A One-Way Delay Metric for
IPPM
Texas A&M
galmes@tamu.edu
Ixia
skalidindi@ixiacom.com
Internet2
matt@internet2.edu
AT&T Labs
200 Laurel Avenue South
Middletown
NJ
07748
USA
+1 732 420 1571
+1 732 368 1192
acmorton@att.com
http://home.comcast.net/~acmacm/
This memo (RFC 2679 bis) defines a metric for one-way delay of
packets across Internet paths. It builds on notions introduced and
discussed in the IPPM Framework document, RFC 2330; the reader is
assumed to be familiar with that document.
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.
The following text constitutes RFC 2769 bis proposed for advancement
on the IETF Standards Track. This section tracks the changes from .
provides the test plan and results
supporting advancement along the standards
track, according to the process in . The
conclusions of list four minor
modifications:
Section 6.2.3 of asserts that the
assumption of post-processing to enforce a constant waiting time
threshold is compliant, and that the text of the RFC should be
revised slightly to include this point (see the last list item of
section 3.6, below).
Section 6.5 of indicates that
Type-P-One-way-Delay-Inverse-Percentile statistic has been ignored
in both implementations, so it is a candidate for removal or
deprecation in RFC2679bis (this small discrepancy does not affect
candidacy for advancement) (see section 5.4, below).
The IETF has reached consensus on guidance for reporting metrics
in , and this memo should be referenced in
RFC2679bis to incorporate recent experience where appropriate (see
the last list item of section 3.6, section 3.8, and section 5
below).
There is currently one erratum with status "Held for document
update" for , and it appears this minor
revision and additional text should be incorporated in RFC2679bis
(see section 5.1).
A number of updates to the text have
been implemented in the text below, to reference key IPPM RFCs that were
approved after , and to address comments on the
IPPM mailing list describing current conditions and experience.
Near the end of section 2.1, update of a network example using
ATM and clarification of TCP's affect on queue occupation and
importance of one-way delay measurement.
Explicit inclusion of the maximum waiting time input parameter in
section 3.2 and 4.2, reflecting recognition of this parameter in
more recent RFCs and ITU-T Recommendation Y.1540.
Addition of reference to RFC6703 in the discussion of packet life
time and application timeouts in section 3.5.
Addition of reference to the default requirement (that packets be
standard-formed) from RFC2330 as a new list item in section 3.5.
GPS-based NTP experience replaces "to be tested" in section
3.5.
Added parenthetical guidance on minimizing interval between
timestamp placement to send time in section 3.6.
Added text recognizing the impending deployment of transport
layer encryption in section 3.6.
Section 3.7.2 notes that some current systems perform host time
stamping on the network interface hardware.
"instrument" replaced by the defined term "host" in sections
3.7.3 and 3.8.3.
Added reference to RFC 3432 Periodic sampling alongside Poisson
sampling in section 4, and also noting that a truncated Poisson
distribution may be needed with modern networks as described in the
IPPM Framework update, RFC7312.
Add reference to RFC 4737 Reordering metric in the related
discussion of section 4.6, Methodologies.
Clarifying the conclusions on two related points on harm to
measurements (recognition of measurement traffic for unexpected
priority treatment and attacker traffic which emulates measurement)
in section 6, Security Considerations.
Section 5.4.4 of suggests a common template
for performance metrics partially derived from previous IPPM and BMWG
RFCs, but also contains some new items. All of the Normative points are covered, but not quite in the
same section names or orientation. Several of the Informative points are
covered. Maintaining the familiar outline of IPPM literature has both
value and minimizes unnecessary differences between this revised RFC and
current/future IPPM RFCs.
The publication of RFC 6921 suggested an area where this memo might
need updating. Packet transfer on Faster-Than-Light (FTL) networks could
result in negative delays and packet reordering, however both are
covered as possibilities in the current text and no revisions are deemed
necessary (we also note that this is an April 1st RFC).
This memo defines a metric for one-way delay of packets across
Internet paths. It builds on notions introduced and discussed in the
IPPM Framework document, ; the reader is assumed
to be familiar with that document.
This memo is intended to be parallel in structure to a companion
document for Packet Loss ("A One-way Packet Loss Metric for IPPM") .
Although was written with protocols in mind,
the key words are used in this document for similar reasons. They are
used to ensure the results of measurements from two different
implementations are comparable, and to note instances when an
implementation could perturb the network.
The structure of the memo is as follows:
+ A 'singleton' analytic metric, called Type-P-One-way-Delay, will be
introduced to measure a single observation of one-way delay.
+ Using this singleton metric, a 'sample', called Type-P-One-way-
Delay-Poisson-Stream, will be introduced to measure a sequence of
singleton delays sent at times taken from a Poisson process.
+ Using this sample, several 'statistics' of the sample will be
defined and discussed. This progression from singleton to sample to
statistics, with clear separation among them, is important.
Whenever a technical term from the IPPM Framework document is first
used in this memo, it will be tagged with a trailing asterisk. For
example, "term*" indicates that "term" is defined in the Framework.
One-way delay of a Type-P* packet from a source host* to a
destination host is useful for several reasons:
+ Some applications do not perform well (or at all) if end-to-end
delay between hosts is large relative to some threshold value.
+ Erratic variation in delay makes it difficult (or impossible) to
support many real-time applications.
+ The larger the value of delay, the more difficult it is for
transport-layer protocols to sustain high bandwidths.
+ The minimum value of this metric provides an indication of the
delay due only to propagation and transmission delay.
+ The minimum value of this metric provides an indication of the
delay that will likely be experienced when the path* traversed is
lightly loaded.
+ Values of this metric above the minimum provide an indication of
the congestion present in the path.
The measurement of one-way delay instead of round-trip delay is
motivated by the following factors:
+ In today's Internet, the path from a source to a destination may
be different than the path from the destination back to the source
("asymmetric paths"), such that different sequences of routers are
used for the forward and reverse paths. Therefore round-trip
measurements actually measure the performance of two distinct paths
together. Measuring each path independently highlights the performance
difference between the two paths which may traverse different Internet
service providers, and even radically different types of networks (for
example, research versus commodity networks, or networks with
asymmetric link capacities, or wireless vs. wireline access).
+ Even when the two paths are symmetric, they may have radically
different performance characteristics due to asymmetric queueing.
+ Performance of an application may depend mostly on the
performance in one direction. For example, a TCP-based communication
may experience reduced throughput if congestion occurs in one
direction of its communication. Trouble shooting may be simplified if
the congested direction of TCP transmission can be identified.
+ In quality-of-service (QoS) enabled networks, provisioning in one
direction may be radically different than provisioning in the reverse
direction, and thus the QoS guarantees differ. Measuring the paths
independently allows the verification of both guarantees.
It is outside the scope of this document to say precisely how delay
metrics would be applied to specific problems.
{Comment: the terminology below differs from that defined by ITU-T
documents (e.g., G.810, "Definitions and terminology for
synchronization networks" and I.356, "B-ISDN ATM layer cell transfer
performance"), but is consistent with the IPPM Framework document. In
general, these differences derive from the different backgrounds; the
ITU-T documents historically have a telephony origin, while the
authors of this document (and the Framework) have a computer systems
background. Although the terms defined below have no direct equivalent
in the ITU-T definitions, after our definitions we will provide a
rough mapping. However, note one potential confusion: our definition
of "clock" is the computer operating systems definition denoting a
time-of-day clock, while the ITU-T definition of clock denotes a
frequency reference.}
Whenever a time (i.e., a moment in history) is mentioned here, it
is understood to be measured in seconds (and fractions) relative to
UTC.
As described more fully in the Framework document, there are four
distinct, but related notions of clock uncertainty:
synchronization*
measures the extent to which two clocks agree on what time it is.
For example, the clock on one host might be 5.4 msec ahead of the
clock on a second host. {Comment: A rough ITU-T equivalent is "time
error".}
accuracy*
measures the extent to which a given clock agrees with UTC. For
example, the clock on a host might be 27.1 msec behind UTC. {Comment:
A rough ITU-T equivalent is "time error from UTC".}
resolution*
measures the precision of a given clock. For example, the clock on
an old Unix host might tick only once every 10 msec, and thus have a
resolution of only 10 msec. {Comment: A very rough ITU-T equivalent is
"sampling period".}
skew*
measures the change of accuracy, or of synchronization, with time.
For example, the clock on a given host might gain 1.3 msec per hour
and thus be 27.1 msec behind UTC at one time and only 25.8 msec an
hour later. In this case, we say that the clock of the given host has
a skew of 1.3 msec per hour relative to UTC, which threatens accuracy.
We might also speak of the skew of one clock relative to another
clock, which threatens synchronization. {Comment: A rough ITU-T
equivalent is "time drift".}
+ Src, the IP address of a host
+ Dst, the IP address of a host
+ T, a time
+ Tmax, a loss threshold waiting time
The value of a Type-P-One-way-Delay is either a real number, or an
undefined (informally, infinite) number of seconds.
For a real number dT, >>the *Type-P-One-way-Delay* from Src
to Dst at T is dT<< means that Src sent the first bit of a
Type-P packet to Dst at wire-time* T and that Dst received the last
bit of that packet at wire-time T+dT.
>>The *Type-P-One-way-Delay* from Src to Dst at T is
undefined (informally, infinite)<< means that Src sent the first
bit of a Type-P packet to Dst at wire-time T and that Dst did not
receive that packet (within the loss threshold waiting time,
Tmax).
Suggestions for what to report along with metric values appear in
Section 3.8 after a discussion of the metric, methodologies for
measuring the metric, and error analysis.
Type-P-One-way-Delay is a relatively simple analytic metric, and
one that we believe will afford effective methods of measurement.
The following issues are likely to come up in practice:
+ Real delay values will be positive. Therefore, it does not make
sense to report a negative value as a real delay. However, an
individual zero or negative delay value might be useful as part of a
stream when trying to discover a distribution of a stream of delay
values.
+ Since delay values will often be as low as the 100 usec to 10
msec range, it will be important for Src and Dst to synchronize very
closely. GPS systems afford one way to achieve synchronization to
within several 10s of usec. Ordinary application of NTP may allow
synchronization to within several msec, but this depends on the
stability and symmetry of delay properties among those NTP agents
used, and this delay is what we are trying to measure. A combination
of some GPS-based NTP servers and a conservatively designed and
deployed set of other NTP servers should yield good results. This was
tested in , where a GPS measurement system's
results compared well with a GPS-based NTP synchronized system for the
same intercontinental path.
+ A given methodology will have to include a way to determine
whether a delay value is infinite or whether it is merely very large
(and the packet is yet to arrive at Dst). As noted by Mahdavi and
Paxson , simple upper bounds (such as the 255
seconds theoretical upper bound on the lifetimes of IP packets ) could be used, but good engineering, including an
understanding of packet lifetimes, will be needed in practice.
{Comment: Note that, for many applications of these metrics, the harm
in treating a large delay as infinite might be zero or very small. A
TCP data packet, for example, that arrives only after several
multiples of the RTT may as well have been lost. See section 4.1.1 of
for examination of unusual packet delays and
application performance estimation.}
+ If the packet is duplicated along the path (or paths) so that
multiple non-corrupt copies arrive at the destination, then the packet
is counted as received, and the first copy to arrive determines the
packet's one-way delay.
+ If the packet is fragmented and if, for whatever reason,
reassembly does not occur, then the packet will be deemed lost.
+ The packet is standard-formed, the default criteria for all
metric definitions defined in Section 15 of ,
otherwise the packet will be deemed lost.
As with other Type-P-* metrics, the detailed methodology will
depend on the Type-P (e.g., protocol number, UDP/TCP port number,
size, precedence).
Generally, for a given Type-P, the methodology would proceed as
follows:
+ Arrange that Src and Dst are synchronized; that is, that they
have clocks that are very closely synchronized with each other and
each fairly close to the actual time.
+ At the Src host, select Src and Dst IP addresses, and form a test
packet of Type-P with these addresses. Any 'padding' portion of the
packet needed only to make the test packet a given size should be
filled with randomized bits to avoid a situation in which the measured
delay is lower than it would otherwise be due to compression
techniques along the path. Note that use of transport layer encryption
will counteract the deployment of network-based analysis and may
reduce the adoption of payload optimizations like compression.
+ At the Dst host, arrange to receive the packet.
+ At the Src host, place a timestamp in the prepared Type-P packet,
and send it towards Dst (ideally minimizing time before sending).
+ If the packet arrives within a reasonable period of time, take a
timestamp as soon as possible upon the receipt of the packet. By
subtracting the two timestamps, an estimate of one-way delay can be
computed. Error analysis of a given implementation of the method must
take into account the closeness of synchronization between Src and
Dst. If the delay between Src's timestamp and the actual sending of
the packet is known, then the estimate could be adjusted by
subtracting this amount; uncertainty in this value must be taken into
account in error analysis. Similarly, if the delay between the actual
receipt of the packet and Dst's timestamp is known, then the estimate
could be adjusted by subtracting this amount; uncertainty in this
value must be taken into account in error analysis. See the next
section, "Errors and Uncertainties", for a more detailed
discussion.
+ If the packet fails to arrive within a reasonable period of time,
Tmax, the one-way delay is taken to be undefined (informally,
infinite). Note that the threshold of 'reasonable' is a parameter of
the metric. These points are examined in detail in , including analysis preferences to assign undefined
delay to packets that fail to arrive with the difficulties emerging
from the informal "infinite delay" assignment, and an estimation of an
upper bound on waiting time for packets in transit. Further, enforcing
a specific constant waiting time on stored singletons of one-way delay
is compliant with this specification and may allow the results to
serve more than one reporting audience.
Issues such as the packet format, the means by which Dst knows when
to expect the test packet, and the means by which Src and Dst are
synchronized are outside the scope of this document. {Comment: We plan
to document elsewhere our own work in describing such more detailed
implementation techniques and we encourage others to as well.}
The description of any specific measurement method should include
an accounting and analysis of various sources of error or uncertainty.
The Framework document provides general guidance on this point, but we
note here the following specifics related to delay metrics:
+ Errors or uncertainties due to uncertainties in the clocks of the
Src and Dst hosts.
+ Errors or uncertainties due to the difference between 'wire time'
and 'host time'.
In addition, the loss threshold may affect the results. Each of
these are discussed in more detail below, along with a section
("Calibration") on accounting for these errors and uncertainties.
The uncertainty in a measurement of one-way delay is related, in
part, to uncertainties in the clocks of the Src and Dst hosts. In
the following, we refer to the clock used to measure when the packet
was sent from Src as the source clock, we refer to the clock used to
measure when the packet was received by Dst as the destination
clock, we refer to the observed time when the packet was sent by the
source clock as Tsource, and the observed time when the packet was
received by the destination clock as Tdest. Alluding to the notions
of synchronization, accuracy, resolution, and skew mentioned in the
Introduction, we note the following:
+ Any error in the synchronization between the source clock and
the destination clock will contribute to error in the delay
measurement. We say that the source clock and the destination clock
have a synchronization error of Tsynch if the source clock is Tsynch
ahead of the destination clock. Thus, if we know the value of Tsynch
exactly, we could correct for clock synchronization by adding Tsynch
to the uncorrected value of Tdest-Tsource.
+ The accuracy of a clock is important only in identifying the
time at which a given delay was measured. Accuracy, per se, has no
importance to the accuracy of the measurement of delay. When
computing delays, we are interested only in the differences between
clock values, not the values themselves.
+ The resolution of a clock adds to uncertainty about any time
measured with it. Thus, if the source clock has a resolution of 10
msec, then this adds 10 msec of uncertainty to any time value
measured with it. We will denote the resolution of the source clock
and the destination clock as Rsource and Rdest, respectively.
+ The skew of a clock is not so much an additional issue as it is
a realization of the fact that Tsynch is itself a function of time.
Thus, if we attempt to measure or to bound Tsynch, this needs to be
done periodically. Over some periods of time, this function can be
approximated as a linear function plus some higher order terms; in
these cases, one option is to use knowledge of the linear component
to correct the clock. Using this correction, the residual Tsynch is
made smaller, but remains a source of uncertainty that must be
accounted for. We use the function Esynch(t) to denote an upper
bound on the uncertainty in synchronization. Thus, |Tsynch(t)| <=
Esynch(t).
Taking these items together, we note that naive computation
Tdest- Tsource will be off by Tsynch(t) +/- (Rsource + Rdest). Using
the notion of Esynch(t), we note that these clock-related problems
introduce a total uncertainty of Esynch(t)+ Rsource + Rdest. This
estimate of total clock-related uncertainty should be included in
the error/uncertainty analysis of any measurement
implementation.
As we have defined one-way delay, we would like to measure the
time between when the test packet leaves the network interface of
Src and when it (completely) arrives at the network interface of
Dst, and we refer to these as "wire times." If the timings are
themselves performed by software on Src and Dst, however, then this
software can only directly measure the time between when Src grabs a
timestamp just prior to sending the test packet and when Dst grabs a
timestamp just after having received the test packet, and we refer
to these two points as "host times".
We note that some systems perform host time stamping on the
network interface hardware, in an attempt to minimize the difference
from wire times.
To the extent that the difference between wire time and host time
is accurately known, this knowledge can be used to correct for host
time measurements and the corrected value more accurately estimates
the desired (wire time) metric.
To the extent, however, that the difference between wire time and
host time is uncertain, this uncertainty must be accounted for in an
analysis of a given measurement method. We denote by Hsource an
upper bound on the uncertainty in the difference between wire time
and host time on the Src host, and similarly define Hdest for the
Dst host. We then note that these problems introduce a total
uncertainty of Hsource+Hdest. This estimate of total wire-vs-host
uncertainty should be included in the error/uncertainty analysis of
any measurement implementation.
Generally, the measured values can be decomposed as follows:
measured value = true value + systematic error + random error
If the systematic error (the constant bias in measured values)
can be determined, it can be compensated for in the reported
results.
reported value = measured value - systematic error
therefore
reported value = true value + random error
The goal of calibration is to determine the systematic and random
error generated by the hosts themselves in as much detail as
possible. At a minimum, a bound ("e") should be found such that the
reported value is in the range (true value - e) to (true value + e)
at least 95 percent of the time. We call "e" the calibration error
for the measurements. It represents the degree to which the values
produced by the measurement host are repeatable; that is, how
closely an actual delay of 30 ms is reported as 30 ms. {Comment: 95
percent was chosen because (1) some confidence level is desirable to
be able to remove outliers, which will be found in measuring any
physical property; (2) a particular confidence level should be
specified so that the results of independent implementations can be
compared; and (3) even with a prototype user-level implementation,
95% was loose enough to exclude outliers.}
From the discussion in the previous two sections, the error in
measurements could be bounded by determining all the individual
uncertainties, and adding them together to form
Esynch(t) + Rsource + Rdest + Hsource + Hdest.
However, reasonable bounds on both the clock-related uncertainty
captured by the first three terms and the host-related uncertainty
captured by the last two terms should be possible by careful design
techniques and calibrating the hosts using a known, isolated,
network in a lab.
For example, the clock-related uncertainties are greatly reduced
through the use of a GPS time source. The sum of Esynch(t) + Rsource
+ Rdest is small, and is also bounded for the duration of the
measurement because of the global time source.
The host-related uncertainties, Hsource + Hdest, could be bounded
by connecting two hosts back-to-back with a high-speed serial link
or isolated LAN segment. In this case, repeated measurements are
measuring the same one-way delay.
If the test packets are small, such a network connection has a
minimal delay that may be approximated by zero. The measured delay
therefore contains only systematic and random error in the
measurement hosts. The "average value" of repeated measurements is
the systematic error, and the variation is the random error.
One way to compute the systematic error, and the random error to
a 95% confidence is to repeat the experiment many times - at least
hundreds of tests. The systematic error would then be the median.
The random error could then be found by removing the systematic
error from the measured values. The 95% confidence interval would be
the range from the 2.5th percentile to the 97.5th percentile of
these deviations from the true value. The calibration error "e"
could then be taken to be the largest absolute value of these two
numbers, plus the clock-related uncertainty. {Comment: as described,
this bound is relatively loose since the uncertainties are added,
and the absolute value of the largest deviation is used. As long as
the resulting value is not a significant fraction of the measured
values, it is a reasonable bound. If the resulting value is a
significant fraction of the measured values, then more exact methods
will be needed to compute the calibration error.}
Note that random error is a function of measurement load. For
example, if many paths will be measured by one host, this might
increase interrupts, process scheduling, and disk I/O (for example,
recording the measurements), all of which may increase the random
error in measured singletons. Therefore, in addition to minimal load
measurements to find the systematic error, calibration measurements
should be performed with the same measurement load that the hosts
will see in the field.
We wish to reiterate that this statistical treatment refers to
the calibration of the host; it is used to "calibrate the meter
stick" and say how well the meter stick reflects reality.
In addition to calibrating the hosts for finite one-way delay,
two checks should be made to ensure that packets reported as losses
were really lost. First, the threshold for loss should be verified.
In particular, ensure the "reasonable" threshold is reasonable: that
it is very unlikely a packet will arrive after the threshold value,
and therefore the number of packets lost over an interval is not
sensitive to the error bound on measurements. Second, consider the
possibility that a packet arrives at the network interface, but is
lost due to congestion on that interface or to other resource
exhaustion (e.g. buffers) in the host.
The calibration and context in which the metric is measured MUST be
carefully considered, and SHOULD always be reported along with metric
results. We now present four items to consider: the Type-P of test
packets, the threshold of infinite delay (if any), error calibration,
and the path traversed by the test packets. This list is not
exhaustive; any additional information that could be useful in
interpreting applications of the metrics should also be reported (see
for extensive discussion of reporting
considerations for different audiences).
As noted in the Framework document , the
value of the metric may depend on the type of IP packets used to
make the measurement, or "type-P". The value of Type-P-One-way-Delay
could change if the protocol (UDP or TCP), port number, size, or
arrangement for special treatment (e.g., IP precedence or RSVP)
changes. The exact Type-P used to make the measurements MUST be
accurately reported.
In addition, the threshold (or methodology to distinguish)
between a large finite delay and loss MUST be reported.
+ If the systematic error can be determined, it SHOULD be removed
from the measured values.
+ You SHOULD also report the calibration error, e, such that the
true value is the reported value plus or minus e, with 95%
confidence (see the last section.)
+ If possible, the conditions under which a test packet with
finite delay is reported as lost due to resource exhaustion on the
measurement host SHOULD be reported.
Finally, the path traversed by the packet SHOULD be reported, if
possible. In general it is impractical to know the precise path a
given packet takes through the network. The precise path may be
known for certain Type-P on short or stable paths. If Type-P
includes the record route (or loose-source route) option in the IP
header, and the path is short enough, and all routers* on the path
support record (or loose-source) route, then the path will be
precisely recorded. This is impractical because the route must be
short enough, many routers do not support (or are not configured
for) record route, and use of this feature would often artificially
worsen the performance observed by removing the packet from
common-case processing. However, partial information is still
valuable context. For example, if a host can choose between two
links* (and hence two separate routes from Src to Dst), then the
initial link used is valuable context. {Comment: For example, with
Merit's NetNow setup, a Src on one NAP can reach a Dst on another
NAP by either of several different backbone networks.}
Given the singleton metric Type-P-One-way-Delay, we now define one
particular sample of such singletons. The idea of the sample is to
select a particular binding of the parameters Src, Dst, and Type-P, then
define a sample of values of parameter T. The means for defining the
values of T is to select a beginning time T0, a final time Tf, and an
average rate lambda, then define a pseudo-random Poisson process of rate
lambda, whose values fall between T0 and Tf. The time interval between
successive values of T will then average 1/lambda.
Note that Poisson sampling is only one way of defining a sample.
Poisson has the advantage of limiting bias, but other methods of
sampling will be appropriate for different situations. For example, a
truncated Poisson distribution may be needed to avoid reactive network
state changes during intervals of inactivity, see section 4.6 of . Sometimes, the goal is sampling with a known bias,
and describes a method for periodic sampling
with random start times.
Type-P-One-way-Delay-Poisson-Stream
+ Src, the IP address of a host
+ Dst, the IP address of a host
+ T0, a time
+ Tf, a time
+ Tmax, a loss threshold waiting time
+ lambda, a rate in reciprocal seconds (or parameters for another
distribution)
A sequence of pairs; the elements of each pair are:
+ T, a time, and
+ dT, either a real number or an undefined number of seconds.
The values of T in the sequence are monotonic increasing. Note that
T would be a valid parameter to Type-P-One-way-Delay, and that dT
would be a valid value of Type-P-One-way-Delay.
Given T0, Tf, and lambda, we compute a pseudo-random Poisson
process beginning at or before T0, with average arrival rate lambda,
and ending at or after Tf. Those time values greater than or equal to
T0 and less than or equal to Tf are then selected. At each of the
times in this process, we obtain the value of Type-P-One-way-Delay at
this time. The value of the sample is the sequence made up of the
resulting <time, delay> pairs. If there are no such pairs, the
sequence is of length zero and the sample is said to be empty.
The reader should be familiar with the in-depth discussion of
Poisson sampling in the Framework document ,
which includes methods to compute and verify the pseudo-random Poisson
process.
We specifically do not constrain the value of lambda, except to
note the extremes. If the rate is too large, then the measurement
traffic will perturb the network, and itself cause congestion. If the
rate is too small, then you might not capture interesting network
behavior. {Comment: We expect to document our experiences with, and
suggestions for, lambda elsewhere, culminating in a "best current
practices" document.}
Since a pseudo-random number sequence is employed, the sequence of
times, and hence the value of the sample, is not fully specified.
Pseudo-random number generators of good quality will be needed to
achieve the desired qualities.
The sample is defined in terms of a Poisson process both to avoid
the effects of self-synchronization and also capture a sample that is
statistically as unbiased as possible. {Comment: there is, of course,
no claim that real Internet traffic arrives according to a Poisson
arrival process.} The Poisson process is used to schedule the delay
measurements. The test packets will generally not arrive at Dst
according to a Poisson distribution, since they are influenced by the
network.
All the singleton Type-P-One-way-Delay metrics in the sequence will
have the same values of Src, Dst, and Type-P.
Note also that, given one sample that runs from T0 to Tf, and given
new time values T0' and Tf' such that T0 <= T0' <= Tf' <= Tf,
the subsequence of the given sample whose time values fall between T0'
and Tf' are also a valid Type-P-One-way-Delay-Poisson-Stream
sample.
The methodologies follow directly from:
+ the selection of specific times, using the specified Poisson
arrival process, and
+ the methodologies discussion already given for the singleton
Type-P-One-way-Delay metric.
Care must, of course, be given to correctly handle out-of-order
arrival of test packets; it is possible that the Src could send one
test packet at TS[i], then send a second one (later) at TS[i+1], while
the Dst could receive the second test packet at TR[i+1], and then
receive the first one (later) at TR[i]. Metrics for reordering may be
found in .
In addition to sources of errors and uncertainties associated with
methods employed to measure the singleton values that make up the
sample, care must be given to analyze the accuracy of the Poisson
process with respect to the wire-times of the sending of the test
packets. Problems with this process could be caused by several things,
including problems with the pseudo-random number techniques used to
generate the Poisson arrival process, or with jitter in the value of
Hsource (mentioned above as uncertainty in the singleton delay
metric). The Framework document shows how to use the Anderson-Darling
test to verify the accuracy of a Poisson process over small time
frames. {Comment: The goal is to ensure that test packets are sent
"close enough" to a Poisson schedule, and avoid periodic
behavior.}
You MUST report the calibration and context for the underlying
singletons along with the stream. (See "Reporting the metric" for
Type-P-One-way-Delay.)
Given the sample metric Type-P-One-way-Delay-Poisson-Stream, we now
offer several statistics of that sample. These statistics are offered
mostly to be illustrative of what could be done. See for additional discussion of statistics that are
relevant to different audiences.
Given a Type-P-One-way-Delay-Poisson-Stream and a percent X between
0% and 100%, the Xth percentile of all the dT values in the Stream. In
computing this percentile, undefined values are treated as infinitely
large. Note that this means that the percentile could thus be
undefined (informally, infinite). In addition, the Type-P-
One-way-Delay-Percentile is undefined if the sample is empty.
Example: suppose we take a sample and the results are:
Stream1 = <
<T1, 100 msec>
<T2, 110 msec>
<T3, undefined>
<T4, 90 msec>
<T5, 500 msec>
>
Then the 50th percentile would be 110 msec, since 90 msec and 100
msec are smaller and 500 msec and 'undefined' are larger. See Section
11.3 of for computing percentiles.
Note that if the possibility that a packet with finite delay is
reported as lost is significant, then a high percentile (90th or 95th)
might be reported as infinite instead of finite.
Given a Type-P-One-way-Delay-Poisson-Stream, the median of all the
dT values in the Stream. In computing the median, undefined values are
treated as infinitely large. As with Type-P-One-way-Delay- Percentile,
Type-P-One-way-Delay-Median is undefined if the sample is empty.
As noted in the Framework document, the median differs from the
50th percentile only when the sample contains an even number of
values, in which case the mean of the two central values is used.
Example: suppose we take a sample and the results are:
Stream2 = < <T1, 100 msec> <T2, 110 msec> <T3,
undefined> <T4, 90 msec> >
Then the median would be 105 msec, the mean of 100 msec and 110
msec, the two central values.
Given a Type-P-One-way-Delay-Poisson-Stream, the minimum of all the
dT values in the Stream. In computing this, undefined values are
treated as infinitely large. Note that this means that the minimum
could thus be undefined (informally, infinite) if all the dT values
are undefined. In addition, the Type-P-One-way-Delay-Minimum is
undefined if the sample is empty.
In the above example, the minimum would be 90 msec.
Note: This statistic is deprecated in this version of the memo
because of lack of use.
Given a Type-P-One-way-Delay-Poisson-Stream and a time duration
threshold, the fraction of all the dT values in the Stream less than
or equal to the threshold. The result could be as low as 0% (if all
the dT values exceed threshold) or as high as 100%. Type-P-One-way-
Delay-Inverse-Percentile is undefined if the sample is empty.
In the above example, the Inverse-Percentile of 103 msec would be
50%.
Conducting Internet measurements raises both security and privacy
concerns. This memo does not specify an implementation of the metrics,
so it does not directly affect the security of the Internet nor of
applications which run on the Internet. However, implementations of
these metrics must be mindful of security and privacy concerns.
There are two types of security concerns: potential harm caused by
the measurements, and potential harm to the measurements. The
measurements could cause harm because they are active, and inject
packets into the network. The measurement parameters MUST be carefully
selected so that the measurements inject trivial amounts of additional
traffic into the networks they measure. If they inject "too much"
traffic, they can skew the results of the measurement, and in extreme
cases cause congestion and denial of service.
The measurements themselves could be harmed by routers giving
measurement traffic a different priority than "normal" traffic, or by an
attacker injecting artificial measurement traffic. If routers can
recognize measurement traffic and treat it separately, the measurements
will not reflect actual user traffic. Therefore, the measurement
methodologies SHOULD include appropriate techniques to reduce the
probability measurement traffic can be distinguished from "normal"
traffic.
If an attacker injects packets emulating traffic that are accepted as
legitimate, the loss ratio or other measured values could be corrupted.
Authentication techniques, such as digital signatures, may be used where
appropriate to guard against injected traffic attacks.
The privacy concerns of network measurement are limited by the active
measurements described in this memo. Unlike passive measurements, there
can be no release of existing user data.
This memo makes no requests of IANA.
Special thanks are due to Vern Paxson of Lawrence Berkeley Labs for
his helpful comments on issues of clock uncertainty and statistics.
Thanks also to Garry Couch, Will Leland, Andy Scherrer, Sean Shapira,
and Roland Wittig for several useful suggestions.
K-sample Anderson-Darling Tests of fit, for continuous and
discrete cases
Boeing Computer
Services
Simon Fraser University