| RFC 0000 | IP Capacity Metrics/Methods | July 2021 |
| Morton, et al. | Standards Track | [Page] |
This memo revisits the problem of Network Capacity metrics first examined in RFC 5136. The memo specifies a more practical Maximum IP-Layer Capacity metric definition catering for measurement purposes and outlines the corresponding methods of measurement.¶
This is an Internet Standards Track document.¶
This document is a product of the Internet Engineering Task Force (IETF). It represents the consensus of the IETF community. It has received public review and has been approved for publication by the Internet Engineering Steering Group (IESG). Further information on Internet Standards is available in Section 2 of RFC 7841.¶
Information about the current status of this document, any errata, and how to provide feedback on it may be obtained at https://www.rfc-editor.org/info/rfc0000.¶
Copyright (c) 2021 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.¶
The IETF's efforts to define Network and Bulk Transport Capacity have been chartered and progressed for over twenty years. Over that time, the performance community has seen the development of Informative definitions in [RFC3148] for the Framework for Bulk Transport Capacity (BTC), [RFC5136] for Network Capacity and Maximum IP-Layer Capacity, and the Experimental metric definitions and methods in "Model-Based Metrics for Bulk Transport Capacity" [RFC8337].¶
This memo revisits the problem of Network Capacity metrics examined first in [RFC3148] and later in [RFC5136]. Maximum IP-Layer Capacity and Bulk Transfer Capacity [RFC3148] (goodput) are different metrics. Maximum IP-Layer Capacity is like the theoretical goal for goodput. There are many metrics in [RFC5136], such as Available Capacity. Measurements depend on the network path under test and the use case. Here, the main use case is to assess the Maximum Capacity of one or more networks where the subscriber receives specific performance assurances, sometimes referred to as Internet access, or where a limit of the technology used on a path is being tested. For example, when a user subscribes to a 1 Gbps service, then the user, the service provider, and possibly other parties want to assure that an appropriate level of performance is delivered. When a test confirms the subscribed performance level, then a tester can seek the location of a bottleneck elsewhere.¶
This memo recognizes the importance of a definition of a Maximum IP-Layer Capacity Metric at a time when Internet subscription speeds have increased dramatically -- a definition that is both practical and effective for the performance community's needs, including Internet users. The metric definition is intended to use Active Methods of Measurement [RFC7799], and a method of measurement is included.¶
The most direct active measurement of IP-Layer Capacity would use IP packets, but in practice a transport header is needed to traverse address and port translators. UDP offers the most direct assessment possibility, and in the measurement study to investigate whether UDP is viable as a general Internet transport protocol [copycat], the authors found that a high percentage of paths tested support UDP transport. A number of liaisons have been exchanged on this topic [LS-SG12-A] [LS-SG12-B], discussing the laboratory and field tests that support the UDP-based approach to IP-Layer Capacity measurement.¶
This memo also recognizes the many updates to the IP Performance Metrics Framework [RFC2330] published over twenty years and makes use of [RFC7312] for the Advanced Stream and Sampling Framework and [RFC8468] for its IPv4, IPv6, and IPv4-IPv6 Coexistence Updates.¶
Appendix A describes the load rate adjustment algorithm in pseudocode. Appendix B discusses the algorithm's compliance with [RFC8085].¶
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.¶
The scope of this memo is to define Active Measurement metrics and corresponding methods to unambiguously determine Maximum IP-Layer Capacity and useful secondary metrics.¶
Another goal is to harmonize the specified metric and method across the industry, and this memo is the vehicle that captures IETF consensus, possibly resulting in changes to the specifications of other Standards Development Organizations (SDOs) (through each SDO's normal contribution process or through liaison exchange).¶
Secondary goals are to add considerations for test procedures and to provide interpretation of the Maximum IP-Layer Capacity results (to identify cases where more testing is warranted, possibly with alternate configurations). Fostering the development of protocol support for this metric and method of measurement is also a goal of this memo (all active testing protocols currently defined by the IPPM WG ("IPPM" stands for "IP Performance Metrics") are UDP based, meeting a key requirement of these methods). The supporting protocol development to measure this metric according to the specified method is a key future contribution to Internet measurement.¶
The load rate adjustment algorithm's scope is limited to helping determine the Maximum IP-Layer Capacity in the context of an infrequent, diagnostic, short-term measurement. It is RECOMMENDED to discontinue non-measurement traffic that shares a subscriber's dedicated resources while testing: measurements may not be accurate, and throughput of competing elastic traffic may be greatly reduced.¶
The primary application of the metric and method of measurement described here is the same as what is described in Section 2 of [RFC7497], where:¶
The access portion of the network is the focus of this problem statement. The user typically subscribes to a service with bidirectional Internet access partly described by rates in bits per second.¶
In addition, the use of the load rate adjustment algorithm described in Section 8.1 has the following additional applicability limitations:¶
Further, the metric and method of measurement are intended for use where specific exact path information is unknown within a range of possible values:¶
Finally, the measurement system's load rate adjustment algorithm SHALL NOT be provided with the exact capacity value to be validated a priori. This restriction fosters a fair result and removes an opportunity for bad actors to operate with knowledge of the "right answer".¶
As with any problem that has been worked on for many years in various SDOs without any special attempts at coordination, various solutions for metrics and methods have emerged.¶
There are five factors that have changed (or begun to change) in the 2013-2019 time frame, and the presence of any one of them on the path requires features in the measurement design to account for the changes:¶
This section lists the REQUIRED input factors to specify a Sender or Receiver metric.¶
A non-Parameter that is required for several metrics is defined below:¶
Note that time stamp format and resolution, sequence numbers, etc. will be established by the chosen test protocol standard or implementation.¶
This section sets requirements for the singleton metric that supports the Maximum IP-Layer Capacity Metric definition in Section 6.¶
Type-P-One-way-IP-Capacity, or informally called IP-Layer Capacity.¶
Note that Type-P depends on the chosen method.¶
This section lists the REQUIRED input factors to specify the metric, beyond those listed in Section 4.¶
No additional Parameters are needed.¶
This section defines the REQUIRED aspects of the measurable IP-Layer Capacity metric (unless otherwise indicated) for measurements between specified Source and Destination hosts:¶
Define the IP-Layer Capacity, C(T,dt,PM), to be the number of IP-Layer bits (including header and data fields) in packets that can be transmitted from the Src host and correctly received by the Dst host during one contiguous sub-interval, dt in length. The IP-Layer Capacity depends on the Src and Dst hosts, the host addresses, and the path between the hosts.¶
The number of these IP-Layer bits is designated n0[dtn,dtn+1] for a specific dt.¶
When the packet size is known and of fixed size, the packet count during a single sub-interval dt multiplied by the total bits in IP header and data fields is equal to n0[dtn,dtn+1].¶
Anticipating a Sample of Singletons, the number of sub-intervals with duration dt MUST be set to a natural number m, so that T+I = T + m*dt with dtn+1 - dtn = dt for 1 <= n <= m.¶
Parameter PM represents other performance metrics (see Section 5.4 below); their measurement results SHALL be collected during measurement of IP-Layer Capacity and associated with the corresponding dtn for further evaluation and reporting. Users SHALL specify the parameter Tmax as required by each metric's reference definition.¶
Mathematically, this definition is represented as (for each n):¶
( n0[dtn,dtn+1] )
C(T,dt,PM) = -------------------------
dt
and:¶
Measurements according to these definitions SHALL use the UDP transport layer. Standard-formed packets are specified in Section 5 of [RFC8468]. The measurement SHOULD use a randomized Source port or equivalent technique, and SHOULD send responses from the Source address matching the test packet destination address.¶
Some compression affects on measurement are discussed in Section 6 of [RFC8468].¶
RTD[dtn,dtn+1] is defined as a Sample of the Round-Trip Delay [RFC2681] between the Src host and the Dst host over the interval [T,T+I] (that contains equal non-overlapping intervals of dt). The "reasonable period of time" in [RFC2681] is the parameter Tmax in this memo. The statistics used to summarize RTD[dtn,dtn+1] MAY include the minimum, maximum, median, and mean, and the range = (maximum - minimum) is referred to below in Section 8.1 for load adjustment purposes.¶
OWL[dtn,dtn+1] is defined as a Sample of the [RFC7680] One-Way Loss between the Src host and the Dst host over the interval [T,T+I] (that contains equal non-overlapping intervals of dt). The statistics used to summarize OWL[dtn,dtn+1] MAY include the lost packet count and the lost packet ratio.¶
Other metrics MAY be measured: one-way reordering, duplication, and delay variation.¶
See the corresponding section for Maximum IP-Layer Capacity.¶
The IP-Layer Capacity SHOULD be reported with at least single Megabit resolution, in units of Megabits per second (Mbps), (which is 1,000,000 bits per second to avoid any confusion).¶
The related One-Way Loss metric and Round-Trip Delay measurements for the same Singleton SHALL be reported, also with meaningful resolution for the values measured.¶
Individual Capacity measurements MAY be reported in a manner consistent with the Maximum IP-Layer Capacity, see Section 9.¶
This section sets requirements for the following components to support the Maximum IP-Layer Capacity Metric.¶
Type-P-One-way-Max-IP-Capacity, or informally called Maximum IP-Layer Capacity.¶
Note that Type-P depends on the chosen method.¶
This section lists the REQUIRED input factors to specify the metric, beyond those listed in Section 4.¶
No additional Parameters or definitions are needed.¶
This section defines the REQUIRED aspects of the Maximum IP-Layer Capacity metric (unless otherwise indicated) for measurements between specified Source and Destination hosts:¶
Define the Maximum IP-Layer Capacity, Maximum_C(T,I,PM), to be the maximum number of IP-Layer bits n0[dtn,dtn+1] divided by dt that can be transmitted in packets from the Src host and correctly received by the Dst host, over all dt length intervals in [T, T+I] and meeting the PM criteria. Equivalently the Maximum of a Sample of size m of C(T,I,PM) collected during the interval [T, T+I] and meeting the PM criteria.¶
The number of sub-intervals with duration dt MUST be set to a natural number m, so that T+I = T + m*dt with dtn+1 - dtn = dt for 1 <= n <= m.¶
Parameter PM represents the other performance metrics (see Section 6.4 below) and their measurement results for the Maximum IP-Layer Capacity. At least one target performance threshold (PM criterion) MUST be defined. If more than one metric and target performance threshold is defined, then the sub-interval with the maximum number of bits transmitted MUST meet all the target performance thresholds. Users SHALL specify the parameter Tmax as required by each metric's reference definition.¶
Mathematically, this definition can be represented as:¶
max ( n0[dtn,dtn+1] )
[T,T+I]
Maximum_C(T,I,PM) = -------------------------
dt
where:
T T+I
_________________________________________
| | | | | | | | | | |
dtn=1 2 3 4 5 6 7 8 9 10 n+1
n=m
and:¶
In this definition, the m sub-intervals can be viewed as trials when the Src host varies the transmitted packet rate, searching for the maximum n0 that meets the PM criteria measured at the Dst host in a test of duration, I. When the transmitted packet rate is held constant at the Src host, the m sub-intervals may also be viewed as trials to evaluate the stability of n0 and metric(s) in the PM list over all dt-length intervals in I.¶
Measurements according to these definitions SHALL use the UDP transport layer.¶
RTD[dtn,dtn+1] and OWL[dtn,dtn+1] are defined in Section 5.4. Here, the test intervals are increased to match the capacity Samples, RTD[T,I] and OWL[T,I].¶
The interval dtn,dtn+1 where Maximum_C[T,I,PM] occurs is the reporting sub-interval within RTD[T,I] and OWL[T,I].¶
Other metrics MAY be measured: one-way reordering, duplication, and delay variation.¶
If traffic conditioning (e.g., shaping, policing) applies along a path for which Maximum_C(T,I,PM) is to be determined, different values for dt SHOULD be picked and measurements be executed during multiple intervals [T, T+I]. Each duration dt SHOULD be chosen so that it is an integer multiple of increasing values k times serialization delay of a path MTU at the physical interface speed where traffic conditioning is expected. This should avoid taking configured burst tolerance singletons as a valid Maximum_C(T,I,PM) result.¶
A Maximum_C(T,I,PM) without any indication of bottleneck congestion, be that an increasing latency, packet loss or ECN marks during a measurement interval I, is likely to underestimate Maximum_C(T,I,PM).¶
The IP-Layer Capacity SHOULD be reported with at least single Megabit resolution, in units of Megabits per second (Mbps) (which is 1,000,000 bits per second to avoid any confusion).¶
The related One-Way Loss metric and Round-Trip Delay measurements for the same Singleton SHALL be reported, also with meaningful resolution for the values measured.¶
When there are demonstrated and repeatable Capacity modes in the Sample, then the Maximum IP-Layer Capacity SHALL be reported for each mode, along with the relative time from the beginning of the stream that the mode was observed to be present. Bimodal Maximum IP-Layer Capacities have been observed with some services, sometimes called a "turbo mode" intending to deliver short transfers more quickly, or reduce the initial buffering time for some video streams. Note that modes lasting less than dt duration will not be detected.¶
Some transmission technologies have multiple methods of operation that may be activated when channel conditions degrade or improve, and these transmission methods may determine the Maximum IP-Layer Capacity. Examples include line-of-sight microwave modulator constellations, or cellular modem technologies where the changes may be initiated by a user moving from one coverage area to another. Operation in the different transmission methods may be observed over time, but the modes of Maximum IP-Layer Capacity will not be activated deterministically as with the "turbo mode" described in the paragraph above.¶
This section sets requirements for the following components to support the IP-Layer Sender Bitrate Metric. This metric helps to check that the sender actually generated the desired rates during a test, and measurement takes place at the Src host to network path interface (or as close as practical within the Src host). It is not a metric for path performance.¶
Type-P-IP-Sender-Bit-Rate, or informally called IP-Layer Sender Bitrate.¶
Note that Type-P depends on the chosen method.¶
This section lists the REQUIRED input factors to specify the metric, beyond those listed in Section 4.¶
S SHALL be longer than I, primarily to account for on-demand activation of the path, or any preamble to testing required, and the delay of the path.¶
st SHOULD be much smaller than the sub-interval dt and on the same order as FT, otherwise the rate measurement will include many rate adjustments and include more time smoothing, thus missing the Maximum IP-Layer Capacity. The st parameter does not have relevance when the Source is transmitting at a fixed rate throughout S.¶
This section defines the REQUIRED aspects of the IP-Layer Sender Bitrate metric (unless otherwise indicated) for measurements at the specified Source on packets addressed for the intended Destination host and matching the required Type-P:¶
Define the IP-Layer Sender Bit Rate, B(S,st), to be the number of IP-Layer bits (including header and data fields) that are transmitted from the Source with address pair Src and Dst during one contiguous sub-interval, st, during the test interval S (where S SHALL be longer than I), and where the fixed-size packet count during that single sub-interval st also provides the number of IP-Layer bits in any interval, [stn,stn+1].¶
Measurements according to these definitions SHALL use the UDP transport layer. Any feedback from Dst host to Src host received by Src host during an interval [stn,stn+1] SHOULD NOT result in an adaptation of the Src host traffic conditioning during this interval (rate adjustment occurs on st interval boundaries).¶
Both the Sender and Receiver or (Source and Destination) bit rates SHOULD be assessed as part of an IP-Layer Capacity measurement. Otherwise, an unexpected sending rate limitation could produce an erroneous Maximum IP-Layer Capacity measurement.¶
The IP-Layer Sender Bit Rate SHALL be reported with meaningful resolution, in units of Megabits per second (which is 1,000,000 bits per second to avoid any confusion).¶
Individual IP-Layer Sender Bit Rate measurements are discussed further in Section 9.¶
The architecture of the method REQUIRES two cooperating hosts operating in the roles of Src (test packet sender) and Dst (receiver), with a measured path and return path between them.¶
The duration of a test, parameter I, MUST be constrained in a production network, since this is an active test method and it will likely cause congestion on the Src to Dst host path during a test.¶
The algorithm described in this section MUST NOT be used as a general Congestion Control Algorithm (CCA). As stated in Section 2 ("Scope, Goals, and Applicability"), the load rate adjustment algorithm's goal is to help determine the Maximum IP-Layer Capacity in the context of an infrequent, diagnostic, short-term measurement. There is a tradeoff between test duration (also the test data volume) and algorithm aggressiveness (speed of ramp-up and down to the Maximum IP-Layer Capacity). The parameter values chosen below strike a well-tested balance among these factors.¶
A table SHALL be pre-built (by the test initiator) defining all the offered load rates that will be supported (R1 through Rn, in ascending order, corresponding to indexed rows in the table). It is RECOMMENDED that rates begin with 0.5 Mbps at index zero, use 1 Mbps at index one, and then continue in 1 Mbps increments to 1 Gbps. Above 1 Gbps, and up to 10 Gbps, it is RECOMMENDED that 100 Mbps increments be used. Above 10 Gbps, increments of 1 Gbps are RECOMMENDED. A higher initial IP-Layer Sender Bitrate might be configured when the test operator is certain that the Maximum IP-Layer Capacity is well above the initial IP-Layer Sender Bitrate and factors such as test duration and total test traffic play an important role. The sending rate table SHOULD backet the Maximum Capacity where it will make measurements, including constrained rates less than 500 kbps if applicable.¶
Each rate is defined as datagrams of size ss, sent as a burst of count cc, each time interval tt (default for tt is 1 ms, a likely system tick-interval). While it is advantageous to use datagrams of as large a size as possible, it may be prudent to use a slightly smaller maximum that allows for secondary protocol headers and/or tunneling without resulting in IP-Layer fragmentation. Selection of a new rate is indicated by a calculation on the current row, Rx. For example:¶
At the beginning of a test, the sender begins sending at rate R1 and the receiver starts a feedback timer of duration FT (while awaiting inbound datagrams). As datagrams are received they are checked for sequence number anomalies (loss, out-of-order, duplication, etc.) and the delay range is measured (one-way or round-trip). This information is accumulated until the feedback timer FT expires and a status feedback message is sent from the receiver back to the sender, to communicate this information. The accumulated statistics are then reset by the receiver for the next feedback interval. As feedback messages are received back at the sender, they are evaluated to determine how to adjust the current offered load rate (Rx).¶
If the feedback indicates that no sequence number anomalies were detected AND the delay range was below the lower threshold, the offered load rate is increased. If congestion has not been confirmed up to this point (see below for the method to declare congestion), the offered load rate is increased by more than one rate (e.g., Rx+10). This allows the offered load to quickly reach a near-maximum rate. Conversely, if congestion has been previously confirmed, the offered load rate is only increased by one (Rx+1). However, if a rate threshold between high and very high sending rates (such as 1 Gbps) is exceeded, the offered load rate is only increased by one (Rx+1) above the rate threshold in any congestion state.¶
If the feedback indicates that sequence number anomalies were detected OR the delay range was above the upper threshold, the offered load rate is decreased. The RECOMMENDED threshold values are 0 for sequence number gaps and 30 ms for lower and 90 ms for upper delay thresholds, respectively. Also, if congestion is now confirmed for the first time by the current feedback message being processed, then the offered load rate is decreased by more than one rate (e.g., Rx-30). This one-time reduction is intended to compensate for the fast initial ramp-up. In all other cases, the offered load rate is only decreased by one (Rx-1).¶
If the feedback indicates that there were no sequence number anomalies AND the delay range was above the lower threshold, but below the upper threshold, the offered load rate is not changed. This allows time for recent changes in the offered load rate to stabilize, and the feedback to represent current conditions more accurately.¶
Lastly, the method for inferring congestion is that there were sequence number anomalies AND/OR the delay range was above the upper threshold for two consecutive feedback intervals. The algorithm described above is also illustrated in Annex B of ITU-T Rec. Y.1540, 2020 version[Y.1540] and is implemented in Appendix A ("Load Rate Adjustment Pseudocode") in this memo.¶
The load rate adjustment algorithm MUST include timers that stop the test when received packet streams cease unexpectedly. The timeout thresholds are provided in the table below, along with values for all other parameters and variables described in this section. Operations of non-obvious parameters appear below:¶
| Parameter | Default | Tested Range or values | Expected Safe Range (not entirely tested, other values NOT RECOMMENDED) |
|---|---|---|---|
| FT, feedback time interval | 50 ms | 20 ms, 50 ms, 100 ms | 20 ms <= FT <= 250 ms Larger values may slow the rate increase and fail to find the max |
| Feedback message timeout (stop test) | L*FT, L=20 (1 sec with FT=50 ms) | L=100 with FT=50 ms (5 sec) | 0.5 sec <= L*FT <= 30 sec Upper limit for very unreliable test paths only |
| load packet timeout (stop test) | 1 sec | 5 sec | 0.250 sec - 30 sec Upper limit for very unreliable test paths only |
| table index 0 | 0.5 Mbps | 0.5 Mbps | when testing <=10 Gbps |
| table index 1 | 1 Mbps | 1 Mbps | when testing <=10 Gbps |
| table index (step) size | 1 Mbps | 1 Mbps<=rate<= 1 Gbps | same as tested |
| table index (step) size, rate>1 Gbps | 100 Mbps | 1 Gbps<=rate<= 10 Gbps | same as tested |
| table index (step) size, rate>10 Gbps | 1 Gbps | untested | >10 Gbps |
| ss, UDP payload size, bytes | none | <=1222 | Recommend max at largest value that avoids fragmentation; use of too-small payload size might result in unexpected sender limitations. |
| cc, burst count | none | 1<=cc<= 100 | same as tested. Vary cc as needed to create the desired maximum sending rate. Sender buffer size may limit cc in implementation. |
| tt, burst interval | 100 microsec | 100 microsec, 1 msec | available range of "tick" values (HZ param) |
| low delay range threshold | 30 ms | 5 ms, 30 ms | same as tested |
| high delay range threshold | 90 ms | 10 ms, 90 ms | same as tested |
| sequence error threshold | 0 | 0, 100 | same as tested |
| consecutive errored status report threshold | 2 | 2 | Use values >1 to avoid misinterpreting transient loss |
| Fast mode increase, in table index steps | 10 | 10 | 2 <= steps <= 30 |
| Fast mode decrease, in table index steps | 3 * Fast mode increase | 3 * Fast mode increase | same as tested |
As a consequence of default parameterization, the Number of table steps in total for rates <10 Gbps is 2000 (excluding index 0).¶
A related sender backoff response to network conditions occurs when one or more status feedback messages fail to arrive at the sender.¶
If no status feedback messages arrive at the sender for the interval greater than the Lost Status Backoff timeout:¶
UDRT + (2+w)*FT = Lost Status Backoff timeout
where:
UDRT = upper delay range threshold (default 90 ms)
FT = feedback time interval (default 50 ms)
w = number of repeated timeouts (w=0 initially, w++ on each
timeout, and reset to 0 when a message is received)
¶
beginning when the last message (of any type) was successfully received at the sender:¶
Then the offered load SHALL be decreased, following the same process as when the feedback indicates presence of one or more sequence number anomalies OR the delay range was above the upper threshold (as described above), with the same load rate adjustment algorithm variables in their current state. This means that rate reduction and congestion confirmation can result from a three-way OR that includes lost status feedback messages, sequence errors, or delay variation.¶
The RECOMMENDED initial value for w is 0, taking a Round-Trip Time (RTT) of less than FT into account. A test with RTT longer than FT is a valid reason to increase the initial value of w appropriately. Variable w SHALL be incremented by 1 whenever the Lost Status Backoff timeout is exceeded. So with FT = 50 ms and UDRT = 90 ms, a status feedback message loss would be declared at 190 ms following a successful message, again at 50 ms after that (240 ms total), and so on.¶
Also, if congestion is now confirmed for the first time by a Lost Status Backoff timeout, then the offered load rate is decreased by more than one rate (e.g., Rx-30). This one-time reduction is intended to compensate for the fast initial ramp-up. In all other cases, the offered load rate is only decreased by one (Rx-1).¶
Appendix B discusses compliance with the applicable mandatory requirements of [RFC8085], consistent with the goals of the IP-Layer Capacity Metric and Method, including the load rate adjustment algorithm described in this section.¶
It is of course necessary to calibrate the equipment performing the IP-Layer Capacity measurement, to ensure that the expected capacity can be measured accurately, and that equipment choices (processing speed, interface bandwidth, etc.) are suitably matched to the measurement range.¶
When assessing a Maximum rate as the metric specifies, artificially high (optimistic) values might be measured until some buffer on the path is filled. Other causes include bursts of back-to-back packets with idle intervals delivered by a path, while the measurement interval (dt) is small and aligned with the bursts. The artificial values might result in an unsustainable Maximum Capacity observed when the method of measurement is searching for the Maximum, and that would not do. This situation is different from the bi-modal service rates (discussed under Reporting), which are characterized by a multi-second duration (much longer than the measured RTT) and repeatable behavior.¶
There are many ways that the Method of Measurement could handle this false-max issue. The default value for measurement of singletons (dt = 1 second) has proven to be of practical value during tests of this method, allows the bimodal service rates to be characterized, and it has an obvious alignment with the reporting units (Mbps).¶
Another approach comes from Section 24 of [RFC2544] and its discussion of Trial duration, where relatively short trials conducted as part of the search are followed by longer trials to make the final determination. In the production network, measurements of Singletons and Samples (the terms for trials and tests of Lab Benchmarking) must be limited in duration because they may be service-affecting. But there is sufficient value in repeating a Sample with a fixed sending rate determined by the previous search for the Maximum IP-Layer Capacity, to qualify the result in terms of the other performance metrics measured at the same time.¶
A qualification measurement for the search result is a subsequent measurement, sending at a fixed 99.x % of the Maximum IP-Layer Capacity for I, or an indefinite period. The same Maximum Capacity Metric is applied, and the Qualification for the result is a Sample without packet loss or a growing minimum delay trend in subsequent singletons (or each dt of the measurement interval, I). Samples exhibiting losses or increasing queue occupation require a repeated search and/or test at reduced fixed sender rate for qualification.¶
Here, as with any Active Capacity test, the test duration must be kept short. 10 second tests for each direction of transmission are common today. The default measurement interval specified here is I = 10 seconds. The combination of a fast and congestion-aware search method and user-network coordination make a unique contribution to production testing. The Maximum IP Capacity metric and method for assessing performance is very different from classic [RFC2544] Throughput metric and methods: it uses near-real-time load adjustments that are sensitive to loss and delay, similar to other congestion control algorithms used on the Internet every day, along with limited duration. On the other hand, [RFC2544] Throughput measurements can produce sustained overload conditions for extended periods of time. Individual trials in a test governed by a binary search can last 60 seconds for each step, and the final confirmation trial may be even longer. This is very different from "normal" traffic levels, but overload conditions are not a concern in the isolated test environment. The concerns raised in [RFC6815] were that [RFC2544] methods would be let loose on production networks, and instead the authors challenged the standards community to develop metrics and methods like those described in this memo.¶
In general, the wide-spread measurements that this memo encourages will encounter wide-spread behaviors. The bimodal IP Capacity behaviors already discussed in Section 6.6 are good examples.¶
In general, it is RECOMMENDED to locate test endpoints as close to the intended measured link(s) as practical (this is not always possible for reasons of scale; there is a limit on number of test endpoints coming from many perspectives, management and measurement traffic for example). The testing operator MUST set a value for the MaxHops parameter, based on the expected path length. This parameter can keep measurement traffic from straying too far beyond the intended path.¶
The path measured may be stateful based on many factors, and the Parameter "Time of day" when a test starts may not be enough information. Repeatable testing may require the time from the beginning of a measured flow, and how the flow is constructed including how much traffic has already been sent on that flow when a state-change is observed, because the state-change may be based on time or bytes sent or both. Both load packets and status feedback messages MUST contain sequence numbers, which helps with measurements based on those packets.¶
Many different types of traffic shapers and on-demand communications access technologies may be encountered, as anticipated in [RFC7312], and play a key role in measurement results. Methods MUST be prepared to provide a short preamble transmission to activate on-demand communications access and to discard the preamble from subsequent test results.¶
Conditions which might be encountered during measurement, where packet losses may occur independently of the measurement sending rate:¶
It is possible to mitigate these conditions using the flexibility of the load rate adjustment algorithm described in Section 8.1 above (tuning specific parameters).¶
If the measurement flow burst duration happens to be on the order of or smaller than the burst size of a shaper or a policer in the path, then the line rate might be measured rather than the bandwidth limit imposed by the shaper or policer. If this condition is suspected, alternate configurations SHOULD be used.¶
In general, results depend on the sending stream characteristics; the measurement community has known this for a long time, and needs to keep it front of mind. Although the default is a single flow (F=1) for testing, use of multiple flows may be advantageous for the following reasons:¶
Each flow would be controlled using its own implementation of the load rate adjustment (search) algorithm.¶
It is obviously counter-productive to run more than one independent and concurrent test (regardless of the number of flows in the test stream) attempting to measure the maximum capacity on a single path. The number of concurrent, independent tests of a path SHALL be limited to one.¶
Tests of a v4-v6 transition mechanism might well be the intended subject of a capacity test. As long as the IPv4 and IPv6 packets sent/received are both standard-formed, this should be allowed (and the change in header size easily accounted for on a per-packet basis).¶
As testing continues, implementers should expect some evolution in the methods. The ITU-T has published a Supplement (60) to the Y-series of Recommendations, "Interpreting ITU-T Y.1540 Maximum IP-Layer Capacity measurements", [Y.Sup60], which is the result of continued testing with the metric, and those results have improved the method described here.¶
RFC Editor: This section is for the benefit of the Document Shepherd's form, and will be deleted prior to publication.¶
Much of the development of the method and comparisons with existing methods conducted at IETF Hackathons and elsewhere have been based on the example udpst Linux measurement tool (which is a working reference for further development) [udpst]. The current project:¶
The singleton IP-Layer Capacity results SHOULD be accompanied by the context under which they were measured.¶
The Maximum IP-Layer Capacity results SHOULD be reported in the format of a table with a row for each of the test Phases and Number of Flows. There SHOULD be columns for the phases with number of flows, and for the resultant Maximum IP-Layer Capacity results for the aggregate and each flow tested.¶
As mentioned in Section 6.6, bi-modal (or multi-modal) maxima SHALL be reported for each mode separately.¶
| Phase, # Flows | Maximum IP-Layer Capacity, Mbps | Loss Ratio | RTT min, max, msec |
|---|---|---|---|
| Search,1 | 967.31 | 0.0002 | 30, 58 |
| Verify,1 | 966.00 | 0.0000 | 30, 38 |
Static and configuration parameters:¶
The sub-interval time, dt, MUST accompany a report of Maximum IP-Layer Capacity results, and the remaining Parameters from Section 4 ("General Parameters and Definitions").¶
The PM list metrics corresponding to the sub-interval where the Maximum Capacity occurred MUST accompany a report of Maximum IP-Layer Capacity results, for each test phase.¶
The IP-Layer Sender Bit Rate results SHOULD be reported in the format of a table with a row for each of the test phases, sub-intervals (st) and number of flows. There SHOULD be columns for the phases with number of flows, and for the resultant IP-Layer Sender Bit Rate results for the aggregate and each flow tested.¶
| Phase, Flow or Aggregate | st, sec | Sender Bitrate, Mbps |
|---|---|---|
| Search,1 | 0.00 - 0.05 | 345 |
| Search,2 | 0.00 - 0.05 | 289 |
| Search,Agg | 0.00 - 0.05 | 634 |
Static and configuration parameters:¶
The subinterval time, st, MUST accompany a report of Sender IP-Layer Bit Rate results.¶
Also, the values of the remaining Parameters from Section 4 ("General Parameters and Definitions") MUST be reported.¶
As a part of the multi-Standards Development Organization (SDO) harmonization of this metric and method of measurement, one of the areas where the Broadband Forum (BBF) contributed its expertise was in the definition of an information model and data model for configuration and reporting. These models are consistent with the metric parameters and default values specified as lists is this memo. [TR-471] provides the Information model that was used to prepare a full data model in related BBF work. The BBF has also carefully considered topics within its purview, such as placement of measurement systems within the Internet access architecture. For example, timestamp resolution requirements that influence the choice of the test protocol are provided in Table 2 of [TR-471].¶
Active metrics and measurements have a long history of security considerations. The security considerations that apply to any active measurement of live paths are relevant here. See [RFC4656] and [RFC5357].¶
When considering privacy of those involved in measurement or those whose traffic is measured, the sensitive information available to potential observers is greatly reduced when using active techniques which are within this scope of work. Passive observations of user traffic for measurement purposes raise many privacy issues. We refer the reader to the privacy considerations described in the Large Scale Measurement of Broadband Performance (LMAP) Framework [RFC7594], which covers active and passive techniques.¶
There are some new considerations for Capacity measurement as described in this memo.¶
The exact specification of these features is left for the future protocol development.¶
This memo makes no requests of IANA.¶
This appendix provides a pseudocode implementation of the algorithm described in Section 8.1, followed by a table of values and descriptions.¶
if ( seqErr == 0 && delay < lowThresh ) {
if ( Rx < hSpeedTresh && slowAdjCount < slowAdjThresh ) {
Rx += highSpeedDelta;
slowAdjCount = 0;
} else {
if ( Rx < maxLoadRates - 1 )
Rx++;
}
} else if ( seqErr > 0 || delay > upperThresh ) {
slowAdjCount++;
if ( Rx < hSpeedTresh && slowAdjCount == slowAdjThresh ) {
if ( Rx > highSpeedDelta * 3 )
Rx -= highSpeedDelta * 3;
else
Rx = 0;
} else {
if ( Rx > 0 )
Rx--;
}
}
¶
| Value | Description |
|---|---|
| Rx = 0 | The current sending rate (equivalent to a row of the table) |
| seqErr = 0 | Measured count of any of Loss or Reordering impairments |
| delay = 0 | Measured Range of Round-Trip Delay (RTD), ms |
| lowThresh = 30 | Low threshold for the Range of RTD, ms |
| upperThresh = 90 | Upper threshold for the Range of RTD, ms |
| hSpeedTresh = 1 Gbps | Threshold for transition between sending rate step sizes (such as 1 Mbps and 100 Mbps) |
| slowAdjCount = 0 | Measured Number of consecutive status reports indicating loss and/or delay variation above upperThresh |
| slowAdjThresh = 2 | Threshold for slowAdjCount used to infer congestion. Use values >1 to avoid misinterpreting transient loss |
| highSpeedDelta = 10 | The number of rows to move in a single adjustment when initially increasing offered load (to ramp-up quickly) |
| maxLoadRates = 2000 | Maximum table index (rows) |
Section 3.1 of [RFC8085] (BCP 145), which provides UDP usage guidelines, focuses primarily on congestion control. The Guidelines appear in mandatory (MUST) and recommendation (SHOULD) categories.¶
The mandatory requirements in Section 3 of [RFC8085] include the following:¶
Internet paths can have widely varying characteristics, ... Consequently, applications that may be used on the Internet MUST NOT make assumptions about specific path characteristics. They MUST instead use mechanisms that let them operate safely under very different path conditions. Typically, this requires conservatively probing the current conditions of the Internet path they communicate over to establish a transmission behavior that it can sustain and that is reasonably fair to other traffic sharing the path.¶
The purpose of the load rate adjustment algorithm in Section 8.1 is to probe the network and enable Maximum IP-Layer Capacity measurements with as few assumptions about the measured path as possible, and within the range application described in Section 2. The degree of probing conservatism is in tension with the need to minimize both the traffic dedicated to testing (especially with Gigabit rate measurements) and the duration of the test (which is one contributing factor to the overall algorithm fairness).¶
The text of Section 3 of [RFC8085] goes on to recommend alternatives to UDP to meet the mandatory requirements, but none are suitable for the scope and purpose of the metrics and methods in this memo. In fact, ad hoc TCP-based methods fail to achieve the measurement accuracy repeatedly proven in comparison measurements with the running code [LS-SG12-A] [LS-SG12-B] [Y.Sup60]. Also, the UDP aspect of these methods is present primarily to support modern Internet transmission where a transport protocol is required [copycat]; the metric is based on the IP-Layer, and UDP allows simple correlation to the IP-Layer.¶
Section 3.1.1 of [RFC8085] discusses protocol timer guidelines:¶
Latency samples MUST NOT be derived from ambiguous transactions. The canonical example is in a protocol that retransmits data, but subsequently cannot determine which copy is being acknowledged.¶
Both load packets and status feedback messages MUST contain sequence numbers, which helps with measurements based on those packets, and there are no retransmissions needed.¶
When a latency estimate is used to arm a timer that provides loss detection -- with or without retransmission -- expiry of the timer MUST be interpreted as an indication of congestion in the network, causing the sending rate to be adapted to a safe conservative rate...¶
The method described in this memo uses timers for sending rate backoff when status feedback messages are lost (Lost Status Backoff timeout), and for stopping a test when connectivity is lost for a longer interval (Feedback message or load packet timeouts).¶
There is no specific benefit foreseen by using Explicit Congestion Notification (ECN) in this memo.¶
Section 3.2 of [RFC8085] discusses message size guidelines:¶
To determine an appropriate UDP payload size, applications MUST subtract the size of the IP header (which includes any IPv4 optional headers or IPv6 extension headers) as well as the length of the UDP header (8 bytes) from the PMTU size.¶
The method uses a sending rate table with a maximum UDP payload size that anticipates significant header overhead and avoids fragmentation.¶
Section 3.3 of [RFC8085] provides reliability guidelines:¶
Applications that do require reliable message delivery MUST implement an appropriate mechanism themselves.¶
The IP-Layer Capacity Metric and Method do not require reliable delivery.¶
Applications that require ordered delivery MUST reestablish datagram ordering themselves.¶
The IP-Layer Capacity Metric and Method does not need to reestablish packet order; it is preferred to measure packet reordering if it occurs [RFC4737].¶
The load rate adjustment algorithm's goal is to determine the Maximum IP-Layer Capacity in the context of an infrequent, diagnostic, short-term measurement. This goal is a global exception to many [RFC8085] SHOULD-level requirements, of which many are intended for long-lived flows that must coexist with other traffic in more-or-less fair way. However, the algorithm (as specified in Section 8.1 and Appendix A above) reacts to indications of congestion in clearly defined ways.¶
A specific recommendation is provided as an example. Section 3.1.5 of [RFC8085] on implications of RTT and Loss Measurements on Congestion Control says:¶
A congestion control designed for UDP SHOULD respond as quickly as possible when it experiences congestion, and it SHOULD take into account both the loss rate and the response time when choosing a new rate.¶
The load rate adjustment algorithm responds to loss and RTT measurements with a clear and concise rate reduction when warranted, and the response makes use of direct measurements (more exact than can be inferred from TCP ACKs).¶
Section 3.1.5 of [RFC8085] goes on to specify the following:¶
The implemented congestion control scheme SHOULD result in bandwidth (capacity) use that is comparable to that of TCP within an order of magnitude, so that it does not starve other flows sharing a common bottleneck.¶
This is a requirement for coexistent streams, and not for diagnostic and infrequent measurements using short durations. The rate oscillations during short tests allow other packets to pass, and don't starve other flows.¶
Ironically, ad hoc TCP-based measurements of "Internet Speed" are also designed to work around this SHOULD-level requirement, by launching many flows (9, for example) to increase the outstanding data dedicated to testing.¶
The load rate adjustment algorithm cannot become a TCP-like congestion control, or it will have the same weaknesses of TCP when trying to make a Maximum IP-Layer Capacity measurement, and will not achieve the goal. The results of the referenced testing [LS-SG12-A] [LS-SG12-B] [Y.Sup60] supported this statement hundreds of times, with comparisons to multi-connection TCP-based measurements.¶
A brief review of some other SHOULD-level requirements follows (marked "Yes" or "N/A" (Not Applicable):¶
| Y? | Recommendation in RFC 8085 | Section |
|---|---|---|
| Yes | MUST tolerate a wide range of Internet path conditions | 3 |
| N/A | SHOULD use a full-featured transport (e.g., TCP) | |
| Yes | SHOULD control rate of transmission | 3.1 |
| N/A | SHOULD perform congestion control over all traffic | |
| For bulk transfers, | 3.1.2 | |
| N/A | SHOULD consider implementing TFRC | |
| N/A | else, SHOULD in other ways use bandwidth similar to TCP | |
| For non-bulk transfers, | 3.1.3 | |
| N/A | SHOULD measure RTT and transmit max. 1 datagram/RTT | 3.1.1 |
| N/A | else, SHOULD send at most 1 datagram every 3 seconds | |
| N/A | SHOULD back-off retransmission timers following loss | |
| Yes | SHOULD provide mechanisms to regulate the bursts of transmission | 3.1.6 |
| N/A | MAY implement ECN; a specific set of application mechanisms are REQUIRED if ECN is used | 3.1.7 |
| Yes | For DiffServ, SHOULD NOT rely on implementation of PHBs | 3.1.8 |
| Yes | For QoS-enabled paths, MAY choose not to use CC | 3.1.9 |
| Yes | SHOULD NOT rely solely on QoS for their capacity | 3.1.10 |
| Non-CC controlled flows SHOULD implement a transport circuit breaker | ||
| MAY implement a circuit breaker for other applications | ||
| For tunnels carrying IP traffic, | 3.1.11 | |
| N/A | SHOULD NOT perform congestion control | |
| N/A | MUST correctly process the IP ECN field | |
| For non-IP tunnels or rate not determined by traffic, | ||
| N/A | SHOULD perform CC or use circuit breaker | 3.1.11 |
| N/A | SHOULD restrict types of traffic transported by the tunnel | |
| Yes | SHOULD NOT send datagrams that exceed the PMTU, i.e., | 3.2 |
| Yes | SHOULD discover PMTU or send datagrams < minimum PMTU | |
| N/A | Specific application mechanisms are REQUIRED if PLPMTUD is used | |
| Yes | SHOULD handle datagram loss, duplication, reordering | 3.3 |
| N/A | SHOULD be robust to delivery delays up to 2 minutes | |
| Yes | SHOULD enable IPv4 UDP checksum | 3.4 |
| Yes | SHOULD enable IPv6 UDP checksum; specific application mechanisms are REQUIRED if a zero IPv6 UDP checksum is used | 3.4.1 |
| N/A | SHOULD provide protection from off-path attacks | 5.1 |
| else, MAY use UDP-Lite with suitable checksum coverage | 3.4.2 | |
| N/A | SHOULD NOT always send middlebox keep-alive messages | 3.5 |
| N/A | MAY use keep-alives when needed (min. interval 15 sec) | |
| Yes | Applications specified for use in limited use (or controlled environments) SHOULD identify equivalent mechanisms and describe their use case | 3.6 |
| N/A | Bulk-multicast apps SHOULD implement congestion control | 4.1.1 |
| N/A | Low-volume multicast apps SHOULD implement congestion control | 4.1.2 |
| N/A | Multicast apps SHOULD use a safe PMTU | 4.2 |
| Yes | SHOULD avoid using multiple ports | 5.1.2 |
| Yes | MUST check received IP source address | |
| N/A | SHOULD validate payload in ICMP messages | 5.2 |
| Yes | SHOULD use a randomized source port or equivalent technique, and, for client/server applications, SHOULD send responses from source address matching request 5.1 | 6 |
| N/A | SHOULD use standard IETF security protocols when needed | 6 |
Thanks to Joachim Fabini, Matt Mathis, J. Ignacio Alvarez-Hamelin, Wolfgang Balzer, Frank Brockners, Greg Mirsky, Martin Duke, Murray Kucherawy, and Benjamin Kaduk for their extensive comments on the memo and related topics. In a second round of reviews, we acknowledge Magnus Westerlund, Lars Eggert, and Zahed Sarkar.¶