Internet Engineering Task Force (IETF) B. Trammell
Request for Comments: 7015 ETH Zurich
Category: Standards Track A. Wagner
ISSN: 2070-1721 Consecom AG
B. Claise
Cisco Systems, Inc.
September 2013
Flow Aggregation for the IP Flow Information Export (IPFIX) Protocol
Abstract
This document provides a common implementation-independent basis for
the interoperable application of the IP Flow Information Export
(IPFIX) protocol to the handling of Aggregated Flows, which are IPFIX
Flows representing packets from multiple Original Flows sharing some
set of common properties. It does this through a detailed
terminology and a descriptive Intermediate Aggregation Process
architecture, including a specification of methods for Original Flow
counting and counter distribution across intervals.
Status of This Memo
This is an Internet Standards Track document.
This document is a product of the Internet Engineering Task Force
(IETF). It represents the consensus of the IETF community. It has
received public review and has been approved for publication by the
Internet Engineering Steering Group (IESG). Further information on
Internet Standards is available in Section 2 of RFC 5741.
Information about the current status of this document, any errata,
and how to provide feedback on it may be obtained at
http://www.rfc-editor.org/info/rfc7015.
Trammell, et al. Standards Track [Page 1]
RFC 7015 IPFIX Aggregation September 2013
Copyright Notice
Copyright (c) 2013 IETF Trust and the persons identified as the
document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal
Provisions Relating to IETF Documents
(http://trustee.ietf.org/license-info) in effect on the date of
publication of this document. Please review these documents
carefully, as they describe your rights and restrictions with respect
to this document. Code Components extracted from this document must
include Simplified BSD License text as described in Section 4.e of
the Trust Legal Provisions and are provided without warranty as
described in the Simplified BSD License.
Table of Contents
1. Introduction ....................................................3
1.1. IPFIX Protocol Overview ....................................4
1.2. IPFIX Documents Overview ...................................5
2. Terminology .....................................................5
3. Use Cases for IPFIX Aggregation .................................7
4. Architecture for Flow Aggregation ...............................8
4.1. Aggregation within the IPFIX Architecture ..................8
4.2. Intermediate Aggregation Process Architecture .............12
4.2.1. Correlation and Normalization ......................14
5. IP Flow Aggregation Operations .................................15
5.1. Temporal Aggregation through Interval Distribution ........15
5.1.1. Distributing Values across Intervals ...............16
5.1.2. Time Composition ...................................18
5.1.3. External Interval Distribution .....................19
5.2. Spatial Aggregation of Flow Keys ..........................19
5.2.1. Counting Original Flows ............................21
5.2.1. Counting Distinct Key Values .......................22
5.3. Spatial Aggregation of Non-key Fields .....................22
5.3.1. Counter Statistics .................................22
5.3.2. Derivation of New Values from Flow Keys and
Non-key fields .....................................23
5.4. Aggregation Combination ...................................23
6. Additional Considerations and Special Cases in Flow
Aggregation ....................................................24
6.1. Exact versus Approximate Counting during Aggregation ......24
6.2. Delay and Loss Introduced by the IAP ......................24
6.3. Considerations for Aggregation of Sampled Flows ...........24
6.4. Considerations for Aggregation of Heterogeneous Flows .....25
7. Export of Aggregated IP Flows Using IPFIX ......................25
7.1. Time Interval Export ......................................25
7.2. Flow Count Export .........................................25
7.2.1. originalFlowsPresent ...............................26
Trammell, et al. Standards Track [Page 2]
RFC 7015 IPFIX Aggregation September 2013
7.2.2. originalFlowsInitiated .............................26
7.2.3. originalFlowsCompleted .............................26
7.2.4. deltaFlowCount .....................................26
7.3. Distinct Host Export ......................................27
7.3.1. distinctCountOfSourceIPAddress .....................27
7.3.2. distinctCountOfDestinationIPAddress ................27
7.3.3. distinctCountOfSourceIPv4Address ...................27
7.3.4. distinctCountOfDestinationIPv4Address ..............28
7.3.5. distinctCountOfSourceIPv6Address ...................28
7.3.6. distinctCountOfDestinationIPv6Address ..............28
7.4. Aggregate Counter Distribution Export .....................28
7.4.1. Aggregate Counter Distribution Options Template ....29
7.4.2. valueDistributionMethod Information Element ........29
8. Examples .......................................................31
8.1. Traffic Time Series per Source ............................32
8.2. Core Traffic Matrix .......................................37
8.3. Distinct Source Count per Destination Endpoint ............42
8.4. Traffic Time Series per Source with Counter Distribution ..44
9. Security Considerations ........................................46
10. IANA Considerations ...........................................46
11. Acknowledgments ...............................................46
12. References ....................................................47
12.1. Normative References .....................................47
12.2. Informative References ...................................47
1. Introduction
The assembly of packet data into Flows serves a variety of different
purposes, as noted in the requirements [RFC3917] and applicability
statement [RFC5472] for the IP Flow Information Export (IPFIX)
protocol [RFC7011]. Aggregation beyond the Flow level, into records
representing multiple Flows, is a common analysis and data reduction
technique as well, with applicability to large-scale network data
analysis, archiving, and inter-organization exchange. This
applicability in large-scale situations, in particular, led to the
inclusion of aggregation as part of the IPFIX Mediation Problem
Statement [RFC5982], and the definition of an Intermediate
Aggregation Process in the Mediator framework [RFC6183].
Aggregation is used for analysis and data reduction in a wide variety
of applications, for example, in traffic matrix calculation,
generation of time series data for visualizations or anomaly
detection, or data reduction for long-term trending and storage.
Depending on the keys used for aggregation, it may additionally have
an anonymizing effect on the data: for example, aggregation
operations that eliminate IP addresses make it impossible to later
directly identify nodes using those addresses.
Trammell, et al. Standards Track [Page 3]
RFC 7015 IPFIX Aggregation September 2013
Aggregation, as defined and described in this document, covers the
applications defined in [RFC5982], including Sections 5.1 "Adjusting
Flow Granularity", 5.4 "Time Composition", and 5.5 "Spatial
Composition". However, Section 4.2 of this document specifies a more
flexible architecture for an Intermediate Aggregation Process than
that envisioned by the original Mediator work [RFC5982]. Instead of
a focus on these specific limited use cases, the Intermediate
Aggregation Process is specified to cover any activity commonly
described as "Flow aggregation". This architecture is intended to
describe any such activity without reference to the specific
implementation of aggregation.
An Intermediate Aggregation Process may be applied to data collected
from multiple Observation Points, as it is natural to use aggregation
for data reduction when concentrating measurement data. This
document specifically does not address the protocol issues that arise
when combining IPFIX data from multiple Observation Points and
exporting from a single Mediator, as these issues are general to
IPFIX Mediation; they are therefore treated in detail in the
Mediation Protocol document [IPFIX-MED-PROTO].
Since Aggregated Flows as defined in the following section are
essentially Flows, the IPFIX protocol [RFC7011] can be used to
export, and the IPFIX File Format [RFC5655] can be used to store,
aggregated data "as is"; there are no changes necessary to the
protocol. This document provides a common basis for the application
of IPFIX to the handling of aggregated data, through a detailed
terminology, Intermediate Aggregation Process architecture, and
methods for Original Flow counting and counter distribution across
intervals. Note that Sections 5, 6, and 7 of this document are
normative.
1.1. IPFIX Protocol Overview
In the IPFIX protocol, { type, length, value } tuples are expressed
in Templates containing { type, length } pairs, specifying which
{ value } fields are present in data records conforming to the
Template, giving great flexibility as to what data is transmitted.
Since Templates are sent very infrequently compared with Data
Records, this results in significant bandwidth savings. Various
different data formats may be transmitted simply by sending new
Templates specifying the { type, length } pairs for the new data
format. See [RFC7011] for more information.
The IPFIX Information Element Registry [IANA-IPFIX] defines a large
number of standard Information Elements that provide the necessary {
type } information for Templates. The use of standard elements
enables interoperability among different vendors' implementations.
Trammell, et al. Standards Track [Page 4]
RFC 7015 IPFIX Aggregation September 2013
Additionally, non-standard enterprise-specific elements may be
defined for private use.
1.2. IPFIX Documents Overview
"Specification of the IP Flow Information Export (IPFIX) Protocol for
the Exchange of Flow Information" [RFC7011] and its associated
documents define the IPFIX protocol, which provides network engineers
and administrators with access to IP traffic Flow information.
IPFIX has a formal description of IPFIX Information Elements, their
names, types, and additional semantic information, as specified in
the IPFIX Information Model [RFC7012]. The IPFIX Information Element
registry [IANA-IPFIX] is maintained by IANA. New Information Element
definitions can be added to this registry subject to an Expert Review
[RFC5226], with additional process considerations described in
[RFC7013].
"Architecture for IP Flow Information Export" [RFC5470] defines the
architecture for the export of measured IP Flow information out of an
IPFIX Exporting Process to an IPFIX Collecting Process and the basic
terminology used to describe the elements of this architecture, per
the requirements defined in "Requirements for IP Flow Information
Export" [RFC3917]. The IPFIX protocol document [RFC7011] covers the
details of the method for transporting IPFIX Data Records and
Templates via a congestion-aware transport protocol from an IPFIX
Exporting Process to an IPFIX Collecting Process.
"IP Flow Information Export (IPFIX) Mediation: Problem Statement"
[RFC5982] introduces the concept of IPFIX Mediators, and defines the
use cases for which they were designed; "IP Flow Information Export
(IPFIX) Mediation: Framework" [RFC6183] then provides an
architectural framework for Mediators. Protocol-level issues (e.g.,
Template and Observation Domain handling across Mediators) are
covered by "Operation of the IP Flow Information Export (IPFIX)
Protocol on IPFIX Mediators" [IPFIX-MED-PROTO].
This document specifies an Intermediate Process for Flow aggregation
that may be applied at an IPFIX Mediator, as well as at an original
Observation Point prior to export, or for analysis and data reduction
purposes after receipt at a Collecting Process.
2. Terminology
Terms used in this document that are defined in the Terminology
section of the IPFIX protocol document [RFC7011] are to be
interpreted as defined there.
Trammell, et al. Standards Track [Page 5]
RFC 7015 IPFIX Aggregation September 2013
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
"SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this
document are to be interpreted as described in [RFC2119].
In addition, this document defines the following terms:
Aggregated Flow: A Flow, as defined by [RFC7011], derived from a set
of zero or more Original Flows within a defined Aggregation
Interval. Note that an Aggregated Flow is defined in the context
of an Intermediate Aggregation Process only. Once an Aggregated
Flow is exported, it is essentially a Flow as in [RFC7011] and can
be treated as such.
Intermediate Aggregation Process: an Intermediate Aggregation
Process (IAP), as in [RFC6183], that aggregates records, based
upon a set of Flow Keys or functions applied to fields from the
record.
Aggregation Interval: A time interval imposed upon an Aggregated
Flow. Intermediate Aggregation Processes may use a regular
Aggregation Interval (e.g., "every five minutes", "every calendar
month"), though regularity is not necessary. Aggregation
intervals may also be derived from the time intervals of the
Original Flows being aggregated.
Partially Aggregated Flow: A Flow during processing within an
Intermediate Aggregation Process; refers to an intermediate data
structure during aggregation within the Intermediate Aggregation
Process architecture detailed in Section 4.2.
Original Flow: A Flow given as input to an Intermediate Aggregation
Process in order to generate Aggregated Flows.
Contributing Flow: An Original Flow that is partially or completely
represented within an Aggregated Flow. Each Aggregated Flow is
made up of zero or more Contributing Flows, and an Original Flow
may contribute to zero or more Aggregated Flows.
Original Exporter: The Exporter from which the Original Flows are
received; meaningful only when an IAP is deployed at a Mediator.
The terminology presented herein improves the precision of, but does
not supersede or contradict the terms related to, Mediation and
aggregation defined in the Mediation Problem Statement [RFC5982] and
the Mediation Framework [RFC6183] documents. Within this document,
the terminology defined in this section is to be considered
normative.
Trammell, et al. Standards Track [Page 6]
RFC 7015 IPFIX Aggregation September 2013
3. Use Cases for IPFIX Aggregation
Aggregation, as a common data reduction method used in traffic data
analysis, has many applications. When used with a regular
Aggregation Interval and Original Flows containing timing
information, it generates time series data from a collection of Flows
with discrete intervals, as in the example in Section 8.1. This time
series data is itself useful for a wide variety of analysis tasks,
such as generating input for network anomaly detection systems or
driving visualizations of volume per time for traffic with specific
characteristics. As a second example, traffic matrix calculation
from Flow data, as shown in Section 8.2 is inherently an aggregation
action, by spatially aggregating the Flow Key down to input or output
interface, address prefix, or autonomous system (AS).
Irregular or data-dependent Aggregation Intervals and key aggregation
operations can also be used to provide adaptive aggregation of
network Flow data. Here, full Flow Records can be kept for Flows of
interest, while Flows deemed "less interesting" to a given
application can be aggregated. For example, in an IPFIX Mediator
equipped with traffic classification capabilities for security
purposes, potentially malicious Flows could be exported directly,
while known-good or probably-good Flows (e.g., normal web browsing)
could be exported simply as time series volumes per web server.
Aggregation can also be applied to final analysis of stored Flow
data, as shown in the example in Section 8.3. All such aggregation
applications in which timing information is not available or not
important can be treated as if an infinite Aggregation Interval
applies.
Note that an Intermediate Aggregation Process that removes
potentially sensitive information as identified in [RFC6235] may tend
to have an anonymizing effect on the Aggregated Flows as well;
however, any application of aggregation as part of a data protection
scheme should ensure that all the issues raised in [RFC6235] are
addressed, specifically Sections 4 ("Anonymization of IP Flow Data"),
7.2 ("IPFIX-Specific Anonymization Guidelines"), and 9 ("Security
Considerations").
While much of the discussion in this document, and all of the
examples, apply to the common case that the Original Flows to be
aggregated are all of the same underlying type (i.e., are represented
with identical Templates or compatible Templates containing a core
set Information Elements that can be freely converted to one
another), and that each packet observed by the Metering Process
associated with the Original Exporter is represented, this is not a
necessary assumption. Aggregation can also be applied as part of a
Trammell, et al. Standards Track [Page 7]
RFC 7015 IPFIX Aggregation September 2013
technique using both aggregation and correlation to pull together
multiple views of the same traffic from different Observation Points
using different Templates. For example, consider a set of
applications running at different Observation Points for different
purposes -- one generating Flows with round-trip times for passive
performance measurement, and one generating billing records. Once
correlated, these Flows could be used to produce Aggregated Flows
containing both volume and performance information together. The
correlation and normalization operation described in Section 4.2.1
handles this specific case of correlation. Flow correlation in the
general case is outside the scope of this document.
4. Architecture for Flow Aggregation
This section specifies the architecture of the Intermediate
Aggregation Process and how it fits into the IPFIX architecture.
4.1. Aggregation within the IPFIX Architecture
An Intermediate Aggregation Process could be deployed at any of three
places within the IPFIX architecture. While aggregation is most
commonly done within a Mediator that collects Original Flows from an
Original Exporter and exports Aggregated Flows, aggregation can also
occur before initial export, or after final collection, as shown in
Figure 1. The presence of an IAP at any of these points is, of
course, optional.
Trammell, et al. Standards Track [Page 8]
RFC 7015 IPFIX Aggregation September 2013
+===========================================+
| IPFIX Exporter +----------------+ |
| | Metering Proc. | |
| +-----------------+ +----------------+ |
| | Metering Proc. | or | IAP | |
| +-----------------+----+----------------+ |
| | Exporting Process | |
| +-|----------------------------------|--+ |
+===|==================================|====+
| |
+===|===========================+ |
| | Aggregating Mediator | |
+ +-V-------------------+ | |
| | Collecting Process | | |
+ +---------------------+ | |
| | IAP | | |
+ +---------------------+ | |
| | Exporting Process | | |
+ +-|-------------------+ | |
+===|===========================+ |
| |
+===|==================================|=====+
| | Collector | |
| +-V----------------------------------V-+ |
| | Collecting Process | |
| +------------------+-------------------+ |
| | IAP | |
| +-------------------+ |
| (Aggregation | File Writer | |
for Storage) +-----------|-------+ |
+================================|===========+
|
+------V-----------+
| IPFIX File |
+------------------+
Figure 1: Potential Aggregation Locations
The Mediator use case is further shown in Figures A and B in
[RFC6183].
Aggregation can be applied for either intermediate or final analytic
purposes. In certain circumstances, it may make sense to export
Aggregated Flows directly after metering, for example, if the
Exporting Process is applied to drive a time series visualization, or
when Flow data export bandwidth is restricted and Flow or packet
sampling is not an option. Note that this case, where the
Aggregation Process is essentially integrated into the Metering
Trammell, et al. Standards Track [Page 9]
RFC 7015 IPFIX Aggregation September 2013
Process, is basically covered by the IPFIX architecture [RFC5470]:
the Flow Keys used are simply a subset of those that would normally
be used, and time intervals may be chosen other than those available
from the cache policies customarily offered by the Metering Process.
A Metering Process in this arrangement MAY choose to simulate the
generation of larger Flows in order to generate Original Flow counts,
if the application calls for compatibility with an Intermediate
Aggregation Process deployed in a separate location.
In the specific case that an Intermediate Aggregation Process is
employed for data reduction for storage purposes, it can take
Original Flows from a Collecting Process or File Reader and pass
Aggregated Flows to a File Writer for storage.
Deployment of an Intermediate Aggregation Process within a Mediator
[RFC5982] is a much more flexible arrangement. Here, the Mediator
consumes Original Flows and produces Aggregated Flows; this
arrangement is suited to any of the use cases detailed in Section 3.
In a Mediator, Original Flows from multiple sources can also be
aggregated into a single stream of Aggregated Flows. The
architectural specifics of this arrangement are not addressed in this
document, which is concerned only with the aggregation operation
itself. See [IPFIX-MED-PROTO] for details.
The data paths into and out of an Intermediate Aggregation Process
are shown in Figure 2.
Trammell, et al. Standards Track [Page 10]
RFC 7015 IPFIX Aggregation September 2013
packets --+ IPFIX Messages IPFIX Files
| | |
V V V
+==================+ +====================+ +=============+
| Metering Process | | Collecting Process | | File Reader |
| | +====================+ +=============+
| (Original Flows | | |
| or direct | | Original Flows |
| aggregation) | V V
+ - - - - - - - - -+======================================+
| Intermediate Aggregation Process (IAP) |
+=========================================================+
| Aggregated Aggregated |
| Flows Flows |
V V
+===================+ +=============+
| Exporting Process | | File Writer |
+===================+ +=============+
| |
V V
IPFIX Messages IPFIX Files
Figure 2: Data Paths through the Aggregation Process
Note that as Aggregated Flows are IPFIX Flows, an Intermediate
Aggregation Process may aggregate already Aggregated Flows from an
upstream IAP as well as Original Flows from an upstream Original
Exporter or Metering Process.
Aggregation may also need to correlate Original Flows from multiple
Metering Processes, each according to a different Template with
different Flow Keys and values. This arrangement is shown in Figure
3; in this case, the correlation and normalization operation
described in Section 4.2.1 handles merging the Original Flows before
aggregation.
Trammell, et al. Standards Track [Page 11]
RFC 7015 IPFIX Aggregation September 2013
packets --+---------------------+------------------+
| | |
V V V
+====================+ +====================+ +====================+
| Metering Process 1 | | Metering Process 2 | | Metering Process n |
+====================+ +====================+ +====================+
| | Original Flows |
V V V
+==================================================================+
| Intermediate Aggregation Process + correlation / normalization |
+==================================================================+
| Aggregated Aggregated |
| Flows Flows |
V V
+===================+ +=============+
| Exporting Process | | File Writer |
+===================+ +=============+
| |
+------------> IPFIX Messages <----------+
Figure 3: Aggregating Original Flows from Multiple Metering Processes
4.2. Intermediate Aggregation Process Architecture
Within this document, an Intermediate Aggregation Process can be seen
as hosting a function composed of four types of operations on
Partially Aggregated Flows, as illustrated in Figure 4: interval
distribution (temporal), key aggregation (spatial), value aggregation
(spatial), and aggregate combination. "Partially Aggregated Flows",
as defined in Section 2, are essentially the intermediate results of
aggregation, internal to the Intermediate Aggregation Process.
Trammell, et al. Standards Track [Page 12]
RFC 7015 IPFIX Aggregation September 2013
Original Flows / Original Flows requiring correlation
+=============|===================|===================|=============+
| | Intermediate | Aggregation | Process |
| | V V |
| | +-----------------------------------------------+ |
| | | (optional) correlation and normalization | |
| | +-----------------------------------------------+ |
| | | |
| V V |
| +--------------------------------------------------------------+ |
| | interval distribution (temporal) | |
| +--------------------------------------------------------------+ |
| | ^ | ^ | |
| | | Partially Aggregated | | | |
| V | Flows V | | |
| +-------------------+ +--------------------+ | |
| | key aggregation |<------| value aggregation | | |
| | (spatial) |------>| (spatial) | | |
| +-------------------+ +--------------------+ | |
| | | | |
| | Partially Aggregated | | |
| V Flows V V |
| +--------------------------------------------------------------+ |
| | aggregate combination | |
| +--------------------------------------------------------------+ |
| | |
+=======================================|===========================+
V
Aggregated Flows
Figure 4: Conceptual Model of Aggregation Operations within an IAP
Interval distribution: a temporal aggregation operation that imposes
an Aggregation Interval on the Partially Aggregated Flow. This
Aggregation Interval may be regular, irregular, or derived from
the timing of the Original Flows themselves. Interval
distribution is discussed in detail in Section 5.1.
Key aggregation: a spatial aggregation operation that results in the
addition, modification, or deletion of Flow Key fields in the
Partially Aggregated Flows. New Flow Keys may be derived from
existing Flow Keys (e.g., looking up an AS number (ASN) for an IP
address), or "promoted" from specific non-key fields (e.g., when
aggregating Flows by packet count per Flow). Key aggregation can
also add new non-key fields derived from Flow Keys that are
deleted during key aggregation: mainly counters of unique reduced
keys. Key aggregation is discussed in detail in Section 5.2.
Trammell, et al. Standards Track [Page 13]
RFC 7015 IPFIX Aggregation September 2013
Value aggregation: a spatial aggregation operation that results in
the addition, modification, or deletion of non-key fields in the
Partially Aggregated Flows. These non-key fields may be "demoted"
from existing key fields, or derived from existing key or non-key
fields. Value aggregation is discussed in detail in Section 5.3.
Aggregate combination: an operation combining multiple Partially
Aggregated Flows having undergone interval distribution, key
aggregation, and value aggregation that share Flow Keys and
Aggregation Intervals into a single Aggregated Flow per set of
Flow Key values and Aggregation Interval. Aggregate combination
is discussed in detail in Section 5.4.
Correlation and normalization: an optional operation that applies
when accepting Original Flows from Metering Processes that export
different views of essentially the same Flows before aggregation.
The details of correlation and normalization are specified in
Section 4.2.1, below.
The first three of these operations may be carried out any number of
times in any order, either on Original Flows or on the results of one
of the operations above, with one caveat: since Flows carry their own
interval data, any spatial aggregation operation implies a temporal
aggregation operation, so at least one interval distribution step,
even if implicit, is required by this architecture. This is shown as
the first step for the sake of simplicity in the diagram above. Once
all aggregation operations are complete, aggregate combination
ensures that for a given Aggregation Interval, set of Flow Key
values, and Observation Domain, only one Flow is produced by the
Intermediate Aggregation Process.
This model describes the operations within a single Intermediate
Aggregation Process, and it is anticipated that most aggregation will
be applied within a single process. However, as the steps in the
model may be applied in any order and aggregate combination is
idempotent, any number of Intermediate Aggregation Processes
operating in series can be modeled as a single process. This allows
aggregation operations to be flexibly distributed across any number
of processes, should application or deployment considerations so
dictate.
4.2.1. Correlation and Normalization
When accepting Original Flows from multiple Metering Processes, each
of which provides a different view of the Original Flow as seen from
the point of view of the IAP, an optional correlation and
normalization operation combines each of these single Flow Records
Trammell, et al. Standards Track [Page 14]
RFC 7015 IPFIX Aggregation September 2013
into a set of unified Partially Aggregated Flows before applying
interval distribution. These unified Flows appear as if they had
been measured at a single Metering Process that used the union of the
set of Flow Keys and non-key fields of all Metering Processes sending
Original Flows to the IAP.
Since, due to export errors or other slight irregularities in Flow
metering, the multiple views may not be completely consistent;
normalization involves applying a set of corrections that are
specific to the aggregation application in order to ensure
consistency in the unified Flows.
In general, correlation and normalization should take multiple views
of essentially the same Flow, as determined by the configuration of
the operation itself, and render them into a single unified Flow.
Flows that are essentially different should not be unified by the
correlation and normalization operation. This operation therefore
requires enough information about the configuration and deployment of
Metering Processes from which it correlates Original Flows in order
to make this distinction correctly and consistently.
The exact steps performed to correlate and normalize Flows in this
step are application, implementation, and deployment specific, and
will not be further specified in this document.
5. IP Flow Aggregation Operations
As stated in Section 2, an Aggregated Flow is simply an IPFIX Flow
generated from Original Flows by an Intermediate Aggregation Process.
Here, we detail the operations by which this is achieved within an
Intermediate Aggregation Process.
5.1. Temporal Aggregation through Interval Distribution
Interval distribution imposes a time interval on the resulting
Aggregated Flows. The selection of an interval is specific to the
given aggregation application. Intervals may be derived from the
Original Flows themselves (e.g., an interval may be selected to cover
the entire time containing the set of all Flows sharing a given Key,
as in Time Composition, described in Section 5.1.2) or externally
imposed; in the latter case the externally imposed interval may be
regular (e.g., every five minutes) or irregular (e.g., to allow for
different time resolutions at different times of day, under different
network conditions, or indeed for different sets of Original Flows).
The length of the imposed interval itself has trade-offs. Shorter
intervals allow higher-resolution aggregated data and, in streaming
applications, faster reaction time. Longer intervals generally lead
Trammell, et al. Standards Track [Page 15]
RFC 7015 IPFIX Aggregation September 2013
to greater data reduction and simplified counter distribution.
Specifically, counter distribution is greatly simplified by the
choice of an interval longer than the duration of longest Original
Flow, itself generally determined by the Original Flow's Metering
Process active timeout; in this case, an Original Flow can contribute
to at most two Aggregated Flows, and the more complex value
distribution methods become inapplicable.
| | | |
| |<--Flow A-->| | | |
| |<--Flow B-->| | |
| |<-------------Flow C-------------->| |
| | | |
| interval 0 | interval 1 | interval 2 |
Figure 5: Illustration of Interval Distribution
In Figure 5, we illustrate three common possibilities for interval
distribution as applies with regular intervals to a set of three
Original Flows. For Flow A, the start and end times lie within the
boundaries of a single interval 0; therefore, Flow A contributes to
only one Aggregated Flow. Flow B, by contrast, has the same duration
but crosses the boundary between intervals 0 and 1; therefore, it
will contribute to two Aggregated Flows, and its counters must be
distributed among these Flows; though, in the two-interval case, this
can be simplified somewhat simply by picking one of the two intervals
or proportionally distributing between them. Only Flows like Flow A
and Flow B will be produced when the interval is chosen to be longer
than the duration of longest Original Flow, as above. More
complicated is the case of Flow C, which contributes to more than two
Aggregated Flows and must have its counters distributed according to
some policy as in Section 5.1.1.
5.1.1. Distributing Values across Intervals
In general, counters in Aggregated Flows are treated the same as in
any Flow. Each counter is independently calculated as if it were
derived from the set of packets in the Original Flow. For example,
delta counters are summed, the most recent total count for each
Original Flow taken then summed across Flows, and so on.
When the Aggregation Interval is guaranteed to be longer than the
longest Original Flow, a Flow can cross at most one Interval
boundary, and will therefore contribute to at most two Aggregated
Flows. Most common in this case is to arbitrarily but consistently
choose to account the Original Flow's counters either to the first or
to the last Aggregated Flow to which it could contribute.
Trammell, et al. Standards Track [Page 16]
RFC 7015 IPFIX Aggregation September 2013
However, this becomes more complicated when the Aggregation Interval
is shorter than the longest Original Flow in the source data. In
such cases, each Original Flow can incompletely cover one or more
time intervals, and apply to one or more Aggregated Flows. In this
case, the Intermediate Aggregation Process must distribute the
counters in the Original Flows across one or more resulting
Aggregated Flows. There are several methods for doing this, listed
here in roughly increasing order of complexity and accuracy; most of
these are necessary only in specialized cases.
End Interval: The counters for an Original Flow are added to the
counters of the appropriate Aggregated Flow containing the end
time of the Original Flow.
Start Interval: The counters for an Original Flow are added to the
counters of the appropriate Aggregated Flow containing the start
time of the Original Flow.
Mid Interval: The counters for an Original Flow are added to the
counters of a single appropriate Aggregated Flow containing some
timestamp between start and end time of the Original Flow.
Simple Uniform Distribution: Each counter for an Original Flow is
divided by the number of time intervals the Original Flow covers
(i.e., of appropriate Aggregated Flows sharing the same Flow
Keys), and this number is added to each corresponding counter in
each Aggregated Flow.
Proportional Uniform Distribution: This is like simple uniform
distribution, but accounts for the fractional portions of a time
interval covered by an Original Flow in the first and last time
interval. Each counter for an Original Flow is divided by the
number of time _units_ the Original Flow covers, to derive a mean
count rate. This rate is then multiplied by the number of time
units in the intersection of the duration of the Original Flow and
the time interval of each Aggregated Flow.
Simulated Process: Each counter of the Original Flow is distributed
among the intervals of the Aggregated Flows according to some
function the Intermediate Aggregation Process uses based upon
properties of Flows presumed to be like the Original Flow. For
example, Flow Records representing bulk transfer might follow a
more or less proportional uniform distribution, while interactive
processes are far more bursty.
Direct: The Intermediate Aggregation Process has access to the
original packet timings from the packets making up the Original
Flow, and uses these to distribute or recalculate the counters.
Trammell, et al. Standards Track [Page 17]
RFC 7015 IPFIX Aggregation September 2013
A method for exporting the distribution of counters across multiple
Aggregated Flows is detailed in Section 7.4. In any case, counters
MUST be distributed across the multiple Aggregated Flows in such a
way that the total count is preserved, within the limits of accuracy
of the implementation. This property allows data to be aggregated
and re-aggregated with negligible loss of original count information.
To avoid confusion in interpretation of the aggregated data, all the
counters in a given Aggregated Flow MUST be distributed via the same
method.
More complex counter distribution methods generally require that the
interval distribution process track multiple "current" time intervals
at once. This may introduce some delay into the aggregation
operation, as an interval should only expire and be available for
export when no additional Original Flows applying to the interval are
expected to arrive at the Intermediate Aggregation Process.
Note, however, that since there is no guarantee that Flows from the
Original Exporter will arrive in any given order, whether for
transport-specific reasons (i.e., UDP reordering) or reasons specific
to the implementation of the Metering Process or Exporting Process,
even simpler distribution methods may need to deal with Flows
arriving in an order other than start time or end time. Therefore,
the use of larger intervals does not obviate the need to buffer
Partially Aggregated Flows within "current" time intervals, to ensure
the IAP can accept Flow time intervals in any arrival order. More
generally, the interval distribution process SHOULD accept Flow start
and end times in the Original Flows in any reasonable order. The
expiration of intervals in interval distribution operations is
dependent on implementation and deployment requirements, and it MUST
be made configurable in contexts in which "reasonable order" is not
obvious at implementation time. This operation may lead to delay and
loss introduced by the IAP, as detailed in Section 6.2.
5.1.2. Time Composition
Time Composition, as in Section 5.4 of [RFC5982] (or interval
combination), is a special case of aggregation, where interval
distribution imposes longer intervals on Flows with matching keys and
"chained" start and end times, without any key reduction, in order to
join long-lived Flows that may have been split (e.g., due to an
active timeout shorter than the actual duration of the Flow). Here,
no Key aggregation is applied, and the Aggregation Interval is chosen
on a per-Flow basis to cover the interval spanned by the set of
Aggregated Flows. This may be applied alone in order to normalize
split Flows, or it may be applied in combination with other
aggregation functions in order to obtain more accurate Original Flow
counts.
Trammell, et al. Standards Track [Page 18]
RFC 7015 IPFIX Aggregation September 2013
5.1.3. External Interval Distribution
Note that much of the difficulty of interval distribution at an IAP
can be avoided simply by configuring the original Exporters to
synchronize the time intervals in the Original Flows with the desired
aggregation interval. The resulting Original Flows would then be
split to align perfectly with the time intervals imposed during
interval imposition, as shown in Figure 6, though this may reduce
their usefulness for non-aggregation purposes. This approach allows
the Intermediate Aggregation Process to use Start Interval or End
Interval distribution, while having equivalent information to that
available to direct interval distribution.
| | | |
|<----Flow D---->|<----Flow E---->|<----Flow F---->|
| | | |
| interval 0 | interval 1 | interval 2 |
Figure 6: Illustration of External Interval Distribution
5.2. Spatial Aggregation of Flow Keys
Key aggregation generates a new set of Flow Key values for the
Aggregated Flows from the Original Flow Key and non-key fields in the
Original Flows or from correlation of the Original Flow information
with some external source. There are two basic operations here.
First, Aggregated Flow Keys may be derived directly from Original
Flow Keys through reduction, or they may be derived by the dropping
of fields or precision in the Original Flow Keys. Second, Aggregated
Flow Keys may be derived through replacement, e.g., by removing one
or more fields from the Original Flow and replacing them with fields
derived from the removed fields. Replacement may refer to external
information (e.g., IP to AS number mappings). Replacement may apply
to Flow Keys as well as non-key fields. For example, consider an
application that aggregates Original Flows by packet count (i.e.,
generating an Aggregated Flow for all one-packet Flows, one for all
two-packet Flows, and so on). This application would promote the
packet count to a Flow Key.
Key aggregation may also result in the addition of new non-key fields
to the Aggregated Flows, namely, Original Flow counters and unique
reduced key counters. These are treated in more detail in Sections
5.2.1 and 5.2.2, respectively.
In any key aggregation operation, reduction and/or replacement may be
applied any number of times in any order. Which of these operations
are supported by a given implementation is implementation and
application dependent.
Trammell, et al. Standards Track [Page 19]
RFC 7015 IPFIX Aggregation September 2013
Original Flow Keys
+---------+---------+----------+----------+-------+-----+
| src ip4 | dst ip4 | src port | dst port | proto | tos |
+---------+---------+----------+----------+-------+-----+
| | | | | |
retain mask /24 X X X X
| |
V V
+---------+-------------+
| src ip4 | dst ip4 /24 |
+---------+-------------+
Aggregated Flow Keys (by source address and destination /24 network)
Figure 7: Illustration of Key Aggregation by Reduction
Figure 7 illustrates an example reduction operation, aggregation by
source address and destination /24 network. Here, the port,
protocol, and type-of-service information is removed from the Flow
Key, the source address is retained, and the destination address is
masked by dropping the lower 8 bits.
Original Flow Keys
+---------+---------+----------+----------+-------+-----+
| src ip4 | dst ip4 | src port | dst port | proto | tos |
+---------+---------+----------+----------+-------+-----+
| | | | | |
V V | | | |
+-------------------+ X X X X
| ASN lookup table |
+-------------------+
| |
V V
+---------+---------+
| src asn | dst asn |
+---------+---------+
Aggregated Flow Keys (by source and destination ASN)
Figure 8: Illustration of Key Aggregation
by Reduction and Replacement
Figure 8 illustrates an example reduction and replacement operation,
aggregation by source and destination Border Gateway Protocol (BGP)
Autonomous System Number (ASN) without ASN information available in
the Original Flow. Here, the port, protocol, and type-of-service
Trammell, et al. Standards Track [Page 20]
RFC 7015 IPFIX Aggregation September 2013
information is removed from the Flow Keys, while the source and
destination addresses are run though an IP address to ASN lookup
table, and the Aggregated Flow Keys are made up of the resulting
source and destination ASNs.
5.2.1. Counting Original Flows
When aggregating multiple Original Flows into an Aggregated Flow, it
is often useful to know how many Original Flows are present in the
Aggregated Flow. Section 7.2 introduces four new Information Elements
to export these counters.
There are two possible ways to count Original Flows, which we call
conservative and non-conservative. Conservative Flow counting has
the property that each Original Flow contributes exactly one to the
total Flow count within a set of Aggregated Flows. In other words,
conservative Flow counters are distributed just as any other counter
during interval distribution, except each Original Flow is assumed to
have a Flow count of one. When a count for an Original Flow must be
distributed across a set of Aggregated Flows, and a distribution
method is used that does not account for that Original Flow
completely within a single Aggregated Flow, conservative Flow
counting requires a fractional representation.
By contrast, non-conservative Flow counting is used to count how many
Contributing Flows are represented in an Aggregated Flow. Flow
counters are not distributed in this case. An Original Flow that is
present within N Aggregated Flows would add N to the sum of non-
conservative Flow counts, one to each Aggregated Flow. In other
words, the sum of conservative Flow counts over a set of Aggregated
Flows is always equal to the number of Original Flows, while the sum
of non-conservative Flow counts is strictly greater than or equal to
the number of Original Flows.
For example, consider Flows A, B, and C as illustrated in Figure 5.
Assume that the key aggregation step aggregates the keys of these
three Flows to the same aggregated Flow Key, and that start interval
counter distribution is in effect. The conservative Flow count for
interval 0 is 3 (since Flows A, B, and C all begin in this interval),
and for the other two intervals is 0. The non-conservative Flow
count for interval 0 is also 3 (due to the presence of Flows A, B,
and C), for interval 1 is 2 (Flows B and C), and for interval 2 is 1
(Flow C). The sum of the conservative counts 3 + 0 + 0 = 3, the
number of Original Flows; while the sum of the non-conservative
counts 3 + 2 + 1 = 6.
Trammell, et al. Standards Track [Page 21]
RFC 7015 IPFIX Aggregation September 2013
Note that the active and inactive timeouts used to generate Original
Flows, as well as the cache policy used to generate those Flows, have
an effect on how meaningful either the conservative or non-
conservative Flow count will be during aggregation. In general,
Original Exporters using the IPFIX Configuration Model SHOULD be
configured to export Flows with equal or similar activeTimeout and
inactiveTimeout configuration values, and the same cacheMode, as
defined in [RFC6728]. Original Exporters not using the IPFIX
Configuration Model SHOULD be configured equivalently.
5.2.2. Counting Distinct Key Values
One common case in aggregation is counting distinct key values that
were reduced away during key aggregation. The most common use case
for this is counting distinct hosts per Flow Key; for example, in
host characterization or anomaly detection, distinct sources per
destination or distinct destinations per source are common metrics.
These new non-key fields are added during key aggregation.
For such applications, Information Elements for distinct counts of
IPv4 and IPv6 addresses are defined in Section 7.3. These are named
distinctCountOf(KeyName). Additional such Information Elements
should be registered with IANA on an as-needed basis.
5.3. Spatial Aggregation of Non-key Fields
Aggregation operations may also lead to the addition of value fields
that are demoted from key fields or are derived from other value
fields in the Original Flows. Specific cases of this are treated in
the subsections below.
5.3.1. Counter Statistics
Some applications of aggregation may benefit from computing different
statistics than those native to each non-key field (e.g., flags are
natively combined via union and delta counters by summing). For
example, minimum and maximum packet counts per Flow, mean bytes per
packet per Contributing Flow, and so on. Certain Information
Elements for these applications are already provided in the IANA
IPFIX Information Elements registry [IANA-IPFIX] (e.g.,
minimumIpTotalLength).
A complete specification of additional aggregate counter statistics
is outside the scope of this document, and should be added in the
future to the IANA IPFIX Information Elements registry on a per-
application, as-needed basis.
Trammell, et al. Standards Track [Page 22]
RFC 7015 IPFIX Aggregation September 2013
5.3.2. Derivation of New Values from Flow Keys and Non-key fields
More complex operations may lead to other derived fields being
generated from the set of values or Flow Keys reduced away during
aggregation. A prime example of this is sample entropy calculation.
This counts distinct values and frequency, so it is similar to
distinct key counting as in Section 5.2.2; however, it may be applied
to the distribution of values for any Flow field.
Sample entropy calculation provides a one-number normalized
representation of the value spread and is useful for anomaly
detection. The behavior of entropy statistics is such that a small
number of keys showing up very often drives the entropy value down
towards zero, while a large number of keys, each showing up with
lower frequency, drives the entropy value up.
Entropy statistics are generally useful for identifier keys, such as
IP addresses, port numbers, AS numbers, etc. They can also be
calculated on Flow length, Flow duration fields, and the like, even
if this generally yields less distinct value shifts when the traffic
mix changes.
As a practical example, one host scanning a lot of other hosts will
drive source IP entropy down and target IP entropy up. A similar
effect can be observed for ports. This pattern can also be caused by
the scan-traffic of a fast Internet worm. A second example would be
a Distributed Denial of Service (DDoS) flooding attack against a
single target (or small number of targets) that drives source IP
entropy up and target IP entropy down.
A complete specification of additional derived values or entropy
Information Elements is outside the scope of this document. Any such
Information Elements should be added in the future to the IANA IPFIX
Information Elements registry on a per-application, as-needed basis.
5.4. Aggregation Combination
Interval distribution and key aggregation together may generate
multiple Partially Aggregated Flows covering the same time interval
with the same set of Flow Key values. The process of combining these
Partially Aggregated Flows into a single Aggregated Flow is called
aggregation combination. In general, non-Key values from multiple
Contributing Flows are combined using the same operation by which
values are combined from packets to form Flows for each Information
Element. Delta counters are summed, flags are unioned, and so on.
Trammell, et al. Standards Track [Page 23]
RFC 7015 IPFIX Aggregation September 2013
6. Additional Considerations and Special Cases in Flow Aggregation
6.1. Exact versus Approximate Counting during Aggregation
In certain circumstances, particularly involving aggregation by
devices with limited resources, and in situations where exact
aggregated counts are less important than relative magnitudes (e.g.,
driving graphical displays), counter distribution during key
aggregation may be performed by approximate counting means (e.g.,
Bloom filters). The choice to use approximate counting is
implementation and application dependent.
6.2. Delay and Loss Introduced by the IAP
When accepting Original Flows in export order from traffic captured
live, the Intermediate Aggregation Process waits for all Original
Flows that may contribute to a given interval during interval
distribution. This is generally dominated by the active timeout of
the Metering Process measuring the Original Flows. For example, with
Metering Processes configured with a five-minute active timeout, the
Intermediate Aggregation Process introduces a delay of at least five
minutes to all exported Aggregated Flows to ensure it has received
all Original Flows. Note that when aggregating Flows from multiple
Metering Processes with different active timeouts, the delay is
determined by the maximum active timeout.
In certain circumstances, additional delay at the original Exporter
may cause an IAP to close an interval before the last Original
Flow(s) accountable to the interval arrives. In this case, the IAP
MAY drop the late Original Flow(s). Accounting of Flows lost at an
Intermediate Process due to such issues is covered in
[IPFIX-MED-PROTO].
6.3. Considerations for Aggregation of Sampled Flows
The accuracy of Aggregated Flows may also be affected by sampling of
the Original Flows, or sampling of packets making up the Original
Flows. At the time of writing, the effect of sampling on Flow
aggregation is still an open research question. However, to maximize
the comparability of Aggregated Flows, aggregation of sampled Flows
should only be applied to Original Flows sampled using the same
sampling rate and sampling algorithm, Flows created from packets
sampled using the same sampling rate and sampling algorithm, or
Original Flows that have been normalized as if they had the same
sampling rate and algorithm before aggregation. For more on packet
sampling within IPFIX, see [RFC5476]. For more on Flow sampling
within the IPFIX Mediator framework, see [RFC7014].
Trammell, et al. Standards Track [Page 24]
RFC 7015 IPFIX Aggregation September 2013
6.4. Considerations for Aggregation of Heterogeneous Flows
Aggregation may be applied to Original Flows from different sources
and of different types (i.e., represented using different, perhaps
wildly different Templates). When the goal is to separate the
heterogeneous Original Flows and aggregate them into heterogeneous
Aggregated Flows, each aggregation should be done at its own
Intermediate Aggregation Process. The Observation Domain ID on the
Messages containing the output Aggregated Flows can be used to
identify the different Processes and to segregate the output.
However, when the goal is to aggregate these Flows into a single
stream of Aggregated Flows representing one type of data, and if the
Original Flows may represent the same original packet at two
different Observation Points, the Original Flows should be correlated
by the correlation and normalization operation within the IAP to
ensure that each packet is only represented in a single Aggregated
Flow or set of Aggregated Flows differing only by aggregation
interval.
7. Export of Aggregated IP Flows Using IPFIX
In general, Aggregated Flows are exported in IPFIX as any other Flow.
However, certain aspects of Aggregated Flow export benefit from
additional guidelines or new Information Elements to represent
aggregation metadata or information generated during aggregation.
These are detailed in the following subsections.
7.1. Time Interval Export
Since an Aggregated Flow is simply a Flow, the existing timestamp
Information Elements in the IPFIX Information Model (e.g.,
flowStartMilliseconds, flowEndNanoseconds) are sufficient to specify
the time interval for aggregation. Therefore, no new aggregation-
specific Information Elements for exporting time interval information
are necessary.
Each Aggregated Flow carrying timing information SHOULD contain both
an interval start and interval end timestamp.
7.2. Flow Count Export
The following four Information Elements are defined to count Original
Flows as discussed in Section 5.2.1.
Trammell, et al. Standards Track [Page 25]
RFC 7015 IPFIX Aggregation September 2013
7.2.1. originalFlowsPresent
Description: The non-conservative count of Original Flows
contributing to this Aggregated Flow. Non-conservative counts
need not sum to the original count on re-aggregation.
Abstract Data Type: unsigned64
Data Type Semantics: deltaCounter
ElementID: 375
7.2.2. originalFlowsInitiated
Description: The conservative count of Original Flows whose first
packet is represented within this Aggregated Flow. Conservative
counts must sum to the original count on re-aggregation.
Abstract Data Type: unsigned64
Data Type Semantics: deltaCounter
ElementID: 376
7.2.3. originalFlowsCompleted
Description: The conservative count of Original Flows whose last
packet is represented within this Aggregated Flow. Conservative
counts must sum to the original count on re-aggregation.
Abstract Data Type: unsigned64
Data Type Semantics: deltaCounter
ElementID: 377
7.2.4. deltaFlowCount
Description: The conservative count of Original Flows contributing
to this Aggregated Flow; may be distributed via any of the methods
expressed by the valueDistributionMethod Information Element.
Abstract Data Type: unsigned64
Data Type Semantics: deltaCounter
ElementID: 3
Trammell, et al. Standards Track [Page 26]
RFC 7015 IPFIX Aggregation September 2013
7.3. Distinct Host Export
The following six Information Elements represent the distinct counts
of source and destination network-layer addresses used to export
distinct host counts reduced away during key aggregation.
7.3.1. distinctCountOfSourceIPAddress
Description: The count of distinct source IP address values for
Original Flows contributing to this Aggregated Flow, without
regard to IP version. This Information Element is preferred to
the IP-version-specific counters, unless it is important to
separate the counts by version.
Abstract Data Type: unsigned64
Data Type Semantics: totalCounter
ElementID: 378
7.3.2. distinctCountOfDestinationIPAddress
Description: The count of distinct destination IP address values for
Original Flows contributing to this Aggregated Flow, without
regard to IP version. This Information Element is preferred to
the version-specific counters below, unless it is important to
separate the counts by version.
Abstract Data Type: unsigned64
Data Type Semantics: totalCounter
ElementID: 379
7.3.3. distinctCountOfSourceIPv4Address
Description: The count of distinct source IPv4 address values for
Original Flows contributing to this Aggregated Flow.
Abstract Data Type: unsigned32
Data Type Semantics: totalCounter
ElementID: 380
Trammell, et al. Standards Track [Page 27]
RFC 7015 IPFIX Aggregation September 2013
7.3.4. distinctCountOfDestinationIPv4Address
Description: The count of distinct destination IPv4 address values
for Original Flows contributing to this Aggregated Flow.
Abstract Data Type: unsigned32
Data Type Semantics: totalCounter
ElementID: 381
7.3.5. distinctCountOfSourceIPv6Address
Description: The count of distinct source IPv6 address values for
Original Flows contributing to this Aggregated Flow.
Abstract Data Type: unsigned64
Data Type Semantics: totalCounter
ElementID: 382
7.3.6. distinctCountOfDestinationIPv6Address
Description: The count of distinct destination IPv6 address values
for Original Flows contributing to this Aggregated Flow.
Abstract Data Type: unsigned64
Data Type Semantics: totalCounter
ElementID: 383
7.4. Aggregate Counter Distribution Export
When exporting counters distributed among Aggregated Flows, as
described in Section 5.1.1, the Exporting Process MAY export an
Aggregate Counter Distribution Option Record for each Template
describing Aggregated Flow records; this Options Template is
described below. It uses the valueDistributionMethod Information
Element, also defined below. Since, in many cases, distribution is
simple, accounting the counters from Contributing Flows to the first
Interval to which they contribute, this is the default situation, for
which no Aggregate Counter Distribution Record is necessary;
Aggregate Counter Distribution Records are only applicable in more
exotic situations, such as using an Aggregation Interval smaller than
the durations of Original Flows.
Trammell, et al. Standards Track [Page 28]
RFC 7015 IPFIX Aggregation September 2013
7.4.1. Aggregate Counter Distribution Options Template
This Options Template defines the Aggregate Counter Distribution
Record, which allows the binding of a value distribution method to a
Template ID. The scope is the Template ID, whose uniqueness, per
[RFC7011], is local to the Transport Session and Observation Domain
that generated the Template ID. This is used to signal to the
Collecting Process how the counters were distributed. The fields are
as below:
+-----------------------------+-------------------------------------+
| IE | Description |
+-----------------------------+-------------------------------------+
| templateId [scope] | The Template ID of the Template |
| | defining the Aggregated Flows to |
| | which this distribution option |
| | applies. This Information Element |
| | MUST be defined as a Scope field. |
| valueDistributionMethod | The method used to distribute the |
| | counters for the Aggregated Flows |
| | defined by the associated Template. |
+-----------------------------+-------------------------------------+
7.4.2. valueDistributionMethod Information Element
Description: A description of the method used to distribute the
counters from Contributing Flows into the Aggregated Flow records
described by an associated scope, generally a Template. The
method is deemed to apply to all the non-Key Information Elements
in the referenced scope for which value distribution is a valid
operation; if the originalFlowsInitiated and/or
originalFlowsCompleted Information Elements appear in the
Template, they are not subject to this distribution method, as
they each infer their own distribution method. This is intended
to be a complete set of possible value distribution methods; it is
encoded as follows:
Trammell, et al. Standards Track [Page 29]
RFC 7015 IPFIX Aggregation September 2013
+-------+-----------------------------------------------------------+
| Value | Description |
+-------+-----------------------------------------------------------+
| 0 | Unspecified: The counters for an Original Flow are |
| | explicitly not distributed according to any other method |
| | defined for this Information Element; use for arbitrary |
| | distribution, or distribution algorithms not described by |
| | any other codepoint. |
| | --------------------------------------------------------- |
| | |
| 1 | Start Interval: The counters for an Original Flow are |
| | added to the counters of the appropriate Aggregated Flow |
| | containing the start time of the Original Flow. This |
| | should be assumed the default if value distribution |
| | information is not available at a Collecting Process for |
| | an Aggregated Flow. |
| | --------------------------------------------------------- |
| | |
| 2 | End Interval: The counters for an Original Flow are added |
| | to the counters of the appropriate Aggregated Flow |
| | containing the end time of the Original Flow. |
| | --------------------------------------------------------- |
| | |
| 3 | Mid Interval: The counters for an Original Flow are added |
| | to the counters of a single appropriate Aggregated Flow |
| | containing some timestamp between start and end time of |
| | the Original Flow. |
| | --------------------------------------------------------- |
| | |
| 4 | Simple Uniform Distribution: Each counter for an Original |
| | Flow is divided by the number of time intervals the |
| | Original Flow covers (i.e., of appropriate Aggregated |
| | Flows sharing the same Flow Key), and this number is |
| | added to each corresponding counter in each Aggregated |
| | Flow. |
| | --------------------------------------------------------- |
| | |
| 5 | Proportional Uniform Distribution: Each counter for an |
| | Original Flow is divided by the number of time units the |
| | Original Flow covers, to derive a mean count rate. This |
| | mean count rate is then multiplied by the number of time |
| | units in the intersection of the duration of the Original |
| | Flow and the time interval of each Aggregated Flow. |
| | This is like simple uniform distribution, but accounts |
| | for the fractional portions of a time interval covered by |
| | an Original Flow in the first and last time interval. |
| | --------------------------------------------------------- |
Trammell, et al. Standards Track [Page 30]
RFC 7015 IPFIX Aggregation September 2013
| | --------------------------------------------------------- |
| 6 | Simulated Process: Each counter of the Original Flow is |
| | distributed among the intervals of the Aggregated Flows |
| | according to some function the Intermediate Aggregation |
| | Process uses based upon properties of Flows presumed to |
| | be like the Original Flow. This is essentially an |
| | assertion that the Intermediate Aggregation Process has |
| | no direct packet timing information but is nevertheless |
| | not using one of the other simpler distribution methods. |
| | The Intermediate Aggregation Process specifically makes |
| | no assertion as to the correctness of the simulation. |
| | --------------------------------------------------------- |
| | |
| 7 | Direct: The Intermediate Aggregation Process has access |
| | to the original packet timings from the packets making up |
| | the Original Flow, and uses these to distribute or |
| | recalculate the counters. |
+-------+-----------------------------------------------------------+
Abstract Data Type: unsigned8
ElementID: 384
8. Examples
In these examples, the same data, described by the same Template,
will be aggregated multiple different ways; this illustrates the
various different functions that could be implemented by Intermediate
Aggregation Processes. Templates are shown in IESpec format as
introduced in [RFC7013]. The source data format is a simplified
Flow: timestamps, traditional 5-tuple, and octet count; the Flow Key
fields are the 5-tuple. The Template is shown in Figure 9.
flowStartMilliseconds(152)[8]
flowEndMilliseconds(153)[8]
sourceIPv4Address(8)[4]{key}
destinationIPv4Address(12)[4]{key}
sourceTransportPort(7)[2]{key}
destinationTransportPort(11)[2]{key}
protocolIdentifier(4)[1]{key}
octetDeltaCount(1)[8]
Figure 9: Input Template for Examples
The data records given as input to the examples in this section are
shown below; timestamps are given in H:MM:SS.sss format. In this and
subsequent figures, flowStartMilliseconds is shown in H:MM:SS.sss
format as 'start time', flowEndMilliseconds is shown in H:MM:SS.sss
Trammell, et al. Standards Track [Page 31]
RFC 7015 IPFIX Aggregation September 2013
format as 'end time', sourceIPv4Address is shown as 'source ip4' with
the following 'port' representing sourceTransportPort,
destinationIPv4Address is shown as 'dest ip4' with the following
'port' representing destinationTransportPort, protocolIdentifier is
shown as 'pt', and octetDeltaCount as 'oct'.
start time |end time |source ip4 |port |dest ip4 |port|pt| oct
9:00:00.138 9:00:00.138 192.0.2.2 47113 192.0.2.131 53 17 119
9:00:03.246 9:00:03.246 192.0.2.2 22153 192.0.2.131 53 17 83
9:00:00.478 9:00:03.486 192.0.2.2 52420 198.51.100.2 443 6 1637
9:00:07.172 9:00:07.172 192.0.2.3 56047 192.0.2.131 53 17 111
9:00:07.309 9:00:14.861 192.0.2.3 41183 198.51.100.67 80 6 16838
9:00:03.556 9:00:19.876 192.0.2.2 17606 198.51.100.68 80 6 11538
9:00:25.210 9:00:25.210 192.0.2.3 47113 192.0.2.131 53 17 119
9:00:26.358 9:00:30.198 192.0.2.3 48458 198.51.100.133 80 6 2973
9:00:29.213 9:01:00.061 192.0.2.4 61295 198.51.100.2 443 6 8350
9:04:00.207 9:04:04.431 203.0.113.3 41256 198.51.100.133 80 6 778
9:03:59.624 9:04:06.984 203.0.113.3 51662 198.51.100.3 80 6 883
9:00:30.532 9:06:15.402 192.0.2.2 37581 198.51.100.2 80 6 15420
9:06:56.813 9:06:59.821 203.0.113.3 52572 198.51.100.2 443 6 1637
9:06:30.565 9:07:00.261 203.0.113.3 49914 198.51.100.133 80 6 561
9:06:55.160 9:07:05.208 192.0.2.2 50824 198.51.100.2 443 6 1899
9:06:49.322 9:07:05.322 192.0.2.3 34597 198.51.100.3 80 6 1284
9:07:05.849 9:07:09.625 203.0.113.3 58907 198.51.100.4 80 6 2670
9:10:45.161 9:10:45.161 192.0.2.4 22478 192.0.2.131 53 17 75
9:10:45.209 9:11:01.465 192.0.2.4 49513 198.51.100.68 80 6 3374
9:10:57.094 9:11:00.614 192.0.2.4 64832 198.51.100.67 80 6 138
9:10:59.770 9:11:02.842 192.0.2.3 60833 198.51.100.69 443 6 2325
9:02:18.390 9:13:46.598 203.0.113.3 39586 198.51.100.17 80 6 11200
9:13:53.933 9:14:06.605 192.0.2.2 19638 198.51.100.3 80 6 2869
9:13:02.864 9:14:08.720 192.0.2.3 40429 198.51.100.4 80 6 18289
Figure 10: Input Data for Examples
8.1. Traffic Time Series per Source
Aggregating Flows by source IP address in time series (i.e., with a
regular interval) can be used in subsequent heavy-hitter analysis and
as a source parameter for statistical anomaly detection techniques.
Here, the Intermediate Aggregation Process imposes an interval,
aggregates the key to remove all key fields other than the source IP
address, then combines the result into a stream of Aggregated Flows.
The imposed interval of five minutes is longer than the majority of
Flows; for those Flows crossing interval boundaries, the entire Flow
is accounted to the interval containing the start time of the Flow.
Trammell, et al. Standards Track [Page 32]
RFC 7015 IPFIX Aggregation September 2013
In this example, the Partially Aggregated Flows after each conceptual
operation in the Intermediate Aggregation Process are shown. These
are meant to be illustrative of the conceptual operations only, and
not to suggest an implementation (indeed, the example shown here
would not necessarily be the most efficient method for performing
these operations). Subsequent examples will omit the Partially
Aggregated Flows for brevity.
The input to this process could be any Flow Record containing a
source IP address and octet counter; consider for this example the
Template and data from the introduction. The Intermediate
Aggregation Process would then output records containing just
timestamps, source IP, and octetDeltaCount, as in Figure 11.
flowStartMilliseconds(152)[8]
flowEndMilliseconds(153)[8]
sourceIPv4Address(8)[4]
octetDeltaCount(1)[8]
Figure 11: Output Template for Time Series per Source
Trammell, et al. Standards Track [Page 33]
RFC 7015 IPFIX Aggregation September 2013
Assume the goal is to get 5-minute (300 s) time series of octet
counts per source IP address. The aggregation operations would then
be arranged as in Figure 12.
Original Flows
|
V
+-----------------------+
| interval distribution |
| * impose uniform |
| 300s time interval |
+-----------------------+
|
| Partially Aggregated Flows
V
+------------------------+
| key aggregation |
| * reduce key to only |
| sourceIPv4Address |
+------------------------+
|
| Partially Aggregated Flows
V
+-------------------------+
| aggregate combination |
| * sum octetDeltaCount |
+-------------------------+
|
V
Aggregated Flows
Figure 12: Aggregation Operations for Time Series per Source
After applying the interval distribution step to the source data in
Figure 10, only the time intervals have changed; the Partially
Aggregated Flows are shown in Figure 13. Note that interval
distribution follows the default Start Interval policy; that is, the
entire Flow is accounted to the interval containing the Flow's start
time.
Trammell, et al. Standards Track [Page 34]
RFC 7015 IPFIX Aggregation September 2013
start time |end time |source ip4 |port |dest ip4 |port|pt| oct
9:00:00.000 9:05:00.000 192.0.2.2 47113 192.0.2.131 53 17 119
9:00:00.000 9:05:00.000 192.0.2.2 22153 192.0.2.131 53 17 83
9:00:00.000 9:05:00.000 192.0.2.2 52420 198.51.100.2 443 6 1637
9:00:00.000 9:05:00.000 192.0.2.3 56047 192.0.2.131 53 17 111
9:00:00.000 9:05:00.000 192.0.2.3 41183 198.51.100.67 80 6 16838
9:00:00.000 9:05:00.000 192.0.2.2 17606 198.51.100.68 80 6 11538
9:00:00.000 9:05:00.000 192.0.2.3 47113 192.0.2.131 53 17 119
9:00:00.000 9:05:00.000 192.0.2.3 48458 198.51.100.133 80 6 2973
9:00:00.000 9:05:00.000 192.0.2.4 61295 198.51.100.2 443 6 8350
9:00:00.000 9:05:00.000 203.0.113.3 41256 198.51.100.133 80 6 778
9:00:00.000 9:05:00.000 203.0.113.3 51662 198.51.100.3 80 6 883
9:00:00.000 9:05:00.000 192.0.2.2 37581 198.51.100.2 80 6 15420
9:00:00.000 9:05:00.000 203.0.113.3 39586 198.51.100.17 80 6 11200
9:05:00.000 9:10:00.000 203.0.113.3 52572 198.51.100.2 443 6 1637
9:05:00.000 9:10:00.000 203.0.113.3 49914 197.51.100.133 80 6 561
9:05:00.000 9:10:00.000 192.0.2.2 50824 198.51.100.2 443 6 1899
9:05:00.000 9:10:00.000 192.0.2.3 34597 198.51.100.3 80 6 1284
9:05:00.000 9:10:00.000 203.0.113.3 58907 198.51.100.4 80 6 2670
9:10:00.000 9:15:00.000 192.0.2.4 22478 192.0.2.131 53 17 75
9:10:00.000 9:15:00.000 192.0.2.4 49513 198.51.100.68 80 6 3374
9:10:00.000 9:15:00.000 192.0.2.4 64832 198.51.100.67 80 6 138
9:10:00.000 9:15:00.000 192.0.2.3 60833 198.51.100.69 443 6 2325
9:10:00.000 9:15:00.000 192.0.2.2 19638 198.51.100.3 80 6 2869
9:10:00.000 9:15:00.000 192.0.2.3 40429 198.51.100.4 80 6 18289
Figure 13: Interval Imposition for Time Series per Source
After the key aggregation step, all Flow Keys except the source IP
address have been discarded, as shown in Figure 14. This leaves
duplicate Partially Aggregated Flows to be combined in the final
operation.
Trammell, et al. Standards Track [Page 35]
RFC 7015 IPFIX Aggregation September 2013
start time |end time |source ip4 |octets
9:00:00.000 9:05:00.000 192.0.2.2 119
9:00:00.000 9:05:00.000 192.0.2.2 83
9:00:00.000 9:05:00.000 192.0.2.2 1637
9:00:00.000 9:05:00.000 192.0.2.3 111
9:00:00.000 9:05:00.000 192.0.2.3 16838
9:00:00.000 9:05:00.000 192.0.2.2 11538
9:00:00.000 9:05:00.000 192.0.2.3 119
9:00:00.000 9:05:00.000 192.0.2.3 2973
9:00:00.000 9:05:00.000 192.0.2.4 8350
9:00:00.000 9:05:00.000 203.0.113.3 778
9:00:00.000 9:05:00.000 203.0.113.3 883
9:00:00.000 9:05:00.000 192.0.2.2 15420
9:00:00.000 9:05:00.000 203.0.113.3 11200
9:05:00.000 9:10:00.000 203.0.113.3 1637
9:05:00.000 9:10:00.000 203.0.113.3 561
9:05:00.000 9:10:00.000 192.0.2.2 1899
9:05:00.000 9:10:00.000 192.0.2.3 1284
9:05:00.000 9:10:00.000 203.0.113.3 2670
9:10:00.000 9:15:00.000 192.0.2.4 75
9:10:00.000 9:15:00.000 192.0.2.4 3374
9:10:00.000 9:15:00.000 192.0.2.4 138
9:10:00.000 9:15:00.000 192.0.2.3 2325
9:10:00.000 9:15:00.000 192.0.2.2 2869
9:10:00.000 9:15:00.000 192.0.2.3 18289
Figure 14: Key Aggregation for Time Series per Source
Aggregate combination sums the counters per key and interval; the
summations of the first two keys and intervals are shown in detail in
Figure 15.
Trammell, et al. Standards Track [Page 36]
RFC 7015 IPFIX Aggregation September 2013
start time |end time |source ip4 |octets
9:00:00.000 9:05:00.000 192.0.2.2 119
9:00:00.000 9:05:00.000 192.0.2.2 83
9:00:00.000 9:05:00.000 192.0.2.2 1637
9:00:00.000 9:05:00.000 192.0.2.2 11538
+ 9:00:00.000 9:05:00.000 192.0.2.2 15420
-----
= 9:00:00.000 9:05:00.000 192.0.2.2 28797
9:00:00.000 9:05:00.000 192.0.2.3 111
9:00:00.000 9:05:00.000 192.0.2.3 16838
9:00:00.000 9:05:00.000 192.0.2.3 119
+ 9:00:00.000 9:05:00.000 192.0.2.3 2973
-----
= 9:00:00.000 9:05:00.000 192.0.2.3 20041
Figure 15: Summation during Aggregate Combination
This can be applied to each set of Partially Aggregated Flows to
produce the final Aggregated Flows that are shown in Figure 16, as
exported by the Template in Figure 11.
start time |end time |source ip4 |octets
9:00:00.000 9:05:00.000 192.0.2.2 28797
9:00:00.000 9:05:00.000 192.0.2.3 20041
9:00:00.000 9:05:00.000 192.0.2.4 8350
9:00:00.000 9:05:00.000 203.0.113.3 12861
9:05:00.000 9:10:00.000 192.0.2.2 1899
9:05:00.000 9:10:00.000 192.0.2.3 1284
9:05:00.000 9:10:00.000 203.0.113.3 4868
9:10:00.000 9:15:00.000 192.0.2.2 2869
9:10:00.000 9:15:00.000 192.0.2.3 20614
9:10:00.000 9:15:00.000 192.0.2.4 3587
Figure 16: Aggregated Flows for Time Series per Source
8.2. Core Traffic Matrix
Aggregating Flows by source and destination ASN in time series is
used to generate core traffic matrices. The core traffic matrix
provides a view of the state of the routes within a network, and it
can be used for long-term planning of changes to network design based
on traffic demand. Here, imposed time intervals are generally much
longer than active Flow timeouts. The traffic matrix is reported in
terms of octets, packets, and flows, as each of these values may have
a subtly different effect on capacity planning.
Trammell, et al. Standards Track [Page 37]
RFC 7015 IPFIX Aggregation September 2013
This example demonstrates key aggregation using derived keys and
Original Flow counting. While some Original Flows may be generated
by Exporting Processes on forwarding devices, and therefore contain
the bgpSourceAsNumber and bgpDestinationAsNumber Information
Elements, Original Flows from Exporting Processes on dedicated
measurement devices without routing data contain only a
destinationIPv[46]Address. For these Flows, the Mediator must look
up a next-hop AS from an IP-to-AS table, replacing source and
destination addresses with ASNs. The table used in this example is
shown in Figure 17. (Note that due to limited example address space,
in this example we ignore the common practice of routing only blocks
of /24 or larger.)
prefix |ASN
192.0.2.0/25 64496
192.0.2.128/25 64497
198.51.100/24 64498
203.0.113.0/24 64499
Figure 17: Example ASN Map
The Template for Aggregated Flows produced by this example is shown
in Figure 18.
flowStartMilliseconds(152)[8]
flowEndMilliseconds(153)[8]
bgpSourceAsNumber(16)[4]
bgpDestinationAsNumber(17)[4]
octetDeltaCount(1)[8]
Figure 18: Output Template for Traffic Matrix
Assume the goal is to get 60-minute time series of octet counts per
source/destination ASN pair. The aggregation operations would then
be arranged as in Figure 19.
Trammell, et al. Standards Track [Page 38]
RFC 7015 IPFIX Aggregation September 2013
Original Flows
|
V
+-----------------------+
| interval distribution |
| * impose uniform |
| 3600s time interval|
+-----------------------+
|
| Partially Aggregated Flows
V
+------------------------+
| key aggregation |
| * reduce key to only |
| sourceIPv4Address + |
| destIPv4Address |
+------------------------+
|
V
+------------------------+
| key aggregation |
| * replace addresses |
| with ASN from map |
+------------------------+
|
| Partially Aggregated Flows
V
+-------------------------+
| aggregate combination |
| * sum octetDeltaCount |
+-------------------------+
|
V
Aggregated Flows
Figure 19: Aggregation Operations for Traffic Matrix
After applying the interval distribution step to the source data in
Figure 10, the Partially Aggregated Flows are shown in Figure 20.
Note that the Flows are identical to those in the interval
distribution step in the previous example, except the chosen interval
(1 hour, 3600 seconds) is different; therefore, all the Flows fit
into a single interval.
Trammell, et al. Standards Track [Page 39]
RFC 7015 IPFIX Aggregation September 2013
start time |end time |source ip4 |port |dest ip4 |port|pt| oct
9:00:00 10:00:00 192.0.2.2 47113 192.0.2.131 53 17 119
9:00:00 10:00:00 192.0.2.2 22153 192.0.2.131 53 17 83
9:00:00 10:00:00 192.0.2.2 52420 198.51.100.2 443 6 1637
9:00:00 10:00:00 192.0.2.3 56047 192.0.2.131 53 17 111
9:00:00 10:00:00 192.0.2.3 41183 198.51.100.67 80 6 16838
9:00:00 10:00:00 192.0.2.2 17606 198.51.100.68 80 6 11538
9:00:00 10:00:00 192.0.2.3 47113 192.0.2.131 53 17 119
9:00:00 10:00:00 192.0.2.3 48458 198.51.100.133 80 6 2973
9:00:00 10:00:00 192.0.2.4 61295 198.51.100.2 443 6 8350
9:00:00 10:00:00 203.0.113.3 41256 198.51.100.133 80 6 778
9:00:00 10:00:00 203.0.113.3 51662 198.51.100.3 80 6 883
9:00:00 10:00:00 192.0.2.2 37581 198.51.100.2 80 6 15420
9:00:00 10:00:00 203.0.113.3 52572 198.51.100.2 443 6 1637
9:00:00 10:00:00 203.0.113.3 49914 197.51.100.133 80 6 561
9:00:00 10:00:00 192.0.2.2 50824 198.51.100.2 443 6 1899
9:00:00 10:00:00 192.0.2.3 34597 198.51.100.3 80 6 1284
9:00:00 10:00:00 203.0.113.3 58907 198.51.100.4 80 6 2670
9:00:00 10:00:00 192.0.2.4 22478 192.0.2.131 53 17 75
9:00:00 10:00:00 192.0.2.4 49513 198.51.100.68 80 6 3374
9:00:00 10:00:00 192.0.2.4 64832 198.51.100.67 80 6 138
9:00:00 10:00:00 192.0.2.3 60833 198.51.100.69 443 6 2325
9:00:00 10:00:00 203.0.113.3 39586 198.51.100.17 80 6 11200
9:00:00 10:00:00 192.0.2.2 19638 198.51.100.3 80 6 2869
9:00:00 10:00:00 192.0.2.3 40429 198.51.100.4 80 6 18289
Figure 20: Interval Imposition for Traffic Matrix
The next steps are to discard irrelevant key fields and to replace
the source and destination addresses with source and destination ASNs
in the map; the results of these key aggregation steps are shown in
Figure 21.
Trammell, et al. Standards Track [Page 40]
RFC 7015 IPFIX Aggregation September 2013
start time |end time |source ASN |dest ASN |octets
9:00:00 10:00:00 AS64496 AS64497 119
9:00:00 10:00:00 AS64496 AS64497 83
9:00:00 10:00:00 AS64496 AS64498 1637
9:00:00 10:00:00 AS64496 AS64497 111
9:00:00 10:00:00 AS64496 AS64498 16838
9:00:00 10:00:00 AS64496 AS64498 11538
9:00:00 10:00:00 AS64496 AS64497 119
9:00:00 10:00:00 AS64496 AS64498 2973
9:00:00 10:00:00 AS64496 AS64498 8350
9:00:00 10:00:00 AS64499 AS64498 778
9:00:00 10:00:00 AS64499 AS64498 883
9:00:00 10:00:00 AS64496 AS64498 15420
9:00:00 10:00:00 AS64499 AS64498 1637
9:00:00 10:00:00 AS64499 AS64498 561
9:00:00 10:00:00 AS64496 AS64498 1899
9:00:00 10:00:00 AS64496 AS64498 1284
9:00:00 10:00:00 AS64499 AS64498 2670
9:00:00 10:00:00 AS64496 AS64497 75
9:00:00 10:00:00 AS64496 AS64498 3374
9:00:00 10:00:00 AS64496 AS64498 138
9:00:00 10:00:00 AS64496 AS64498 2325
9:00:00 10:00:00 AS64499 AS64498 11200
9:00:00 10:00:00 AS64496 AS64498 2869
9:00:00 10:00:00 AS64496 AS64498 18289
Figure 21: Key Aggregation for Traffic Matrix:
Reduction and Replacement
Finally, aggregate combination sums the counters per key and
interval. The resulting Aggregated Flows containing the traffic
matrix, shown in Figure 22, are then exported using the Template in
Figure 18. Note that these Aggregated Flows represent a sparse
matrix: AS pairs for which no traffic was received have no
corresponding record in the output.
start time end time source ASN dest ASN octets
9:00:00 10:00:00 AS64496 AS64497 507
9:00:00 10:00:00 AS64496 AS64498 86934
9:00:00 10:00:00 AS64499 AS64498 17729
Figure 22: Aggregated Flows for Traffic Matrix
The output of this operation is suitable for re-aggregation: that is,
traffic matrices from single links or Observation Points can be
aggregated through the same interval imposition and aggregate
combination steps in order to build a traffic matrix for an entire
network.
Trammell, et al. Standards Track [Page 41]
RFC 7015 IPFIX Aggregation September 2013
8.3. Distinct Source Count per Destination Endpoint
Aggregating Flows by destination address and port, and counting
distinct sources aggregated away, can be used as part of passive
service inventory and host characterization. This example shows
aggregation as an analysis technique, performed on source data stored
in an IPFIX File. As the Transport Session in this File is bounded,
removal of all timestamp information allows summarization of the
entire time interval contained within the interval. Removal of
timing information during interval imposition is equivalent to an
infinitely long imposed time interval. This demonstrates both how
infinite intervals work, and how unique counters work. The
aggregation operations are summarized in Figure 23.
Trammell, et al. Standards Track [Page 42]
RFC 7015 IPFIX Aggregation September 2013
Original Flows
|
V
+-----------------------+
| interval distribution |
| * discard timestamps |
+-----------------------+
|
| Partially Aggregated Flows
V
+----------------------------+
| value aggregation |
| * discard octetDeltaCount |
+----------------------------+
|
| Partially Aggregated Flows
V
+----------------------------+
| key aggregation |
| * reduce key to only |
| destIPv4Address + |
| destTransportPort, |
| * count distinct sources |
+----------------------------+
|
| Partially Aggregated Flows
V
+----------------------------------------------+
| aggregate combination |
| * no-op (distinct sources already counted) |
+----------------------------------------------+
|
V
Aggregated Flows
Figure 23: Aggregation Operations for Source Count
The Template for Aggregated Flows produced by this example is shown
in Figure 24.
destinationIPv4Address(12)[4]
destinationTransportPort(11)[2]
distinctCountOfSourceIPAddress(378)[8]
Figure 24: Output Template for Source Count
Trammell, et al. Standards Track [Page 43]
RFC 7015 IPFIX Aggregation September 2013
Interval distribution, in this case, merely discards the timestamp
information from the Original Flows in Figure 10, and as such is not
shown. Likewise, the value aggregation step simply discards the
octetDeltaCount value field. The key aggregation step reduces the
key to the destinationIPv4Address and destinationTransportPort,
counting the distinct source addresses. Since this is essentially
the output of this aggregation function, the aggregate combination
operation is a no-op; the resulting Aggregated Flows are shown in
Figure 25.
dest ip4 |port |dist src
192.0.2.131 53 3
198.51.100.2 80 1
198.51.100.2 443 3
198.51.100.67 80 2
198.51.100.68 80 2
198.51.100.133 80 2
198.51.100.3 80 3
198.51.100.4 80 2
198.51.100.17 80 1
198.51.100.69 443 1
Figure 25: Aggregated Flows for Source Count
8.4. Traffic Time Series per Source with Counter Distribution
Returning to the example in Section 8.1, note that our source data
contains some Flows with durations longer than the imposed interval
of five minutes. The default method for dealing with such Flows is
to account them to the interval containing the Flow's start time.
In this example, the same data is aggregated using the same
arrangement of operations and the same output Template as in
Section 8.1, but using a different counter distribution policy,
Simple Uniform Distribution, as described in Section 5.1.1. In order
to do this, the Exporting Process first exports the Aggregate Counter
Distribution Options Template, as in Figure 26.
templateId(12)[2]{scope}
valueDistributionMethod(384)[1]
Figure 26: Aggregate Counter Distribution Options Template
This Template is followed by an Aggregate Counter Distribution Record
described by this Template; assuming the output Template in Figure 11
has ID 257, this record would appear as in Figure 27.
Trammell, et al. Standards Track [Page 44]
RFC 7015 IPFIX Aggregation September 2013
template ID | value distribution method
257 4 (simple uniform)
Figure 27: Aggregate Counter Distribution Record
Following metadata export, the aggregation steps follow as before.
However, two long Flows are distributed across multiple intervals in
the interval imposition step, as indicated with "*" in Figure 28.
Note the uneven distribution of the three-interval, 11200-octet Flow
into three Partially Aggregated Flows of 3733, 3733, and 3734 octets;
this ensures no cumulative error is injected by the interval
distribution step.
start time |end time |source ip4 |port |dest ip4 |port|pt| oct
9:00:00.000 9:05:00.000 192.0.2.2 47113 192.0.2.131 53 17 119
9:00:00.000 9:05:00.000 192.0.2.2 22153 192.0.2.131 53 17 83
9:00:00.000 9:05:00.000 192.0.2.2 52420 198.51.100.2 443 6 1637
9:00:00.000 9:05:00.000 192.0.2.3 56047 192.0.2.131 53 17 111
9:00:00.000 9:05:00.000 192.0.2.3 41183 198.51.100.67 80 6 16838
9:00:00.000 9:05:00.000 192.0.2.2 17606 198.51.100.68 80 6 11538
9:00:00.000 9:05:00.000 192.0.2.3 47113 192.0.2.131 53 17 119
9:00:00.000 9:05:00.000 192.0.2.3 48458 198.51.100.133 80 6 2973
9:00:00.000 9:05:00.000 192.0.2.4 61295 198.51.100.2 443 6 8350
9:00:00.000 9:05:00.000 203.0.113.3 41256 198.51.100.133 80 6 778
9:00:00.000 9:05:00.000 203.0.113.3 51662 198.51.100.3 80 6 883
9:00:00.000 9:05:00.000 192.0.2.2 37581 198.51.100.2 80 6 7710*
9:00:00.000 9:05:00.000 203.0.113.3 39586 198.51.100.17 80 6 3733*
9:05:00.000 9:10:00.000 203.0.113.3 52572 198.51.100.2 443 6 1637
9:05:00.000 9:10:00.000 203.0.113.3 49914 197.51.100.133 80 6 561
9:05:00.000 9:10:00.000 192.0.2.2 50824 198.51.100.2 443 6 1899
9:05:00.000 9:10:00.000 192.0.2.3 34597 198.51.100.3 80 6 1284
9:05:00.000 9:10:00.000 203.0.113.3 58907 198.51.100.4 80 6 2670
9:05:00.000 9:10:00.000 192.0.2.2 37581 198.51.100.2 80 6 7710*
9:05:00.000 9:10:00.000 203.0.113.3 39586 198.51.100.17 80 6 3733*
9:10:00.000 9:15:00.000 192.0.2.4 22478 192.0.2.131 53 17 75
9:10:00.000 9:15:00.000 192.0.2.4 49513 198.51.100.68 80 6 3374
9:10:00.000 9:15:00.000 192.0.2.4 64832 198.51.100.67 80 6 138
9:10:00.000 9:15:00.000 192.0.2.3 60833 198.51.100.69 443 6 2325
9:10:00.000 9:15:00.000 192.0.2.2 19638 198.51.100.3 80 6 2869
9:10:00.000 9:15:00.000 192.0.2.3 40429 198.51.100.4 80 6 18289
9:10:00.000 9:15:00.000 203.0.113.3 39586 198.51.100.17 80 6 3734*
Figure 28: Distributed Interval Imposition for Time Series per Source
Subsequent steps are as in Section 8.1; the results, to be exported
using the Template shown in Figure 11, are shown in Figure 29, with
Aggregated Flows differing from the example in Section 8.1 indicated
by "*".
Trammell, et al. Standards Track [Page 45]
RFC 7015 IPFIX Aggregation September 2013
start time |end time |source ip4 |octets
9:00:00.000 9:05:00.000 192.0.2.2 21087*
9:00:00.000 9:05:00.000 192.0.2.3 20041
9:00:00.000 9:05:00.000 192.0.2.4 8350
9:00:00.000 9:05:00.000 203.0.113.3 5394*
9:05:00.000 9:10:00.000 192.0.2.2 9609*
9:05:00.000 9:10:00.000 192.0.2.3 1284
9:05:00.000 9:10:00.000 203.0.113.3 8601*
9:10:00.000 9:15:00.000 192.0.2.2 2869
9:10:00.000 9:15:00.000 192.0.2.3 20614
9:10:00.000 9:15:00.000 192.0.2.4 3587
9:10:00.000 9:15:00.000 203.0.113.3 3734*
Figure 29: Aggregated Flows for Time Series per Source
with Counter Distribution
9. Security Considerations
This document specifies the operation of an Intermediate Aggregation
Process with the IPFIX protocol; the Security Considerations for the
protocol itself in Section 11 of [RFC7011] therefore apply. In the
common case that aggregation is performed on a Mediator, the Security
Considerations for Mediators in Section 9 of [RFC6183] apply as well.
As mentioned in Section 3, certain aggregation operations may tend to
have an anonymizing effect on Flow data by obliterating sensitive
identifiers. Aggregation may also be combined with anonymization
within a Mediator, or as part of a chain of Mediators, to further
leverage this effect. In any case in which an Intermediate
Aggregation Process is applied as part of a data anonymization or
protection scheme, or is used together with anonymization as
described in [RFC6235], the Security Considerations in Section 9 of
[RFC6235] apply.
10. IANA Considerations
This document specifies the creation of new IPFIX Information
Elements in the IPFIX Information Element registry [IANA-IPFIX], as
defined in Section 7 above. IANA has assigned Information Element
numbers to these Information Elements, and entered them into the
registry.
11. Acknowledgments
Special thanks to Elisa Boschi for early work on the concepts laid
out in this document. Thanks to Lothar Braun, Christian Henke, and
Rahul Patel for their reviews and valuable feedback, with special
Trammell, et al. Standards Track [Page 46]
RFC 7015 IPFIX Aggregation September 2013
thanks to Paul Aitken for his multiple detailed reviews. This work
is materially supported by the European Union Seventh Framework
Programme under grant agreement 257315 (DEMONS).
12. References
12.1. Normative References
[RFC2119] Bradner, S., "Key words for use in RFCs to Indicate
Requirement Levels", BCP 14, RFC 2119, March 1997.
[RFC5226] Narten, T. and H. Alvestrand, "Guidelines for Writing an
IANA Considerations Section in RFCs", BCP 26, RFC 5226,
May 2008.
[RFC7011] Claise, B., Ed., Trammell, B., Ed., and P. Aitken,
"Specification of the IP Flow Information Export (IPFIX)
Protocol for the Exchange of Flow Information", STD 77,
RFC 7011, September 2013.
12.2. Informative References
[RFC3917] Quittek, J., Zseby, T., Claise, B., and S. Zander,
"Requirements for IP Flow Information Export (IPFIX)", RFC
3917, October 2004.
[RFC5470] Sadasivan, G., Brownlee, N., Claise, B., and J. Quittek,
"Architecture for IP Flow Information Export", RFC 5470,
March 2009.
[RFC5472] Zseby, T., Boschi, E., Brownlee, N., and B. Claise, "IP
Flow Information Export (IPFIX) Applicability", RFC 5472,
March 2009.
[RFC5476] Claise, B., Johnson, A., and J. Quittek, "Packet Sampling
(PSAMP) Protocol Specifications", RFC 5476, March 2009.
[RFC5655] Trammell, B., Boschi, E., Mark, L., Zseby, T., and A.
Wagner, "Specification of the IP Flow Information Export
(IPFIX) File Format", RFC 5655, October 2009.
[RFC5982] Kobayashi, A. and B. Claise, "IP Flow Information Export
(IPFIX) Mediation: Problem Statement", RFC 5982, August
2010.
[RFC6183] Kobayashi, A., Claise, B., Muenz, G., and K. Ishibashi,
"IP Flow Information Export (IPFIX) Mediation: Framework",
RFC 6183, April 2011.
Trammell, et al. Standards Track [Page 47]
RFC 7015 IPFIX Aggregation September 2013
[RFC6235] Boschi, E. and B. Trammell, "IP Flow Anonymization
Support", RFC 6235, May 2011.
[RFC6728] Muenz, G., Claise, B., and P. Aitken, "Configuration Data
Model for the IP Flow Information Export (IPFIX) and
Packet Sampling (PSAMP) Protocols", RFC 6728, October
2012.
[RFC7012] Claise, B., Ed. and B. Trammell, Ed., "Information Model
for IP Flow Information Export (IPFIX)", RFC 7012,
September 2013.
[RFC7013] Trammell, B. and B. Claise, "Guidelines for Authors and
Reviewers of IP Flow Information Export (IPFIX)
Information Elements", BCP 184, RFC 7013, September 2013.
[RFC7014] D'Antonio, S., Zseby, T., Henke, C., and L. Peluso, "Flow
Selection Techniques", RFC 7014, September 2013.
[IANA-IPFIX]
IANA, "IP Flow Information Export (IPFIX) Entities",
<http://www.iana.org/assignments/ipfix>.
[IPFIX-MED-PROTO]
Claise, B., Kobayashi, A., and B. Trammell, "Operation of
the IP Flow Information Export (IPFIX) Protocol on IPFIX
Mediators", Work in Progress, July 2013.
Trammell, et al. Standards Track [Page 48]
RFC 7015 IPFIX Aggregation September 2013
Authors' Addresses
Brian Trammell
Swiss Federal Institute of Technology Zurich
Gloriastrasse 35
8092 Zurich
Switzerland
Phone: +41 44 632 70 13
EMail: trammell@tik.ee.ethz.ch
Arno Wagner
Consecom AG
Bleicherweg 64a
8002 Zurich
Switzerland
EMail: arno@wagner.name
Benoit Claise
Cisco Systems, Inc.
De Kleetlaan 6a b1
1831 Diegem
Belgium
Phone: +32 2 704 5622
EMail: bclaise@cisco.com
Trammell, et al. Standards Track [Page 49]