless <linux-wireless@vger.kernel.org>,
	netdev@vger.kernel.org,
	bloat-devel <bloat-devel@lists.bufferbloat.net>,
	bloat <bloat@lists.bufferbloat.net>
Date: Wed, 07 Dec 2011 12:53:19 +0100
In-Reply-To: <9dfc3c5eb811f2774b378fce0158b3e7@localhost>
References: <CAA93jw7kQYONQiAozatRmdRRJeK0GB9eNDWtRZ3AL+OVKn0OpA@mail.gmail.com>
	 <1323082774.2670.40.camel@edumazet-HP-Compaq-6005-Pro-SFF-PC>
	 <9dfc3c5eb811f2774b378fce0158b3e7@localhost>
Content-Type: text/plain; charset="UTF-8"
Mime-Version: 1.0
Sender: linux-wireless-owner@vger.kernel.org

Le mercredi 07 décembre 2011 à 11:15 +0100, Hagen Paul Pfeifer a écrit :
> On Mon, 05 Dec 2011 11:59:34 +0100, Eric Dumazet wrote:
> 
> 
> > Adding a time limit is possible, all we need is a proper design and
> > implementation :)
> > 
> > Here is my suggestion :
> > 
> > Design a new tfifo/tred qdisc, with following properties :
> > 
> > Adaptative RED, (ECN enabled + head drop), but instead of using
> > bytes/packet qlen average, use time_in_queue average.
> 
> Question one: is anything wrong with sfb or choke as the basis, instead of
> RED?
> 

RED is the module to bring EWMA stuff, it seems natural to start with
it. Please note that choke has a RED module too.

Then later, we can add time limit stuff to other Qdisc if needed, its a
plug anyway. But is there any meaning to compute a global EWMA after
SFB/SFQ packet classification ?

> Question two: I submitted pfast_head_drop to drop more outdated data
> instead of new data. Back in time I thought TCP _may_ experience benefits
> because more up-to-date SACK data packets are saved. Are there any other
> TCP advantages with head drop policy?
> 

Note that head drop is a consequence of time limit idea on top of FIFO,
since only at dequeue time, we compute the delta between current time
and enqueue time, and we drop/mark the (head) packet if time exceeds our
current limit.

In general, being able to drop/mark firsts packets in queue instead of
last ones can let TCP sender be notified of congestion much earlier than
a tail drop. (We gain the time to transmit whole packets in queue before
receiver can report in its ACK the congestion back to sender)