Packet Delay Variation Applicability
StatementAT&T Labs200 Laurel Avenue SouthMiddletown,NJ07748USA+1 732 420 1571+1 732 368 1192acmorton@att.comhttp://home.comcast.net/~acmacm/Cisco Systems, Inc.De Kleetlaan 6a b1Diegem1831Belgium+32 2 704 5622bclaise@cisco.comPacket delay variation metrics appear in many different standards
documents. The metric definition in RFC 3393 has considerable
flexibility, and it allows multiple formulations of delay variation
through the specification of different packet selection functions.Although flexibility provides wide coverage and room for new ideas,
it can make comparisons of independent implementations more difficult.
Two different formulations of delay variation have come into wide use in
the context of active measurements. This memo examines a range of
circumstances for active measurements of delay variation and their uses,
and recommends which of the two forms is best matched to particular
conditions and tasks.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 RFC 2119.There are many ways to formulate packet delay variation metrics for
the Internet and other packet-based networks. The IETF itself has
several specifications for delay variation , sometimes called jitter or even inter-arrival jitter , and these have achieved wide adoption. The
International Telecommunication Union - Telecommunication
Standardization Sector (ITU-T) has also recommended several delay
variation metrics (called parameters in their terminology) , and some of these
are widely cited and used. Most of the standards above specify more than
one way to quantify delay variation, so one can conclude that
standardization efforts have tended to be inclusive rather than
selective.This memo uses the term "delay variation" for metrics that quantify a
path's ability to transfer packets with consistent delay. and both prefer
this term. Some refer to this phenomenon as "jitter" (and the buffers
that attempt to smooth the variations as de-jitter buffers).
Applications of the term "jitter" are much broader than packet transfer
performance, with "unwanted signal variation" as a general definition.
"Jitter" has been used to describe frequency or phase variations, such
as data stream rate variations or carrier signal phase noise. The phrase
"delay variation" is almost self-defining and more precise, so it is
preferred in this memo.Most (if not all) delay variation metrics are derived metrics, in
that their definitions rely on another fundamental metric. In this case,
the fundamental metric is one-way delay, and variation is assessed by
computing the difference between two individual one-way delay
measurements, or a pair of singletons. One of the delay singletons is
taken as a reference, and the result is the variation with respect to
the reference. The variation is usually summarized for all packets in a
stream using statistics.The industry has predominantly implemented two specific formulations
of delay variation (for one survey of the situation, see):Inter-Packet Delay Variation, IPDV, where the reference is the
previous packet in the stream (according to sending sequence), and
the reference changes for each packet in the stream. Properties of
variation are coupled with packet sequence in this formulation. This
form was called Instantaneous Packet Delay Variation in early IETF
contributions.Packet Delay Variation, PDV, where a single reference is chosen
from the stream based on specific criteria. The most common
criterion for the reference is the packet with the minimum delay in
the sample. This term derives its name from a similar definition for
Cell Delay Variation, an ATM performance metric.It is important to note that the authors of relevant standards for
delay variation recognized there are many different users with varying
needs, and allowed sufficient flexibility to formulate several metrics
with different properties. Therefore, the comparison is not so much
between standards bodies or their specifications as it is between
specific formulations of delay variation. Both Inter-Packet Delay
Variation and Packet Delay Variation are compliant with , because different packet selection functions
will produce either form.With more people joining the measurement community every day, it is
possible this memo is the first from the IP Performance Metrics (IPPM)
Working Group that the reader has consulted. This section provides a
brief roadmap and background on the IPPM literature, and the published
specifications of other relevant standards organizations.The IPPM framework provides a
background for this memo and other IPPM RFCs. Key terms such as
singleton, sample, and statistic are defined there, along with methods
of collecting samples (Poisson streams), time related issues, and the
"packet of Type-P" convention.There are two fundamental and related metrics that can be applied
to every packet transfer attempt: one-way loss and one-way delay . Lost and delayed packets are separated by a
waiting time threshold. Packets that arrive at the measurement
destination within their waiting time have finite delay and are not
lost. Otherwise, packets are designated lost and their delay is
undefined. Guidance on setting the waiting time threshold may be found
in and .Another fundamental metric is packet reordering as specified in
. The reordering metric was defined to
be "orthogonal" to packet loss. In other words, the gap in a packet
sequence caused by loss does not result in reordered packets, but a
re-arrangement of packet arrivals from their sending order constitutes
reordering.Derived metrics are based on the fundamental metrics. The metric of
primary interest here is delay variation , a metric which is derived from one-way delay
. Another derived metric is the loss
patterns metric , which is derived from
loss.The measured values of all metrics (both fundamental and derived)
depend to great extent on the stream characteristics used to collect
them. Both Poisson streams and Periodic
streams have been used with the IPDV
and PDV metrics. The choice of stream specifications for active
measurement will depend on the purpose of the characterization and the
constraints of the testing environment. Periodic streams are
frequently chosen for use with IPDV and PDV, because the application
streams that are most sensitive to delay variation exhibit
periodicity. Additional details that are method-specific are discussed
the section on Measurement Considerations.In the ITU-T, the framework, fundamental metrics and derived
metrics for IP performance are specified in Recommendation Y.1540
. defines
additional delay variation metrics, analyses the operation of fixed
and adaptive de-jitter buffers, and describes an example adaptive
de-jitter buffer emulator. Appendix II of describes the models for network impairments
(including delay variation) that are part of standardized IP network
emulator which may be useful when evaluating measurement
techniques.The Purpose and Scope follows in Section 2. We then give a summary
of the main tasks for delay variation metrics in section 3. Section 4
defines the two primary forms of delay variation, and section 5
presents summaries of four earlier comparisons. Section 6 adds new
comparisons to the analysis, and section 7 reviews the applicability
and recommendations for each form of delay variation. Section 8 then
looks at many important delay variation measurement considerations.
Following the IANA and Security Considerations, there is an Appendix
on the calculation of the minimum delay for the PDV form.The IPDV and PDV formulations have certain features that make them
more suitable for one circumstance and less so for another. The purpose
of this memo is to compare two forms of delay variation, so that it will
be evident which of the two is better suited for each of many possible
uses and their related circumstances.The scope of this memo is limited to the two forms of delay variation
briefly described above (Inter-Packet Delay Variation and Packet Delay
Variation), circumstances related to active measurement, and uses that
are deemed relevant and worthy of inclusion here through IPPM Working
Group consensus.It is entirely possible that the analysis and conclusions drawn here
are applicable beyond the intended scope, but the reader is cautioned to
fully appreciate the circumstances of active measurement on IP networks
before doing so.The scope excludes assessment of delay variation for packets with
undefined delay. This is accomplished by conditioning the delay
distribution on arrival within a reasonable waiting time based on an
understanding of the path under test and packet lifetimes. The waiting
time is sometimes called the loss threshold : if a packet arrives beyond this threshold, it
may as well have been lost because it is no longer useful. This is
consistent with , where the
Type-P-One-way-ipdv is undefined when the destination fails to receive
one or both packets in the selected pair. Furthermore, it is consistent
with application performance analysis to consider only arriving packets,
because a finite waiting time-out is a feature of many protocols.This section presents a set of tasks that call for delay variation
measurements. Here, the memo provides several answers to the question,
"How will the results be used?" for the delay variation metric.As packets travel along the path from source to destination, they
pass through many network elements, including a series of router
queues. Some types of the delay sources along the path are constant,
such as links between two locations. But the latency encountered in
each queue varies, depending on the number of packets in the queue
when a particular packet arrives. If one assumes that at least one of
the packets in a test stream encounters virtually empty queues all
along the path (and the path is stable), then the additional delay
observed on other packets can be attributed to the time spent in one
or more queues. Otherwise, the delay variation observed is the
variation in queue time experienced by the test stream.It is worth noting that delay variation can occur beyond IP router
queues, in other communication components. Examples include media
contention: DOCSIS, IEEE 802.11 and some mobile radio technologies.
However, delay variation from all sources at the IP layer and below
will be quantified using the two formulations discussed here.Note - while this memo and other IPPM literature prefer the term
delay variation, the terms "jitter buffer" and the more accurate
"de-jitter buffer" are widely adopted names for a component of packet
communication systems, and they will be used here to designate that
system component.Most Isochronous applications (a.k.a. real-time applications)
employ a buffer to smooth out delay variation encountered on the path
from source to destination. The buffer must be big enough to
accommodate the expected variation of delay, or packet loss will
result. However, if the buffer is too large, then some of the desired
spontaneity of communication will be lost and conversational dynamics
will be affected. Therefore, application designers need to know the
range of delay variation they must accommodate, whether they are
designing fixed or adaptive buffer systems.Network service providers also attempt to constrain delay variation
to ensure the quality of real-time applications, and monitor this
metric (possibly to compare with a numerical objective or Service
Level Agreement).De-jitter buffer size can be expressed in units of octets of
storage space for the packet stream, or in units of time that the
packets are stored. It is relatively simple to convert between octets
and time when the buffer read rate (in octets per second) is
constant:read_rate * storage_time = storage_octetsUnits of time are used in the discussion below.The objective of a de-jitter buffer is to compensate for all prior
sources of delay variation and produce a packet stream with constant
delay. Thus, a packet experiencing the minimum transit delay from
source to destination, D_min, should spend the maximum time in a
de-jitter buffer, B_max. The sum of D_min and B_max should equal the
sum of the maximum transit delay (D_max) and the minimum buffer time
(B_min). We haveConstant = D_min + B_max = D_max + B_min,after rearranging terms,B_max - B_min = D_max - D_min = range(B) = range(D)where range(B) is the range of packet buffering times, and range(D)
is the range of packet transit delays from source to destination.Packets with transit delay between the max and min spend a
complimentary time in the buffer and also see the constant delay.In practice, the minimum buffer time, B_min, may not be zero, and
the maximum transit delay, D_max may be a high percentile (99.9%-ile)
instead of the maximum.Note that B_max - B_min = range(B) is the range of buffering times
needed to compensate for delay variation. The actual size of the
buffer may be larger (where B_min > 0) or smaller than
range(B).There must be a process to align the de-jitter buffer time with
packet transit delay. This is a process to identify the packets with
minimum delay and schedule their play-out time so that they spend the
maximum time in the buffer. The error in the alignment process can be
accounted for by a factor, A. In the equation below, the range of
buffering times *available* to the packet stream, range(b), depends on
buffer alignment with the actual arrival times of D_min and D_max.range(b) = b_max - b_min = D_max - D_min + AWhen A is positive, the de-jitter buffer applies more delay than
necessary (where Constant = D_max+b_min+A represents one possible
alignment). When A is negative, there is insufficient buffer time
available to compensate for range(D) because of mis-alignment. Packets
with D_min may be arriving too early and encountering a full buffer,
or packets with D_max may be arriving too late, and in either case the
packets would be discarded.In summary, the range of transit delay variation is a critical
factor in the determination of de-jitter buffer size.In Spatial Composition, the tasks are similar to those described
above, but with the additional complexity of a multiple network path
where several sub-paths are measured separately and no source to
destination measurements are available. In this case, the source to
destination performance must be estimated, using Composed Metrics as
described in
and . Note that determining the composite
delay variation is not trivial: simply summing the sub-path variations
is not accurate.IP performance measurements are often used as the basis for
agreements (or contracts) between service providers and their
customers. The measurement results must compare favorably with the
performance levels specified in the agreement.Packet delay variation is usually one of the metrics specified in
these agreements. In principle, any formulation could be specified in
the Service Level Agreement (SLA). However, the SLA is most useful
when the measured quantities can be related to ways in which the
communication service will be utilized by the customer, and this can
usually be derived from one of the tasks described above.The design of application-layer Forward Error Correction (FEC)
components is closely related to the design of a de-jitter buffer in
several ways. The FEC designer must choose a protection interval (time
to send/receive a block of packets in a constant packet rate system)
consistent with the packet loss characteristics, but also mindful of
the extent of delay variation expected. Further, the system designer
must decide how long to wait for "late" packets to arrive. Again, the
range of delay variation is the relevant expression delay variation
for these tasks.This section presents the formulations of IPDV and PDV, and provides
some illustrative examples. We use the basic singleton definition in
(which itself is based on ):"Type-P-One-way-ipdv is defined for two packets from Src to Dst
selected by the selection function F, as the difference between the
value of the Type-P-One-way-delay from Src to Dst at T2 and the value of
the Type-P-One-Way-Delay from Src to Dst at T1."If we have packets in a stream consecutively numbered i = 1,2,3,...
falling within the test interval, then IPDV(i) = D(i)-D(i-1) where
D(i) denotes the one-way-delay of the ith packet of a stream.One-way delays are the difference between timestamps applied at the
ends of the path, or the receiver time minus the transmission time. So
D(2) = R2-T2. With this timestamp notation, it can be shown that IPDV
also represents the change in inter-packet spacing between
transmission and reception:IPDV(2) = D(2) - D(1) = (R2-T2) - (R1-T1) = (R2-R1) - (T2-T1)An example selection function given in is "Consecutive Type-P packets within the
specified interval." This is exactly the function needed for IPDV. The
reference packet in the pair is always the previous packet in the
sending sequence.Note that IPDV can take on positive and negative values (and zero).
One way to analyze the IPDV results is to concentrate on the positive
excursions. However, approach has limitations that are discussed in
more detail below (see section 5.3).The mean of all IPDV(i) for a stream is usually zero. However, a
slow delay change over the life of the stream, or a frequency error
between the measurement system clocks, can result in a non-zero
mean.The name Packet Delay Variation is used in and its predecessors, and refers to a
performance parameter equivalent to the metric described below.The Selection Function for PDV requires two specific roles for the
packets in the pair. The first packet is any Type-P packet within the
specified interval. The second, or reference packet is the Type-P
packet within the specified interval with the minimum
one-way-delay.Therefore, PDV(i) = D(i)-D(min) (using the nomenclature introduced
in the IPDV section). D(min) is the delay of the packet with the
lowest value for delay (minimum) over the current test interval.
Values of PDV may be zero or positive, and quantiles of the PDV
distribution are direct indications of delay variation.PDV is a version of the one-way delay distribution, shifted to the
origin by normalizing to the minimum delay.Both IPDV and PDV are derived from the one-way delay metric. One
way delay requires knowledge of time at two points, e.g., the source
and destination of an IP network path in end-to-end measurement.
Therefore, both IPDV and PDV can be categorized as 2-point metrics
because they are derived from one-way delay. Specific methods of
measurement may make assumptions or have a priori knowledge about one
of the measurement points, but the metric definitions themselves are
based on information collected at two measurement points.Note: This material originally presented in slides 2 and 3 of .The Figure below gives a sample of packet delays and calculates
IPDV and PDV values and depicts a histogram for each one.The sample of packets contains three packets with "typical" delays
of 20ms, one packet with a low delay of 10ms (the minimum of the
sample) and one packet with 25ms delay.As noted above, this example illustrates that IPDV may take on
positive and negative values, while the PDV values are greater than or
equal to zero. The Histograms of IPDV and PDV are quite different in
general shape, and the ranges are different, too (IPDV range = 20ms,
PDV range = 15 ms). Note that the IPDV histogram will change if the
sequence of delays is modified, but the PDV histogram will stay the
same. PDV normalizes the one-way delay distribution to the minimum
delay and emphasizes the variation independent from the sequence of
delays.This section summarizes previous work to compare these two forms of
delay variation.In , Demichelis compared the early
draft versions of two forms of delay variation. Although the IPDV form
would eventually see widespread use, the ITU-T work-in-progress he
cited did not utilize the same reference packets as PDV. Demichelis
compared IPDV with the alternatives of using the delay of the first
packet in the stream and the mean delay of the stream as the PDV
reference packet. Neither of these alternative references were used in
practice, and they are now deprecated in favor of the minimum delay of
the stream .Active measurements of a transcontinental path (Torino to Tokyo)
provided the data for the comparison. The Poisson test stream had
0.764 second average inter-packet interval, with more than 58 thousand
packets over 13.5 hours. Among Demichelis' observations about IPDV are
the following:IPDV is a measure of the network's ability to preserve the
spacing between packets.The distribution of IPDV is usually symmetrical about the
origin, having a balance of negative and positive values (for the
most part). The mean is usually zero, unless some long-term delay
trend is present.IPDV singletons distinguish quick delay variations (short-term,
on the order of the interval between packets) from longer term
variations.IPDV places reduced demands on the stability and skew of
measurement clocks.He also notes these features of PDV:The PDV distribution does not distinguish short-term variation
from variation over the complete test interval. (Comment: PDV can
be determined over any sub-intervals when the singletons are
stored.)The location of the distribution is very sensitive to the delay
of the first packet, IF this packet is used as the reference. This
would be a new formulation that differs from the PDV definition in
this memo (PDV references the packet with minimum delay, so it
does not have this drawback).The shape of the PDV distribution is identical to the delay
distribution, but shifted by the reference delay.Use of a common reference over measurement intervals that are
longer than a typical session length may indicate more PDV than
would be experienced by streams that support such sessions.
(Ideally, the measurement interval should be aligned with the
session length of interest, and this influences determination of
the reference delay, D(min).)The PDV distribution characterizes the range of queue
occupancies along the measurement path (assuming the path is
fixed), but the range says nothing about how the variation took
place.The summary metrics used in this comparison were the number
of values exceeding a +/-50ms range around the mean, the Inverse
Percentiles, and the Inter-Quartile Range.In , the authors compared IPDV and PDV
(referred to as delta) using a periodic packet stream conforming to
with inter-packet interval of 20
ms.One of the comparisons between IPDV and PDV involves a laboratory
set-up where a queue was temporarily congested by a competing packet
burst. The additional queuing delay was 85ms to 95ms, much larger than
the inter-packet interval. The first packet in the stream that follows
the competing burst spends the longest time queued, and others
experience less and less queuing time until the queue is drained.The authors observed that PDV reflects the additional queuing time
of the packets affected by the burst, with values of 85, 65, 45, 25,
and 5ms. Also, it is easy to determine (by looking at the PDV range)
that a de-jitter buffer of >85 ms would have been sufficient to
accommodate the delay variation. Again, the measurement interval is a
key factor in the validity of such observations (it should have
similar length to the session interval of interest).The IPDV values in the congested queue example are very different:
85, -20, -20, -20, -20, -5ms. Only the positive excursion of IPDV
gives an indication of the de-jitter buffer size needed. Although the
variation exceeds the inter-packet interval, the extent of negative
IPDV values is limited by that sending interval. This preference for
information from the positive IPDV values has prompted some to ignore
the negative values, or to take the absolute value of each IPDV
measurement (sacrificing key properties of IPDV in the process, such
as its ability to distinguish delay trends).Note that this example illustrates a case where the IPDV
distribution is asymmetrical, because the delay variation range (85ms)
exceeds the inter-packet spacing (20ms). We see that the IPDV values
85, -20, -20, -20, -20, -5ms have zero mean, but the left side of the
distribution is truncated at -20ms.Elsewhere, the authors considered the range as a summary statistic
for IPDV, and the 99.9%-ile minus the minimum delay as a summary
statistic for delay variation, or PDV.Mike Pierce made many comments in the context of the 05 version of
draft-ietf-ippm-ipdv. One of his main points was that a delay
histogram is a useful approach to quantifying variation. Another point
was that the time duration of evaluation is a critical aspect.Carlo Demichelis then mailed his comparison paper to the IPPM list,
as discussed in more detail
above.Ruediger Geib observed that both IPDV and the delay histogram (PDV)
are useful, and suggested that they might be applied to different
variation time scales. He pointed out that loss has a significant
effect on IPDV, and encouraged that the loss information be retained
in the arrival sequence.Several example delay variation scenarios were discussed,
including:Clearly, the range of PDV values is 50 ms in both cases above, and
this is the statistic that determines the size of a de-jitter buffer.
The IPDV range is minimal in response to the smooth variation in
Example A (20 ms). However, IPDV responds to the faster variations in
Example B (60 ms range from 40 to -20). Here the IPDV range is larger
than the PDV range, and over-estimates the buffer size
requirements.A heuristic method to estimate buffer size using IPDV is to sum the
consecutive positive or zero values as an estimate of PDV range.
However, this is more complicated to assess than the PDV range, and
has strong dependence on the actual sequence of IPDV values (any
negative IPDV value stops the summation, and again causes an
underestimate).IPDV values can be viewed as the adjustments that an adaptive
de-jitter buffer would make, IF it could make adjustments on a
packet-by-packet basis. However, adaptive de-jitter buffers don't make
adjustments this frequently, so the value of this information is
unknown. The short-term variations may be useful to know in some other
cases.Appendix II of describes a secondary
terminology for delay variation. It compares IPDV, PDV (referred to as
2-point PDV), and 1-point packet delay variation (which assumes a
periodic stream and assesses variation against an ideal arrival
schedule constructed at a single measurement point). This early
comparison discusses some of the same considerations raised in section
6 below.Alan Clark's contribution to ITU-T Study Group 12 in January 2003,
provided an analysis of the root causes of delay variation and
investigated different techniques for measurement and modeling of
"jitter" . Clark compared a metric
closely related to IPDV, Mean Packet-to-Packet Delay Variation, MPPDV
= mean(abs(D(i)-D(i-1))) to the newly proposed Mean Absolute Packet
Delay Variation (MAPDV2, see ). One of
the tasks for this study was to estimate the number of packet discards
in a de-jitter buffer. Clark concluded that MPPDV did not track the
ramp delay variation he associated access link congestion (similar to
Figure 2, Example A above), but MAPDV2 did.Clark also briefly looked at PDV (as described in the 2002 version
of). He concluded that if PDV was applied
to a series of very short measurement intervals (e.g., 200ms), it
could be used to determine the fraction of intervals with high packet
discard rates.This section treats some of the earlier comparison areas in more
detail, and introduces new areas for comparison.The measurement packet loss is of great influence for the delay
variation results, as displayed in the figures 3 and 4 (L means Lost
and U means undefined). Figure 3 shows that in the extreme case of
every other packet loss, the IPDV doesn't produce any results, while
the PDV produces results for all arriving packets.In case of a burst of packet loss, as displayed in figure 3, both
the IPDV and PDV produces some results. Note that PDV still produces
more values than IPDV.In conclusion, the PDV results are affected by the packet loss
ratio. The IPDV results are affected by both the packet loss ratio and
the packet loss distribution. In the extreme case of loss of every
other packet, IPDV doesn't provide any results.When there is little or no stability in the network under test,
then the devices that attempt to characterize the network are equally
stressed, especially if the results displayed are used to make
inferences which may not be valid.Sometimes the path characteristics change during a measurement
interval. The change may be due to link or router failure,
administrative changes prior to maintenance (e.g., link cost change),
or re-optimization of routing using new information. All these causes
are usually infrequent, and network providers take appropriate
measures to ensure this. Automatic restoration to a back-up path is
seen as a desirable feature of IP networks.Frequent path changes and prolonged congestion with substantial
packet loss clearly make delay variation measurements challenging.
Path changes are usually accompanied by a sudden, persistent increase
or decrease in one-way-delay. gives one
such example. We assume that a restoration path either accepts a
stream of packets, or is not used for that particular stream (e.g., no
multi-path for flows).In any case, a change in the TTL (or Hop Limit) of the received
packets indicates that the path is no longer the same. Transient
packet reordering may also be observed with path changes, due to use
of non-optimal routing while updates propagate through the network
(see and
)Many, if not all, packet streams experience packet loss in
conjunction with a path change. However, it is certainly possible that
the active measurement stream does not experience loss. This may be
due to use of a long inter-packet sending interval with respect to the
restoration time, and it becomes more likely as "fast restoration"
techniques see wider deployment (e.g., .Thus, there are two main cases to consider, path changes
accompanied by loss, and those that are lossless from the point of
view of the active measurement stream. The subsections below examine
each of these cases.In the lossless case, a path change will typically affect only
one IPDV singleton. For example, the delay sequence in the Figure
below always produces IPDV=0 except in the one case where the value
is 5 (U, 0, 0, 0, 5, 0, 0, 0, 0).However, if the change in delay is negative and larger than the
inter-packet sending interval, then more than one IPDV singleton may
be affected because packet reordering is also likely to occur.The use of the new path and its delay variation can be quantified
by treating the PDV distribution as bi-modal, and characterizing
each mode separately. This would involve declaring a new path within
the sample, and using a new local minimum delay as the PDV reference
delay for the sub-sample (or time interval) where the new path is
present.The process of detecting a bi-modal delay distribution is made
difficult if the typical delay variation is larger than the delay
change associated with the new path. However, information on TTL (or
Hop Limit) change or the presence of transient reordering can assist
in an automated decision.The effect of path changes may also be reduced by making PDV
measurements over short intervals (minutes, as opposed to hours).
This way, a path change will affect one sample and its PDV values.
Assuming that the mean or median one-way-delay changes appreciably
on the new path, then subsequent measurements can confirm a path
change and trigger special processing on the interval to revise the
PDV result.Alternatively, if the path change is detected, by monitoring the
test packets TTL or Hop Limit, or monitoring the change in the IGP
link-state database, the results of measurement before and after the
path change could be kept separated, presenting two different
distributions. This avoids the difficult task of determining the
different modes of a multi-modal distribution.If the path change is accompanied by loss, such that the are no
consecutive packet pairs that span the change, then no IPDV
singletons will reflect the change. This may or may not be
desirable, depending on the ultimate use of the delay variation
measurement. Figure 6, in which L means Lost and U means undefined,
illustrates this case.PDV will again produce a bimodal distribution. But here, the
decision process to define sub-intervals associated with each path
is further assisted by the presence of loss, in addition to TTL,
reordering information, and use of short measurement intervals
consistent with the duration of user sessions. It is reasonable to
assume that at least loss and delay will be measured simultaneously
with PDV and/or IPDV.IPDV does not help to detect path changes when accompanied by
loss, and this is a disadvantage for those who rely solely on IPDV
measurements.Low cost or low complexity measurement systems may be embedded in
communication devices that do not have access to high stability
clocks, and time errors will almost certainly be present. However,
larger time-related errors (~1ms) may offer an acceptable trade-off
for monitoring performance over a large population (the accuracy
needed to detect problems may be much less than required for a
scientific study, ~0.01ms for example).Maintaining time accuracy <<1ms has typically required access
to dedicated time receivers at all measurement points. Global
positioning system (GPS) receivers have often been installed to
support measurements. The GPS installation conditions are fairly
restrictive, and many prospective measurement efforts have found the
deployment complexity and system maintenance too difficult.As mentioned above, observed that
PDV places greater demands on clock synchronization than for IPDV.
This observation deserves more discussion. Synchronization errors have
two components: time of day errors and clock frequency errors
(resulting in skew).Both IPDV and PDV are sensitive to time-of-day errors when
attempting to align measurement intervals at the source and
destination. Gross mis-alignment of the measurement intervals can lead
to lost packets, for example if the receiver is not ready when the
first test packet arrives. However, both IPDV and PDV assess delay
differences, so the error present in any two one-way-delay singletons
will cancel as long as the error is constant. So, the demand for NTP
or GPS synchronization comes primarily from one-way delay measurement
time-of-day accuracy requirements. Delay variation and measurement
interval alignment are relatively less demanding.Skew is a measure of the change in clock time over an interval
w.r.t. a reference clock. Both IPDV and PDV are affected by skew, but
the error sensitivity in IPDV singletons is less because the intervals
between consecutive packets are rather small, especially when compared
to the overall measurement interval. Since PDV computes the difference
between a single reference delay (the sample minimum) and all other
delays in the measurement interval, the constraint on skew error is
greater to attain the same accuracy as IPDV. Again, use of short PDV
measurement intervals (on the order of minutes, not hours) provides
some relief from the effects of skew error. Thus, the additional
accuracy demand of PDV can be expressed as a ratio of the measurement
interval to the inter-packet spacing.A practical example is a measurement between two hosts, one with a
synchronized clock and the other with a free-running clock having 50
part per million (ppm) long term accuracy.If IPDV measurements are made on packets with a 1 second
spacing, the maximum singleton error will be 1 x 5 x 10^-5
seconds, or 0.05ms.If PDV measurements are made on the same packets over a 60
second measurement interval, then the delay variation due to the
max free-running clock error will be 60 x 5 x 10-5 seconds, or 3ms
delay variation error from the first packet to the last.Therefore, the additional accuracy required for equivalent
PDV error under these conditions is a factor of 60 more than for IPDV.
This is a rather extreme scenario, because time-of-day error of 1
second would accumulate in ~5.5 hours, potentially causing the
measurement interval alignment issue described above.If skew is present in a sample of one-way-delays, its symptom is
typically a nearly linear growth or decline over all the one-way-delay
values. As a practical matter, if the same slope appears consistently
in the measurements, then it may be possible to fit the slope and
compensate for the skew in the one-way-delay measurements, thereby
avoiding the issue in the PDV calculations that follow. See for additional information on compensating
for skew.Values for IPDV may have non-zero mean over a sample when clock
skew is present. This tends to complicate IPDV analysis when using the
assumptions of a zero mean and a symmetric distribution.There is a third factor related to clock error and stability: this
is the presence of a clock synchronization protocol (e.g., NTP) and
the time adjustment operations that result. When a time error is
detected (typically on the order of a few milliseconds), the host
clock frequency is continuously adjusted to reduce the time error. If
these adjustments take place during a measurement interval, they may
appear as delay variation when none was present, and therefore are a
source of error (regardless of the DV form considered).ITU-T Recommendation gives a
provisional method to compose a PDV metric using PDV measurement
results from two or more sub-paths. Additional methods are considered
in .PDV has a clear advantage at this time, since there is no validated
method to compose an IPDV metric. In addition, IPDV results depend
greatly on the exact sequence of packets and may not lend themselves
easily to the composition problem, where segments must be assumed to
have independent delay distributions.Despite the risk of over-summarization, measurements must often be
displayed for easy consumption. If the right summary report is
prepared, then the "dashboard" view correctly indicates whether there
is something different and worth investigating further, or that the
status has not changed. The dashboard model restricts every instrument
display to a single number. The packet network dashboard could have
different instruments for loss, delay, delay variation, reordering,
etc., and each must be summarized as a single number for each
measurement interval. The single number summary statistic is a key
component of SLAs, where a threshold on that number must be met x% of
the time.The simplicity of the PDV distribution lends itself to this
summarization process (including use of the percentiles, median or
mean). An SLA of the form "no more than x% of packets in a measurement
interval shall have PDV >= y ms, for no less than z% of time" is
relatively straightforward to specify and implement. introduced the notion of a pseudo-range when
setting an objective for the 99.9%-ile of PDV. The conventional range
(max-min) was avoided for several reasons, including stability of the
maximum delay. The 99.9%-ile of PDV is helpful to performance planners
(seeking to meet some user-to-user objective for delay) and in design
of de-jitter buffer sizes, even those with adaptive capabilities.IPDV does not lend itself to summarization so easily. The mean IPDV
is typically zero. As the IPDV distribution will have two tails
(positive and negative) the range or pseudo-range would not match the
needed de-jitter buffer size. Additional complexity may be introduced
when the variation exceeds the inter-packet sending interval, as
discussed above (in sections 5.2 and 6.2.1). Should the Inter-Quartile
Range be used? Should the singletons beyond some threshold be counted
(e.g., mean +/- 50ms)? A strong rationale for one of these summary
statistics has yet to emerge.When summarizing IPDV, some prefer the simplicity of the
single-sided distribution created by taking the absolute value of each
singleton result, abs(D(i)-D(i-1)). This approach sacrifices the
two-sided inter-arrival spread information in the distribution. It
also makes the evaluation using percentiles more confusing, because a
single late packet that exceeds the variation threshold will cause two
pairs of singletons to fail the criteria (one positive, the other
negative converted to positive). The single-sided PDV distribution is
an advantage in this category. gives the calculation of the
inter-arrival Jitter field for the RTCP report, with a sample
implementation in an Appendix.The RTCP Jitter value can be calculated using IPDV singletons. If
there is packet reordering, as defined in , then estimates of Jitter based on IPDV may
vary slightly, because specifies the
use of receive packet order.Just as there is no simple way to convert PDV singletons to IPDV
singletons without returning to the original sample of delay
singletons, there is no clear relationship between PDV and Jitter.MAPDV2 stands for Mean Absolute Packet Delay Variation (version) 2,
and is specified in . The MAPDV2
algorithm computes a smoothed running estimate of the mean delay using
the one-way delays of 16 previous packets. It compares the current
one-way-delay to the estimated mean, separately computes the means of
positive and negative deviations, and sums these deviation means to
produce MAPVDV2. In effect, there is a MAPDV2 singleton for every
arriving packet, so further summarization is usually warranted.Neither IPDV or PDV forms assist in the computation of MAPDV2.Network traffic load balancing is a process to divide packet
traffic in order to provide a more even distribution over two or more
equally viable paths. The paths chosen are based on the IGP cost
metrics, while the delay depends on the path's physical layout.
Usually, the balancing process is performed on a per-flow basis to
avoid delay variation experienced when packets traverse different
physical paths.If the sample includes test packets with different characteristics
such as IP addresses/ports, there could be multi-modal delay
distributions present. The PDV form makes the identification of
multiple modes possible. IPDV may also reveal that multiple paths are
in use with a mixed flow sample, but the different delay modes are not
easily divided and analyzed separately.Should the delay singletons using multiple addresses/ports be
combined in the same sample? Should we characterize each mode
separately? (This question also applies to the Path Change case.) It
depends on the task to be addressed by the measurement.For the task of de-jitter buffer sizing or assessing queue
occupation, the modes should be characterized separately because flows
will experience only one mode on a stable path. Use of a single flow
description (address/port combination) in each sample simplifies this
analysis. Multiple modes may be identified by collecting samples with
different flow attributes, and characterization of multiple paths can
proceed with comparison of the delay distributions from each
sample.For the task of capacity planning and routing optimization,
characterizing the modes separately could offer an advantage. Network
wide capacity planning (as opposed to link capacity planning) takes as
input the core traffic matrix, which corresponds to a matrix of
traffic transferred from every source to every destination in the
network. Applying the core traffic matrix along with the routing
information (typically the link state database of a routing protocol)
in a capacity planning tool offers the possibility to visualize the
paths where the traffic flows and to optimize the routing based on the
link utilization. In the case where equal cost multiple paths (ECMP)
are used, the traffic will be load balanced onto multiple paths. If
each mode of the IP delay multi-modal distribution can be associated
with a specific path, the delay performance offers an extra
optimization parameter, i.e. the routing optimization based on the IP
delay variation metric. As an example, the load balancing across ECMPs
could be suppressed so that the VoIP calls would only be routed via
the path with the lower IP delay variation. Clearly, any modifications
can result in new delay performance measurements, so there must be a
verification step to ensure the desired outcome.Based on the comparisons of IPDV and PDV presented above, this
section matches the attributes of each form with the tasks described
earlier. We discuss the more general circumstances first.The PDV distribution is anchored at the minimum delay observed in
the measurement interval. When the sample minimum coincides with the
true minimum delay of the path, then the PDV distribution is
equivalent to the queuing time distribution experienced by the test
stream. If the minimum delay is not the true minimum, then the PDV
distribution captures the variation in queuing time and some
additional amount of queuing time is experienced, but unknown. One
can summarize the PDV distribution with the mean, median, and other
statistics.IPDV can capture the difference in queuing time from one packet
to the next, but this is a different distribution from the queue
occupancy revealed by PDV.This task is complimentary to the problem of inferring queue
occupancy through measurement. Again, use of the sample minimum as
the reference delay for PDV yields a distribution that is very
relevant to de-jitter buffer size. This is because the minimum delay
is an alignment point for the smoothing operation of de-jitter
buffers. A de-jitter buffer that is ideally aligned with the delay
variation adds zero buffer time to packets with the longest
accommodated network delay (any packets with longer delays are
discarded). Thus, a packet experiencing minimum network delay should
be aligned to wait the maximum length of the de-jitter buffer. With
this alignment, the stream is smoothed with no unnecessary delay
added. illustrates the ideal
relationship between network delay variation and buffer time.The PDV distribution is also useful for this task, but different
statistics are preferred. The range (max-min) or the 99.9%-ile of
PDV (pseudo-range) are closely related to the buffer size needed to
accommodate the observed network delay variation.The PDV distribution directly addresses the FEC waiting time
question. When the PDV distribution has a 99th percentile of 10ms,
then waiting 10ms longer than the FEC protection interval will allow
99% of late packets to arrive and be used in the FEC block.In some cases, the positive excursions (or series of positive
excursions) of IPDV may help to approximate the de-jitter buffer
size, but there is no guarantee that a good buffer estimate will
emerge, especially when the delay varies as a positive trend over
several test packets.PDV has a clear advantage at this time, since there is no
validated method to compose an IPDV metric.The one-sided PDV distribution can be constrained with a single
statistic, such as an upper percentile, so it is preferred. The IPDV
distribution is two-sided, usually has zero mean, and no universal
summary statistic that relates to a physical quantity has emerged in
years of experience.Note that measurement of delay variation may not be the primary
concern under unstable and unreliable circumstances.When appreciable skew is present between measurement system
clocks, then IPDV has an advantage because PDV would require
processing over the entire sample to remove the skew error. However,
significant skew can invalidate IPDV analysis assumptions, such as
the zero mean and symmetric distribution characteristics. Small skew
may well be within the error tolerance, and both PDV and IPDV
results will be usable. There may be a portion of the skew,
measurement interval, and required accuracy 3-D space where IPDV has
an advantage, depending on the specific measurement
specifications.Neither form of delay variation is more suited than the other to
on-the-fly summarization without memory, and this may be one of the
reasons that RTCP Jitter and MAPDV2
in have attained deployment in
low-cost systems.If the network under test exhibits frequent path changes, on the
order of several new routes per minute, then IPDV appears to isolate
the delay variation on each path from the transient effect of path
change (especially if there is packet loss at the time of path
change). However, if one intends to use IPDV to indicate path
changes, it cannot do this when the change is accompanied by loss.
It is possible to make meaningful PDV measurements when paths are
unstable, but great importance would be placed on the algorithms
that infer path change and attempt to divide the sample on path
change boundaries.When path changes are frequent and cause packet loss, delay
variation is probably less important than the loss episodes and
attention should be turned to the loss metric instead.If the network under test exhibits frequent loss, then PDV may
produce a larger set of singletons for the sample than IPDV. This is
due to IPDV requiring consecutive packet arrivals to assess delay
variation, compared to PDV where any packet arrival is useful. The
worst case is when no consecutive packets arrive, and the entire
IPDV sample would be undefined. PDV would successfully produce a
sample based on the arriving packets.PDV distributions offer the most straightforward way to identify
that a sample of packets have traversed multiple paths. The tasks of
de-jitter buffer sizing or assessing queue occupation with PDV
should be use a sample with a single flow because flows will
experience only one mode on a stable path, and it simplifies the
analysis.Comparison AreaPDVIPDVChallenging CircumstancesLess sensitive to packet loss, and simplifies analysis when load
balancing or multiple paths are presentPreferred when path changes are frequent or when measurement
clocks exhibit some skewSpatial Composition of DV metricAll validated methods use this formHas sensitivity to sequence and spacing changes, which tends to
break the requirement for independent distributions between path
segmentsDetermine De-Jitter Buffer Size Required"Pseudo-range" reveals this property by anchoring the
distribution at the minimum delayNo reliable relationship, but some heuristicsEstimate of Queuing Time and VariationDistribution has one-to-one relationship on a stable path,
especially when sample min = true minNo reliable relationshipSpecification Simplicity: Single Number SLSOne constraint needed for single-sided distribution, and easily
related to quantities aboveDistribution is two-sided, usually has zero mean, and no
universal summary statistic that relates to a physical quantityThis section discusses the practical aspects of delay variation
measurement, with special attention to the two formulations compared in
this memo.As stated in the background section, there is a strong dependency
between the active measurement stream characteristics and the results.
The IPPM literature includes two primary methods for collecting
samples: Poisson sampling described in ,
and Periodic sampling in. The Poisson
method was intended to collect an unbiased sample of performance,
while the Periodic method addresses a "known bias of interest".
Periodic streams are required to have random start times and limited
stream duration, in order to avoid unwanted synchronization with some
other periodic process, or cause congestion-aware senders to
synchronize with the stream and produce atypical results. The random
start time should be different for each new stream.It is worth noting that was
developed in parallel with . As a
result, all the stream metrics defined in specify the Poisson sampling method.Periodic sampling is frequently used in measurements of delay
variation. Several factors foster this choice:Many application streams that are sensitive to delay variation
also exhibit periodicity, and so exemplify the bias of interest.
If the application has a constant packet spacing, this constant
spacing can be the inter-packet gap for the test stream. VoIP
streams often use 20ms spacing, so this is an obvious choice for
an Active stream. This applies to both IPDV and PDV forms.The spacing between packets in the stream will influence
whether the stream experiences short-range dependency, or only
long-range dependency, as investigated in . The packet spacing also influences the
IPDV distribution and the stream's sensitivity to reordering. For
example, with a 20 ms spacing the IPDV distribution cannot go
below -20ms without packet reordering.The measurement process may make several simplifying
assumptions when the send spacing and send rate are constant. For
example, the inter-arrival times at the destination can be
compared with an ideal sending schedule, and allowing a one-point
measurement of delay variation (described in ) that approximates the IPDV form.
Simplified methods that approximate PDV are possible as well (some
are discussed in Appendix II of ).Analysis of truncated, or non-symmetrical IPDV distributions is
simplified. Delay variations in excess of the periodic sending
interval can cause multiple singleton values at the negative limit
of the packet spacing (see section 5.2 and ). Only packet reordering can cause the
negative spacing limit to be exceeded.Despite the emphasis on inter-packet delay differences with
IPDV, both Poisson and Periodic
streams have been used, and these
references illustrate the different analyses that are possible.The advantages of using a Poisson distribution are discussed in
. The main properties are to avoid
predicting the sample times, avoid synchronization with periodic
events that are present in networks, and avoid inducing
synchronization with congestion-aware senders. When a Poisson stream
is used with IPDV, the distribution will reflect inter-packet delay
variation on many different time scales (or packet spacings). The
unbiased Poisson sampling brings a new layer of complexity in the
analysis of IPDV distributions.One key aspect of measurement devices is their ability to store
singletons (or individual measurements). This feature usually is
closely related to local calculation capabilities. For example, an
embedded measurement device with limited storage will like provide
only a few statistics on the delay variation distribution, while
dedicated measurement systems store all the singletons and allow
detailed analysis (later calculation of either form of delay variation
is possible with the original singletons).Therefore, systems with limited storage must choose their metrics
and summary statistics in advance. If both IPDV and PDV statistics are
desired, the supporting information must be collected as packets
arrive. For example, the PDV range and high percentiles can be
determined later if the minimum and several of the largest delays are
stored while the measurement is in-progress.Both IPDV and PDV can be summarized as a range in milliseconds.With IPDV, it is interesting to report on a positive percentile,
and an inter-quantile range is appropriate to reflect both positive
and negative tails (e.g., 5% to 95%). If the IPDV distribution is
symmetric around a mean of zero, then it is sufficient to report on
the positive side of the distribution.With PDV, it is sufficient to specify the upper percentile (e.g.,
99.9%).At several points in this memo, we have recommended use of test
intervals on the order of minutes. In their paper examining the
stability of Internet path properties, Zhang et al. concluded that consistency was present on the
order of minutes for the performance metrics considered (loss, delay,
and throughput) for the paths they measured.The topic of temporal aggregation of performance measured in small
intervals to estimate some larger interval is described in the Metric
Composition Framework .The primary recommendation here is to test using durations that are
similar in length to the session time of interest. This applies to
both IPDV and PDV, but is possibly more relevant for PDV since the
duration determines how often the D_min will be determined, and the
size of the associated sample.As with one-way delay measurements, local clock synchronization is
an important matter for delay variation measurements.There are several options available:Global Positioning System receiversIn some parts of the world, Cellular Code Division Multiple
Access (CDMA) systems distribute timing signals that are derived
from GPS and traceable to UTC.Network Time Protocol is a
convenient choice in many cases, but usually offers lower accuracy
than the options above.When clock synchronization is inconvenient or subject to
appreciable errors, then round-trip measurements may give a cumulative
indication of the delay variation present on both directions of the
path. However, delay distributions are rarely symmetrical, so it is
difficult to infer much about the one-way delay variation from
round-trip measurements. Also, measurements on asymmetrical paths add
complications for the one-way delay metric.Lost and delayed packets are separated by a waiting time threshold.
Packets that arrive at the measurement destination within their
waiting time have finite delay and are not lost. Otherwise, packets
are designated lost and their delay is undefined. Guidance on setting
the waiting time threshold may be found in and .In essence, suggests to use a
long waiting time to serve network characterization and revise results
for specific application delay thresholds as needed.Packet reordering, defined in , is
essentially an extreme form of delay variation where the packet stream
arrival order differs from the sending order.PDV results are not sensitive to packet arrival order, and are not
affected by reordering other than to reflect the more extreme
variation.IPDV results will change if reordering is present because they are
sensitive to the sequence of delays of arriving packets. The main
example of this sensitivity is in the truncation of the negative tail
of the distribution.When there is no reordering, the negative tail is limited by
the sending time spacing between packets.If reordering occurs (and the reordered packets are not
discarded), the negative tail can take on any value (in
principal).In general, measurement systems should have the capability to
detect when sequence has changed. If IPDV measurements are made
without regard to packet arrival order, the IPDV will be
under-reported when reordering occurs.All of the references that discuss or define delay variation
suggest ways to represent or report the results, and interested
readers should review the various possibilities.For example, suggests to report
a pseudo range of delay variation based on calculating the difference
between a high percentile of delay and the minimum delay. The
99.9%-ile minus the minimum will give a value that can be compared
with objectives in .This document makes no request of IANA.Note to RFC Editor: this section may be removed on publication as an
RFC.The security considerations that apply to any active measurement of
live networks are relevant here as well. See The authors would like to thank Phil Chimento for his suggestion to
employ the convention of conditional distributions for Delay to deal
with packet loss, and his encouragement to "write the memo" after
hearing "the talk" on this topic at IETF-65. We also acknowledge
constructive comments from Alan Clark, Loki Jorgenson, Carsten Schmoll,
and Robert Holley.Practitioners have raised questions several questions that this
section intends to answer:- how is this D_min calculated? Is it DV(99%) as mentioned in ?- do we need to keep all the values from the interval, then take the
minimum? Or do we keep the minimum from previous intervals?The value of D_min used as the reference delay for PDV calculations
is simply the minimum delay of all packets in the current sample. The
usual single value summary of the PDV distribution is D_99.9%-ile minus
D_min.It may be appropriate to segregate sub-sets and revise the minimum
value during a sample. For example, if it can be determined with
certainty that the path has changed by monitoring the Time to Live or
Hop Count of arriving packets, this may be sufficient justification to
reset the minimum for packets on the new path. There is also a simpler
approach to solving this problem: use samples collected over short
evaluation intervals (on the order of minutes). Intervals with path
changes may be more interesting from the loss or one-way delay
perspective (possibly failing to meet one or more SLAs), and it may not
be necessary to conduct delay variation analysis. Short evaluation
intervals are preferred for measurements that serve as a basis for
troubleshooting, since the results are available to report soon after
collection.It is not necessary to store all delay values in a sample when
storage is a major concern. D_min can be found by comparing each new
singleton value with the current value and replacing it when required.
In a sample with 5000 packets, evaluation of the 99.9%-ile can also be
achieved with limited storage. One method calls for storing the top 50
delay singletons and revising the top value list each time 50 more
packets arrive.A Fine-Grained View of High Performance Networking, NANOG 22
Conf.; http://www.nanog.org/mtg-0105/agenda.htmlStandardized Active Measurements on a Tier 1 IP Backbone,
IEEE Communications Mag., pp 90-97.Packet Delay Variation Comparison between ITU-T and IETF
Draft DefinitionsCSELTJitter Definitions: What is What?"A Brief Jitter Metrics Comparison, and not the last word, by
any means…", Slide Presentation at IETF-65, IPPM
Session,AT&T Labs"Performance parameter definitions for the quality of speech
and other voiceband applications utilizing IP networks""Network model for evaluating multimedia transmission
performance over Internet Protocol"Internet protocol data communication service - IP packet
transfer and availability performance parametersNetwork Performance Objectives for IP-Based ServicesITU-T Delayed Contribution COM 12 - D98, "Analysis,
measurement and modelling of Jitter"Telchemy Inc."The Implications of Short-Range Dependency on Delay
Variation Measurement", Second IEEE Symposium on Network Computing
and ApplicationsPhilipsUniversity of Delaware"On the Constancy of Internet Path Properties", Proceedings
of ACM SIGCOMM Internet Measurement Workshop,AT&T LabsDuffielAT&T LabsACIRIACIRI