[Docs] [txt|pdf] [Tracker] [WG] [Email] [Diff1] [Diff2] [Nits]

Versions: 00 01 02 03 04 05 06 07 08 RFC 2722

Internet Engineering Task Force                    Brownlee, Mills, Ruth
INTERNET-DRAFT                                The University of Auckland
Nov 1997
Expires Mar 1998

Traffic Flow Measurement:  Architecture

<draft-ietf-rtfm-architecture-01.txt>

Status of this Memo

This document is an Internet-Draft.  Internet-Drafts are working
documents of the Internet Engineering Task Force (IETF), its Areas, and
its Working Groups.  Note that other groups may also distribute working
documents as Internet-Drafts.  This Internet Draft is a product of the
Realtime Traffic Flow Measurement Working Group of the IETF.

Internet Drafts are draft documents valid for a maximum of six months.
Internet Drafts may be updated, replaced, or obsoleted by other
documents at any time.  It is not appropriate to use Internet Drafts as
reference material or to cite them other than as a "working draft" or
"work in progress."

To view the entire list of current Internet-Drafts, please check the
"1id-abstracts.txt" listing contained in the Internet-Drafts Shadow
Directories on ftp.is.co.za (Africa), ftp.nordu.net (Europe),
munnari.oz.au (Pacific Rim), ds.internic.net (US East Coast), or
ftp.isi.edu (US West Coast).

Abstract

This document describes an architecture for the measurement and
reporting of network traffic flows, discusses how this relates to an
overall network traffic flow architecture, and describes how it can be
used within the Internet.  It is intended to provide a starting point
for the Realtime Traffic Flow Measurement Working Group.

Contents

1 Statement of Purpose and Scope                                      3
1.1 Changes Introduced Since RFC 2063  . . . . . . . . . . . . . .  4

2 Traffic Flow Measurement Architecture                               5
2.1 Meters and Traffic Flows . . . . . . . . . . . . . . . . . . .  5
2.2 Interaction Between METER and METER READER . . . . . . . . . .  7



INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

2.3 Interaction Between MANAGER and METER  . . . . . . . . . . . .  8
2.4 Interaction Between MANAGER and METER READER . . . . . . . . .  8
2.5 Multiple METERs or METER READERs . . . . . . . . . . . . . . .  9
2.6 Interaction Between MANAGERs (MANAGER - MANAGER) . . . . . . . 10
2.7 METER READERs and APPLICATIONs . . . . . . . . . . . . . . . . 10

3 Traffic Flows and Reporting Granularity                            10
3.1 Flows and their Attributes . . . . . . . . . . . . . . . . . . 11
3.2 Granularity of Flow Measurements . . . . . . . . . . . . . . . 13
3.3 Rolling Counters, Timestamps, Report-in-One-Bucket-Only. . . . 15

4 Meters                                                             17
4.1 Meter Structure  . . . . . . . . . . . . . . . . . . . . . . . 17
4.2 Flow Table . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.3 Packet Handling, Packet Matching . . . . . . . . . . . . . . . 19
4.4 Rules and Rule Sets  . . . . . . . . . . . . . . . . . . . . . 23
4.5 Maintaining the Flow Table . . . . . . . . . . . . . . . . . . 27
4.6 Handling Increasing Traffic Levels . . . . . . . . . . . . . . 28

5.1 Identifying Flows in Flow Records  . . . . . . . . . . . . . . 28
5.2 Usage Records, Flow Data Files . . . . . . . . . . . . . . . . 29
5.3 Meter to Meter Reader:  Usage Record Transmission  . . . . . . 30

6 Managers                                                           30
6.1 Between Manager and Meter:  Control Functions  . . . . . . . . 31
6.2 Between Manager and Meter Reader:  Control Functions . . . . . 31
6.3 Exception Conditions . . . . . . . . . . . . . . . . . . . . . 33
6.4 Standard Rule Sets . . . . . . . . . . . . . . . . . . . . . . 34

7 Security Considerations                                            35
7.1 Threat Analysis  . . . . . . . . . . . . . . . . . . . . . . . 35
7.2 Countermeasures  . . . . . . . . . . . . . . . . . . . . . . . 35

8 APPENDICES                                                         37
8.1 Appendix A: Network Characterisation . . . . . . . . . . . . . 37
8.2 Appendix B: Recommended Traffic Flow Measurement Capabilities .38
8.3 Appendix C: List of Defined Flow Attributes  . . . . . . . . . 39
8.4 Appendix D: List of Meter Control Variables  . . . . . . . . . 40

9 Acknowledgments                                                    41

10 References                                                         41

Brownlee, Mills, Ruth                                           [Page 2]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

1 Statement of Purpose and Scope

This document describes an architecture for traffic flow measurement and
reporting for data networks which has the following characteristics:

- The traffic flow model can be consistently applied to any
protocol/application at any network layer (e.g.  network,
transport, application layers).

- Traffic flow attributes are defined in such a way that they are
valid for multiple networking protocol stacks, and that traffic
flow measurement implementations are useful in MULTI-PROTOCOL
environments.

- Users may specify their traffic flow measurement requirements by
writing 'rule sets,' allowing them to collect the flow data they
need while ignoring other traffic.

- The data reduction effort to produce requested traffic flow
information is placed as near as possible to the network
measurement point.  This minimises the volume of data to be
obtained (and transmitted across the network for storage), and
reduces the amount of processing required in traffic flow analysis
applications.

The architecture specifies common metrics for measuring traffic flows.
By using the same metrics, traffic flow data can be exchanged and
compared across multiple platforms.  Such data is useful for:

- Understanding the behaviour of existing networks,

- Planning for network development and expansion,

- Quantification of network performance,

- Verifying the quality of network service, and

- Attribution of network usage to users.

The traffic flow measurement architecture is deliberately structured so
that specific protocol implementations may extend coverage to
multi-protocol environments and to other protocol layers, such as usage
measurement for application-level services.  Use of the same model for
both network- and application-level measurement may simplify the
development of generic analysis applications which process and/or

Brownlee, Mills, Ruth                                           [Page 3]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

correlate any or all levels of traffic and usage information.  Within
this docuemt the term 'usage data' is used as a generic term for the
data obtained using the traffic flow measurement architecture.

This document is not a protocol specification.  It specifies and
structures the information that a traffic flow measurement system needs
to collect, describes requirements that such a system must meet, and

For performance reasons, it may be desirable to use traffic information
gathered through traffic flow measurement in lieu of network statistics
obtained in other ways.  Although the quantification of network
performance is not the primary purpose of this architecture, the
measured traffic flow data may be used as an indication of network
performance.

A cost recovery structure decides "who pays for what." The major issue
here is how to construct a tariff (who gets billed, how much, for which
things, based on what information, etc).  Tariff issues include
fairness, predictability (how well can subscribers forecast their
network charges), practicality (of gathering the data and administering
the tariff), incentives (e.g.  encouraging off-peak use), and cost
recovery goals (100% recovery, subsidisation, profit making).  Issues
such as these are not covered here.

Background information explaining why this approach was selected is
provided by the 'Internet Accounting:  Background' RFC [1].

1.1 Changes Introduced Since RFC 2063

The first version of the Traffic Flow Measurement Architecture was
published as RFC 2063 in January 1997.  The most significant changes
made since then are summarised below.

- A Traffic Meter should be able to run multiple rule sets
concurrently.  This makes a meter much more useful, and required
only minimal changes to the architecture.

- 'NoMatch' replaces 'Fail' as an action.  This name was agreed to at
the Working Group 1996 meeting in Montreal; it better indicates
that although a particular match has failed, it may be tried again

- The 'MatchingStoD' attribute has been added.  This is a Packet
Matching Engine (PME) attribute indicating that addresses are being
matched in StoD (i.e.  'wire') order.  It can be used to perform
different actions when the match is retried, thereby simplifying

Brownlee, Mills, Ruth                                           [Page 4]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

some kinds of rule sets.  It was discussed and agreed to at the San
Jose meeting in 1996.

- Computed attributes (Class and Kind) may now be tested within a
rule set.  This lifts an unneccessary earlier restriction.

- The list of attribute numbers has been extended to define ranges
for 'basic' attributes (in this document) and 'extended' attributes
(currently being developed by the RTFM Working Group).

- The 'Security Considerations' section has been completely
rewritten.  It provides an evaluation of traffic measurement
security risks and their countermeasures.

2 Traffic Flow Measurement Architecture

A traffic flow measurement system is used by Network Operations
personnel to aid in managing and developing a network.  It provides a
tool for measuring and understanding the network's traffic flows.  This
information is useful for many purposes, as mentioned in section 1
(above).

The following sections outline a model for traffic flow measurement,
which draws from working drafts of the OSI accounting model [2].  Future
protocol layers.

2.1 Meters and Traffic Flows

At the heart of the traffic measurement model are network entities
called traffic METERS. Meters observe packets as they pass by a single
point on their way through the network and classify them into certain
groups.  For each such group a meter will accumulate certain attributes,
for example the numbers of packets and bytes observed for the group.
These METERED TRAFFIC GROUPS may correspond to a user, a host system,
a network, a group of networks, a particular transport address (e.g. an
IP port number), any combination of the above, etc, depending on the
meter's configuration.

We assume that routers or traffic monitors throughout a network are
instrumented with meters to measure traffic.  Issues surrounding the
choice of meter placement are discussed in the 'Traffic Flow
Measurement:  Background' RFC [1].  An important aspect of meters is
that they provide a way of succinctly aggregating traffic information.

Brownlee, Mills, Ruth                                           [Page 5]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

For the purpose of traffic flow measurement we define the concept of a
TRAFFIC FLOW, which is like an artificial logical equivalent to a call
or connection.  A flow is a portion of traffic, delimited by a start and
stop time, that belongs to one of the metered traffic groups mentioned
above.  Attribute values (source/destination addresses, packet counts,
byte counts, etc.)  associated with a flow are aggregate quantities
reflecting events which take place in the DURATION between the start and
stop times.  The start time of a flow is fixed for a given flow; the
stop time may increase with the age of the flow.

For connectionless network protocols such as IP there is by definition
no way to tell whether a packet with a particular source/destination
combination is part of a stream of packets or not - each packet is
completely independent.  A traffic meter has, as part of its
configuration, a set of 'rules' which specify the flows of interest, in
terms of the values of their attributes.  It derives attribute values
from each observed packet, and uses these to decide which flow they
belong to.  Classifying packets into 'flows' in this way provides an
economical and practical way to measure network traffic and subdivide it
into well-defined groups.

Usage information which is not derivable from traffic flows may also be
of interest.  For example, an application may wish to record accesses to
various different information resources or a host may wish to record the
username (subscriber id) for a particular network session.  Provision is
made in the traffic flow architecture to do this.  In the future the
measurement model may be extended to gather such information from
applications and hosts so as to provide values for higher-layer flow
attributes.

As well as FLOWS and METERS, the traffic flow measurement model includes
MANAGERS, METER READERS and ANALYSIS APPLICAIONS, which are explained in
following sections.  The relationships between them are shown by the
diagram below.  Numbers on the diagram refer to sections in this
document.

MANAGER
/       \
2.3 /         \ 2.4
/           \
/             \                       ANALYSIS
METER   <----->   METER READER  <----->   APPLICATION
2.2                     2.7

- MANAGER: A traffic measurement manager is an application which
configures 'meter' entities and controls 'meter reader' entities.
It sends configuration commands to the meters, and supervises the
proper operation of each meter and meter reader.  It may well be
convenient to combine the functions of meter reader and manager
within a single network entity.

Brownlee, Mills, Ruth                                           [Page 6]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

- METER: Meters are placed at measurement points determined by
Network Operations personnel.  Each meter selectively records
network activity as directed by its configuration settings.  It can
also aggregate, transform and further process the recorded activity
before the data is stored.  The processed and stored results are
called the 'usage data.'

meters so that it is available to analysis applications.

- ANALYSIS APPLICATION: An analysis application processes the usage
data so as to provide information and reports which are useful for
network engineering and management purposes.  Examples include:

-  TRAFFIC FLOW MATRICES, showing the total flow rates for many of
the possible paths within an internet.

-  FLOW RATE FREQUENCY DISTRIBUTIONS, indicating how flow rates
vary with time.

-  USAGE DATA showing the total traffic volumes sent and received
by particular hosts.

The operation of the traffic measurement system as a whole is best
understood by considering the interactions between its components.
These are described in the following sections.

2.2 Interaction Between METER and METER READER

The information which travels along this path is the usage data itself.
A meter holds usage data in an array of flow data records known as the
FLOW TABLE. A meter reader may collect the data in any suitable manner.
For example it might upload a copy of the whole flow table using a file
transfer protocol, or read the records in the current flow set one at a
time using a suitable data transfer protocol.  Note that the meter
reader need not read complete flow data records, a subset of their
attribute values may well be sufficient.

A meter reader may collect usage data from one or more meters.  Data may
be collected from the meters at any time.  There is no requirement for
collections to be synchronized in any way.

Brownlee, Mills, Ruth                                           [Page 7]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

2.3 Interaction Between MANAGER and METER

A manager is responsible for configuring and controlling one or more
meters.  Each meter's configuration includes information such as:

- Flow specifications, e.g.  which traffic flows are to be measured,
how they are to be aggregated, and any data the meter is required
to compute for each flow being measured.

- Meter control parameters, e.g.  the 'inactivity' time for flows (if
no packets belonging to a flow are seen for this time the flow is
considered to have ended, i.e.  to have become idle).

- Sampling rate.  Normally every packet will be observed.  It may
sometimes be necessary to use sampling techniques to observe only
some of the packets.  (Sampling algorithms are not prescribed by
the architecture; it should be noted that before using sampling one
should verify the statistical validity of the algorithm used).
Current experience with the measurement architecture shows that a
carefully-designed and implemented meter compresses the data such
that in normal LANs and WANs of today sampling is really not
needed.

A meter may run several rule sets concurrently on behalf of one or more
(i.e.  a 'rule set') to a meter.  Control parameters which apply to an
individual rule set should be set by the manager after it downloads that
rule set.

One manager should be designated as the 'master' for a meter.
Parameters such as sampling rate, which affect the overall operation of
the meter, should only be set by the master manager.

2.4 Interaction Between MANAGER and METER READER

A manager is responsible for configuring and controlling one or more
meter readers.  A meter reader may only be controlled by a single
manager.  A meter reader needs to know at least the following for every
meter it is collecting usage data from:

- The meter's unique identity, i.e.  its network name or address.

- How often usage data is to be collected from the meter.

Brownlee, Mills, Ruth                                           [Page 8]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

- Which flow records are to be collected (e.g. flows for a particular
rule set, whole flow table, flows seen since a given time, etc.).

- Which attribute values are to be collected for the required flow
records (e.g.  all attributes, or a small subset of them)

Since redundant reporting may be used in order to increase the
reliability of usage data, exchanges among multiple entities must be
considered as well.  These are discussed below.

2.5 Multiple METERs or METER READERs

/         |          \
/          |           \
=====METER 1     METER 2=====METER 3    METER 4=====
\           |          /
\          |         /

Several uniquely identified meters may report to one or more meter
readers.  The diagram above gives an example of how multiple meters and

In the diagram above meter 1 is read by meter reader A, and meter 4 is
read by meter reader B. Meters 1 and 4 have no redundancy; if either
meter fails, usage data for their network segments will be lost.

Meters 2 and 3, however, measure traffic on the same network segment.
One of them may fail leaving the other collecting the segment's usage
If one meter reader fails, the other will continue collecting usage data
from both meters.

The architecture does not require multiple meter readers to be
synchronized.  In the situation above meter readers A and B could both
collect usage data at the same intervals, but not neccesarily at the
same times.  Note that because collections are asynchronous it is
unlikely that usage records from two different meter readers will agree
exactly.

If precisely synchronized collections are required this can be achieved
by having one manager request each meter to begin running a new set of
rules at the same time, then allowing all meter readers to collect the
usage data from the flows recorded by the old rule sets.

Brownlee, Mills, Ruth                                           [Page 9]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

If there is only one meter reader and it fails, the meters continue to
run.  When the meter reader is restarted it can collect all of the
accumulated flow data.  Should this happen, time resolution will be lost
(because of the missed collections) but overall traffic flow information
will not.  The only exception to this would occur if the traffic volume
was sufficient to 'roll over' counters for some flows during the
failure; this is addressed in the section on 'Rolling Counters.'

2.6 Interaction Between MANAGERs (MANAGER - MANAGER)

Synchronization between multiple management systems is the province of
network management protocols.  This traffic flow measurement
architecture specifies only the network management controls necessary to
perform the traffic flow measurement function and does not address the
more global issues of simultaneous or interleaved (possibly conflicting)
commands from multiple network management stations or the process of
transferring control from one network management station to another.

Once a collection of usage data has been assembled by a meter reader it
can be processed by an analysis application.  Details of analysis
applications - such as the reports they produce and the data they
require - are outside the scope of this architecture.

It should be noted, however, that analysis applications will often
require considerable amounts of input data.  An important part of
running a traffic flow measurement system is the storage and regular
reduction of flow data so as to produce daily, weekly or monthly summary
files for further analysis.  Again, details of such data handling are
outside the scope of this architecture.

3 Traffic Flows and Reporting Granularity

A flow was defined in section 2.1 above in abstract terms as follows:

"A TRAFFIC FLOW is an artifical logical equivalent to a call or
connection, belonging to a (user-specieied) METERED TRAFFIC
GROUP."

Brownlee, Mills, Ruth                                          [Page 10]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

In practical terms, a flow is a stream of packets passing across a
network between two end points (or being sent from a single end point),
which have been summarized by a traffic meter for analysis purposes.

3.1 Flows and their Attributes

Every traffic meter maintains a table of 'flow records' for flows seen
by the meter.  A flow record holds the values of the ATTRIBUTES of
interest for its flow.  These attributes might include:

- ADDRESSES for the flow's source and destination.  These comprise
the protocol type, the source and destination addresses at various
network layers (extracted from the packet header), and the number
of the interface on which the packet was observed.

- First and last TIMES when packets were seen for this flow, i.e.
the 'creation' and 'last activity' times for the flow.

- COUNTS for 'forward' (source to destination) and 'backward'
(destination to source) components (e.g.  packets and bytes) of the
flow's traffic.  The specifying of 'source' and 'destination' for
flows is discussed in the section on packet matching below.

- OTHER attributes, e.g.  the index of the flow's record in the flow
table and the rule set id for the rules which the meter was running
while the flow was observed.  The values of these attributes
provide a way of distinguishing flows observed by a meter at
different times.

The attributes listed in this document (Appendix C) provide a basic
(i.e.  useful minimum) set; they are assigned attribute numbers in the
range 0 to 63.  The RTFM working group is working on an extended set of
attributes, which will have numbers in the range 64 to 127.
Implementors wishing to experiment with further new attributes should
use attribute numbers 128 and above.

A flow's METERED TRAFFIC GROUP is specified by the values of its ADDRESS
attributes.  For example, if a flow's address attributes specified only
that "source address = IP address 10.1.0.1," then all IP packets from
and to that address would be counted in that flow.  If a flow's address
10.1.0.1 and 26.1.0.1 would be counted in that flow.

more of the following types:

Brownlee, Mills, Ruth                                          [Page 11]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

- The INTERFACE NUMBER for the flow, i.e.  the interface on which the
meter measured the traffic.  Together with a unique address for the
meter this uniquely identifies a particular physical-level port.

immediate source or destination on the path of the packet.  For
example, if flow measurement is being performed at the IP layer on
six-octet Media Access Control (MAC) address.  For a host connected
to the same LAN segment as the meter the adjacent address will be
the MAC address of that host.  For hosts on other LAN segments it
router carrying the traffic flow.

- The PEER ADDRESS, which identifies the source or destination of the
PEER-LEVEL packet.  The form of a peer address will depend on the
network-layer protocol in use, and the network layer [n] at which
traffic measurement is being performed.

- The TRANSPORT ADDRESS, which identifies the source or destination
port for the packet, i.e.  its [n+1] layer address.  For example,
if flow measurement is being performed at the IP layer a transport
address is a two-octet UDP or TCP port number.

The four definitions above specify addresses for each of the four lowest
layers of the OSI reference model, i.e.  Physical layer, Link layer,
Network layer and Transport layer.  A FLOW RECORD stores both the VALUE
for each of its addresses (as described above) and a MASK specifying
which bits of the address value are being used and which are ignored.
Note that if address bits are being ignored the meter will set them to
zero, however their actual values are undefined.

One of the key features of the traffic measurement architecture is that
attributes have essentially the same meaning for different protocols, so
that analysis applications can use the same reporting formats for all
protocols.  This is straightforward for peer addresses; although the
form of addresses differs for the various protocols, the meaning of a
'peer address' remains the same.  It becomes harder to maintain this
correspondence at higher layers - for example, at the Network layer IP,
Novell IPX and AppleTalk all use port numbers as a 'transport address,'
but CLNP and DECnet have no notion of ports.  Further work is needed
here, particularly in selecting attributes which will be suitable for
the higher layers of the OSI reference model.

Reporting by adjacent intermediate sources and destinations or simply by
meter interface (most useful when the meter is embedded in a router)
supports hierarchical Internet reporting schemes as described in the
'Internet Accounting:  Background' RFC [1].  That is, it allows backbone
and regional networks to measure usage to just the next lower level of
granularity (i.e.  to the regional and stub/enterprise levels,

Brownlee, Mills, Ruth                                          [Page 12]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

respectively), with the final breakdown according to end user (e.g.  to
source IP address) performed by the stub/enterprise networks.

In cases where network addresses are dynamically allocated (e.g.  mobile
subscribers), further subscriber identification will be necessary if
flows are to ascribed to individual users.  Provision is made to further
specify the metered traffic group through the use of an optional
SUBSCRIBER ID as part of the flow id.  A subscriber ID may be associated
with a particular flow either through the current rule set or by
proprietary means within a meter, for example via protocol exchanges
with one or more (multi-user) hosts.  At this time a subscriber ID is an
arbitrary text string; later versions of the architecture may specify
its contents in more detail.

3.2 Granularity of Flow Measurements

GRANULARITY is the 'control knob' by which an application and/or the
reporting against the level of detail supplied.  A coarser granularity
means a greater level of aggregation; finer granularity means a greater
level of detail.  Thus, the number of flows measured (and stored) at a
meter can be regulated by changing the granularity of the metered
traffic group, the attributes, or the inactivity time interval.  Flows
are like an adjustable pipe - many fine-granularity streams can carry
the data with each stream measured individually, or data can be bundled
in one coarse-granularity pipe.

Flow granularity is controlled by adjusting the level of detail at which
the following are determined:

- The metered traffic group (address attributes, discussed above).

- The categorisation of packets (other attributes, discussed below).

- The lifetime/duration of flows (the reporting interval needs to be
short enough to measure them with sufficient precision).

The set of rules controlling the determination of each packet's metered
traffic group is known as the meter's CURRENT RULE SET. As will be
shown, the meter's current rule set forms an integral part of the
reported information, i.e.  the recorded usage information cannot be
properly interpreted without a definition of the rules used to collect
that information.

Settings for these granularity factors may vary from meter to meter.
They are determined by the meter's current rule set, so they will change

Brownlee, Mills, Ruth                                          [Page 13]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

if network Operations personnel reconfigure the meter to use a new rule
set.  It is expected that the collection rules will change rather
infrequently; nonetheless, the rule set in effect at any time must be
identifiable via a RULE SET ID. Granularity of metered traffic groups is
further specified by additional ATTRIBUTES. These attributes include:

- Attributes which record information derived from other attribute
values.  Six of these are defined (SourceClass, DestClass,
FlowClass, SourceKind, DestKind, FlowKind), and their meaning is
determined by the meter's rule set.  For example, one could have a
subroutine in the rule set which determined whether a source or
destination peer address was a member of an arbitrary list of
networks, and set SourceClass/DestClass to one if the source/dest
peer address was in the list or to zero otherwise.

- Administratively specified attributes such as Quality Of Service
and Priority, etc.  These are not defined at this time.

- Higher-layer (especially application-level) attributes.  These are
not defined at this time.

Settings for these granularity factors may vary from meter to meter.
They are determined by the meter's current rule set, so they will change
if Network Operations personnel reconfigure the meter to use a new rule
set.

The LIFETIME of a flow is the time interval which began when the meter
observed the first packet belonging to the flow and ended when it saw
the last packet.  Flow lifetimes are very variable, but many - if not
most - are rather short.  A meter cannot measure lifetimes directly;
instead a meter reader collects usage data for flows which have been
active since the last collection, and an analysis application may
compare the data from each collection so as to determine when each flow
actually stopped.

The meter does, however, need to reclaim memory (i.e.  records in the
flow table) being held by idle flows.  The meter configuration includes
a variable called InactivityTimeout, which specifies the minimum time a
meter must wait before recovering the flow's record.  In addition,
before recovering a flow record the meter must be sure that the flow's
data has been collected by all meter readers which registered to collect
it.

These 'lifetime' issues are considered further in the section on meter
readers (below).  A complete list of the attributes currently defined is
given in Appendix C later in this document.

Brownlee, Mills, Ruth                                          [Page 14]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

3.3 Rolling Counters, Timestamps, Report-in-One-Bucket-Only

Once a usage record is sent, the decision needs to be made whether to
clear any existing flow records or to maintain them and add to their
counts when recording subsequent traffic on the same flow.  The second
method, called rolling counters, is recommended and has several
reliability - the system can now often survive the loss of some usage
records, such as might occur if a meter reader failed and later
restarted.  The next usage record will very often contain yet another
reading of many of the same flow buckets which were in the lost usage
record.  The 'continuity' of data provided by rolling counters can also
supply information used for "sanity" checks on the data itself, to guard
against errors in calculations.

The use of rolling counters does introduce a new problem:  how to
distinguish a follow-on flow record from a new flow record.  Consider
the following example.

CONTINUING FLOW        OLD FLOW, then NEW FLOW

start time = 1            start time = 1
Usage record N:       flow count = 2000      flow count = 2000 (done)

start time = 1            start time = 5
Usage record N+1:     flow count = 3000      new flow count = 1000

Total count:                 3000                    3000

In the continuing flow case, the same flow was reported when its count
was 2000, and again at 3000:  the total count to date is 3000.  In the
OLD/NEW case, the old flow had a count of 2000.  Its record was then
stopped (perhaps because of temporary idleness), but then more traffic
with the same characteristics arrived so a new flow record was started
and it quickly reached a count of 1000.  The total flow count from both
the old and new records is 3000.

The flow START TIMESTAMP attribute is sufficient to resolve this.  In
the example above, the CONTINUING FLOW flow record in the second usage
record has an old FLOW START timestamp, while the NEW FLOW contains a
recent FLOW START timestamp.

Each packet is counted in at most one flow for each running ruleset, so
as to avoid multiple counting of a single packet.  The record of a
single flow is informally called a "bucket." If multiple, sometimes
overlapping, records of usage information are required (aggregate,
individual, etc), the network manager should collect the counts in
sufficiently detailed granularity so that aggregate and combination

Brownlee, Mills, Ruth                                          [Page 15]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

counts can be reconstructed in post-processing of the raw usage data.
Alternatively, multiple rulesets could be used to collect data at
different granularities.

For example, consider a meter from which it is required to record both
'total packets coming in interface #1' and 'total packets arriving from
any interface sourced by IP address = a.b.c.d,' using a single rule set.
Although a bucket can be declared for each case, it is not clear how to
handle a packet which satisfies both criteria.  It must only be counted
once.  By default it will be counted in the first bucket for which it
qualifies, and not in the other bucket.  Further, it is not possible to
reconstruct this information by post-processing.  The solution in this
case is to define not two, but THREE buckets, each one collecting a
unique combination of the two criteria:

Bucket 1:  Packets which came in interface 1,
AND were sourced by IP address a.b.c.d

Bucket 2:  Packets which came in interface 1,
AND were NOT sourced by IP address a.b.c.d

Bucket 3:  Packets which did NOT come in interface 1,
AND were sourced by IP address a.b.c.d

(Bucket 4:  Packets which did NOT come in interface 1,
AND NOT sourced by IP address a.b.c.d)

The desired information can now be reconstructed by post-processing.
"Total packets coming in interface 1" can be found by adding buckets 1 &
2, and "Total packets sourced by IP address a.b.c.d" can be found by
adding buckets 1 & 3.  Note that in this case bucket 4 is not explicitly
required since its information is not of interest, but it is supplied
here in parentheses for completeness.

Alternatively, the above could be achieved by running two rule sets (A
and B), as follows:

Bucket 1:  Packets which came in interface 1;
counted by rule set A.

Bucket 2:  Packets which were sourcede by IP address a.b.c.d;
counted by rule set B.

Brownlee, Mills, Ruth                                          [Page 16]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

4 Meters

A traffic flow meter is a device for collecting data about traffic flows
at a given point within a network; we will call this the METERING POINT.
The header of every packet passing the network metering point is offered
to the traffic meter program.

A meter could be implemented in various ways, including:

- A dedicated small host, connected to a LAN (so that it can see all
packets as they pass by) and running a 'traffic meter' program.
The metering point is the LAN segment to which the meter is
attached.

- A multiprocessing system with one or more network interfaces, with
drivers enabling a traffic meter program to see packets.  In this
case the system provides multiple metering points - traffic flows
on any subset of its network interfaces can be measured.

- A packet-forwarding device such as a router or switch.  This is
similar to (b) except that every received packet should also be
forwarded, usually on a different interface.

4.1 Meter Structure

An outline of the meter's structure is given in the following diagram:

Briefly, the meter works as follows:

- Incoming packet headers arrive at the top left of the diagram and
are passed to the PACKET PROCESSOR.

- The packet processor passes them to the Packet Matching Engine
(PME) where they are classified.

- The PME is a Virtual Machine running a pattern matching program
contained in the CURRENT RULE SET. It is invoked by the Packet
Processor, and returns instructions on what to do with the packet.

- Some packets are classified as 'to be ignored.'  They are discarded
by the Packet Processor.

- Other packets are matched by the PME, which returns a FLOW KEY
describing the flow to which the packet belongs.

Brownlee, Mills, Ruth                                          [Page 17]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

- The flow key is used to locate the flow's entry in the FLOW TABLE;
a new entry is created when a flow is first seen.  The entry's data
fields (e.g.  packet and byte counters) are updated.

- A meter reader may collect data from the flow table at any time.
It may use the 'collect' index to locate the flows to be collected
within the flow table.

packet                +------------------+
header                | Current Rule Set |
|                   +--------+---------+
|                            |
+--------*---------+       +----------*-------------+
| Packet Processor |<----->| Packet Matching Engine |
+--+------------+--+       +------------------------+
|            |
Ignore *            | Count via flow key
|
+--*--------------+
| 'Search' index  |
+--------+--------+
|
+--------*--------+
|                 |
|   Flow Table    |
|                 |
+--------+--------+
|
+--------*--------+
| 'Collect' index |
+--------+--------+
|
*

The discussion above assumes that a meter will only be running a single
rule set.  A meter may, however, run several rule sets concurrently.  To
do this the meter maintains a table of current rulesets.  The packet
processor matches each packet against every current ruleset, producing a
single flow table with flows from all the rule sets.  The overall effect
of doing this is somewhat similar to running several independent meters,
one for each rule set.

Brownlee, Mills, Ruth                                          [Page 18]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

4.2 Flow Table

Every traffic meter maintains a table of TRAFFIC FLOW RECORDS for flows
seen by the meter.  A flow record contains attribute values for its
flow, including:

- Addresses for the flow's source and destination.  These include
packet), and the number of the interface on which the packet was
observed.

- First and last times when packets were seen for this flow.

- Counts for 'forward' (source to destination) and 'backward'
(destination to source) components of the flow's traffic.

- Other attributes, e.g.  state of the flow record (discussed below).

The state of a flow record may be:

- INACTIVE: The flow record is not being used by the meter.

- CURRENT: The record is in use and describes a flow which belongs to
the 'current flow set,' i.e.  the set of flows recently seen by the
meter.

- IDLE: The record is in use and the flow which it describes is part
of the current flow set.  In addition, no packets belonging to this
flow have been seen for a period specified by the meter's
InactivityTime variable.

4.3 Packet Handling, Packet Matching

Each packet header received by the traffic meter program is processed as
follows:

- Extract attribute values from the packet header and use them to
create a MATCH KEY for the packet.

- Match the packet's key against the current rule set, as explained
in detail below.

Brownlee, Mills, Ruth                                          [Page 19]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

The rule set specifies whether the packet is to be counted or ignored.
If it is to be counted the matching process produces a FLOW KEY for the
flow to which the packet belongs.  This flow key is used to find the
flow's record in the flow table; if a record does not yet exist for this
flow, a new flow record may be created.  The data for the matching flow
record can then be updated.

For example, the rule set could specify that packets to or from any host
in IP network 130.216 are to be counted.  It could also specify that
flow records are to be created for every pair of 24-bit (Class C)
subnets within network 130.216.

Each packet's match key is passed to the meter's PATTERN MATCHING ENGINE
(PME) for matching.  The PME is a Virtual Machine which uses a set of
instructions called RULES, i.e.  a RULE SET is a program for the PME. A
packet's match key contains an interface number, source address (S) and
destination address (D) values.  It does not, however, contain any
attribute masks for its attributes, only their values.

If measured flows were unidirectional, i.e.  only counted packets
travelling in one direction, the matching process would be simple.  The
PME would be called once to match the packet.  Any flow key produced by
a successful match would be used to find the flow's record in the flow
table, and that flow's counters would be updated.

Flows are, however, bidirectional, reflecting the forward and reverse
packets of a protocol interchange or 'session.'  Maintaining two sets of
counters in the meter's flow record makes the resulting flow data much
simpler to handle, since analysis programs do not have to gather
together the 'forward' and 'reverse' components of sessions.
Implementing bi-directional flows is, of course, more difficult for the
meter, since it must decide whether a packet is a 'forward' packet or a
'reverse' one.  To make this decision the meter will often need to
invoke the PME twice, once for each possible packet direction.

The diagram below describes the algorithm used by the traffic meter to
process each packet.  Flow through the diagram is from left to right and
top to bottom, i.e.  from the top left corner to the bottom right
corner.  S indicates the flow's source address (i.e.  its set of source

There are several cases to consider.  These are:

- The packet is recognised as one which is TO BE IGNORED.

- The packet would MATCH IN EITHER DIRECTION. One situation in which
this could happen would be a rule set which matches flows within
network X (Source = X, Dest = X) but specifies that flows are to be
created for each subnet within network X, say subnets y and z.  If,
for example a packet is seen for y->z, the meter must check that
flow z->y is not already current before creating y->z.

Brownlee, Mills, Ruth                                          [Page 20]


INTERNET-DRAFT      Traffic Flow Measurement: Architecture      Nov 1997

- The packet MATCHES IN ONE DIRECTION ONLY. If its flow is already
current, its forward or reverse counters are incremented.
Otherwise it is added to the flow table and then counted.

Ignore
--- match(S->D) -------------------------------------------------+
| Suc   | NoMatch                                           |
|       |          Ignore                                   |
|      match(D->S) -----------------------------------------+
|       | Suc   | NoMatch                                   |
|       |       |                                           |
|       |       +-------------------------------------------+
|       |                                                   |
|       |             Suc                                   |
|      current(D->S) ---------- count(D->S,r) --------------+
|       | Fail                                              |
|       |                                                   |
|      create(D->S) ----------- count(D->S,r) --------------+
|                                                           |
|             Suc                                           |
current(S->D) ------------------ count(S->D,f) --------------+
| Fail                                                      |
|             Suc                                           |
current(D->S) ------------------ count(D->S,r) --------------+
| Fail                                                      |
|                                                           |
create(S->D) ------------------- count(S->D,f) --------------+
|
*

The algorithm uses four functions, as follows:

match(A->B) implements the PME.  It uses the meter's current rule set
to match the attribute values in the packet's match key.  A->B means
that the assumed source address is A and destination address B, i.e.
that the packet was travelling from A to B.  match() returns one of
three results:

'Ignore' means that the packet was matched but this flow is not
to be counted.

'NoMatch' means that the packet did not match.  It might, however
match with its direction reversed, i.e. from B to A.

'Suc'  means that the packet did match, i.e. it belongs to a flow
which is to be counted.

Brownlee, Mills, Ruth                                          [Page 21]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

current(A->B) succeeds if the flow A-to-B is current - i.e. has
a record in the flow table whose state is Current - and fails
otherwise.

create(A->B) adds the flow A-to-B to the flow table, setting the
value for attributes - such as addresses - which remain constant,
and zeroing the flow's counters.

count(A->B,f) increments the 'forward' counters for flow A-to-B.
count(A->B,r) increments the 'reverse' counters for flow A-to-B.
'Forward' here means the counters for packets travelling from
A to B.  Note that count(A->B,f) is identical to count(B->A,r).

When writing rule sets one must remember that the meter will normally
try to match each packet in the reverse direction if the forward match
does not succeed.  It is particularly important that the rule set does
not contain inconsistencies which will upset this process.

Consider, for example, a rule set which counts packets from source
network A to destination network B, but which ignores packets from
source network B. This is an obvious example of an inconsistent rule
set, since packets from network B should be counted as reverse packets
for the A-to-B flow.

This problem could be avoided by devising a language for specifying rule
files and writing a compiler for it, thus making it much easier to
produce correct rule sets.  Another approach would be to write a 'rule
set consistency checker' program, which could detect problems in
hand-written rule sets.

Normally, the best way to avoid these problems is to write rule sets
which only classify flows in the forward direction, and rely on the
meter to handle reverse-travelling packets.

Occasionally there can be situations when a rule set needs to know the
direction in which a packet is being matched.  Consider, for example, a
rule set which wants to save some attribute values (source and
destination addresses perhaps) for any 'unusual' packets.  The rule set
will contain a sequence of tests for all the 'usual' source addresses;
follwed by a rule which will execute a 'NoMatch' action.  If the match
fails in the S->D direction, the NoMatch action will cause it to be
retried.  If it fails in the D->S direction, the packet can be counted
as an 'unusual' packet.

To count such an 'unusual' packet we need to know the matching
direction:  the MatchingStoD attribute provides this.  To use it, one
follows the source address tests with a rule which tests whether the
matching direction is S->D (MatchingStoD value is 1).  If so, a

Brownlee, Mills, Ruth                                          [Page 22]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

'NoMatch' action is executed.  Otherwise, the packet has failed to match
in both directions; we can Push whatever attribute values are of
interest and count the 'unusual' packet.

4.4 Rules and Rule Sets

A rule set is an array of rules.  Rule sets are held within a meter as
entries in an array of rule sets.

Rule set 1 (the first entry in the rule set table) is built in to the
meter and cannot be changed.  It is run when the meter is started up,
and provides a very coarse reporting granularity; it is mainly useful
for verifying that the meter is running, before a 'useful' rule set is

A meter also maintains an array of 'tasks,' which specify what rule sets
the meter is running.  Each task has a 'current' rule set (the one which
it normally uses), and a 'standby' rule set (which will be used when the
overall traffic level is unusually high).  If a task is instructed to
use rule set 0, it will cease measuring; all packets will be ignored
until another (non-zero) rule set is made current.

Each rule in a rule set is structured as follows:

+-------- test ---------+    +---- action -----+
attribute & mask = value:    opcode,  parameter;

Opcodes contain two flags:  'goto' and 'test.'  The PME maintains a
Boolean indicator called the 'test indicator,' which is initially set
(true).  Execution begins with rule 1, the first in the rule set.  It
proceeds as follows:

Brownlee, Mills, Ruth                                          [Page 23]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

If the test indicator is true:
Perform the test, i.e. AND the attribute value with the
mask and compare it with the value.
If these are equal the test has succeeded; perform the
rule's action (below).
If the test fails execute the next rule in the rule set.
If there are no more rules in the rule set, return from the
match() function indicating NoMatch.

If the test indicator is false, or the test (above) succeeded:
Set the test indicator to this opcode's test flag value.
Determine the next rule to execute.
If the opcode has its goto flag set, its parameter value
specifies the number of the next rule.
Opcodes which don't have their goto flags set either
determine the next rule in special ways (Return),
or they terminate execution (Ignore, NoMatch, Count,
CountPkt).
Perform the action.

The PME maintains two 'history' data structures.  The first, the
'return' stack, simply records the index (i.e.  1-origin rule number) of
each Gosub rule as it is executed; Return rules pop their Gosub rule
index.  The second, the 'pattern' queue, is used to save information for
later use in building a flow key.  A flow key is built by zeroing all
its attribute values, then copying attribute and mask information from
the pattern queue in the order it was enqueued.

The opcodes are:

opcode         goto    test

1  Ignore           0       -
2  NoMatch          0       -
3  Count            0       -
4  CountPkt         0       -
5  Return           0       0
6  Gosub            1       1
7  GosubAct         1       0
8  Assign           1       1
9  AssignAct        1       0
10  Goto             1       1
11  GotoAct          1       0
12  PushRuleTo       1       1
13  PushRuleToAct    1       0
14  PushPktTo        1       1
15  PushPktToAct     1       0

Brownlee, Mills, Ruth                                          [Page 24]


INTERNET-DRAFT      Traffic Flow Measurement: Architecture      Nov 1997

The actions they perform are:

Ignore:         Stop matching, return from the match() function
indicating that the packet is to be ignored.

NoMatch:        Stop matching, return from the match() function
indicating failure.

Count:          Stop matching.  Save this rule's attribute name,
mask and value in the PME's pattern queue, then
construct a flow key for the flow to which this
packet belongs.  Return from the match() function
indicating success.  The meter will use the flow
key to search for the flow record for this
packet's flow.

CountPkt:       As for Count, except that the masked value from
the packet is saved in the PME's pattern queue

Gosub:          Call a rule-matching subroutine.  Push the current
rule number on the PME's return stack, set the
test indicator then goto the specified rule.

GosubAct:       Same as Gosub, except that the test indicator is
cleared before going to the specified rule.

Return:         Return from a rule-matching subroutine.  Pop the
number of the calling gosub rule from the PME's
'return' stack and add this rule's parameter value
to it to determine the 'target' rule.  Clear the
test indicator then goto the target rule.

A subroutine call appears in a rule set as a Gosub
rule followed by a small group of following rules.
Since a Return action clears the test flag, the
action of one of these 'following' rules will be
executed; this allows the subroutine to return a
result (in addition to any information it may save
in the PME's pattern queue).

Assign:         Set the attribute specified in this rule to the
value specified in this rule.  Set the test
indicator then goto the specified rule.

AssignAct:      Same as Assign, except that the test indicator
is cleared before going to the specified rule.

Goto:           Set the test indicator then goto the
specified rule.

Brownlee, Mills, Ruth                                          [Page 25]


INTERNET-DRAFT      Traffic Flow Measurement: Architecture      Nov 1997

GotoAct:        Clear the test indicator then goto the specified
rule.

PushRuleTo:     Save this rule's attribute name, mask and value
in the PME's pattern queue. Set the test
indicator then goto the specified rule.

PushRuleToAct:  Same as PushRuleTo, except that the test indicator
is cleared before going to the specified rule.

PushRuleTo actions may be used to save the value
and mask used in a test, or (if the test is not
performed) to save an arbitrary value and mask.

PushPktTo:      Save this rule's attribute name, mask, and the
pattern queue.  Set the test indicator then goto
the specified rule.

PushPktToAct:   Same as PushPktTo, except that the test indicator
is cleared before going to the specified rule.

PushPktTo actions may be used to save a value from
the packet using a specified mask.  The test in
PushPktTo rules will almost never be executed.

As well as the attributes applying directly to packets (such as
further attribtes.  These are:

Null:       Tests performed on the Null attribute always succeed.

MatchingStoD:  Indicates whether the PME is matching the packet
with its addresses in 'wire order' or with its
addresses reversed.  MatchingStoD's value is 1 if the
addresses are in wire order (StoD), and != 1 otherwise.

v1 .. v5:   v1, v2, v3, v4 and v5 are 'meter variables.'  They
provide a way to pass parameters into rule-matching
subroutines.  Each may hold the number of a normal
attribute; its value is set by an Assign action.
When a meter variable appears as the attribute of a
rule, its value specifies the actual attribute to be
tested.  For example, if v1 had been assigned
SourcePeerAddress as its value, a rule with v1 as its

SourceClass, DestClass, FlowClass,

Brownlee, Mills, Ruth                                          [Page 26]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

SourceKind, DestKind, FlowKind:
These six attributes may be set by executing PushRuleto
actions.  They allow the PME to save (in flow records)
information which has been built up during matching.
Their values may be tested in rules; this allows one
to set them early in a rule set, and test them later.

4.5 Maintaining the Flow Table

The flow table may be thought of as a 1-origin array of flow records.
(A particular implementation may, of course, use whatever data structure
is most suitable).  When the meter starts up there are no known flows;
all the flow records are in the 'inactive' state.

Each time a packet is matched for a flow which is not in a current flow
set a flow record is created for it; the state of such a record is
'current.'  When selecting a record for the new flow the meter searches
the flow table for an 'inactive' record.  If no inactive records are
available it will search for an 'idle' one instead.  Note that there is
no particular significance in the ordering of records within the table.

Flow data may be collected by a 'meter reader' at any time.  There is no
requirement for collections to be synchronized.  The reader may collect
the data in any suitable manner, for example it could upload a copy of
the whole flow table using a file transfer protocol, or it could read
the records in the current flow set row by row using a suitable data
transfer protocol.

The meter keeps information about collections, in particular it
maintains ReaderLastTime variables which remember the time the last
specifies the minimum time the meter will wait before considering that a
flow is idle.

The meter must recover records used for idle flows, if only to prevent
it running out of flow records.  Recovered flow records are returned to
the 'inactive' state.  A variety of recovery strategies are possible,
including the following:

One possible recovery strategy is to recover idle flow records as soon
as possible after their data has been collected by all readers which
have registered to do so.  To implement this the meter could run a
background process which scans the flow table looking for 'current'
flows whose 'last packet' time is earlier than the meter's
LastCollectTime.

Another recovery strategy is to leave idle flows alone as long as
possible, which would be suitable if one was only interested in

Brownlee, Mills, Ruth                                          [Page 27]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

measuring total traffic volumes.  It could be implemented by having the
meter search for collected idle flows only when it ran low on 'inactive'
flow records.

One further factor a meter should consider before recovering a flow is
the number of meter readers which have collected the flow's data.  If
flow's data before its memory is recovered.

4.6 Handling Increasing Traffic Levels

Under normal conditions the meter reader specifies which set of usage
records it wants to collect, and the meter provides them.

If memory usage rises above the high-water mark the meter should switch
to a STANDBY RULE SET so as to decrease the rate at which new flows are
created.  When the manager, usually as part of a regular poll, becomes
aware that the meter is using its standby rule set, it could decrease
the interval between collections.  The meter should also increase its
efforts to recover flow memory so as to reduce the number of idle flows
in memory.  When the situation returns to normal, the manager may
request the meter to switch back to its normal rule set.

Usage data is accumulated by a meter (e.g.  in a router) as memory
permits.  It is collected at regular reporting intervals by meter
readers, as specified by a manager.  The collected data is recorded in a
disk file called a FLOW DATA FILE, as a sequence of USAGE RECORDS.

The following sections describe the contents of usage records and flow
data files.  Note, however, that at this stage the details of such
records and files is not specified in the architecture.  Specifying a
common format for them would be a worthwhile future development.

5.1 Identifying Flows in Flow Records

Once a packet has been classified and is ready to be counted, an
appropriate flow data record must already exist in the flow table;
otherwise one must be created.  The flow record has a flexible format
where unnecessary identification attributes may be omitted.  The
determination of which attributes of the flow record to use, and of what
values to put in them, is specified by the current rule set.

Brownlee, Mills, Ruth                                          [Page 28]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

Note that the combination of start time, rule set id and subscript (row
number in the flow table) provide a unique flow identifier, regardless
of the values of its other attributes.

The current rule set may specify additional information, e.g.  a
computed attribute value such as FlowKind, which is to be placed in the
attribute section of the usage record.  That is, if a particular flow is
matched by the rule set, then the corresponding flow record should be
marked not only with the qualifying identification attributes, but also
with the additional information.  Using this feature, several flows may
each carry the same FlowKind value, so that the resulting usage records
can be used in post-processing or between meter reader and meter as a
criterion for collection.

5.2 Usage Records, Flow Data Files

The collected usage data will be stored in flow data files on the meter
reader, one file for each meter.  As well as containing the measured
usage data, flow data files must contain information uniquely
identifiying the meter from which it was collected.

A USAGE RECORD contains the descriptions of and values for one or more
flows.  Quantities are counted in terms of number of packets and number
of bytes per flow.  Each usage record contains the metered traffic group
identifier of the meter (a set of network addresses), a time stamp and a
list of reported flows (FLOW DATA RECORDS). A meter reader will build up
a file of usage records by regularly collecting flow data from a meter,
using this data to build usage records and concatenating them to the
tail of a file.  Such a file is called a FLOW DATA FILE.

A usage record contains the following information in some form:

+-------------------------------------------------------------------+
|    RECORD IDENTIFIERS:                                            |
|      Meter Id (& digital signature if required)                   |
|      Timestamp                                                    |
|      Collection Rules ID                                          |
+-------------------------------------------------------------------+
|    FLOW IDENTIFIERS:            |    COUNTERS                     |
|      Address List               |       Packet Count              |
|      Subscriber ID (Optional)   |       Byte Count                |
|      Attributes (Optional)      |    Flow Start/Stop Time         |
+-------------------------------------------------------------------+

Brownlee, Mills, Ruth                                          [Page 29]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

5.3 Meter to Meter Reader:  Usage Record Transmission

The usage record contents are the raison d'etre of the system.  The
accuracy, reliability, and security of transmission are the primary
concerns of the meter/meter reader exchange.  Since errors may occur on
networks, and Internet packets may be dropped, some mechanism for
ensuring that the usage information is transmitted intact is needed.

Flow data is moved from meter to meter reader via a series of protocol
exchanges between them.  This may be carried out in various ways, moving
individual attribute values, complete flows, or the entire flow table
(i.e.  all the active and idle flows).  One possible method of achieving
this transfer is to use SNMP; the 'Traffic Flow Measurement:  Meter MIB'
document [4] gives details.  Note that this is simply one example; the
transfer of flow data from meter to meter reader is not specified in
this document.

The reliability of the data transfer method under light, normal, and
extreme network loads should be understood before selecting among
collection methods.

In normal operation the meter will be running a rule file which provides
the required degree of flow reporting granularity, and the meter
reader(s) will collect the flow data often enough to allow the meter's
garbage collection mechanism to maintain a stable level of memory usage.

In the worst case traffic may increase to the point where the meter is
in danger of running completely out of flow memory.  The meter
implementor must decide how to handle this, for example by switching to
a default (extremely coarse granularity) rule set, by sending a trap
message to the manager, or by attempting to dump flow data to the meter

Users of the Traffic Flow Measurement system should analyse their
requirements carefully and assess for themselves whether it is more
important to attempt to collect flow data at normal granularity
(increasing the collection frequency as needed to keep up with traffic
volumes), or to accept flow data with a coarser granularity.  Similarly,
it may be acceptable to lose flow data for a short time in return for
being sure that the meter keeps running properly, i.e.  is not
overwhelmed by rising traffic levels.

6 Managers

A manager configures meters and controls meter readers.  It does this
via the interactions described below.

Brownlee, Mills, Ruth                                          [Page 30]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

6.1 Between Manager and Meter:  Control Functions

- DOWNLOAD RULE SET: A meter may hold an array of rule sets.  One of
these, the 'default' rule set, is built in to the meter and cannot
manager may use any suitable protocol exchange to achieve this, for
example an FTP file transfer or a series of SNMP SETs, one for each
row of the rule set.

manager must instruct the meter which rule sets will be the
'current' and 'standby' ones for each task the meter is to perform.

- SET HIGH WATER MARK: A percentage value interpreted by the meter
which tells the meter when to switch to its standby rule set, so as
to increase the granularity of the flows and conserve the meter's
flow memory.  Once this has happened, the manager may also change
the polling frequency or the meter's control parameters (so as to
increase the rate at which the meter can recover memory from idle
flows).

If the high traffic levels persist, the meter's normal rule set may
have to be rewritten to permanently reduce the reporting
granularity.

- SET FLOW TERMINATION PARAMETERS: The meter should have the good
sense in situations where lack of resources may cause data loss to
purge flow records from its tables.  Such records may include:

-  Flows that have already been reported to all registered meter
readers, and show no activity since the last report,

-  Oldest flows, or

-  Flows with the smallest number of observed packets.

- SET INACTIVITY TIMEOUT: This is a time in seconds since the last
packet was seen for a flow.  Flow records may be reclaimed if they
have been idle for at least this amount of time, and have been
collected in accordance with the current collection criteria.

6.2 Between Manager and Meter Reader:  Control Functions

Because there are a number of parameters that must be set for traffic
flow measurement to function properly, and viable settings may change as

Brownlee, Mills, Ruth                                          [Page 31]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

a result of network traffic characteristics, it is desirable to have
dynamic network management as opposed to static meter configurations.
Many of these operations have to do with space tradeoffs - if memory at
the meter is exhausted, either the collection interval must be decreased
or a coarser granularity of aggregation must be used so that the flow
data fits into less space.

Increasing the collection interval effectively stores data in the meter;
usage data in transit is limited by the effective bandwidth of the
virtual link between the meter and the meter reader, and since these
limited network resources are usually also used to carry user data (the
purpose of the network), the level of traffic flow measurement traffic
should be kept to an affordable fraction of the bandwidth.
("Affordable" is a policy decision made by the Network Operations
personnel).  At any rate, it must be understood that the operations
below do not represent the setting of independent variables; on the
contrary, each of the values set has a direct and measurable effect on
the behaviour of the other variables.

Network management operations follow:

- MANAGER and METER READER IDENTIFICATION: The manager should ensure
that meters are read by the correct set of meter readers, and take
meter readers so identified should be prepared to poll if necessary
and accept data from the appropriate meters.  Alternate meter
readers may be identified in case both the primary manager and the
primary meter reader are unavailable.  Similarly, alternate
managers may be identified.

- REPORTING INTERVAL CONTROL: The usual reporting interval should be
selected to cope with normal traffic patterns.  However, it may be
possible for a meter to exhaust its memory during traffic spikes
even with a correctly set reporting interval.  Some mechanism
should be available for the meter to tell the manager that it is in
danger of exhausting its memory (by declaring a 'high water'
condition), and for the manager to arbitrate (by decreasing the
polling interval, letting nature take its course, or by telling the
meter to ask for help sooner next time).

- GRANULARITY CONTROL: Granularity control is a catch-all for all the
parameters that can be tuned and traded to optimise the system's
ability to reliably measure and store information on all the
traffic (or as close to all the traffic as an administration
requires).  Granularity

-  Controls flow-id granularities for each interface, and

-  Determines the number of buckets into which user traffic will
be lumped together.

Brownlee, Mills, Ruth                                          [Page 32]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

Since granularity is controlled by the meter's current rule set,
the manager can only change it by requesting the meter to switch to
a different rule set.  The new rule set could be downloaded when
required, or it could have been downloaded as part of the meter's
initial configuration.

- FLOW LIFETIME CONTROL: Flow termination parameters include timeout
parameters for obsoleting inactive flows and removing them from
tables, and maximum flow lifetimes.  This is intertwined with
reporting interval and granularity, and must be set in accordance
with the other parameters.

6.3 Exception Conditions

Exception conditions must be handled, particularly occasions when the
meter runs out of space for flow data.  Since, to prevent counting any
packet twice, packets can only be counted in a single flow at any given
time, discarding records will result in the loss of information.  The
mechanisms to deal with this are as follows:

- METER OUTAGES: In case of impending meter outages (controlled
restarts, etc.)  the meter could send a trap to the manager.  The
manager could then request one or more meter readers to pick up the
data from the meter.

Following an uncontrolled meter outage such as a power failure, the
meter could send a trap to the manager indicating that it has
rule set and advise the meter reader(s) that the meter is running
again.  Alternatively, the meter reader may discover from its
regular poll that a meter has failed and restarted.  It could then
advise the manager of this, instead of relying on a trap from the
meter.

- METER READER OUTAGES: If the collection system is down or isolated,
the meter should try to inform the manager of its failure to
communicate with the collection system.  Usage data is maintained
in the flows' rolling counters, and can be recovered when the meter

- MANAGER OUTAGES: If the manager fails for any reason, the meter
should continue measuring and the meter reader(s) should keep
gathering usage records.

Brownlee, Mills, Ruth                                          [Page 33]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

- BUFFER PROBLEMS: The network manager may realise that there is a
'low memory' condition in the meter.  This can usually be
attributed to the interaction between the following controls:

-  The reporting interval is too infrequent,

-  The reporting granularity is too fine, or

-  The throughput/bandwidth of circuits carrying the usage data is
too low.

The manager may change any of these parameters in response to the
meter (or meter reader's) plea for help.

6.4 Standard Rule Sets

Although the rule table is a flexible tool, it can also become very
complex.  It may be helpful to develop some rule sets for common
applications:

- PROTOCOL TYPE: The meter records packets by protocol type.  This
will be the default rule table for Traffic Flow Meters.

- ADJACENT SYSTEMS: The meter records packets by the MAC address of
the Adjacent Systems (neighbouring originator or next-hop).
(Variants on this table are "report source" or "report sink" only.)
This strategy might be used by a regional or backbone network which
wants to know how much aggregate traffic flows to or from its
subscriber networks.

- END SYSTEMS: The meter records packets by the IP address pair
contained in the packet.  (Variants on this table are "report
source" or "report sink" only.)  This strategy might be used by an
End System network to get detailed host traffic matrix usage data.

- TRANSPORT TYPE: The meter records packets by transport address; for
IP packets this provides usage information for the various IP
services.

- HYBRID SYSTEMS: Combinations of the above, e.g.  for one interface
This strategy might be used by an enterprise network to learn
detail about local usage and use an aggregate count for the shared
regional network.

Brownlee, Mills, Ruth                                          [Page 34]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

7 Security Considerations

7.1 Threat Analysis

A traffic flow measurement system may be subject to the following kinds
of attacks:

- UNAUTHORIZED USE OF SYSTEM RESOURCES: An attacker may wish to gain
advantage or cause mischief (e.g.  denial of service) by subverting
any of the system elements - meters, meter readers or managers.

- UNAUTHORIZED DISCLOSURE OF DATA: Any data that is sensitive to
disclosure can be read through active or passive attacks unless it
is suitably protected.  Usage data may or may not be of this type.
Control messages, traps, etc.  are not likely to be considered
sensitive to disclosure.

- UNAUTHORIZED ALTERATION, REPLACEMENT OR DESTRUCTION OF DATA:
Similarly, any data whose integrity is sensitive can be altered,
replaced/injected or deleted through active or passive attacks
unless it is suitably protected.  Attackers may modify message
streams to falsify usage data or interfere with the proper
operation of the traffic flow measurement system.  Therefore, all
messages, both those containing usage data and those containing
control data, should be considered vulnerable to such attacks.

7.2 Countermeasures

The following countermeasures are recommended to address the possible
threats enumerated above:

- UNAUTHORIZED USE OF SYSTEM RESOURCES is countered through the use
of authentication and access control services.

- UNAUTHORIZED DISCLOSURE OF DATA is countered through the use of a
confidentiality (encryption) service.

- UNAUTHORIZED ALTERATION, REPLACEMENT OR DESTRUCTION OF DATA is
countered through the use of an integrity service.

An Internet Accounting system must address all of these concerns.  Since
a high degree of protection is required, the use of strong cryptographic

Brownlee, Mills, Ruth                                          [Page 35]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

methodologies is recommended.  The security requirements for
communication between pairs of accounting system elements are summarized
in the table below.  It is assumed that meters do not communicate with
other meters, and that meter readers do not communicate directly with
other meter readers (if synchronization is required, it is handled by
the manager, see Section 2.5).  Each entry in the table indicates which
kinds of security services are required.  Basically, the requirements
are as follows:

Security Service Requirements for RTFM elements

+------------------------------------------------------------------+
| from\to |    meter     | meter reader | application |  manager   |
|---------+--------------+--------------+-------------+------------|
| meter   |     N/A      |  authent     |     N/A     |  authent   |
|         |              |  acc ctrl    |             |  acc ctrl  |
|         |              |  integrity   |             |            |
|         |              |  confid **   |             |            |
|---------+--------------+--------------+-------------+------------|
| meter   |   authent    |     N/A      |  authent    |  authent   |
| reader  |   acc ctrl   |              |  acc ctrl   |  acc ctrl  |
|         |              |              |  integrity  |            |
|         |              |              |  confid **  |            |
|---------+--------------+--------------+-------------+------------|
| appl    |     N/A      |  authent     |             |            |
|         |              |  acc ctrl    |     ##      |    ##      |
|---------+--------------+--------------+-------------+------------|
| manager |  authent     |  authent     |     N/A     |  authent   |
|         |  acc ctrl    |  acc ctrl    |             |  acc ctrl  |
|         |  integrity   |  integrity   |             |  integrity |
+------------------------------------------------------------------+

N/A = Not Applicable    ** = optional    ## = outside RTFM scope

- When any two elements intercommunicate they should mutually
authenticate themselves to one another.  This is indicated by
'authent' in the table.  Once authentication is complete, an
element should check that the requested type of access is allowed;
this is indicated on the table by 'acc ctrl.'

- Whenever there is a transfer of information its integrity should be
protected.

- Whenever there is a transfer of usage data it should be possible to
ensure its confidentiality if it is deemed sensitive to disclosure.
This is indicated by 'confid' in the table.

Brownlee, Mills, Ruth                                          [Page 36]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

Security protocols are not specified in this document.  The system
elements' management and collection protocols are responsible for
providing sufficient data integrity, confidentiality, authentication and
access control services.

8 APPENDICES

8.1 Appendix A: Network Characterisation

Internet users have extraordinarily diverse requirements.  Networks
differ in size, speed, throughput, and processing power, among other
factors.  There is a range of traffic flow measurement capabilities and
requirements.  For traffic flow measurement purposes, the Internet may
be viewed as a continuum which changes in character as traffic passes
through the following representative levels:

International                    |
Backbones/National        ---------------
/                \
Regional/MidLevel     ----------   ----------
/     \     \ /    /     \
Stub/Enterprise     ---   ---   ---   ----   ----
|||   |||   |||   ||||   ||||
End-Systems/Hosts   xxx   xxx   xxx   xxxx   xxxx

Note that mesh architectures can also be built out of these components,
and that these are merely descriptive terms.  The nature of a single
network may encompass any or all of the descriptions below, although
some networks can be clearly identified as a single type.

BACKBONE networks are typically bulk carriers that connect other
networks.  Individual hosts (with the exception of network management
devices and backbone service hosts) typically are not directly connected
to backbones.

REGIONAL networks are closely related to backbones, and differ only in
size, the number of networks connected via each port, and geographical
coverage.  Regionals may have directly connected hosts, acting as hybrid
backbone/stub networks.  A regional network is a SUBSCRIBER to the
backbone.

STUB/ENTERPRISE networks connect hosts and local area networks.
STUB/ENTERPRISE networks are SUBSCRIBERS to regional and backbone
networks.

Brownlee, Mills, Ruth                                          [Page 37]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

END SYSTEMS, colloquially HOSTS, are SUBSCRIBERS to any of the above
networks.

Providing a uniform identification of the SUBSCRIBER in finer
granularity than that of end-system, (e.g.  user/account), is beyond the
scope of the current architecture, although an optional attribute in the
traffic flow measurement record may carry system-specific "accountable
(billable) party" labels so that meters can implement proprietary or
non-standard schemes for the attribution of network traffic to
responsible parties.

8.2 Appendix B: Recommended Traffic Flow Measurement Capabilities

Initial recommended traffic flow measurement conventions are outlined
here according to the following Internet building blocks.  It is
important to understand what complexity reporting introduces at each
network level.  Whereas the hierarchy is described top-down in the
previous section, reporting requirements are more easily addressed
bottom-up.

End-Systems
Stub Networks
Enterprise Networks
Regional Networks
Backbone Networks

END-SYSTEMS are currently responsible for allocating network usage to
end-users, if this capability is desired.  From the Internet Protocol
perspective, end-systems are the finest granularity that can be
identified without protocol modifications.  Even if a meter violated
protocol boundaries and tracked higher-level protocols, not all packets
could be correctly allocated by user, and the definition of user itself
varies widely from operating system to operating system (e.g.  how to
trace network usage back to users from shared processes).

STUB and ENTERPRISE networks will usually collect traffic data either by
is required in the local area network.  If no local reporting is
required, they may record usage information in the exit router to track
external traffic only.  (These are the only networks which routinely use
attributes to perform reporting at granularities finer than end-system

REGIONAL networks are intermediate networks.  In some cases, subscribers
will be enterprise networks, in which case the intermediate system
network address is sufficient to identify the regional's immediate

Brownlee, Mills, Ruth                                          [Page 38]


INTERNET-DRAFT      Traffic Flow Measurement: Architecture      Nov 1997

subscriber.  In other cases, individual hosts or a disjoint group of
hosts may constitute a subscriber.  Then end-system network address
pairs need to be tracked for those subscribers.  When the source may be
an aggregate entity (such as a network, or adjacent router representing
traffic from a world of hosts beyond) and the destination is a singular
entity (or vice versa), the meter is said to be operating as a HYBRID
system.

At the regional level, if the overhead is tolerable it may be

BACKBONE networks are the highest level networks operating at higher
link speeds and traffic levels.  The high volume of traffic will in most
cases preclude detailed traffic flow measurement.  Backbone networks

8.3 Appendix C: List of Defined Flow Attributes

This Appendix provides a checklist of the attributes defined to date;
others will be added later as the Traffic Measurement Architecture is
further developed.

0  Null
1  Flow Subscript                Integer    Flow table info

4  Source Interface              Integer    Source Address
8  Source Peer Type              Integer
11  Source Trans Type             Integer

14  Destination Interface         Integer    Destination Address
18  Destination PeerType          Integer
21  Destination TransType         Integer

Brownlee, Mills, Ruth                                          [Page 39]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

26  Rule Set Number               Integer    Meter attribute

27  Forward Bytes                 Counter    Source-to-Dest counters
28  Forward Packets               Counter
29  Reverse Bytes                 Counter    Dest-to-Source counters
30  Reverse Packets               Counter
31  First Time                    TimeTicks  Activity times
32  Last Active Time              TimeTicks
33  Source Subscriber ID          String     Session attributes
34  Destination Subscriber ID     String
35  Session ID                    String

36  Source Class                  Integer    'Computed' attributes
37  Destination Class             Integer
38  Flow Class                    Integer
39  Source Kind                   Integer
40  Destination Kind              Integer
41  Flow Kind                     Integer

50  MatchingStoD                  Integer    PME variable

65
..  'Extended' attributes (to be defined by the RTFM working group)
127

8.4 Appendix D: List of Meter Control Variables

Meter variables:
Flood Mark                    Percentage
Inactivity Timeout (seconds)  Integer

Current Rule Set Number       Integer
Standby Rule Set Number       Integer
High Water Mark               Percentage

Brownlee, Mills, Ruth                                          [Page 40]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

9 Acknowledgments

An initial draft of this document was produced under the auspices of the
IETF's Internet Accounting Working Group with assistance from SNMP, RMON
and SAAG working groups.  This version documents the implementation work
done by the Internet Accounting Working Group, and is intended to
provide a starting point for the Realtime Traffic Flow Measurement
Working Group.  Particular thanks are due to Stephen Stibler (IBM
Research) for his patient and careful comments during the preparation of
this draft.

10 References

[1] Mills, C., Hirsch, G. and Ruth, G., "Internet Accounting
Background", RFC 1272, Bolt Beranek and Newman Inc., Meridian
Technology Corporation, November 1991.

[2] International Standards Organisation (ISO), "Management
Framework," Part 4 of Information Processing Systems Open
Systems Interconnection Basic Reference Model, ISO 7498-4,
1994.

[3] IEEE 802.3/ISO 8802-3 Information Processing Systems -
Local Area Networks - Part 3:  Carrier sense multiple access
with collision detection (CSMA/CD) access method and physical
layer specifications, 2nd edition, September 21, 1990.

[4] Brownlee, N., "Traffic Flow Measurement:  Meter MIB,"
Internet Draft, 'Working draft' to become an experimental RFC.

Brownlee, Mills, Ruth                                          [Page 41]


INTERNET-DRAFT      Traffic Flow Measurement:  Architecture     Nov 1997

Nevil Brownlee
Information Technology Systems & Services
The University of Auckland

Phone: +64 9 373 7599 x8941
E-mail: n.brownlee@auckland.ac.nz

Cyndi Mills
GTE Laboratories, Inc

Phone: +1 617 466 4278
E-mail: cmills@gte.com

Greg Ruth
GTE Laboratories, Inc

Phone: +1 617 466 2448
E-mail: gruth@gte.com

Expires Mar 1998

Brownlee, Mills, Ruth                                          [Page 42]



Html markup produced by rfcmarkup 1.109, available from https://tools.ietf.org/tools/rfcmarkup/