[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [dccp] Faster Restart: An Issue..



You are neither wrong nor stupid, but I'm not sure that your suggested algorithm is the best possible fix.

Part of my confusion stems from the fact that I can't quite follow the example. Where is the idle period? Exactly which packets are sent when? Also are you measuring X in bytes per second? Because if you are, then the quadrupled rate (=> 888 B/s) is more than 4 160-byte packets per second, even with headers.

I share your concern about Vlad's "first adjustment algorithm", namely ignoring the first feedback packet after an idle period might cause problems if a feedback packet covered multiple idle periods. But we aren't quite there yet.

So please explain the scenario in more detail.

Thanks!
Eddie




Arjuna Sathiaseelan wrote:
Dear Eddie,
I just managed to implement the receiver rate adjustment algorithm
(the second adjustment algorithm) mentioned in the draft. I did some
tests with bursty traffic and found that the using second algorithm is
not being able to maintain the minimum sending rate ( equal to the
initial sending rate) in certain cases. One case is this:

RTT = 600ms, packet size = 160 bytes:

One packet is sent by the application in 1.44 s, which means the
receiver reports X_recv_in = 111, and using the second adjustment
algorithm, we get X_recv = 222. So using the FR algorithm we could
quadruple to 888 which is less than the minimum rate = initial rate of
4 packets per RTT.

Now in this case its wise to use the first adjustment algorithm by
ignoring this feedback packet.

But what happens if the same thing happens again? :). Another 1 packet
in 1.44s? Do we ignore that feedback packet too?

I havFrom dccp-bounces at ietf.org Wed Sep 27 13:21:28 2006
Received: from [127.0.0.1] (helo=stiedprmman1.va.neustar.com)
	by megatron.ietf.org with esmtp (Exim 4.43)
	id 1GSd5w-0001i2-W2; Wed, 27 Sep 2006 13:21:16 -0400
Received: from [10.91.34.44] (helo=ietf-mx.ietf.org)
	by megatron.ietf.org with esmtp (Exim 4.43) id 1GSd5v-0001hx-G8
	for dccp at ietf.org; Wed, 27 Sep 2006 13:21:15 -0400
Received: from smtp-8.smtp.ucla.edu ([169.232.47.138])
	by ietf-mx.ietf.org with esmtp (Exim 4.43) id 1GSd5q-0007up-1d
	for dccp at ietf.org; Wed, 27 Sep 2006 13:21:15 -0400
Received: from mail.ucla.edu (mail.ucla.edu [169.232.48.150])
	by smtp-8.smtp.ucla.edu (8.13.6/8.13.6) with ESMTP id k8RHL3S1008359;
	Wed, 27 Sep 2006 10:21:03 -0700
Received: from [131.179.33.137] (Cs-33-137.CS.UCLA.EDU [131.179.33.137])
	(authenticated bits=0)
	by mail.ucla.edu (8.13.6/8.13.6) with ESMTP id k8RHL2Qm008919
	(version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=NOT);
	Wed, 27 Sep 2006 10:21:02 -0700
Message-ID: <451AA790.3010703 at cs.ucla.edu>
Date: Wed, 27 Sep 2006 09:32:16 -0700
From: Eddie Kohler <kohler at cs.ucla.edu>
User-Agent: Thunderbird 1.5.0.5 (X11/20060808)
MIME-Version: 1.0
To: Arjuna Sathiaseelan <arjuna.sathiaseelan at gmail.com>
Subject: Re: [dccp] Faster Restart: An Issue..
References: <1ef225900609231327u2e7c2970pbed591bf995d8aeb at mail.gmail.com>	
	<4515C0B2.7030406 at cs.ucla.edu>	
	<1ef225900609232255q373b56bnec888638eb3b1879 at mail.gmail.com>
	<1ef225900609240820v68ccd840nf9bb932838061d47 at mail.gmail.com>
In-Reply-To: <1ef225900609240820v68ccd840nf9bb932838061d47 at mail.gmail.com>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit
X-Probable-Spam: no
X-Spam-Report: none
X-Scanned-By: smtp.ucla.edu on 169.232.47.138
X-Spam-Score: 0.0 (/)
X-Scan-Signature: a7d2e37451f7f22841e3b6f40c67db0f
Cc: Gorry Fairhurst <gorry at erg.abdn.ac.uk>,
	'dccp' working group <dccp at ietf.org>
X-BeenThere: dccp at ietf.org
X-Mailman-Version: 2.1.5
Precedence: list
List-Id: Datagram Congestion Control Protocol  <dccp.ietf.org>
List-Unsubscribe: <https://www1.ietf.org/mailman/listinfo/dccp>,
	<mailto:dccp-request at ietf.org?subject=unsubscribe>
List-Post: <mailto:dccp at ietf.org>
List-Help: <mailto:dccp-request at ietf.org?subject=help>
List-Subscribe: <https://www1.ietf.org/mailman/listinfo/dccp>,
	<mailto:dccp-request at ietf.org?subject=subscribe>
Errors-To: dccp-bounces at ietf.org

You are neither wrong nor stupid, but I'm not sure that your suggested algorithm is the best possible fix.

Part of my confusion stems from the fact that I can't quite follow the example. Where is the idle period? Exactly which packets are sent when? Also are you measuring X in bytes per second? Because if you are, then the quadrupled rate (=> 888 B/s) is more than 4 160-byte packets per second, even with headers.

I share your concern about Vlad's "first adjustment algorithm", namely ignoring the first feedback packet after an idle period might cause problems if a feedback packet covered multiple idle periods. But we aren't quite there yet.

So please explain the scenario in more detail.

Thanks!
Eddie




Arjuna Sathiaseelan wrote:
Dear Eddie,
I just managed to implement the receiver rate adjustment algorithm
(the second adjustment algorithm) mentioned in the draft. I did some
tests with bursty traffic and found that the using second algorithm is
not being able to maintain the minimum sending rate ( equal to the
initial sending rate) in certain cases. One case is this:

RTT = 600ms, packet size = 160 bytes:

One packet is sent by the application in 1.44 s, which means the
receiver reports X_recv_in = 111, and using the second adjustment
algorithm, we get X_recv = 222. So using the FR algorithm we could
quadruple to 888 which is less than the minimum rate = initial rate of
4 packets per RTT.

Now in this case its wise to use the first adjustment algorithm by
ignoring this feedback packet.

But what happens if the same thing happens again? :). Another 1 packet
in 1.44s? Do we ignore that feedback packet too?

I have a feeling that we could have something like this:

X_recv_limit := max(2*X_recv,s/R).

or  something like this:

X_recv = max(X_recv_in*(T2 - T1 + I)/(T2 - T1), s/R)

So that we make sure that the receiver reported rate doesnt go below
one packet per RTT.

My few cents and pardon me if I am wrong and stupid :)

-Arjuna

On 9/24/06, Arjuna Sathiaseelan <arjuna.sathiaseelan at gmail.com> wrote:
Dear Eddie,
 Thanks for your reply :).

> I think you have got it wrong. FR is used whenever a feedback packet is
> received -- the nofeedback timer is not relevant. I would personally approve
> of the use of FR you describe. And I think the draft totally encourages it.


That sounds good to me :)


> In periods of extreme congestion it is possible that a flow's fair share is
> less than 8 pkt/RTT. The initial sending rate is NOT a minimum sending rate.
> HOWEVER, flows should NEVER drop below the initial sending rate EXCEPT in
> periods of extreme congestion. Idle periods, for example, should not cause
> the sending rate to drop below 8 p/RTT. Hopefully the current FR draft
> contains sufficient mechanism to enforce this. If it doesn't let us know!


I guess the current draft allows us to maintain the initial sending
rate. One occurrence or a possibility of the flows dropping below the
low initial sending rate is "say" when the application sends one
packet every 2 RTTs, but the Receive Rate Adjustment algorithm should
solve this issue. So I feel the algorithm holds good. :)

Thanks for your clarifications :).

Arjuna


-- Dr.Arjuna Sathiaseelan Electronics Research Group University of Aberdeen Aberdeen AB24 3UE Web: www.erg.abdn.ac.uk/users/arjuna Phone : +44-1224-272780 Fax : +44-1224-272497