Network Working Group M. Luby
Request for Comments: 3738 Digital Fountain
Category: Experimental V. Goyal
M.I.T.
April 2004
Wave and Equation Based Rate Control (WEBRC) Building Block
Status of this Memo
This memo defines an Experimental Protocol for the Internet
community. It does not specify an Internet standard of any kind.
Discussion and suggestions for improvement are requested.
Distribution of this memo is unlimited.
Copyright Notice
Copyright (C) The Internet Society (2004). All Rights Reserved.
Abstract
This document specifies Wave and Equation Based Rate Control (WEBRC),
which provides rate and congestion control for data delivery. WEBRC
is specifically designed to support protocols using IP multicast. It
provides multiple-rate, congestion-controlled delivery to receivers,
i.e., different receivers joined to the same session may be receiving
packets at different rates depending on the bandwidths of their
individual connections to the sender and on competing traffic along
these connections. WEBRC requires no feedback from receivers to the
sender, i.e., it is a completely receiver-driven congestion control
protocol. Thus, it is designed to scale to potentially massive
numbers of receivers attached to a session from a single sender.
Furthermore, because each individual receiver adjusts to the
available bandwidth between the sender and that receiver, there is
the potential to deliver data to each individual receiver at the
fastest possible rate for that receiver, even in a highly
heterogeneous network architecture, using a single sender.
Luby & Goyal Experimental [Page 1]RFC 3738 WEBRC Building Block April 2004
Table of Contents
1. Introduction. . . . . . . . . . . . . . . . . . . . . . . . . 3
2. Rationale . . . . . . . . . . . . . . . . . . . . . . . . . . 5
3. Functionality . . . . . . . . . . . . . . . . . . . . . . . . 6
3.1. Sender Operation . . . . . . . . . . . . . . . . . . . . 9
3.1.1. Sender inputs and initialization. . . . . . . . . 9
3.1.2. Sending packets to the session. . . . . . . . . . 10
3.2. Receiver Operation . . . . . . . . . . . . . . . . . . . 12
3.2.1. Receiver inputs and initialization. . . . . . . . 12
3.2.2. Receiver measurements and calculations. . . . . . 13
3.2.2.1. Average loss probability . . . . . . . . 13
3.2.2.2. Average round-trip time. . . . . . . . . 16
3.2.2.3. Rate Equation. . . . . . . . . . . . . . 16
3.2.2.4. Epochs . . . . . . . . . . . . . . . . . 17
3.2.2.5. Average reception rate . . . . . . . . . 17
3.2.2.6. Slow start . . . . . . . . . . . . . . . 19
3.2.2.7. Target rate. . . . . . . . . . . . . . . 20
3.2.3. Receiver events . . . . . . . . . . . . . . . . . 20
3.2.3.1. Packet reception . . . . . . . . . . . . 20
3.2.3.2. First packet after join. . . . . . . . . 20
3.2.3.3. Time slot change . . . . . . . . . . . . 20
3.2.3.4. Loss event . . . . . . . . . . . . . . . 21
3.2.3.5. Epoch change . . . . . . . . . . . . . . 21
3.2.3.6. Join the next higher layer . . . . . . . 21
3.2.3.7. Join timeout . . . . . . . . . . . . . . 23
3.2.3.8. Exceptional timeouts . . . . . . . . . . 23
4. Applicability Statement . . . . . . . . . . . . . . . . . . . 23
4.1. Environmental Requirements and Considerations. . . . . . 23
5. Packet Header Fields. . . . . . . . . . . . . . . . . . . . . 25
5.1. Short Format Congestion Control Information. . . . . . . 26
5.2. Long Format Congestion Control Information . . . . . . . 27
6. Requirements From Other Building Blocks . . . . . . . . . . . 28
7. Security Considerations . . . . . . . . . . . . . . . . . . . 28
8. References. . . . . . . . . . . . . . . . . . . . . . . . . . 29
8.1. Normative References . . . . . . . . . . . . . . . . . . 29
8.2. Informative References . . . . . . . . . . . . . . . . . 30
9. Authors' Addresses. . . . . . . . . . . . . . . . . . . . . . 31
10. Full Copyright Statement. . . . . . . . . . . . . . . . . . . 32
Luby & Goyal Experimental [Page 2]RFC 3738 WEBRC Building Block April 2004
1. Introduction
This document specifies Wave and Equation Based Rate Control (WEBRC).
WEBRC is a congestion control building block that is designed to be
massively scalable when used with the IP multicast network service.
WEBRC is also suitable as the basis for unicast congestion control,
but this is outside the scope of this document. WEBRC is designed to
compete fairly with TCP and similar congestion-controlled sessions.
WEBRC can be used as a congestion control protocol for any type of
data delivery, including reliable content delivery and streaming
delivery.
WEBRC is a receiver-driven congestion control protocol in the spirit
of [5] and [18]. This means that all measurements and decisions to
raise or lower the reception rate are made by each individual
receiver, and these decisions are acted upon by sending join and
leave messages for channels to the network. A receiver using WEBRC
adjusts its reception rate without regard for other concerns such as
reliability. This is different from TCP, where the congestion
control protocol and the reliability protocol are intricately
interwoven.
WEBRC takes the same basic equation-based approach as TFRC [9]. In
particular, each WEBRC receiver measures parameters that are plugged
into a TCP-like equation to compute the receiver target reception
rate and adjusts its reception rate up and down to closely
approximate the target reception rate. The sender sends packets to
multiple channels; one channel is called the base channel and the
remaining channels are called wave channels. Each wave channel
follows the same pattern of packet rate transmission spread out over
equally-spaced intervals of time. The pattern of each wave is that
it starts at a high rate and the rate decreases gradually and
continually over a long period of time. (Picture an infinite
sequence of waves.) The receiver increases its reception rate by
joining the next wave channel earlier in the descent of the wave than
it joined the previous wave channel, and the receiver decreases its
reception rate by joining the next wave channel later in the descent
of the wave than it joined the previous wave channel.
The wave channels are ordered at each point in time from a lowest
layer to a highest layer. At each point in time, the lowest layer is
the wave channel that among all active wave channels is nearest to
the end of its active period; the highest layer is the wave channel
that is furthest from the end of its active period. Because waves
are dynamically becoming active and quiescent over time, the
designation of which wave channel is at which layer changes
dynamically over time. In addition to being joined to the base
channel, at each point in time a receiver is joined to a consecutive
Luby & Goyal Experimental [Page 3]
RFC 3738 WEBRC Building Block April 2004
set of layers starting at the lowest layer and proceeding towards the
highest.
WEBRC introduces a natural notion of a multicast round-trip time
(MRTT). An MRTT is measured individually by each receiver and
averaged as a substitute for conventional unicast round-trip time
(RTT). Because the throughput of a TCP session depends strongly on
RTT, having some measure of RTT is essential in making the WEBRC
equation-based rate control protocol "TCP-friendly". The use of the
MRTT also helps to coordinate and equalize the reception rates of
proximate receivers joined to a session behind a bottleneck link.
This implies that packets for the session that flow through the
bottleneck link are on average sent to almost all downstream
receivers, and thus the efficiencies of multicast are realized.
Furthermore, WEBRC is designed to be massively scalable in the sense
that the sender is insensitive to the number of receivers joined to a
multicast session.
WEBRC is designed for applications that use a fixed packet size and
vary their packet reception rates in response to congestion. WEBRC
is designed to be reasonably fair when competing for bandwidth with
TCP flows, where a flow is "reasonably fair" if its reception rate is
generally within a factor of two of the reception rate of a TCP flow
under the same conditions. However WEBRC has a much lower variation
of throughput over time compared to TCP, which makes it more suitable
for applications such as telephony or streaming media where a
relatively smooth rate is of importance. The penalty of having
smoother throughput than TCP while competing fairly for bandwidth is
that WEBRC responds more slowly than TCP to changes in available
bandwidth.
The receiver measures and performs the calculation of congestion
control parameters (e.g., the average loss probability, the average
MRTT) and makes decisions on how to increase or decrease its rate
based on these parameters. The receiver-based approach is well
suited to an application where the sender is handling many concurrent
connections and therefore WEBRC is suitable as a building block for
multicast congestion control.
The paper [16] and technical report [15] provide much of the
rationale and intuition for the WEBRC design and describe some
preliminary simulations.
This document describes a building block as defined in RFC 3048 [4].
This document describes a congestion control building block that
conforms to RFC 2357 [3]. This document is a product of the IETF RMT
WG and follows the general guidelines provided in RFC 3269 [2]. The
key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
Luby & Goyal Experimental [Page 4]
RFC 3738 WEBRC Building Block April 2004
"SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this
document are to be interpreted as described in BCP 14, RFC 2119 [1].
Statement of Intent
This memo contains part of the definitions necessary to fully specify
a Reliable Multicast Transport protocol in accordance with RFC 2357.
As per RFC 2357, the use of any reliable multicast protocol in the
Internet requires an adequate congestion control scheme. This
document specifies an experimental congestion control scheme. While
waiting for initial deployment and experience to show this scheme to
be effective and scalable, the IETF publishes this scheme in the
"Experimental" category.
It is the intent of the Reliable Multicast Transport (RMT) Working
Group to re-submit the specification as an IETF Proposed Standard as
soon as the scheme is deemed adequate.
2. Rationale
WEBRC provides congestion control for massively scalable protocols
using the IP multicast network service. The congestion control that
WEBRC provides is common to a variety of applications, including
reliable content delivery and streaming applications.
WEBRC is designed to provide congestion control for all packets that
are sent to a session. A session comprises multiple channels
originating at a single sender that are used for some period of time
to carry packets pertaining to the transmission of one or more
objects that can be of interest to receivers. The logic behind
defining a session as originating from a single sender is that this
is the right granularity to regulate packet traffic via congestion
control. The rationale for providing congestion control that uses
multiple channels within the same session is that this allows the
data on the channels to be layered, which in turn allows each
receiver to control its reception rate by joining and leaving
channels during its participation in the session. There are
advantages to layered data for streaming, where the most important
data can be sent to the lower layers and incrementally valuable data
to the higher layers. For reliable content delivery, as described in
[13], an application can send in packets encoded data generated from
an object in such a way that the arrival of enough packets by a
receiver is sufficient to reliably reconstruct the original object.
A primary advantage of WEBRC is that each receiver controls it
reception rate independent of other receivers. Thus, for example, a
receiver with a slow connection to the sender does not slow down the
receivers with faster connections.
Luby & Goyal Experimental [Page 5]
RFC 3738 WEBRC Building Block April 2004
There are coding techniques that provide massively scalable
reliability and asynchronous delivery which are compatible with
WEBRC, e.g., as described in [11]. When combined the result is a
massively scalable, reliable, asynchronous content delivery protocol
that is network friendly. WEBRC also provides congestion control
that is suitable for streaming applications.
WEBRC avoids using techniques that are not massively scalable. For
example, WEBRC does not provide any mechanisms for sending
information from receivers to senders, although this does not rule
out protocols that both use WEBRC and that send information from
receivers to senders.
WEBRC provides congestion control that can be tuned for different
applications that may have differing application requirements. For
example, a content delivery protocol may aggressively strive to use
all available bandwidth between receivers and the sender, and thus to
maintain fairness it must drastically reduce its rate when there is
competing traffic. On the other hand, a streaming delivery protocol
may strive to maintain a constant rate instead of trying to use all
available bandwidth, and thus it may not reduce its rate as fast when
there is competing traffic.
WEBRC does not provide any support beyond congestion control, and
thus WEBRC is to be combined with other building blocks to provide a
complete protocol instantiation. For example, WEBRC does not provide
any means that can be used to identify which session each received
packet belongs to. As another example, WEBRC does not provide
support for identifying which object each packet is carrying
information about.
3. Functionality
A WEBRC session comprises a logically related set of channels
originating from a single sender that are used for some period of
time to carry data packets with a header carrying WEBRC Congestion
Control Information. When packets are received, they are first
checked to see that they belong to the appropriate session before
WEBRC is applied. A session label defined by a protocol
instantiation may be carried in each packet to identify to which
session the packet belongs. For example, if LCT [12] is being used
with the session, then the sender IP address together with the
Transport Session Identifier supported by LCT would be used to
determine which session a received packet belongs to. The particular
details of how this filtering is performed is outside the scope of
this document. In the remainder of this document, references to
channels are always within the scope of a single session.
Luby & Goyal Experimental [Page 6]
RFC 3738 WEBRC Building Block April 2004
A channel can be uniquely identified at the network layer by a
(sender IP address, multicast group address) pair, and this is the
address to which the receiver sends messages to join and leave the
channel. The channels used by a WEBRC session are mapped uniquely to
consecutive channel numbers. In each packet sent to a channel, the
channel number that corresponds to the channel is carried in the
WEBRC Congestion Control Information. A WEBRC receiver uses the
channel number to determine which channel within a session a packet
is received from.
At the sender, time is partitioned into time slots, each of duration
TSD seconds. There is a fixed number T of time slot indices
associated with a session. As time progresses, the current time slot
index increases by one modulo T each TSD seconds. The current time
slot index CTSI is carried in the WEBRC Congestion Control
Information. This allows receivers to perform very coarse-grained
synchronization within a session.
WEBRC congestion control is achieved by having the sender send
packets associated with a given session to several different
channels. Individual receivers dynamically join and leave these
channels according to the network congestion they experience. These
congestion control adjustments are performed at each receiver
independently of all other receivers, without any impact on the
sender. A packet sequence number is carried in the WEBRC Congestion
Control Information. The packet sequence numbers are consecutively
numbered per channel and are used by receivers to measure packet
loss.
The channels associated with a session consist of one base channel
and T wave channels. The packet rate for each channel varies over
time. For the base channel, packets are sent to the channel at a low
rate BCR_P at the beginning of a time slot and this rate gradually
decreases to P*BCR_P at the end of the time slot, where P < 1 is a
constant defined later. This pattern for the base channel repeats
over each time slot. For each wave channel i, packets are sent to
channel i at a rate that first increases very quickly to a high rate
and then decreases over time by a fixed fraction P per time slot
until a rate of BCR_P is reached at the end of time slot i. Then,
for a period of time called the quiescent period, no packets are sent
to wave channel i, and thereafter the whole cycle repeats itself,
where the duration of the cycle is T*TSD seconds. Thus, the wave
channels are going through the same cyclic pattern of packet rate
transmission spaced out evenly by TSD seconds.
Before joining a session, the receivers MUST obtain enough of the
session description to start the session. This MUST include the
relevant session parameters needed by a receiver to participate in
Luby & Goyal Experimental [Page 7]
RFC 3738 WEBRC Building Block April 2004
the session and perform WEBRC congestion control. The session
description is determined by the sender and is typically communicated
to the receivers out of band. How receivers obtain the session
description is outside the scope of this document.
When a receiver initiates a session, it first joins the base channel.
The packets in the base channel help the receiver orient itself in
terms of what the current time slot index is, which in turn allows
the receiver to know the relative rates on the wave channels. The
receiver remains joined to the base channel for the duration of its
participation in the session.
At each point in time the active (non-quiescent) wave channels are
ordered into layers, where the lowest layer is the active wave
channel whose wave is nearest to completion and the highest layer is
the active wave channel whose wave is furthest from completion.
(This is almost the same as saying that the lowest layer has the
lowest rate and the highest layer has the highest rate. The possible
deviation from this is due to the optional non-exponential beginnings
of the waves as described in [8].) Each time a wave channel becomes
active, it is the highest layer. At the end of each time slot the
lowest-layer wave channel becomes quiescent, and thus all active wave
channels move down a layer at this point in time. At each point in
time a receiver is joined to the base channel and a consecutive set
of layers starting with the lowest. Each time a receiver joins a
wave channel it joins the lowest layer not yet joined. A receiver
always leaves the lowest layer when it becomes quiescent.
After joining a session the receiver adjusts its rate upwards by
joining wave channels in sequence, starting with the lowest layer and
moving towards the highest. The rates on the active wave channels
are decreasing with time, so the receiver adjusts its rate downwards
simply by refraining from joining additional wave channels. Since
the layer ordering among the channels changes dynamically over time
depending on the current time slot index, it is important that the
receiver continually monitor the current time slot index contained in
received packets. The reception rate at the receiver is determined
by how early each wave channel is joined by the receiver: the earlier
the receiver joins a channel with respect to when its wave started,
the higher the reception rate.
Once the receiver is joined to a wave channel, the receiver remains
joined to the wave channel until the channel goes quiescent, at which
point the receiver MUST leave the channel.
The way the receiver adjusts its reception rate is inspired by TFRC
[9]. The receiver at all points in time maintains a target reception
rate, and the receiver is allowed to join the next wave channel if
Luby & Goyal Experimental [Page 8]
RFC 3738 WEBRC Building Block April 2004
after joining its anticipated reception rate from all the layers it
is joined to would be at most its target reception rate. The target
rate is continually updated based on a set of measured parameters.
The primary parameters are an estimate LOSSP of the average loss
probability and an estimate ARTT of the average multicast round-trip
time.
In the remainder of this document, log(X) denotes the natural
logarithm of X, i.e., the logarithm base 2.71828459... of X.
3.1. Sender Operation
The sender operation is by design much simpler than the receiver
operation.
3.1.1. Sender inputs and initialization
The primary input to the sender for the session is SR_b. SR_b is an
upper bound to the sender transmission rate in bits per second at any
point in time (with some reasonable granularity) in aggregate to all
channels. Naturally, this is then also the maximum rate in bits per
second that any receiver could receive data from the session at any
point in time. It is RECOMMENDED that the sender transmission rate
in aggregate to all channels be made constant as described in [8].
It is also RECOMMENDED that the session description indicate whether
the aggregate transmission rate is constant, unless there is no
ambiguity.
The secondary inputs to the sender are listed below. These inputs
are secondary because their values will generally be fixed to default
values that will not change, because they will be derived from SR_b,
or because they are chosen based on non-WEBRC considerations.
o LENP_B is the length of packets in bytes sent to the session. The
value of LENP_B depends on the complete protocol, but in general
this SHOULD be set to as high a value as possible without
exceeding the MTU size for the network that would cause
fragmentation.
o BCR_P is the transmission rate on the base channel at the
beginning of a time slot in packets per second. The default value
for BCR_P is 1.
o TSD is the time slot duration measured in seconds. The
RECOMMENDED value for TSD is 10.
o QD is the minimum quiescent period duration measured in seconds.
The RECOMMENDED value for QD is 300.
Luby & Goyal Experimental [Page 9]RFC 3738 WEBRC Building Block April 2004
o P is the multiplicative drop in every channel rate over each time
slot. The default value for P is 0.75.
o N is the duration in time slots for each wave. N is also the
number of wave channels active at any time. (A wave channel is
called active when it is not quiescent.) A sender may choose any
value that allows it to produce waves that substantially follow
the required exponential shape described in Section 3.1.2. A
RECOMMENDED mechanism for relating N to SR_b, BCR_P and P is
described in [8].
From these inputs the following fixed sender parameters can be
derived as follows.
o SR_P = SR_b/(8*LENP_B) is the sender transmission rate in packets
per second.
o BCR_b = 8*LENP_B*BCR_P is the rate of the base channel at the
beginning of a time slot in bits per second.
o L = ceil(BCR_P*TSD*(P-1)/log(P)) is the number of base channel
packets sent in each time slot.
o Q = ceil(QD/TSD) is the number of quiescent time slots per cycle
for a wave channel.
o T = N + Q is the total number of time slots in a cycle. T is also
the total number of wave channels.
o For the base channel CN = T and for the wave channels CN =
0,1,...,T-1. The sender has the description of the channels
assigned to the session and the mapping between the channels and
the CNs.
o C = TSD*T is the total duration of a cycle in seconds.
3.1.2. Sending packets to the session
The sender keeps track of the current time slot index CTSI. The
value of CTSI is incremented by 1 modulo T each TSD seconds. The
value of CTSI is placed into each packet in the format described in
Section 5. For each packet sent to the session, the sender also
places the channel number CN of the channel into the packets in the
format described in Section 5. Recall that CN = T for the base
channel and CN = 0,1,...,T-1 for the wave channels.
Luby & Goyal Experimental [Page 10]RFC 3738 WEBRC Building Block April 2004
For each packet sent to the session, the sender calculates a packet
sequence number PSN and places it into the packet. The value of PSN
is scoped by CN, and the value of PSN is consecutively increasing
within each channel. Furthermore, for each wave channel, the last
packet sent before the channel becomes quiescent must have the
maximum possible PSN value. When the short format for Congestion
Control Information is used (see Section 5.1), this implies that for
any wave channel the last PSN value sent to the channel just before
the channel becomes quiescent is 2^16-1 = 65,535. Similarly, when
the long format for Congestion Control Information is used (see
Section 5.2), the PSN for the final packet of any wave is 2^32-1 =
4,294,967,295. The PSN of the initial packet of a wave thus depends
on TSD, P, BCR_P and SR_P. For the base channel, the first packet of
each time slot has a PSN congruent to zero modulo L. Hence, instead
of 2^16 - 1 or 2^32 - 1 being the highest PSN used (depending on the
choice of short format or long format Congestion Control
Information), the highest PSN is one less than the largest multiple
of L that does not exceed 2^16 (short format) or 2^32 (long format).
The format for the PSN within packets is described in Section 5.
The rate at which packets are sent to the base channel starts at
BCR_P packets per second at the beginning of each time slot and
decreases exponentially to P*BCR_P at the end of that time slot.
The packet rate for the wave channels is more complicated. Each wave
channel carries a sequence of waves separated by quiescent periods.
On each wave channel each wave is active during N time slots followed
by a quiescent period of Q time slots. The waves on wave channel i
end at the ends of time slots with CTSI i. Therefore wave channel i
is active during time slots i-N+1 modulo T, i-N+2 modulo T, ..., i
and is quiescent for time slots i+1 modulo T, i+2 modulo T, ..., i+Q
modulo T. Wave channel i first becomes active within time slot i-N+1
modulo T at a point in time that may depend on the value of SR_b.
Except for at most the first two time slots after a wave becomes
active, the packet rate of the wave MUST decrease exponentially by a
factor of P per TSD seconds, down to a rate of BCR_P at the end of
the last active time slot. At the beginning of each wave, i.e., for
at most the first two time slots when the wave becomes active, the
rate MAY deviate from this exponential form so that the total sending
rate in aggregate to all of the channels is constant. A RECOMMENDED
design for the beginnings of waves to achieve this goal is described
in [8].
Luby & Goyal Experimental [Page 11]
RFC 3738 WEBRC Building Block April 2004
3.2. Receiver Operation
The bulk of the complexity in WEBRC is in the receiver operation.
For ease of explanation, suppose for the moment that during the
reception there is no packet loss and packets are arriving at exactly
the rate at which they were sent. The sender transmission rate to
the channels is designed so that the receiver reception rate behaves
as follows.
Upon entering a session, the receiver immediately joins the base
channel. When the receiver wants to increase its rate, it joins
consecutive layers starting with the lowest and moving towards the
highest. (Recall that the designations of lowest to highest change
as waves become active and quiescent.) When the receiver wants to
maintain its current reception rate and it is already joined to the
lowest NWC layers, if the receiver joins channel i-1+NWC modulo T
sometime during time slot i then the receiver joins channel i+NWC
modulo T TSD seconds later in time slot i+1. When the lowest layer
becomes quiescent the receiver leaves the channel.
Suppose the receiver wants to decrease its rate till it is joined to
just the base channel. Assume that a receiver is joined to the
lowest NWC < N-2 layers at the beginning of time slot i, i.e., wave
channels i, i+1 modulo T,..., i+NWC-1 modulo T. Then, the aggregate
packet reception rate of the receiver over the next NWC time slots
will behave as follows if the receiver does not join any wave
channels during this time. At the beginning of time slot i the
receiver reception rate is BCR_P*(1 + (1/P) + (1/P)^2 + ... +
(1/P)^NWC). Then the receiver reception rate decreases by a factor
of P over the duration of each time slot, and at the end of each time
slot the reception rate decreases by an additive amount of P*BCR_P.
At the end of time slot i+NWC-1 mod T, the receiver reception rate is
BCR_P*(1+P), and at the beginning of time slot i+NWC mod T the
receiver is joined only to the base channel and its reception rate is
BCR_P.
3.2.1. Receiver inputs and initialization
Before joining a session the receiver MUST know the mapping between
the CNs and the channels. Upon joining the session or shortly
thereafter, it SHOULD have the values of LENP_B, BCR_P, TSD, P, N, L,
Q and T. Some of these values may be computed or measured once the
receiver has joined the session. For example, the receiver MAY
obtain LENP_B and T from the first packet received from the base
channel, and the receiver MAY measure BCR_P once it is joined to the
base channel. The values of P, Q and TSD MAY be fixed to default
values built into the receiver that do not change from session to
session, and the value of N MAY be computed as T-Q. The receiver
Luby & Goyal Experimental [Page 12]
RFC 3738 WEBRC Building Block April 2004
SHOULD know whether the sender is employing a technique to produce
constant aggregate rate as described in [8].
When a receiver first joins a session, it MUST first join just the
base channel and start receiving packets to determine the current
time slot index. If during the course of the session the receiver
continually loses a high fraction of the packets from the base
channel even when the receiver is only joined to the base channel,
the receiver SHOULD leave the session.
The receiver MAY also have other individually set parameters that may
be used to determine its behavior. One such parameter is MRR_b:
o MRR_b is the maximum receiver reception rate in bits per second.
This may be used to determine the maximum reception rate this
receiver is willing to reach. Thus, the maximum reception rate
that the receiver can possibly achieve in the session is the
minimum of SR_b and MRR_b. A recommended value of MRR_b for a
receiver is the bandwidth capacity of the last link to the
receiver. MRR_P is the maximum receiver reception rate in packets
per second, i.e., MRR_P = MRR_b/(8*LENP_B).
3.2.2. Receiver measurements and calculations
As outlined in the introduction, the way a receiver adjusts its
reception rate is inspired by TFRC [9]. The receiver at all points in
time maintains a target reception rate, and the receiver is allowed
to join the next wave channel if joining would increase its reception
rate to at most its target reception rate. The target rate is
continually updated based on a set of measured parameters.
Two primary parameters are the estimate LOSSP of the average loss
probability and the estimate ARTT of the average MRTT. Both LOSSP
and ARTT are moving averages of measurements based on discrete
events. For many of the other estimates calculated by WEBRC, using
an exponentially weighted moving average (EWMA) with a fixed
averaging fraction is sufficient. However, the calculations of LOSSP
and ARTT require a more general and sophisticated filtering approach.
3.2.2.1. Average loss probability
The design of TFRC [9] reflects that, because the average packet loss
probability can vary by orders of magnitude, any estimate of the
average loss probability based on either a fixed number of packets or
on a fixed period of time with a fixed averaging fraction will be
poor. In TFRC the average is estimated from the numbers of packets
between beginnings of loss events, and the number of loss events used
is fixed.
Luby & Goyal Experimental [Page 13]RFC 3738 WEBRC Building Block April 2004
The estimate LOSSP of the average loss probability of the receiver is
maintained in a manner somewhat similar to that described in TFRC
[9]. The WEBRC receiver estimates the inverse of the average loss
probability by applying two EWMA filters to the packet reception
measurements, a time-based filter with smoothing constant 0 < Nu < 1
and a loss-based filter with smoothing constant 0 < Delta < 1. The
recommended values for the smoothing constants are Nu = 0.3 and Delta
= 0.3. The reason for the time-based filter is that the loss events
in WEBRC are bursty; they typically occur just after a new wave has
been joined. To smooth out this burstiness, the time-based filter is
applied to the packet reception measurements at the end of each epoch
to smooth out the bursty loss events over a few time slot durations.
Intuitively, the time-based filter averages packet reception events
such that the events are smoothed out over an interval of time
proportional to TSD/Nu seconds. The loss-based filter, similar to
what is suggested in TFRC, is applied to the output of the time-based
filter to produce the estimate of the inverse of the average loss
probability. Intuitively, the loss-based filter averages loss events
such that each loss event is averaged in with weight Delta.
As described later, LOSSP is initialized at the end of slow start and
occasionally reset due to other events. Let W and X be counts of
packets, let Y be a count of loss events and let Z be the long-term
estimate of the inverse of the average loss probability. Whenever
the value of LOSSP is initialized or reset, the values of W, X, Y and
Z are also initialized or reset.
Recall that TSD is the duration of a time slot. The epoch length EL
is the duration of time between decisions to adjust the reception
rate. Generally EL is much smaller than TSD, and the RECOMMENDED
values are EL = 0.5 seconds and TSD = 10 seconds.
Define G = Nu*EL/TSD as the amount of time-based smoothing to perform
at the end of each epoch. The update rules for W, X, Y, Z, and LOSSP
are the following:
o At the end of each epoch, adjust X, Y and Z and compute LOSSP as
follows:
Z = Z*(1-Delta)^(G*Y) + G*X/(G*Y+1)*(1-(1-Delta)^(G*Y+1))
X = X*(1-G)
Y = Y*(1-G)
Z1 = Z*(1-Delta)^Y + X/(Y+1)*(1-(1-Delta)^(Y+1))
Z2 = Z*(1-Delta)^(Y+1) + (X+W+1)/(Y+2)*(1-(1-Delta)^(Y+2))
Luby & Goyal Experimental [Page 14]RFC 3738 WEBRC Building Block April 2004
LOSSP = 1/max{Z1,Z2,1}
o For each packet event (whether it is a received packet or a lost
packet), W = W + 1
o At the beginning of each loss event, update W, X, and Y as
follows:
X = X + W
W = 0
Y = Y + 1
The intuition behind these update rules is the following. If just
loss-filtering were used to update Z, then Z would be decreased by a
multiplicative amount 1 - Delta for each loss event and Z would be
increased by an additive amount Delta for each packet. To smooth out
loss events over more than one time slot, these adjustments are
filtered into Z over time, at the rate of a fraction G at the end of
each epoch. Thus, the variables X and Y are counts of the portions
of the packets and loss events, respectively, that have not yet been
filtered into the long-term memory Z. W is the count of packets
since the last loss event started. This explains why W is increased
by one for each packet and Y is increased by one for each loss event.
At the end of each epoch a fraction G of both X and Y are filtered
into Z according to the loss-filter rule described above, and then
the same fraction G is removed from both X and Y to account for the
fact that this portion has been filtered into Z. The LOSSP
calculation combines the short-term history (X,Y) with the long-term
history Z and also allows the arrivals since the last loss W to have
some influence. The value of Z2 is what Z1 would become were the
next packet to be lost.
To reset the loss calculation to a value LOSSP = a, the state
variables are set as follows:
W = 0
X = 0
Y = 0
Z = 1/a
Luby & Goyal Experimental [Page 15]RFC 3738 WEBRC Building Block April 2004
3.2.2.2. Average round-trip time
The receiver maintains an average round-trip time, ARTT, as a
measurement-based filter of MRTT measurements using a smoothing
constant 0 < Alpha < 1. The RECOMMENDED value for Alpha is 0.25.
Each time the receiver joins a channel (either the base channel upon
entering a session or wave channels continually), it makes a
measurement of the multicast round-trip time MRTT as follows. Let V
be an auxiliary variable that is used that keep track of the average
of the square of the MRTT measurements. When the receiver sends the
join for the channel it records the current time JoinTime and sets a
Boolean variable JOINING to true. When the first packet is received
from the channel the receiver records the current time FirstTime and
resets the value of JOINING to false. If it is the base channel that
has been joined, ARTT is set to FirstTime-JoinTime and V is set to
ARTT*ARTT. Otherwise, the value of MRTT is set to (FirstTime -
JoinTime) - log(1/P)/2/(1-P)/BCR_P * P^NWC. (Note that this value
can be negative.) Then, ARTT is updated as follows. Let Omega =
Alpha*ARTT*ARTT/V, and at the Kth MRTT measurement let Rho =
Omega/(1-(1-Omega)^(K+1)). (Note that as K grows Rho approaches
Omega.) Then, V is updated to (1-Rho)*V+Rho*MRTT*MRTT and ARTT is
updated to max{P*ARTT,(1-Rho)*ARTT+Rho*MRTT}.
Usually ARTT is updated to the second term in the max, and in this
case ARTT is the EWMA of the previous value of ARTT and the new MRTT,
with a weighting on the new MRTT that as K grows is proportional to
the square of the previous ARTT divided by the previous average V of
the square of the MRTT. Thus, if there is not much variance in the
previous MRTTs relative to the square of their average then the new
MRTT will be filtered into ARTT with a high weight, whereas if there
is a lot of variance in the previous MRTTs relative to the square of
their average then the new MRTT will be filtered into ARTT with a low
weight. The intuitive rationale for this is that in general the
number of measurements needed to compute a meaningful average for a
random variable is proportional to its variance divided by the square
of its average; see, e.g., [6]. By making the weight factor depend on
previous measurements in this way, the appropriate weight to use to
average the new MRTT into the ARTT self-adjusts automatically to the
variability in the measurements.
3.2.2.3. Rate Equation
The receiver calculates the reception rate REQN based on the TCP
equation as follows: REQN = 1/(ARTT*sqrt{LOSSP}(0.816 +
7.35*LOSSP*(1+32*LOSSP^2))). This equation comes from TFRC [9].
Luby & Goyal Experimental [Page 16]RFC 3738 WEBRC Building Block April 2004
3.2.2.4. Epochs
The receiver makes decisions on whether or not to join another wave
channel at equally-spaced units of time called epochs. The duration
of an epoch in seconds, EL, is set to be a small fraction of TSD, so
that decisions to join a channel can be made at a much finer
granularity than TSD. A standard setting is EL = TSD/20. Thus, with
the recommended setting of TSD = 10, it is RECOMMENDED that EL = 0.5.
3.2.2.5. Average reception rate
There are two averaged reception rates maintained by the receiver:
TRR_P, the true reception rate, and ARR_P, the anticipated reception
rate. These are used for different purposes and thus are calculated
quite differently. Recommended values for the filtering weights Beta
and Zeta are provided at the end of this subsection.
In start-up mode, the true reception rate TRR_P is used to ensure
that the receiver does not increase its reception rate too quickly
above its current reception rate. In the transition from start-up
mode to normal operation and in normal operation, TRR_P is used in
setting the slow start rate. TRR_P is calculated based on the
measurement of RR_P, where RR_P is the receiver reception rate in
packets per second measured at the beginning of an epoch averaged
over the epoch that just ended. TRR_P is initialized to BCR_P +
k*log(P)/TSD when the first base channel packet of the session
arrives, where k is the PSN of the packet reduced modulo L. TRR_P is
updated to (1-Zeta)*TRR_P + Zeta*RR_P at the beginning of each epoch
after RR_P is measured for the previous epoch.
The anticipated reception rate ARR_P is the receiver's estimate of
the total instantaneous rate of the currently joined channels. It is
used to compare against the target rate to decide whether or not the
receiver should increase its reception rate by joining the next
higher unjoined layer. ARR_P is calculated based on a measurement
IRR_P and on the number of joined wave channels NWC. The ideal
reception rate IRR_P is the reception rate in packets per second
including both received and lost packets; like RR_P, it is measured
at the beginning of the epoch and averaged over the previous epoch.
ARR_P, IRR_P and NWC are updated as follows:
o NWC is initialized to 0.
o When the first base channel packet arrives, ARR_P is set to BCR_P
+ k*log(P)/TSD, where k is the PSN of the packet reduced modulo L.
Luby & Goyal Experimental [Page 17]RFC 3738 WEBRC Building Block April 2004
o At the beginning of each epoch, IRR_P is measured over the
previous epoch and then ARR_P is updated to
P^(EL/TSD)*(1-Beta)*ARR_P + Beta*IRR_P. Then if ARR_P exceeds
ARR_P_max = ((1/P)^(NWC+1)-1)/((1/P)-1)*BCR_P, ARR_P is updated to
ARR_P_max.
o When a join is made to the next higher unjoined layer, NWC is
updated to NWC+1 and then ARR_P is multiplicatively increased by
the factor ((1/P)^(NWC+1)-1)/((1/P)^NWC-1). (Joins happen at
epoch boundaries; this adjustment is in addition to the adjustment
above.)
o Each time a next time slot index is detected, ARR_P is additively
increased by (1-P)*BCR_P to account for the change in rate on the
base channel. In addition, the bottom layer in the previous time
slot has just gone quiescent and thus a message to leave this
layer has been sent, ARR_P is additively decreased by BCR_P and
NWC is decremented by 1. Thus, the combination of these effects
on ARR_P is that it is additively decreased by P*BCR_P.
Consider for the moment what happens if Beta = 0 and ARR_P is an
accurate estimate of the total rate of the joined channels. The
adjustments to ARR_P upon joining and leaving wave channels, with the
passage of epochs, and with the detection of time slot changes will
then cause ARR_P to remain an accurate estimate. In practice, Beta
MUST be positive; allowing an influence of IRR_P prevents ARR_P from
drifting away from being an accurate estimate of the total joined
rate.
The motivation for separate estimates TRR_P and ARR_P is as follows.
ARR_P is needed for comparison with the TFRC-inspired target rate
because there is no lag before it reflects the potential rate
increase resulting from joining the next higher layer and because it
measures the total possible impact on the network since it also
includes lost packets. TRR_P is needed because it reflects the rate
of data arriving at the receiver and this is used to ensure that
there is not a large gap between the joined rate and the receiving
rate.
The recommended values for Beta and Zeta depend on whether the
receiver is in start-up mode (SSR_P = infinity). In start-up mode,
it is RECOMMENDED that Beta = (1 - P^(0.25))/2 and Zeta = sqrt(P)/(1
+ sqrt(P)). In normal operation, it is RECOMMENDED that Beta = 1 -
(P/(1+P))^(EL/TSD) and Zeta = 2*EL/(4+TSD).
Luby & Goyal Experimental [Page 18]RFC 3738 WEBRC Building Block April 2004
3.2.2.6. Slow start
WEBRC uses a slow start mechanism to quickly ramp up its rate at both
the beginning of the session and in the middle of a session when the
rate drops precipitously. To enact this, the receiver maintains the
following parameters:
o SSMINR_P is the minimum allowed slow start threshold rate in
packets per second. The recommended value for SSMINR_P is
BCR_P*(1+1/P+1/P^2).
o SSR_P is the slow start threshold rate in packets per second. It
is adjusted at the beginning of loss events as described in
Section 3.2.3.4. SSR_P is initialized to infinity and is first set
to a finite value when the receiver leaves the initial start-up
period as described below.
At the beginning of a session, the receiver cannot compute a
meaningful target rate from its measurements. Thus, it uses SSR_P =
infinity until one of the following events causes an end to this
start-up mode:
o A packet loss is detected. In this case the value of SSR_P is
updated to max{SSMINR_P, P*TRR_P} as with the beginning of any
other loss event.
o A sharp increase in MRTT is detected. While SSR_P = infinity the
receiver MUST compute, in the notation of Section 3.2.2.2,
differences in successive measurements of (FirstTime-JoinTime)
from successive waves and MUST set SSR_P to max{SSMINR_P, P*TRR_P}
when a large increase in (FirstTime-JoinTime) is observed. It is
RECOMMENDED that an increase in (FirstTime-JoinTime) be considered
large if it exceeds (P^(NWC+1)-1)/(P*log(P)) / ARR_P.
o The maximum reception rate is reached. When SSR_P = infinity, if
(P^(-NWC-2)-1)/(P^(-NWC-1)-1)*ARR_P exceeds MRR_P or SR_P, the
receiver MUST set SSR_P to max{SSMINR_P, TRR_P}.
o TRR_P is not increasing consistent with the last join of a wave
channel. While SSR_P = infinity, it is RECOMMENDED that the
receiver wait at least one full epoch after the first packet of a
wave is received before joining the next wave. If the TRR_P after
that full epoch is greatly below ARR_P the receiver SHOULD NOT
join and SHOULD then set SSR_P to max{SSMINR_P, TRR_P}. It is
RECOMMENDED that TRR_P be considered greatly below ARR_P if TRR_P
< c * ARR_P - 2/EL, where c = Zeta + (1-Zeta)*(P^(-EL/TSD))*(Zeta
+ (1-Zeta)*sqrt(P)*(P^(-EL/TSD)))/g with g = (P^(-NWC-1)-1)/(P^(-
NWC)-1).
Luby & Goyal Experimental [Page 19]RFC 3738 WEBRC Building Block April 2004
In any of these four cases, the variables associated with LOSSP are
reset to make REQN, calculated as in Section 3.2.2.3 with the current
value of ARTT, equal TRR_P.
3.2.2.7. Target rate
In typical operation, SSR_P has a finite value and the target rate
TRATE is computed as TRATE = min{max{SSR_P, REQN}, MRR_P}. When
SSR_P = infinity, TRATE is computed as TRATE = min{4*TRR_P, MRR_P}.
3.2.3. Receiver events
There are various receiver events, some of which are triggered by the
passing of time on the receiver, and others by events such as packet
reception, detection of packet loss, reception of a first packet from
a channel, and exceptional time-outs.
3.2.3.1. Packet reception
Most packet reception events require the receiver to merely register
the reception for later calculation of RR_P and IRR_P (see Section
3.2.2.5) and increment W for later calculation of LOSSP (see Section
3.2.2.1).
Additional actions, described in the following three subsections, are
required if the packet is the first packet received in response to a
join operation, the CTSI of the packet indicates a time slot change,
or the CN and PSN of the packet indicate a packet loss.
3.2.3.2. First packet after join
When channel i is the most recently joined channel and the Boolean
variable JOINING is true, the reception of a packet with PSN = i is a
special event because it is the first packet received in response to
the most recent join. MRTT is calculated and ARTT and V are updated
as described in Section 3.2.2.2, and JOINING is set to false. The
first received packet of the session furthermore necessitates
initialization of ARR_P and TRR_P as described in Section 3.2.2.5.
3.2.3.3. Time slot change
This is an event that is triggered by the reception of a packet with
a CTSI value that is one larger modulo T than the previous CTSI
value. When a packet with a new CTSI = i is received, a leave is
sent for the lowest layer in the previous time slot, i.e., wave
channel i-1 modulo T, NWC is updated to NWC-1, and ARR_P is updated
Luby & Goyal Experimental [Page 20]RFC 3738 WEBRC Building Block April 2004
to ARR_P - P*BCR_P as described in Section 3.2.2.5. If the channel
for which the leave is sent is also the most recently joined wave
channel and JOINING is true, then JOINING is set to false.
It is possible due to packet reordering for some packets from the
previous time slot to be received after packets from the current time
slot. It is RECOMMENDED that measures be put into place to handle
this situation appropriately, i.e., to not trigger a time slot change
in this situation. One simple mechanism for this is as follows:
Compute the difference i-j modulo T, where i is the CTSI of the
received packet and j is the current CTSI of the receiver. A
difference of zero is, of course, not a time slot change. In
addition, a very large difference, for example a difference larger
than T-Q/2, should also not trigger a time slot change.
3.2.3.4. Loss event
Each time the receiver detects a lost packet (based on the sequence
numbers in the packets scoped by the channel number), the receiver
records the start of a new loss event and sets a Boolean variable
LOSS_EVENT to true that will automatically reset to false after ARTT
seconds. All subsequent packet loss for a period of ARTT seconds is
considered as part of the same loss event. When a start of a loss
event is detected, the value of SSR_P is updated to max{SSMINR_P,
P*TRR_P}.
It is RECOMMENDED that the receiver account for simple misordering of
packets without inferring a loss.
3.2.3.5. Epoch change
This is an event that is triggered by the passage of time at the
receiver, which occurs each EL seconds. When this happens, TRR_P and
ARR_P are computed as described in Section 3.2.2.5. Immediately after
these updates, a decision is made about whether to join the next
higher layer as described in Section 3.2.3.6.
3.2.3.6. Join the next higher layer
At the beginning of each epoch, after updating the values of ARR_P
and TRR_P as described in Section 3.2.2.5, the receiver decides
whether or not to join the next higher layer as follows:
o If the first base channel packet has not yet arrived the receiver
MUST not join.
o If there is a loss event in progress (LOSS_EVENT = true) the
receiver MUST not join.
Luby & Goyal Experimental [Page 21]RFC 3738 WEBRC Building Block April 2004
o If a join of a channel is in progress (JOINING = true) the
receiver MUST not join.
o If NWC = N the receiver MUST not join.
o If the receiver is employing the OPTIONAL rule described in
Section 3.2.2.6, SSR_P = infinity, and a full epoch has not passed
since the first packet arrival on the most recently joined wave
channel then the receiver MUST not join.
o If the receiver is employing the OPTIONAL rule described in
Section 3.2.2.6, SSR_P = infinity, and a full epoch has passed
since the first packet arrival on the most recently joined wave
channel, then the receiver checks if TRR_P is greatly below ARR_P
as described in Section 3.2.2.6. If TRR_P is greatly below ARR_P
the receiver MUST not join.
o The receiver calculates REQN as described in Section 3.2.2.3.
o The receiver calculates TRATE as described in Section 3.2.2.7.
o If the sender is not sending at constant aggregate rate and TRATE
< ARR_P*((1/P)^{NWC+2}-1)/((1/P)^{NWC+1}-1), the receiver MUST not
join. If the sender is sending at constant aggregate rate and
TRATE < ARR_P*((1/P)^{NWC+2}-1)/((1/P)^{NWC+1}-1) and TRATE <
SR_P, the receiver MUST not join.
o If SSR_P is finite and the sender is not sending at constant
aggregate rate or SSR_P is finite and the sender is sending at
constant aggregate rate and TRATE < SR_P then the receiver MAY
apply one additional OPTIONAL check before deciding to join.
It is RECOMMENDED that the receiver not join if the value of RR_P
is not sufficiently lower than the maximum value of RR_P observed
since the last join. It is RECOMMENDED that RR_P is sufficiently
low to allow a join if RR_P <= max{RRmax-2/EL,P*RRmax}, where
RRmax is the maximum measured RR_P since the last join.
If the receiver does not join because RR_P is not sufficiently
small then a value of LOSSP is calculated so as to make the value
of the REQN equation given in Section 3.2.2.3 evaluate to
ARR_P*((1/P)^(NWC+2)-1)/((1/P)^(NWC+1)-1) with respect to the
current value of ARR_P. Then, the variables associated with LOSSP
are reset based on this calculated value of LOSSP as described at
the end of Section 3.2.2.1.
o Otherwise, the receiver MAY join the next higher layer.
Luby & Goyal Experimental [Page 22]RFC 3738 WEBRC Building Block April 2004
Suppose the receiver has decided to join and CTSI = i. The receiver
joins the next higher wave channel, i.e., the wave channel with CN =
i+NWC modulo T, increments NWC by 1, and then updates ARR_P to
ARR_P*((1/P)^{NWC+1}-1)/((1/P)^NWC-1) as described in Section
3.2.2.5. The time of the join is recorded for use in updating ARTT
as described in Section 3.2.2.2.
3.2.3.7. Join timeout
When no packet arrives in response to the join of channel for a long
period of time, the join times out. The receiver sets JOINING to
false, updates ARR_P to ARR_P*((1/P)^NWC-1)/((1/P)^{NWC+1}-1), and
then decrements NWC by 1.
The RECOMMENDED threshold for a join timeout is max{2*V/ARTT,10*ARTT}
seconds.
3.2.3.8. Exceptional timeouts
These are timeouts when the packet reception behavior is far from
what it should be and these MUST trigger the receiver to leave the
session. Exceptional timeouts include
o No packets are received for a long period. A RECOMMENDED
threshold is max{10,TSD} seconds.
o There is no change in time slot index for a long period. A
RECOMMENDED threshold is max{20,2*TSD} seconds.
4. Applicability Statement
WEBRC is intended to be a congestion control scheme that can be used
in a complete protocol instantiation that delivers objects and
streams (both reliable content delivery and streaming of multimedia
information). WEBRC is most applicable for delivery of objects or
streams of substantial length, i.e., objects or streams that range in
length from hundreds of kilobytes to many gigabytes, and whose
transfer time is on the order of tens of seconds or more.
4.1. Environmental Requirements and Considerations
WEBRC can be used with both multicast and unicast networks. However,
the scope of this document is limited to multicast. WEBRC requires
connectivity between a sender and receivers, but does not require
connectivity from receivers to the sender.
WEBRC inherently works with all types of networks, including LANs,
WANs, Intranets, the Internet, asymmetric networks, wireless
Luby & Goyal Experimental [Page 23]RFC 3738 WEBRC Building Block April 2004
networks, and satellite networks. Thus, the inherent raw scalability
of WEBRC is unlimited. However, in some network environments varying
reception rates to receivers may not be advantageous. For example,
the network may have dedicated a fixed amount of bandwidth to the
session or there may be no effective way for receivers to dynamically
vary the set of channels they are joined to, as in a satellite
network.
Receivers join and leave channels using the appropriate multicast
join and leave messages. For IPv4 multicast, IGMP messages are used
by receivers to join and leave channels. For IPv6, MLDv2 messages
are used by receivers to join and leave channels. This is the only
dependency of WEBRC on the IP version.
WEBRC requires receivers to be able to uniquely identify and
demultiplex packets associated with a session in order to effectively
perform congestion control over all packets associated with the
session. How receivers achieve this is outside the scope of this
document.
WEBRC is presumed to be used with an underlying network or transport
service that is a "best effort" service that does not guarantee
packet reception, packet reception order, and which does not have any
support for flow or congestion control. For example, the Any-Source
Multicast (ASM) model of IP multicast as defined in RFC 1112 [7] is
such a best effort network service. While the basic service provided
by RFC 1112 is largely scalable, providing congestion control or
reliability should be done carefully to avoid severe scalability
limitations, especially in the presence of heterogeneous sets of
receivers.
There are currently two models of multicast delivery, the Any-Source
Multicast (ASM) model as defined in RFC 1112 [7] and the Source-
Specific Multicast (SSM) model as defined in [10]. WEBRC works with
both multicast models, but in a slightly different way with somewhat
different environmental concerns. When using ASM, a sender S sends
packets to a multicast group G, and the WEBRC channel address
consists of the pair (S,G), where S is the IP address of the sender
and G is a multicast group address. When using SSM, a sender S sends
packets to an SSM channel (S,G), and the WEBRC channel address
coincides with the SSM channel address.
A sender can locally allocate unique SSM channel addresses, and this
makes allocation of channel addresses easy with SSM. To allocate
channel addresses using ASM, the sender must uniquely chose the ASM
multicast group address across the scope of the group, and this makes
allocation of WEBRC channel addresses more difficult with ASM. This
is an issue for WEBRC because several channels are used per session.
Luby & Goyal Experimental [Page 24]
RFC 3738 WEBRC Building Block April 2004
WEBRC channels and SSM channels coincide, and thus the receiver will
only receive packets sent to the requested WEBRC channel. With ASM,
the receiver joins a channel by joining a multicast group G, and all
packets sent to G, regardless of the sender, may be received by the
receiver. Thus, SSM has compelling security advantages over ASM for
prevention of denial of service attacks. In either case, receivers
SHOULD use mechanisms to filter out packets from unwanted sources.
WEBRC assumes that the packet route between the sender and a
particular receiver is the same for all channels associated with a
session. For SSM this assumption is true because the multicast tree
is a shortest path tree from each receiver to the sender and
generally this path changes infrequently. For ASM there are some
issues that if not properly considered may invalidate this
assumption. With ASM, the packet route between the sender and
receivers may initially be through the Rendezvous Point (RP) and then
switch over to the shortest path to the sender as packets start
flowing in a channel. The first issue is that the RP may not be the
same for all channels associated with a session, and thus the first
packets sent to the channels may follow a route that depends on the
RP of the channel. This depends on the RP configuration for the
sender. If the sender registers all channels associated with the
session with the same RP then the assumption is true, but if the
sender registers different channels with different RPs then the
assumption may not be true. Thus, it is RECOMMENDED that the sender
register all channels associated with a session with the same RP.
Another issue is that when the channel switches over from the RP to
the sender-based tree then the route to the receivers may vary within
a channel. Furthermore, this may cause either the receipt of
duplicate packets at receivers or loss of packets depending on the
smoothness of the switchover. Thus, it is RECOMMENDED that the RP be
placed as close as possible to the sender. The best location for the
RP is that it be the first-hop router closest to the sender, in which
case the path to the sender and the path to the RP is the same for
each receiver and the problems mentioned above are eliminated. The
consequences of this assumption not being true are that the receiver
reaction to congestion may not be appropriate. Generally, the WEBRC
receiver will act conservatively and reduce its reception rate too
much if this assumption is not true, but there can be cases where the
receivers will act inappropriately.
5. Packet Header Fields
Packets sent to a session using WEBRC MUST include Congestion Control
Information fields as specified in this section. This document
specifies short and long formats for the Congestion Control
Information, and it is RECOMMENDED that protocol instantiations use
one of these two formats. Other formats for the Congestion Control
Luby & Goyal Experimental [Page 25]
RFC 3738 WEBRC Building Block April 2004
Information fields MAY be used by protocol instantiations, but all
protocol instantiations are REQUIRED to use these fields in a format
that is compatible with the interpretations of these fields. Thus,
if a protocol does use a different format for the fields in the
Congestion Control Information then it MUST specify the lengths and
positions of these fields within the packet header.
All integer fields are carried in "big-endian" or "network order"
format, that is, most significant byte (octet) first. All constants,
unless otherwise specified, are expressed in base ten.
5.1. Short Format Congestion Control Information
The short format for the Congestion Control Information is shown in
Fig. 1. The total length of the short format is 32-bits.
0 1 2 3
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
| CTSI | Channel Number| Packet Sequence Number |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
Fig. 1 - Short format for Congestion Control Information
The function of each field in the Congestion Control Information is
the following.
Current Time Slot Index (CTSI): 8 bits
CTSI indicates the index of the current time slot. This must
be sent in each packet within the session. The Current Time
Slot Index increases by one modulo T each TSD seconds at the
sender, where T is the number of time slots associated with the
session and TSD is the time slot duration. Note that T is also
the number of wave channels associated with the session, and
thus T MUST be at most 255.
Channel Number (CN): 8 bits
CN is the channel number that this packet belongs to. CN for
the base channel is T, and the CNs for the wave channels are 0
through T-1. Thus, T+1 channels in total are used, and thus T
MUST be at most 255.
Packet Sequence Number (PSN): 16 bits
The PSN of each packet is scoped by its CN value. The sequence
numbers of consecutive packets sent to the base channel are
Luby & Goyal Experimental [Page 26]RFC 3738 WEBRC Building Block April 2004
numbered consecutively modulo 2^16. The same sequence of PSNs
are used for each wave channel in each cycle. The sequence
numbers of consecutive packets sent to a wave channel are
numbered consecutively modulo 2^16 within each cycle, ending
with the last packet sent to the channel before the channel
goes quiescent with PSN = 2^16-1.
5.2. Long Format Congestion Control Information
The long format for the Congestion Control Information is shown in
Fig. 2. The total length of the long format is 64-bits.
0 1 2 3
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
| CTSI | Channel Number |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
| Packet Sequence Number |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
Fig. 2 - Long format for Congestion Control Information
The meaning of each field for the long format is the same as for the
short format, the only difference is that each field is twice as
long.
Current Time Slot Index (CTSI): 16 bits
CTSI indicates the index of the current time slot. This must
be sent in each packet within the session. The Current Time
Slot Index increases by one modulo T each TSD seconds at the
sender, where T is the number of time slots associated with the
session and TSD is the time slot duration. Note that T is also
the number of wave channels associated with the session, and
thus T MUST be at most 65,535.
Channel Number (CN): 16 bits
CN is the channel number that this packet belongs to. CN for
the base channel is T, and the CNs for the wave channels are 0
through T-1. Thus, T+1 channels in total are used, and thus T
MUST be at most 65,535.
Packet Sequence Number (PSN): 32 bits
The PSN of each packet is scoped by its CN value. The sequence
numbers of consecutive packets sent to the base channel are
numbered consecutively modulo 2^32. The same sequence of PSNs
Luby & Goyal Experimental [Page 27]RFC 3738 WEBRC Building Block April 2004
are used for each wave channel in each cycle. The sequence
numbers of consecutive packets sent to a wave channel are
numbered consecutively modulo 2^32 within each cycle, ending
with the last packet sent to the channel before the channel
goes quiescent with PSN = 2^32-1.
6. Requirements From Other Building Blocks
As described in RFC 3048 [4], WEBRC is a building block that is
intended to be used, in conjunction with other building blocks, to
help specify a protocol instantiation.
WEBRC does not provide higher level session support, e.g., how
receivers obtain the necessary session description and how the
receivers demultiplex received packets based on their session. There
is support provided by other building blocks that can be used in
conjunction with WEBRC to provide some of this support. For example,
LCT [12] can provide some of the higher level in-band session support
that may be needed by receivers, and the WEBRC Congestion Control
Information (CCI) required in each packet can be carried in the CCI
field of the LCT header [12].
WEBRC does not provide any type of reliability, and in particular
does not provide support for retransmission of loss packets.
Reliability can be added by independent means, such as by the use of
FEC codes as described in [13] and specified in the FEC building
block [14].
7. Security Considerations
WEBRC can be subject to denial-of-service attacks by attackers that
try to confuse the congestion control mechanism for receivers by
injecting forged packets into the multicast stream. This attack most
adversely affects network elements and receivers downstream of the
attack, and much less significantly the rest of the network and other
receivers. Because of this and because of the potential attacks due
to the use of FEC described above, it is RECOMMENDED that Reverse
Path Forwarding checks be enabled in all network routers and switches
along the path from the sender to receivers to limit the possibility
of a bad agent injecting forged packets into the multicast tree data
path.
It is also RECOMMENDED that packet authentication be used to
authenticate each packet immediately upon receipt before the receiver
performs any WEBRC actions based upon its receipt. Unfortunately,
there are currently no practical multicast packet authentication
schemes that offer instant packet authentication upon receipt.
However, TESLA [17] can be used to authenticate each packet a few
Luby & Goyal Experimental [Page 28]RFC 3738 WEBRC Building Block April 2004
seconds after receipt. Thus, TESLA could be used in conjunction with
WEBRC to authenticate packets and for example terminate the session
upon detection of a forged packet. However, it is RECOMMENDED that
the normal WEBRC receiver responses to received packets occur
immediately -- not delayed by the TESLA authentication process. This
is because the overall WEBRC performance would be greatly degraded if
the receiver delayed its WEBRC response to packet receipt for several
seconds.
A receiver with an incorrect or corrupted implementation of WEBRC may
affect health of the network in the path between the sender and the
receiver, and may also affect the reception rates of other receivers
joined to the session. It is therefore RECOMMENDED that receivers be
required to identify themselves as legitimate before they receive the
session description needed to join the session.
Another vulnerability of WEBRC is the potential of receivers
obtaining an incorrect session description for the session. The
consequences of this could be that legitimate receivers with the
wrong session description are unable to correctly receive the session
content, or that receivers inadvertently try to receive at a much
higher rate than they are capable of, thereby disrupting traffic in
portions of the network. To avoid these problems, it is RECOMMENDED
that measures be taken to prevent receivers from accepting incorrect
session descriptions, e.g., by using source authentication to ensure
that receivers only accept legitimate session descriptions from
authorized senders.
8. References
8.1. Normative References
[1] Bradner, S., "Key words for use in RFCs to Indicate Requirement
Levels", BCP 14, RFC 2119, March 1997.
[2] Kermode, R. and L. Vicisano, "Author Guidelines for Reliable
Multicast Transport (RMT) Building Blocks and Protocol
Instantiation documents", RFC 3269, April 2002.
[3] Mankin, A., Romanow, A., Bradner, S. and V. Paxson, "IETF
Criteria for Evaluating Reliable Multicast Transport and
Application Protocols", RFC 2357, June 1998.
[4] Whetten, B., Vicisano, L., Kermode, R., Handley, M., Floyd, S.
and M. Luby, "Reliable Multicast Transport Building Blocks for
One-to-Many Bulk-Data Transfer", RFC 3048, January 2001.
Luby & Goyal Experimental [Page 29]RFC 3738 WEBRC Building Block April 2004
8.2. Informative References
[5] Byers, J., Horn, G., Luby, M., Mitzenmacher, M. and W. Shaver.
"FLID-DL: Congestion control for layered multicast," IEEE J. on
Selected Areas in Communications, Special Issue on Network
Support for Multicast Communication, Vol. 20, No. 8, October
2002, pp. 1558-1570.
[6] Dagum, P., Karp, R., Luby, M. and S. Ross, "An optimal algorithm
for Monte Carlo estimation," SIAM J. Comput., 29(5):1484-1496,
April 2000.
[7] Deering, S., "Host Extensions for IP Multicasting", STD 5, RFC
1112, August 1989.
[8] Goyal, V., "On WEBRC Wave Design and Server Implementation",
Digital Fountain Technical Report no. DF2002-09-001, September
2002, available at http://www.digitalfountain.com/technology/.
[9] Handley, M., Floyd, S., Padhye, J. and J. Widmer, "TCP Friendly
Rate Control (TFRC): Protocol Specification", RFC 3448, January
2003.
[10] Holbrook, H., "A Channel Model for Multicast", Ph.D.
Dissertation, Stanford University, Department of Computer
Science, Stanford, California, August 2001.
[11] Luby, M., Gemmell, J., Vicisano, L., Rizzo, L. and J. Crowcroft,
"Asynchronous Layered Coding (ALC) Protocol Instantiation", RFC
3450, December 2002.
[12] Luby, M., Gemmell, J., Vicisano, L., Rizzo, L., Handley, M. and
J. Crowcroft, "Layered Coding Transport (LCT) Building Block",
RFC 3451, December 2002.
[13] Luby, M., Vicisano, L., Gemmell, J., Rizzo, L., Handley, M. and
J. Crowcroft, "The Use of Forward Error Correction (FEC) in
Reliable Multicast", RFC 3453, December 2002.
[14] Luby, M., Vicisano, L., Gemmell, J., Rizzo, L., Handley, M. and
J. Crowcroft, "Forward Error Correction (FEC) Building Block",
RFC 3452, December 2002.
[15] Luby, M. and V. Goyal, "Wave and Equation Based Rate Control
Using Multicast Round Trip Time: Extended Report", Digital
Fountain Technical Report no. DF2002-07-001, September 2002,
available at http://www.digitalfountain.com/technology/.
Luby & Goyal Experimental [Page 30]RFC 3738 WEBRC Building Block April 2004
[16] Luby, M., Goyal, V., Skaria, S. and G. Horn, "Wave and Equation
Based Rate Control Using Multicast Round Trip Time", Proc. ACM
SIGCOMM 2002, Pittsburgh, PA, August 2002, pp. 191-214.
[17] Perrig, A., Canetti, R., Song, D. and J. Tygar, "Efficient and
Secure Source Authentication for Multicast", Network and
Distributed System Security Symposium, NDSS 2001, pp. 35-46,
February 2001.
[18] Vicisano, L., Rizzo, L. and J. Crowcroft, "TCP-like Congestion
Control for Layered Multicast Data Transfer", Proc. IEEE Infocom
'98, San Francisco, CA, March-April 1998, pp. 996-1003.
9. Authors' Addresses
Michael Luby
Digital Fountain
39141 Civic Center Drive, Suite 300
Fremont, CA, USA, 94538
EMail: luby@digitalfountain.com
Vivek K Goyal
Massachusetts Institute of Technology
Room 36-690
77 Massachusetts Avenue
Cambridge, MA, USA, 02139
EMail: v.goyal@ieee.org
Luby & Goyal Experimental [Page 31]RFC 3738 WEBRC Building Block April 2004
10. Full Copyright Statement
Copyright (C) The Internet Society (2004). This document is subject
to the rights, licenses and restrictions contained in BCP 78 and
except as set forth therein, the authors retain all their rights.
This document and the information contained herein are provided on an
"AS IS" basis and THE CONTRIBUTOR, THE ORGANIZATION HE/SHE REPRESENTS
OR IS SPONSORED BY (IF ANY), THE INTERNET SOCIETY AND THE INTERNET
ENGINEERING TASK FORCE DISCLAIM ALL WARRANTIES, EXPRESS OR IMPLIED,
INCLUDING BUT NOT LIMITED TO ANY WARRANTY THAT THE USE OF THE
INFORMATION HEREIN WILL NOT INFRINGE ANY RIGHTS OR ANY IMPLIED
WARRANTIES OF MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE.
Intellectual Property
The IETF takes no position regarding the validity or scope of any
Intellectual Property Rights or other rights that might be claimed to
pertain to the implementation or use of the technology described in
this document or the extent to which any license under such rights
might or might not be available; nor does it represent that it has
made any independent effort to identify any such rights. Information
on the procedures with respect to rights in RFC documents can be
found in BCP 78 and BCP 79.
Copies of IPR disclosures made to the IETF Secretariat and any
assurances of licenses to be made available, or the result of an
attempt made to obtain a general license or permission for the use of
such proprietary rights by implementers or users of this
specification can be obtained from the IETF on-line IPR repository at
http://www.ietf.org/ipr.
The IETF invites any interested party to bring to its attention any
copyrights, patents or patent applications, or other proprietary
rights that may cover technology that may be required to implement
this standard. Please address the information to the IETF at ietf-
ipr@ietf.org.
Acknowledgement
Funding for the RFC Editor function is currently provided by the
Internet Society.
Luby & Goyal Experimental [Page 32]
