Internet DRAFT - draft-ohanlon-rmcat-dflow
draft-ohanlon-rmcat-dflow
Network Working Group P. O'Hanlon
Internet-Draft University of Oxford
Intended status: Standards Track K. Carlberg
Expires: October 26, 2013 G11
April 24, 2013
Congestion control algorithm for lower latency and lower loss media
transport
draft-ohanlon-rmcat-dflow-02
Abstract
This memo provides a design for a congestion control algorithm, for
media transport, which aims to provide for lower delay and lower loss
communications.
Status of This Memo
This Internet-Draft is submitted in full conformance with the
provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering
Task Force (IETF). Note that other groups may also distribute
working documents as Internet-Drafts. The list of current Internet-
Drafts is at http://datatracker.ietf.org/drafts/current/.
Internet-Drafts are draft documents valid for a maximum of six months
and may be updated, replaced, or obsoleted by other documents at any
time. It is inappropriate to use Internet-Drafts as reference
material or to cite them other than as "work in progress."
This Internet-Draft will expire on October 26, 2013.
Copyright Notice
Copyright (c) 2013 IETF Trust and the persons identified as the
document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal
Provisions Relating to IETF Documents
(http://trustee.ietf.org/license-info) in effect on the date of
publication of this document. Please review these documents
carefully, as they describe your rights and restrictions with respect
to this document. Code Components extracted from this document must
include Simplified BSD License text as described in Section 4.e of
the Trust Legal Provisions and are provided without warranty as
described in the Simplified BSD License.
O'Hanlon & Carlberg Expires October 26, 2013 [Page 1]
Internet-Draft DFlow media congestion control April 2013
Table of Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 2
2. Conventions, Definitions and Acronyms . . . . . . . . . . . . 3
3. Background . . . . . . . . . . . . . . . . . . . . . . . . . 3
3.1. TFRC . . . . . . . . . . . . . . . . . . . . . . . . . . 3
3.2. Delay-Based schemes . . . . . . . . . . . . . . . . . . . 4
4. Objectives . . . . . . . . . . . . . . . . . . . . . . . . . 5
5. Design Outline . . . . . . . . . . . . . . . . . . . . . . . 6
5.1. Delay Composition . . . . . . . . . . . . . . . . . . . . 6
5.2. Delay Measurement . . . . . . . . . . . . . . . . . . . . 6
5.3. Congestion Detection . . . . . . . . . . . . . . . . . . 7
5.4. Slow Start . . . . . . . . . . . . . . . . . . . . . . . 7
5.5. Loss-mode . . . . . . . . . . . . . . . . . . . . . . . . 7
6. Further Work . . . . . . . . . . . . . . . . . . . . . . . . 8
7. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 8
8. Security Considerations . . . . . . . . . . . . . . . . . . . 8
9. References . . . . . . . . . . . . . . . . . . . . . . . . . 8
9.1. Normative References . . . . . . . . . . . . . . . . . . 8
9.2. Informative References . . . . . . . . . . . . . . . . . 8
Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 9
1. Introduction
This memo outlines DFlow, a congestion control algorithm that aims to
minimise delay and loss by using delay-based techniques. The scheme
is based upon TCP Friendly Rate Control (TFRC) [RFC5348], and adds a
delay-based congestion detection scheme which feeds into a
'congestion event history' mechanism based upon TFRC's loss history.
This then provides for a 'congestion event rate' which drives the TCP
equation.
Congestion control that aims to minimise the delay is important for
real-time streams as high delay can render the communication
unacceptable [ITU.G114.2003]. On today's Internet a number of paths
have an excess of buffering which can lead to persistent high
latencies, which has become known as the Bufferbloat phenomenon.
These problems are particularly apparent with loss-based congestion
control schemes such as TCP, as they operate by filling the queues on
a path till loss occurs, thus maximising the delay. The unfortunate
consequence is that loss-based approaches not only lead to high delay
for their own packets but also introduce delays and losses for all
other flows that traverse those same filled queues.
Thus when competing with TCP, without the widespread deployment of
Active Queue Management that aims to minimise delay, (e.g. Codel
[I-D.nichols-tsvwg-codel]), it is not possible to maintain low delay
O'Hanlon & Carlberg Expires October 26, 2013 [Page 2]
Internet-Draft DFlow media congestion control April 2013
as TCP will do its best to keep the queues full and maximise the
delay.
However there are many paths where the flows are not competing
directly with TCP and where delay may be minimised.
The DFlow scheme can transport media with low delay and loss on paths
where there is no direct competition with TCP in the same queue.
Though we are currently testing some techniques to enable it compete
with loss-based schemes (at the expense of delay) but they will be
included in a later version of the draft. In simulations it has been
seen to be reasonably fair when competing with other DFlow streams.
2. Conventions, Definitions and Acronyms
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
"SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and
"OPTIONAL" in this document are to be interpreted as described in RFC
2119 [RFC2119].
3. Background
Whilst the existing standard for media transport, Real-time Transport
Protocol (RTP) [RFC3550], suggests that congestion control should be
employed, in practice many systems tend to use fixed or variable bit
rate UDP and do very little or no adaptation to their network
environment. Most of the existing work on real-time congestion
control algorithms has been rooted in TCP-friendly approaches but
with smoother adaptation cycles. TCP congestion control is
unsuitable for interactive media for a number of reasons including
the fact that it is loss-based so it maximises the latency on a path,
it changes its transmit rate to quickly for multimedia, and favours
reliability over timeliness. Various TCP-friendly congestion control
algorithms such as TFRC [RFC5348], Sisalem's LDA+ [SisalemLDA.2000],
and Choi's TCP Friendly Window Control (TFWC) [ChoiTFWC.2007] have
been devised for media transport, that attempt to smooth the short-
term variation in sending rate. More recently there have been
development of some delay-based schemes which aim to provide for low
delay.
3.1. TFRC
TFRC is a rate based receiver driven congestion control algorithm
which utilises the Padhye TCP equation to provide a smoothed TCP-
friendly rate. The sender explicitly sets the transmission rate,
using the TCP equation driven by the loss event rate which is
measured and fed back by the receiver, where a loss event consists of
one or more packet losses within a single RTT. It utilises a
O'Hanlon & Carlberg Expires October 26, 2013 [Page 3]
Internet-Draft DFlow media congestion control April 2013
weighted smoothed loss event rate, and EWMA smoothed RTT, as input to
the TCP equation which enables it to achieve a smoother rate
adaptation that provides for a more suitable transport for
multimedia. TFRC was primarily aimed at streaming media delivery
where a smooth rate and TCP-friendliness are more important than low
latency operation.
However there are number of issues with TFRC as regards real-time
media transport:
Loss-based operation: Firstly since it is a loss-based based scheme
the latency is maximised which is a problem for real-time
transport over heavily buffered paths. The other problem with
loss-based protocols is that they rely on a certain level of
packet loss which can be an issue for media traffic since lost
media packet cannot usually be retransmitted in time. This
problem becomes more of a concern at lower transmission rates
since the TCP equation requires a corresponding increase in loss
rates.
Bursty media flows: Many media flows exhibit bursty behaviour due to
a number of factors. Firstly there may be negative bursts (i.e.
gaps) due to silence or low motion which can lead oscillatory
behaviours due to the data-limited and/or idle behaviours.
Secondly there may be positive bursts (i.e. larger than normal)
can also be due to the bursty nature of the media and codec (e.g.
I-frames) which can be lead to drops or increased latency. Whilst
the current version of TFRC [RFC5348] has attempted to address
some of these issues, they are still a concern.
Small RTT environments: When operating in low RTT environments
(<5ms), such as a LAN, systems implementing TFRC can have problems
with scheduling packet transmissions as inter-packet timings can
be lower than application level clock granularity. Whilst the
current version of TFRC [RFC5348] has attempted to address these
issues, they can still be a concern in some low RTT environments.
Variable packet sizes: As originally designed TFRC will only operate
correctly when packet sizes are close to MTU size, and when the
packet sizes are much smaller fairness issues arise. Although
there have been attempts to address this problem for small packets
[RFC4828], it is not clear how to deal with flows that do vary
their packet sizes substantially. However this issue is only
really a marked problem with lower bit rate video flows or
variable packet rate audio.
3.2. Delay-Based schemes
O'Hanlon & Carlberg Expires October 26, 2013 [Page 4]
Internet-Draft DFlow media congestion control April 2013
In the last few years there has been a renewed interest in the use of
delay based congestion control for media, with a slightly different
emphasis to that of the history of TCP based approaches such as
Jain's CARD, Wang and Crowcroft's Tri-S, Brakmo's Vegas, Tan et al's
Compound TCP, and more recently Budzisz's CxTCP [BudziszCxTCP.2011].
Where the primary goal with media based transports is to actually
minimise the latency of the flow, as opposed to just using delay as
an early indication of loss. This is of particular relevance on
paths with large queues, as is the case with a number of today's
Internet paths. In 2007 Ghanbari et al [GhanbariFuzzy.2007] did some
pioneering work on delay-based video congestion control using fuzzy
logic based systems. Recently there has been on going activity in
the IETF as part of the Low Extra Delay Background Transport (LEDBAT)
Working Group which aims to provide a less than best effort delay-
based transport with lower delay. However [RFC6817] specifies a one-
way queuing delay target of 100ms which is quite a high baseline for
interactive media, considering the recommended total one-way delay
limit for a VoIP call should be less than 150ms [ITU.G114.2003].
4. Objectives
The objectives of DFlow are to provide for low delay and low loss
media transport when possible. We also aim to provide (in a future
version of the draft) mechanisms to provide for better burst
management, and loss-mode operation.
Lower Delay: The one-way delay should be kept well within the
acceptable levels of 150ms, and MUST NOT exceed 400ms
[ITU.G114.2003].
Lower Loss: For media transport it is important to minimise loss as
it is usually not possible to retransmit within the delay budget
for many connections. Whilst modern codecs can tolerate some loss
it is beneficial to avoid it. The advantage of low delay
congestion control is that since it aims to operate within the
queuing boundaries it generally avoids loss.
Smoothness: The media rate should aim to be smooth within the
constraints of the media, codec, and the network path. A smooth
rate generally provides for more palatable media consumption.
Fairness: The system should aim to be reasonably fair with itself
and TCP flows. Initially we aim for self fairness, and we will
aim to tackle TCP fairness when we have sufficiently robust loss-
mode operation.
O'Hanlon & Carlberg Expires October 26, 2013 [Page 5]
Internet-Draft DFlow media congestion control April 2013
[Burst Management]: [Due in later rev] We are working on mechanisms
to manage the bursty nature of media allowing it maintain a
smoother quality.
[Loss-based mode]: [Due in later rev] We are working on mechanisms
to allow the system to compete with loss-based congestion control
and maintain throughput, though without additional network support
it is understood that the delay (and loss) would be largely beyond
control.
5. Design Outline
The DFlow scheme aims to primarily utilise delay measurements to
drive the congestion control. It currently utilises some of the core
aspects of TFRC, such as its rate based operation, utilisation of the
TCP equation, and its rate smoothing. It also employs similar
signalling mechanisms. However as the design evolves we expect that
DFlow may diverge further from TFRC.
5.1. Delay Composition
The total end-to-end one-way delay (OWD) a packet incurs may be
considered to consist of four elements; transmission (or
serialisation), propagation, processing, and queuing delays. For our
purposes the first three elements may be considered together as a
largely static component, termed the base delay. The base delay
generally does not change significantly unless the node is mobile or
the underlying link alters due to something like a route change. The
main dynamic element of the delay, which DFlow aims to utilise, is
the queuing delay. Taken together with the base delay, the queuing
delay provides an indication of the actual path latency and also
provides an insight as to the level of congestion on the path.
5.2. Delay Measurement
The notional one-way delay is measured for each packet by comparing
the sender and receiver timestamps. Whilst the clocks on the sender
and receiver are unlikely to be synchronised, it is assumed that
their offset is relatively constant as the clock skew is generally
quite small. Thus the notional OWD may only be used in a relative
context. The notional OWD is measured for each packet over two
sampling periods; Firstly over the longer base_period (typically
10*RTT) from which the minima are stored as the base_delay. And
secondly it is sampled over a shorter period current_period
(typically 50ms), which is also filtered, usually also using a minima
filter, and stored as current_delay. The minima of the OWDs are used
to reduce noise of the measurements, which can be beneficial in the
case of variable link types such as wireless.
O'Hanlon & Carlberg Expires October 26, 2013 [Page 6]
Internet-Draft DFlow media congestion control April 2013
5.3. Congestion Detection
The delay-based detection algorithm, outlined in Figure 1, operates
by comparing the current_delay to the base_delay, which gives an
indication of the queuing delay. If it exceeds a set congestion
detection threshold, cd_thresh, then the packet is considered for the
next stage of detection. The cd_thresh sets the limit for the
queuing delay incurred by the flow, and is typically set at 50ms (we
are also investigating automated thresholds). Once a flow has
exceeded its cd_thresh then it undergoes a second test which is based
upon the gradient of the delay change over two current_period's,
indicating that delay is on the increase, if it is positive then a
'congestion event' is flagged.
If ((base_delay - current_delay) > cd_thresh AND
(current_delay - prev_current_delay) > 0)
DelayCongestionEvent = True
Figure 1: Congestion Detection pseudo-code
This algorithm then provides input to the 'congestion interval
history' mechanism (based on TFRC's 'loss interval history'), which
is combined with normal input from the TFRC packet loss detection
mechanisms, from which a 'congestion event rate' is derived which is
then fed into the TCP equation to determine the send rate.
Note that we currently disable TFRC's oscillation reduction mechanism
from [RFC5348] (Section 4.5) as it adversely affects the delay-based
operation.
We have performed a number of simulations of the above mechanism in
operation and have found it to be reasonably fair to itself,
providing for smooth rates at suitable RTTs.
5.4. Slow Start
The delay based congestion detection is not only used during normal
the congestion avoidance phase of the protocol but it also employed
during slow start allowing for rapid, lower loss, attainment of the
operating rate.
5.5. Loss-mode
We are actively investigating techniques to enable competitive
behaviours with loss-based protocol such as TCP. We aim to develop a
solution that provides for automatic fallback between loss and delay
modes.
O'Hanlon & Carlberg Expires October 26, 2013 [Page 7]
Internet-Draft DFlow media congestion control April 2013
6. Further Work
The design is still under active development and there is more work
to be done. We are seeking feedback on these ideas and future
directions.
7. IANA Considerations
This document makes no requests of IANA.
8. Security Considerations
With a congestion control algorithm an attacker can attempt to
interfere with the protocol to cause rate changes. However
encryption of the protocol will largely protect it against such
threats.
9. References
9.1. Normative References
[RFC2119] Bradner, S., "Key words for use in RFCs to Indicate
Requirement Levels", BCP 14, RFC 2119, March 1997.
[RFC4828] Floyd, S. and E. Kohler, "TCP Friendly Rate Control
(TFRC): The Small-Packet (SP) Variant", RFC 4828, April
2007.
[RFC5348] Floyd, S., Handley, M., Padhye, J., and J. Widmer, "TCP
Friendly Rate Control (TFRC): Protocol Specification", RFC
5348, September 2008.
9.2. Informative References
[BudziszCxTCP.2011]
Budzisz, L., Stanojevic, R., Schlote, A., Shorten, R., and
F. Baker, "On the Fair Coexistence of Loss- and Delay-
Based TCP", July 2011.
[ChoiTFWC.2007]
Choi, S. and M. Handley, "Fairer TCP-friendly congestion
control protocol for multimedia streaming applications",
Dec 2007.
[GhanbariFuzzy.2007]
Jammeh, E., Fleury, M., and M. Ghanbari, "Delay-based
congestion avoidance for video communication with fuzzy
logic control", Nov 2007.
O'Hanlon & Carlberg Expires October 26, 2013 [Page 8]
Internet-Draft DFlow media congestion control April 2013
[I-D.nichols-tsvwg-codel]
Nichols, K. and V. Jacobson, "Controlled Delay Active
Queue Management", draft-nichols-tsvwg-codel-01 (work in
progress), February 2013.
[ITU.G114.2003]
International Telecommunications Union, "One-way
transmission time", ITU-T Recommendation G.707, May 2003.
[RFC3550] Schulzrinne, H., Casner, S., Frederick, R., and V.
Jacobson, "RTP: A Transport Protocol for Real-Time
Applications", STD 64, RFC 3550, July 2003.
[RFC6817] Shalunov, S., Hazel, G., Iyengar, J., and M. Kuehlewind,
"Low Extra Delay Background Transport (LEDBAT)", RFC 6817,
December 2012.
[SisalemLDA.2000]
Sisalem, D. and A. Wolisz, "LDA+: A TCP-Friendly
Adaptation Scheme for Multimedia Communication", May 2000.
Authors' Addresses
Piers O'Hanlon
University of Oxford
Oxford Internet Institute
1 St Giles
Oxford OX1 3JS
United Kingdom
Email: piers.ohanlon@oii.ox.ac.uk
Ken Carlberg
G11
1600 Clarendon Blvd
Arlington VA
USA
Email: carlberg@g11.org.uk
O'Hanlon & Carlberg Expires October 26, 2013 [Page 9]