Network Working Group S. Floyd, Ed.
Request for Comments: 5166 March 2008
Category: Informational
Metrics for the Evaluation of Congestion Control Mechanisms
Status of This Memo
This memo provides information for the Internet community. It does
not specify an Internet standard of any kind. Distribution of this
memo is unlimited.
IESG Note
This document is not an IETF Internet Standard. It represents the
individual opinion(s) of one or more members of the TMRG Research
Group of the Internet Research Task Force. It may be considered for
standardization by the IETF or adoption as an IRTF Research Group
consensus document in the future.
Abstract
This document discusses the metrics to be considered in an evaluation
of new or modified congestion control mechanisms for the Internet.
These include metrics for the evaluation of new transport protocols,
of proposed modifications to TCP, of application-level congestion
control, and of Active Queue Management (AQM) mechanisms in the
router. This document is the first in a series of documents aimed at
improving the models that we use in the evaluation of transport
protocols.
This document is a product of the Transport Modeling Research Group
(TMRG), and has received detailed feedback from many members of the
Research Group (RG). As the document tries to make clear, there is
not necessarily a consensus within the research community (or the
IETF community, the vendor community, the operations community, or
any other community) about the metrics that congestion control
mechanisms should be designed to optimize, in terms of trade-offs
between throughput and delay, fairness between competing flows, and
the like. However, we believe that there is a clear consensus that
congestion control mechanisms should be evaluated in terms of trade-
offs between a range of metrics, rather than in terms of optimizing
for a single metric.
Table of Contents
1. Introduction ....................................................2
2. Metrics .........................................................3
2.1. Throughput, Delay, and Loss Rates ..........................4
2.1.1. Throughput ..........................................5
2.1.2. Delay ...............................................6
2.1.3. Packet Loss Rates ...................................6
2.2. Response Times and Minimizing Oscillations .................7
2.2.1. Response to Changes .................................7
2.2.2. Minimizing Oscillations .............................8
2.3. Fairness and Convergence ...................................9
2.3.1. Metrics for Fairness between Flows .................10
2.3.2. Metrics for Fairness between Flows with
Different Resource Requirements ....................10
2.3.3. Comments on Fairness ...............................12
2.4. Robustness for Challenging Environments ...................13
2.5. Robustness to Failures and to Misbehaving Users ...........14
2.6. Deployability .............................................14
2.7. Metrics for Specific Types of Transport ...................15
2.8. User-Based Metrics ........................................15
3. Metrics in the IP Performance Metrics (IPPM) Working Group .....15
4. Comments on Methodology ........................................16
5. Security Considerations ........................................16
6. Acknowledgements ...............................................16
7. Informative References .........................................17
1. Introduction
As a step towards improving our methodologies for evaluating
congestion control mechanisms, in this document we discuss some of
the metrics to be considered. We also consider the relationship
between metrics, e.g., the well-known trade-off between throughput
and delay. This document doesn't attempt to specify every metric
that a study might consider; for example, there are domain-specific
metrics that researchers might consider that are over and above the
metrics laid out here.
We consider metrics for aggregate traffic (taking into account the
effect of flows on competing traffic in the network) as well as the
heterogeneous goals of different applications or transport protocols
(e.g., of high throughput for bulk data transfer, and of low delay
for interactive voice or video). Different transport protocols or
AQM mechanisms might have goals of optimizing different sets of
metrics, with one transport protocol optimized for per-flow
throughput and another optimized for robustness over wireless links,
and with different degrees of attention to fairness with competing
traffic. We hope this document will be used as a step in evaluating
proposed congestion control mechanisms for a wide range of metrics,
for example, noting that Mechanism X is good at optimizing Metric A,
but pays the price with poor performance for Metric B. The goal
would be to have a broad view of both the strengths and weaknesses of
newly proposed congestion control mechanisms.
Subsequent documents are planned to present sets of simulation and
testbed scenarios for the evaluation of transport protocols and of
congestion control mechanisms, based on the best current practice of
the research community. These are not intended to be complete or
final benchmark test suites, but simply to be one step of many to be
used by researchers in evaluating congestion control mechanisms.
Subsequent documents are also planned on the methodologies in using
these sets of scenarios.
This document is a product of the Transport Modeling Research Group
(TMRG), and has received detailed feedback from many members of the
Research Group (RG). As the document tries to make clear, there is
not necessarily a consensus within the research community (or the
IETF community, the vendor community, the operations community, or
any other community) about the metrics that congestion control
mechanisms should be designed to optimize, in terms of trade-offs
between throughput and delay, fairness between competing flows, and
the like. However, we believe that there is a clear consensus that
congestion control mechanisms should be evaluated in terms of
trade-offs between a range of metrics, rather than in terms of
optimizing for a single metric.
2. Metrics
The metrics that we discuss are the following:
o Throughput;
o Delay;
o Packet loss rates;
o Response to sudden changes or to transient events;
o Minimizing oscillations in throughput or in delay;
o Fairness and convergence times;
o Robustness for challenging environments;
o Robustness to failures and to misbehaving users;
o Deployability;
o Metrics for specific types of transport;
o User-based metrics.
We consider each of these below. Many of the metrics have
network-based, flow-based, and user-based interpretations. For
example, network-based metrics can consider aggregate bandwidth and
aggregate drop rates, flow-based metrics can consider end-to-end
transfer times for file transfers or end-to-end delay and packet drop
rates for interactive traffic, and user-based metrics can consider
user wait time or user satisfaction with the multimedia experience.
Our main goal in this document is to explain the set of metrics that
can be relevant, and not to legislate on the most appropriate
methodology for using each general metric.
For some of the metrics, such as fairness, there is not a clear
agreement in the network community about the desired goals. In these
cases, the document attempts to present the range of approaches.
2.1. Throughput, Delay, and Loss Rates
Because of the clear trade-offs between throughput, delay, and loss
rates, it can be useful to consider these three metrics together.
The trade-offs are most clear in terms of queue management at the
router; is the queue configured to maximize aggregate throughput, to
minimize delay and loss rates, or some compromise between the two?
An alternative would be to consider a separate metric such as power,
defined in this context as throughput over delay, that combines
throughput and delay. However, we do not propose in this document a
clear target in terms of the trade-offs between throughput and delay;
we are simply proposing that the evaluation of transport protocols
include an exploration of the competing metrics.
Using flow-based metrics instead of router-based metrics, the
relationship between per-flow throughput, delay, and loss rates can
often be given by the transport protocol itself. For example, in
TCP, the sending rate s in packets per second is given as:
s = 1.2/(RTT*sqrt(p)),
for the round-trip time RTT and loss rate p [FF99].
2.1.1. Throughput
Throughput can be measured as a router-based metric of aggregate link
utilization, as a flow-based metric of per-connection transfer times,
and as user-based metrics of utility functions or user wait times.
It is a clear goal of most congestion control mechanisms to maximize
throughput, subject to application demand and to the constraints of
the other metrics.
Throughput is sometimes distinguished from goodput, where throughput
is the link utilization or flow rate in bytes per second; goodput,
also measured in bytes per second, is the subset of throughput
consisting of useful traffic. That is, 'goodput' excludes duplicate
packets, packets that will be dropped downstream, packet fragments or
ATM cells that are dropped at the receiver because they can't be
re-assembled into complete packets, and the like. In general, this
document doesn't distinguish between throughput and goodput, and uses
the general term "throughput".
We note that maximizing throughput is of concern in a wide range of
environments, from highly-congested networks to under-utilized ones,
and from long-lived flows to very short ones. As an example,
throughput has been used as one of the metrics for evaluating
Quick-Start, a proposal to allow flows to start up faster than
slow-start, where throughput has been evaluated in terms of the
transfer times for connections with a range of transfer sizes
[RFC4782] [SAF06].
In some contexts, it might be sufficient to consider the aggregate
throughput or the mean per-flow throughput [DM06], while in other
contexts it might be necessary to consider the distribution of
per-flow throughput. Some researchers evaluate transport protocols
in terms of maximizing the aggregate user utility, where a user's
utility is generally defined as a function of the user's throughput
[KMT98].
Individual applications can have application-specific needs in terms
of throughput. For example, real-time video traffic can have highly
variable bandwidth demands; Voice over IP (VoIP) traffic is sensitive
to the amount of bandwidth received immediately after idle periods.
Thus, user metrics for throughput can be more complex than simply the
per-connection transfer time.
2.1.2. Delay
Like throughput, delay can be measured as a router-based metric of
queueing delay over time, or as a flow-based metric in terms of
per-packet transfer times. Per-packet delay can also include delay
at the sender waiting for the transport protocol to send the packet.
For reliable transfer, the per-packet transfer time seen by the
application includes the possible delay of retransmitting a lost
packet.
Users of bulk data transfer applications might care about per-packet
transfer times only in so far as they affect the per-connection
transfer time. On the other end of the spectrum, for users of
streaming media, per-packet delay can be a significant concern. Note
that in some cases the average delay might not capture the metric of
interest to the users; for example, some users might care about the
worst-case delay, or about the tail of the delay distribution.
Note that queueing delay at a router is shared by all flows at a FIFO
(First-In First-Out) queue. Thus, the router-based queueing delay
induced by bulk data transfers can be important even if the bulk data
transfers themselves are not concerned with per-packet transfer
times.
2.1.3. Packet Loss Rates
Packet loss rates can be measured as a network-based or as a
flow-based metric.
When evaluating the effect of packet losses or ECN marks (Explicit
Congestion Notification) [RFC3168] on the performance of a congestion
control mechanism for an individual flow, researchers often use both
the packet loss/mark rate for that connection and the congestion
event rate (also called the loss event rate), where a congestion
event or loss event consists of one or more lost or marked packets in
one round-trip time [RFC3448].
Some users might care about the packet loss rate only in so far as it
affects per-connection transfer times, while other users might care
about packet loss rates directly. RFC 3611, RTP Control Protocol
Extended Reports, describes a VoIP performance-reporting standard
called RTP Control Protocol Extended Reports (RTCP XR), which
includes a set of burst metrics. In RFC 3611, a burst is defined as
the maximal sequence starting and ending with a lost packet, and not
including a sequence of Gmin or more packets that are not lost
[RFC3611]. The burst metrics in RFC 3611 consist of the burst
density (the fraction of packets in bursts), gap density (the
fraction of packets in the gaps between bursts), burst duration (the
mean duration of bursts in seconds), and gap duration (the mean
duration of gaps in seconds). RFC 3357 derives metrics for "loss
distance" and "loss period", along with statistics that capture loss
patterns experienced by packet streams on the Internet [RFC3357].
In some cases, it is useful to distinguish between packets dropped at
routers due to congestion and packets lost in the network due to
corruption.
One network-related reason to avoid high steady-state packet loss
rates is to avoid congestion collapse in environments containing
paths with multiple congested links. In such environments, high
packet loss rates could result in congested links wasting scarce
bandwidth by carrying packets that will only be dropped downstream
before being delivered to the receiver [RFC2914]. We also note that
in some cases, the retransmit rate can be high, and the goodput
correspondingly low, even with a low packet drop rate [AEO03].
2.2. Response Times and Minimizing Oscillations
In this section, we consider response times and oscillations
together, as there are well-known trade-offs between improving
response times and minimizing oscillations. In addition, the
scenarios that illustrate the dangers of poor response times are
often quite different from the scenarios that illustrate the dangers
of unnecessary oscillations.
2.2.1. Response to Changes
One of the key concerns in the design of congestion control
mechanisms has been the response times to sudden congestion in the
network. On the one hand, congestion control mechanisms should
respond reasonably promptly to sudden congestion from routing or
bandwidth changes or from a burst of competing traffic. At the same
time, congestion control mechanisms should not respond too severely
to transient changes, e.g., to a sudden increase in delay that will
dissipate in less than the connection's round-trip time.
Congestion control mechanisms also have to contend with sudden
changes in the bandwidth-delay product due to mobility. Such
bandwidth-delay product changes are expected to become more frequent
and to have greater impact than path changes today. As a result of
both mobility and of the heterogeneity of wireless access types
(802.11b,a,g, WIMAX, WCDMA, HS-WCDMA, E-GPRS, Bluetooth, etc.), both
the bandwidth and the round-trip delay can change suddenly, sometimes
by several orders of magnitude.
Evaluating the response to sudden or transient changes can be of
particular concern for slowly responding congestion control
mechanisms such as equation-based congestion control [RFC3448] and
AIMD (Additive Increase Multiplicative Decrease) or for related
mechanisms using parameters that make them more slowly-responding
than TCP [BB01] [BBFS01].
In addition to the responsiveness and smoothness of aggregate
traffic, one can consider the trade-offs between responsiveness,
smoothness, and aggressiveness for an individual connection [FHP00]
[YKL01]. In this case, smoothness can be defined by the largest
reduction in the sending rate in one round-trip time, in a
deterministic environment with a packet drop exactly every 1/p
packets. The responsiveness is defined as the number of round-trip
times of sustained congestion required for the sender to halve the
sending rate; aggressiveness is defined as the maximum increase in
the sending rate in one round-trip time, in packets per second, in
the absence of congestion. This aggressiveness can be a function of
the mode of the transport protocol; for TCP, the aggressiveness of
slow-start is quite different from the aggressiveness of congestion
avoidance mode.
2.2.2. Minimizing Oscillations
One goal is that of stability, in terms of minimizing oscillations of
queueing delay or of throughput. In practice, stability is
frequently associated with rate fluctuations or variance. Rate
variations can result in fluctuations in router queue size and
therefore of queue overflows. These queue overflows can cause loss
synchronizations across coexisting flows and periodic
under-utilization of link capacity, both of which are considered to
be general signs of network instability. Thus, measuring the rate
variations of flows is often used to measure the stability of
transport protocols. To measure rate variations, [JWL04], [RX05],
and [FHPW00] use the coefficient of variation (CoV) of per-flow
transmission rates, and [WCL05] suggests the use of standard
deviations of per-flow rates. Since rate variations are a function
of time scales, it makes sense to measure these rate variations over
various time scales.
Measuring per-flow rate variations, however, is only one aspect of
transport protocol stability. A realistic experimental setting
always involves multiple flows of the transport protocol being
observed, along with a significant amount of cross traffic, with
rates varying over time on both the forward and reverse paths. As a
congestion control protocol must adapt its rate to the varying rates
of competing traffic, just measuring the per-flow statistics of a
subset of the traffic could be misleading because it measures the
rate fluctuations due in part to the adaptation to competing traffic
on the path. Thus, per-flow statistics are most meaningful if they
are accompanied by the statistics measured at the network level. As
a complementary metric to the per-flow statistics, [HKLRX06] uses
measurements of the rate variations of the aggregate flows observed
in bottleneck routers over various time scales.
Minimizing oscillations in queueing delay or throughput has related
per-flow metrics of minimizing jitter in round-trip times and loss
rates.
An orthogonal goal for some congestion control mechanisms, e.g., for
equation-based congestion control, is to minimize the oscillations in
the sending rate for an individual connection, given an environment
with a fixed, steady-state packet drop rate. (As is well known, TCP
congestion control is characterized by a pronounced oscillation in
the sending rate, with the sender halving the sending rate in
response to congestion.) One metric for the level of oscillations is
the smoothness metric given in Section 2.2.1 above.
As discussed in [FK07], the synchronization of loss events can also
affect convergence times, throughput, and delay.
2.3. Fairness and Convergence
Another set of metrics is that of fairness and convergence times.
Fairness can be considered between flows of the same protocol and
between flows using different protocols (e.g., TCP-friendliness,
referring to fairness between TCP and a new transport protocol).
Fairness can also be considered between sessions, between users, or
between other entities.
There are a number of different fairness measures. These include
max-min fairness [HG86], proportional fairness [KMT98] [K01], the
fairness index proposed in [JCH84], and the product measure, a
variant of network power [BJ81].
2.3.1. Metrics for Fairness between Flows
This section discusses fairness metrics that consider the fairness
between flows, but that don't take into account different
characteristics of flows (e.g., the number of links in the path or
the round-trip time). For the discussion of fairness metrics, let
x_i be the throughput for the i-th connection.
Jain's fairness index: The fairness index in [JCH84] is:
(( sum_i x_i )^2) / (n * sum_i ( (x_i)^2 )),
where there are n users. This fairness index ranges from 0 to 1, and
it is maximum when all users receive the same allocation. This index
is k/n when k users equally share the resource, and the other n-k
users receive zero allocation.
The product measure: The product measure:
product_i x_i ,
the product of the throughput of the individual connections, is also
used as a measure of fairness. (In some contexts x_i is taken as the
power of the i-th connection, and the product measure is referred to
as network power.) The product measure is particularly sensitive to
segregation; the product measure is zero if any connection receives
zero throughput. In [MS91], it is shown that for a network with many
connections and one shared gateway, the product measure is maximized
when all connections receive the same throughput.
Epsilon-fairness: A rate allocation is defined as epsilon-fair if
(min_i x_i) / (max_i x_i) >= 1 - epsilon.
Epsilon-fairness measures the worst-case ratio between any two
throughput rates [ZKL04]. Epsilon-fairness is related to max-min
fairness, defined later in this document.
2.3.2. Metrics for Fairness between Flows with Different Resource
Requirements
This section discusses metrics for fairness between flows with
different resource requirements, that is, with different utility
functions, round-trip times, or number of links on the path. Many of
these metrics can be described as solutions to utility maximization
problems [K01]; the total utility quantifies both the fairness and
the throughput.
Max-min fairness: In order to satisfy the max-min fairness criteria,
the smallest throughput rate must be as large as possible. Given
this condition, the next-smallest throughput rate must be as large as
possible, and so on. Thus, the max-min fairness gives absolute
priority to the smallest flows. (Max-min fairness can be explained
by the progressive filling algorithm, where all flow rates start at
zero, and the rates all grow at the same pace. Each flow rate stops
growing only when one or more links on the path reach link capacity.)
Proportional fairness: In contrast, a feasible allocation, x, is
defined as proportionally fair if, for any other feasible allocation
x*, the aggregate of proportional changes is zero or negative:
sum_i ( (x*_i - x_i)/x_i ) <= 0.
"This criterion favours smaller flows, but less emphatically than
max-min fairness" [K01]. (Using the language of utility functions,
proportional fairness can be achieved by using logarithmic utility
functions, and maximizing the sum of the per-flow utility functions;
see [KMT98] for a fuller explanation.)
Minimum potential delay fairness: Minimum potential delay fairness
has been shown to model TCP [KS03], and is a compromise between
max-min fairness and proportional fairness. An allocation, x, is
defined as having minimum potential delay fairness if:
sum_i (1/x_i)
is smaller than for any other feasible allocation. That is, it would
minimize the average download time if each flow was an equal-sized
file.
In [CRM05], Colussi, et al. propose a new definition of TCP fairness,
that "a set of TCP fair flows do not cause more congestion than a set
of TCP flows would cause", where congestion is defined in terms of
queueing delay, queueing delay variation, the congestion event rate
[e.g., loss event rate], and the packet loss rate.
Chiu and Tan in [CT06] argue for redefining the notion of fairness
when studying traffic controls for inelastic traffic, proposing that
inelastic flows adopt other traffic controls such as admission
control.
The usefulness of flow-rate fairness has been challenged in [B07] by
Briscoe, and defended in [FA08] by Floyd and Allman. In [B07],
Briscoe argues that flow-rate fairness should not be a desired goal,
and that instead "we should judge fairness mechanisms on how they
share out the 'cost' of each user's actions on others". Floyd and
Allman in [FA08] argue that the current system based on TCP
congestion control and flow-rate fairness has been useful in the real
world, posing minimal demands on network and economic infrastructure
and enabling users to get a share of the network resources.
2.3.3. Comments on Fairness
Trade-offs between fairness and throughput: The fairness measures in
the section above generally measure both fairness and throughput,
giving different weights to each. Potential trade-offs between
fairness and throughput are also discussed by Tang, et al. in
[TWL06], for a framework where max-min fairness is defined as the
most fair. In particular, [TWL06] shows that in some topologies,
throughput is proportional to fairness, while in other topologies,
throughput is inversely proportional to fairness.
Fairness and the number of congested links: Some of these fairness
metrics are discussed in more detail in [F91]. We note that there is
not a clear consensus for the fairness goals, in particular for
fairness between flows that traverse different numbers of congested
links [F91]. Utility maximization provides one framework for
describing this trade-off in fairness.
Fairness and round-trip times: One goal cited in a number of new
transport protocols has been that of fairness between flows with
different round-trip times [KHR02] [XHR04]. We note that there is
not a consensus in the networking community about the desirability of
this goal, or about the implications and interactions between this
goal and other metrics [FJ92] (Section 3.3). One common argument
against the goal of fairness between flows with different round-trip
times has been that flows with long round-trip times consume more
resources; this aspect is covered by the previous paragraph.
Researchers have also noted the difference between the RTT-unfairness
of standard TCP, and the greater RTT-unfairness of some proposed
modifications to TCP [LLS05].
Fairness and packet size: One fairness issue is that of the relative
fairness for flows with different packet sizes. Many file transfer
applications will use the maximum packet size possible; in contrast,
low-bandwidth VoIP flows are likely to send small packets, sending a
new packet every 10 to 40 ms., to limit delay. Should a small-packet
VoIP connection receive the same sending rate in *bytes* per second
as a large-packet TCP connection in the same environment, or should
it receive the same sending rate in *packets* per second? This
fairness issue has been discussed in more detail in [RFC3714], with
[RFC4828] also describing the ways that packet size can affect the
packet drop rate experienced by a flow.
Convergence times: Convergence times concern the time for convergence
to fairness between an existing flow and a newly starting one, and
are a special concern for environments with high-bandwidth long-delay
flows. Convergence times also concern the time for convergence to
fairness after a sudden change such as a change in the network path,
the competing cross-traffic, or the characteristics of a wireless
link. As with fairness, convergence times can matter both between
flows of the same protocol, and between flows using different
protocols [SLFK03]. One metric used for convergence times is the
delta-fair convergence time, defined as the time taken for two flows
with the same round-trip time to go from shares of 100/101-th and
1/101-th of the link bandwidth, to having close to fair sharing with
shares of (1+delta)/2 and (1-delta)/2 of the link bandwidth [BBFS01].
A similar metric for convergence times measures the convergence time
as the number of round-trip times for two flows to reach epsilon-
fairness, when starting from a maximally-unfair state [ZKL04].
2.4. Robustness for Challenging Environments
While congestion control mechanisms are generally evaluated first
over environments with static routing in a network of two-way
point-to-point links, some environments bring up more challenging
problems (e.g., corrupted packets, reordering, variable bandwidth,
and mobility) as well as new metrics to be considered (e.g., energy
consumption).
Robustness for challenging environments: Robustness needs to be
explored for paths with reordering, corruption, variable bandwidth,
asymmetric routing, router configuration changes, mobility, and the
like [GF04]. In general, the Internet architecture has valued
robustness over efficiency, e.g., when there are trade-offs between
robustness and the throughput, delay, and fairness metrics described
above.
Energy consumption: In mobile environments, the energy consumption
for the mobile end-node can be a key metric that is affected by the
transport protocol [TM02].
The goodput ratio: For wireless networks, the goodput ratio can be a
useful metric, where the goodput ratio can be defined as the useful
data delivered to users as a fraction of the total amount of data
transmitted on the network. A high goodput ratio indicates an
efficient use of the radio spectrum and lower interference with other
users.
2.5. Robustness to Failures and to Misbehaving Users
One goal is for congestion control mechanisms to be robust to
misbehaving users, such as receivers that 'lie' to data senders about
the congestion experienced along the path or otherwise attempt to
bypass the congestion control mechanisms of the sender [SCWA99].
Another goal is for congestion control mechanisms to be as robust as
possible to failures, such as failures of routers in using explicit
feedback to end-nodes or failures of end-nodes to follow the
prescribed protocols.
2.6. Deployability
One metric for congestion control mechanisms is their deployability
in the current Internet. Metrics related to deployability include
the ease of failure diagnosis and the overhead in terms of packet
header size or added complexity at end-nodes or routers.
One key aspect of deployability concerns the range of deployment
needed for a new congestion control mechanism. Consider the
following possible deployment requirements:
* Only at the sender (e.g., NewReno in TCP [RFC3782]);
* Only at the receiver (e.g., delayed acknowledgements in TCP);
* Both the sender and receiver (e.g., Selective Acknowledgment
(SACK) TCP [RFC2018]);
* At a single router (e.g., Random Early Detection (RED) [FJ93]);
* All of the routers along the end-to-end path;
* Both end-nodes and all routers along the path (e.g., Explicit
Control Protocol (XCP) [KHR02]).
Some congestion control mechanisms (e.g., XCP [KHR02], Quick-Start
[RFC4782]) may also have deployment issues with IPsec, IP in IP,
MPLS, other tunnels, or with non-router queues such as those in
Ethernet switches.
Another deployability issue concerns the complexity of the code. How
complex is the code required to implement the mechanism in software?
Is floating point math required? How much new state must be kept to
implement the new mechanism, and who holds that state, the routers or
the end-nodes? We note that we don't suggest these questions as ways
to reduce the deployability metric to a single number; we suggest
them as issues that could be considered in evaluating the
deployability of a proposed congestion control mechanism.
2.7. Metrics for Specific Types of Transport
In some cases, modified metrics are needed for evaluating transport
protocols intended for quality-of-service (QoS)-enabled environments
or for below-best-effort traffic [VKD02] [KK03].
2.8. User-Based Metrics
An alternate approach that has been proposed for the evaluation of
congestion control mechanisms would be to evaluate in terms of user
metrics, such as user satisfaction or in terms of
application-specific utility functions. Such an approach would
require the definition of a range of user metrics or of
application-specific utility functions for the range of applications
under consideration (e.g., FTP, HTTP, VoIP).
3. Metrics in the IP Performance Metrics (IPPM) Working Group
The IPPM Working Group [IPPM] was established to define performance
metrics to be used by network operators, end users, or independent
testing groups. The metrics include metrics for connectivity
[RFC2678], delay and loss [RFC2679], [RFC2680], and [RFC2681], delay
variation [RFC3393], loss patterns [RFC3357], packet reordering
[RFC4737], bulk transfer capacity [RFC3148], and link capacity
[RFC5136]. The IPPM documents give concrete, well-defined metrics,
along with a methodology for measuring the metric. The metrics
discussed in this document have a different purpose from the IPPM
metrics; in this document, we are discussing metrics as used in
analysis, simulations, and experiments for the evaluation of
congestion control mechanisms. Further, we are discussing these
metrics in a general sense, rather than looking for specific concrete
definitions for each metric. However, there are many cases where the
metric definitions from IPPM could be useful, for specific issues of
how to measure these metrics in simulations, or in testbeds, and for
providing common definitions for talking about these metrics.
4. Comments on Methodology
The types of scenarios that are used to test specific metrics, and
the range of parameters that it is useful to consider, will be
discussed in separate documents, e.g., along with specific scenarios
for use in evaluating congestion control mechanisms.
We note that it can be important to evaluate metrics over a wide
range of environments, with a range of link bandwidths, congestion
levels, and levels of statistical multiplexing. It is also important
to evaluate congestion control mechanisms in a range of scenarios,
including typical ranges of connection sizes and round-trip times
[FK02]. It is also useful to compare metrics for new or modified
transport protocols with those of the current standards for TCP.
As an example from the literature on evaluating transport protocols,
Li, et al. in "Experimental Evaluation of TCP Protocols for High-
Speed Networks" [LLS05] focus on the performance of TCP in high-
speed networks, and consider metrics for aggregate throughput, loss
rates, fairness (including fairness between flows with different
round-trip times), response times (including convergence times), and
incremental deployment. More general references on methodology
include [J91]. Papers that discuss the range of metrics for
evaluating congestion control include [MTZ04].
5. Security Considerations
Section 2.5 discusses the robustness of congestion control mechanisms
to failures and to misbehaving users. Transport protocols also have
other security concerns that are unrelated to congestion control
mechanisms; these are not discussed in this document.
6. Acknowledgements
Thanks to Lachlan Andrew, Mark Allman, Armando Caro, Dah Ming Chiu,
Eric Coe, Dado Colussi, Wesley Eddy, Aaron Falk, Nelson Fonseca,
Janardhan Iyengar, Doug Leith, Sara Landstrom, Tony Li, Saverio
Mascolo, Sean Moore, Injong Rhee, David Ros, Juergen Schoenwaelder,
Andras Veres, Michael Welzl, and Damon Wischik, and members of the
Transport Modeling Research Group for feedback and contributions.
7. Informative References
[AEO03] M. Allman, W. Eddy, and S. Ostermann, Estimating Loss Rates
With TCP, ACM Performance Evaluation Review, 31(3),
December 2003.
[BB01] D. Bansal and H. Balakrishnan, Binomial Congestion Control
Algorithms, IEEE Infocom, April 2001.
[BBFS01] D. Bansal, H. Balakrishnan, S. Floyd, and S. Shenker,
Dynamic Behavior of Slowly-Responsive Congestion Control
Algorithms, SIGCOMM 2001.
[BJ81] K. Bharath-Kumar and J. Jeffrey, A New Approach to
Performance-Oriented Flow Control, IEEE Transactions on
Communications, Vol.29 N.4, April 1981.
[B07] B. Briscoe, "Flow Rate Fairness: Dismantling a Religion",
Computer Communications Review, V.37 N.2, April 2007.
[CRM05] D. Colussi, A New Approach to TCP-Fairness, Report C-2005-
49, University of Helsinki, Finland, 2005.
[CT06] D. Chiu and A. Tam, Redefining Fairness in the Study of
TCP-friendly Traffic Controls, Technical Report, 2006.
[DM06] N. Dukkipati and N. McKeown, Why Flow-Completion Time is
the Right Metric for Congestion Control, ACM SIGCOMM,
January 2006.
[F91] S. Floyd, Connections with Multiple Congested Gateways in
Packet-Switched Networks Part 1: One-way Traffic, Computer
Communication Review, Vol.21 No.5, October 1991, p. 30-47.
[FA08] S. Floyd and M. Allman, Comments on the Usefulness of
Simple Best-Effort Traffic, Work in Progress, January 2007.
[FF99] Floyd, S., and Fall, K., "Promoting the Use of End-to-End
Congestion Control in the Internet", IEEE/ACM Transactions
on Networking, August 1999.
[FHP00] S. Floyd, M. Handley, and J. Padhye, A Comparison of
Equation-Based and AIMD Congestion Control, May 2000. URL
http://www.icir.org/tfrc/.
[FHPW00] S. Floyd, M. Handley, J. Padhye, and J. Widmer, Equation-
Based Congestion Control for Unicast Applications, SIGCOMM
2000, August 2000.
[FJ92] S. Floyd and V. Jacobson, On Traffic Phase Effects in
Packet-Switched Gateways, Internetworking: Research and
Experience, V.3 N.3, September 1992, p.115-156.
[FJ93] S. Floyd and V. Jacobson, Random Early Detection gateways
for Congestion Avoidance, IEEE/ACM Transactions on
Networking, V.1 N.4, August 1993,
[FK02] S. Floyd and E. Kohler, Internet Research Needs Better
Models, Hotnets-I. October 2002.
[FK07] S. Floyd and E. Kohler, "Tools for the Evaluation of
Simulation and Testbed Scenarios", Work in Progress,
February 2008.
[GF04] A. Gurtov and S. Floyd, Modeling Wireless Links for
Transport Protocols, ACM CCR, 34(2):85-96, April 2004.
[HKLRX06] S. Ha, Y. Kim, L. Le, I. Rhee, and L. Xu, A Step Toward
Realistic Evaluation of High-speed TCP Protocols, technical
report, North Carolina State University, January 2006.
[HG86] E. Hahne and R. Gallager, Round Robin Scheduling for Fair
Flow Control in Data Communications Networks, IEEE
International Conference on Communications, June 1986.
[IPPM] IP Performance Metrics (IPPM) Working Group, URL
http://www.ietf.org/html.charters/ippm-charter.html.
[J91] R. Jain, The Art of Computer Systems Performance Analysis:
Techniques for Experimental Design, Measurement,
Simulation, and Modeling, John Wiley & Sons, 1991.
[JCH84] R. Jain, D.M. Chiu, and W. Hawe, A Quantitative Measure of
Fairness and Discrimination for Resource Allocation in
Shared Systems, DEC TR-301, Littleton, MA: Digital
Equipment Corporation, 1984.
[JWL04] C. Jin, D. Wei, and S. Low, FAST TCP: Motivation,
Architecture, Algorithms, Performance, IEEE INFOCOM, March
2004.
[K01] F. Kelly, Mathematical Modelling of the Internet,
"Mathematics Unlimited - 2001 and Beyond" (Editors B.
Engquist and W. Schmid), Springer-Verlag, Berlin, pp.
685-702, 2001.
[KHR02] D. Katabi, M. Handley, and C. Rohrs, Congestion Control for
High Bandwidth-Delay Product Networks, ACM Sigcomm, 2002.
[KK03] A. Kuzmanovic and E. W. Knightly, TCP-LP: A Distributed
Algorithm for Low Priority Data Transfer, IEEE INFOCOM
2003, April 2003.
[KMT98] F. Kelly, A. Maulloo and D. Tan, Rate Control in
Communication Networks: Shadow Prices, Proportional
Fairness and Stability. Journal of the Operational
Research Society 49, pp. 237-252, 1998.
[KS03] S. Kunniyur and R. Srikant, End-to-end Congestion Control
Schemes: Utility Functions, Random Losses and ECN Marks,
IEEE/ACM Transactions on Networking, 11(5):689-702, October
2003.
[LLS05] Y-T. Li, D. Leith, and R. Shorten, Experimental Evaluation
of TCP Protocols for High-Speed Networks, Hamilton
Institute, 2005. URL
http://www.hamilton.ie/net/eval/results_HI2005.pdf.
[MS91] D. Mitra and J. Seery, Dynamic Adaptive Windows for High
Speed Data Networks with Multiple Paths and Propagation
Delays, INFOCOM '91, pp 39-48.
[MTZ04] L. Mamatas, V. Tsaoussidis, and C. Zhang, Approaches to
Congestion Control in Packet Networks, 2004.
[RFC2018] Mathis, M., Mahdavi, J., Floyd, S., and A. Romanow, "TCP
Selective Acknowledgment Options", RFC 2018, October 1996.
[RFC2680] Almes, G., Kalidindi, S., and M. Zekauskas, "A One-way
Packet Loss Metric for IPPM", RFC 2680, September 1999.
[RFC2678] Mahdavi, J. and V. Paxson, "IPPM Metrics for Measuring
Connectivity", RFC 2678, September 1999.
[RFC2679] Almes, G., Kalidindi, S., and M. Zekauskas, "A One-way
Delay Metric for IPPM", RFC 2679, September 1999.
[RFC2681] Almes, G., Kalidindi, S., and M. Zekauskas, "A Round-trip
Delay Metric for IPPM", RFC 2681, September 1999.
[RFC2914] Floyd, S., "Congestion Control Principles", BCP 41, RFC
2914, September 2000.
[RFC3148] Mathis, M. and M. Allman, "A Framework for Defining
Empirical Bulk Transfer Capacity Metrics", RFC 3148, July
2001.
[RFC3168] Ramakrishnan, K., Floyd, S., and D. Black, "The Addition of
Explicit Congestion Notification (ECN) to IP", RFC 3168,
September 2001.
[RFC3357] Koodli, R. and R. Ravikanth, "One-way Loss Pattern Sample
Metrics", RFC 3357, August 2002.
[RFC3393] Demichelis, C. and P. Chimento, "IP Packet Delay Variation
Metric for IP Performance Metrics (IPPM)", RFC 3393,
November 2002.
[RFC3448] Handley, M., Floyd, S., Padhye, J., and J. Widmer, "TCP
Friendly Rate Control (TFRC): Protocol Specification", RFC
3448, January 2003.
[RFC3611] Friedman, T., Ed., Caceres, R., Ed., and A. Clark, Ed.,
"RTP Control Protocol Extended Reports (RTCP XR)", RFC
3611, November 2003.
[RFC3714] Floyd, S., Ed., and J. Kempf, Ed., "IAB Concerns Regarding
Congestion Control for Voice Traffic in the Internet", RFC
3714, March 2004.
[RFC3782] Floyd, S., Henderson, T., and A. Gurtov, "The NewReno
Modification to TCP's Fast Recovery Algorithm", RFC 3782,
April 2004.
[RFC4737] Morton, A., Ciavattone, L., Ramachandran, G., Shalunov, S.,
and J. Perser, "Packet Reordering Metrics", RFC 4737,
November 2006.
[RFC4782] Floyd, S., Allman, M., Jain, A., and P. Sarolahti, "Quick-
Start for TCP and IP", RFC 4782, January 2007.
[RFC4828] Floyd, S. and E. Kohler, "TCP Friendly Rate Control (TFRC):
The Small-Packet (SP) Variant", RFC 4828, April 2007.
[RFC5136] Chimento, P. and J. Ishac, "Defining Network Capacity", RFC
5136, February 2008.
[RX05] I. Rhee and L. Xu, CUBIC: A New TCP-Friendly High-Speed TCP
Variant, PFLDnet 2005.
[SAF06] P. Sarolahti, M. Allman, and S. Floyd, Determining an
Appropriate Sending Rate Over an Underutilized Network
Path, Computer Networks, September 2006.
[SLFK03] R.N. Shorten, D.J. Leith, J. Foy, and R. Kilduff, Analysis
and Design of Congestion Control in Synchronised
Communication Networks. Proc. 12th Yale Workshop on
Adaptive & Learning Systems, May 2003.
[SCWA99] S. Savage, N. Cardwell, D. Wetherall, and T. Anderson, TCP
Congestion Control with a Misbehaving Receiver, ACM
Computer Communications Review, October 1999.
[TM02] V. Tsaoussidis and I. Matta, Open Issues of TCP for Mobile
Computing, Journal of Wireless Communications and Mobile
Computing: Special Issue on Reliable Transport Protocols
for Mobile Computing, February 2002.
[TWL06] A. Tang, J. Wang and S. H. Low. Counter-Intuitive
Throughput Behaviors in Networks Under End-to-End Control,
IEEE/ACM Transactions on Networking, 14(2):355-368, April
2006.
[WCL05] D. X. Wei, P. Cao and S. H. Low, Time for a TCP Benchmark
Suite?, Technical Report, Caltech CS, Stanford EAS,
Caltech, 2005.
[VKD02] A. Venkataramani, R. Kokku, and M. Dahlin, TCP Nice: A
Mechanism for Background Transfers, Fifth USENIX Symposium
on Operating System Design and Implementation (OSDI), 2002.
[XHR04] L. Xu, K. Harfoush, and I. Rhee, Binary Increase Congestion
Control for Fast, Long Distance Networks, Infocom 2004.
[YKL01] Y. Yang, M. Kim, and S. Lam, Transient Behaviors of TCP-
friendly Congestion Control Protocols, Infocom 2001.
[ZKL04] Y. Zhang, S.-R. Kang, and D. Loguinov, Delayed Stability
and Performance of Distributed Congestion Control, ACM
SIGCOMM, August 2004.
Author's Address
Sally Floyd
ICSI Center for Internet Research
1947 Center Street, Suite 600
Berkeley, CA 94704
USA
EMail: floyd@icir.org
Full Copyright Statement
Copyright (C) The IETF Trust (2008).
This document is subject to the rights, licenses and restrictions
contained in BCP 78 and at http://www.rfc-editor.org/copyright.html,
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, THE IETF TRUST 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.
|
Comment about this RFC, ask questions, or add new information about this topic: