RFC 8337
Model-Based Metrics for Bulk Transport Capacity
Pages: 55
Experimental
Experimental
Part 3 of 3 – Pages 34 to 55
8. IP Diagnostic Tests
The IP diagnostic tests below are organized according to the technique used to generate the test stream as described in Section 6. All of the results are evaluated in accordance with Section 7, possibly with additional test-specific criteria. We also introduce some combined tests that are more efficient when networks are expected to pass but conflate diagnostic signatures when they fail.8.1. Basic Data Rate and Packet Transfer Tests
We propose several versions of the basic data rate and packet transfer statistics test that differ in how the data rate is controlled. The data can be paced on a timer or window controlled (and self-clocked). The first two tests implicitly confirm that
sub_path has sufficient raw capacity to carry the target_data_rate. They are recommended for relatively infrequent testing, such as an installation or periodic auditing process. The third test, Background Packet Transfer Statistics, is a low-rate test designed for ongoing monitoring for changes in subpath quality.8.1.1. Delivery Statistics at Paced Full Data Rate
This test confirms that the observed run length is at least the target_run_length while relying on timer to send data at the target_rate using the procedure described in Section 6.1 with a burst size of 1 (single packets) or 2 (packet pairs). The test is considered to be inconclusive if the packet transmission cannot be accurately controlled for any reason. RFC 6673 [RFC6673] is appropriate for measuring packet transfer statistics at full data rate.8.1.2. Delivery Statistics at Full Data Windowed Rate
This test confirms that the observed run length is at least the target_run_length while sending at an average rate approximately equal to the target_data_rate, by controlling (or clamping) the window size of a conventional transport protocol to test_window. Since losses and ECN CE marks cause transport protocols to reduce their data rates, this test is expected to be less precise about controlling its data rate. It should not be considered inconclusive as long as at least some of the round trips reached the full target_data_rate without incurring losses or ECN CE marks. To pass this test, the network must deliver target_window_size packets in target_RTT time without any losses or ECN CE marks at least once per two target_window_size round trips, in addition to meeting the run length statistical test.8.1.3. Background Packet Transfer Statistics Tests
The Background Packet Transfer Statistics Test is a low-rate version of the target rate test above, designed for ongoing lightweight monitoring for changes in the observed subpath run length without disrupting users. It should be used in conjunction with one of the above full-rate tests because it does not confirm that the subpath can support raw data rate. RFC 6673 [RFC6673] is appropriate for measuring background packet transfer statistics.
8.2. Standing Queue Tests
These engineering tests confirm that the bottleneck is well behaved across the onset of packet loss, which typically follows after the onset of queuing. Well behaved generally means lossless for transient queues, but once the queue has been sustained for a sufficient period of time (or reaches a sufficient queue depth), there should be a small number of losses or ECN CE marks to signal to the transport protocol that it should reduce its window or data rate. Losses that are too early can prevent the transport from averaging at the target_data_rate. Losses that are too late indicate that the queue might not have an appropriate AQM [RFC7567] and, as a consequence, be subject to bufferbloat [wikiBloat]. Queues without AQM have the potential to inflict excess delays on all flows sharing the bottleneck. Excess losses (more than half of the window) at the onset of loss make loss recovery problematic for the transport protocol. Non-linear, erratic, or excessive RTT increases suggest poor interactions between the channel acquisition algorithms and the transport self-clock. All of the tests in this section use the same basic scanning algorithm, described here, but score the link or subpath on the basis of how well it avoids each of these problems. Some network technologies rely on virtual queues or other techniques to meter traffic without adding any queuing delay, in which case the data rate will vary with the window size all the way up to the onset of load-induced packet loss or ECN CE marks. For these technologies, the discussion of queuing in Section 6.3 does not apply, but it is still necessary to confirm that the onset of losses or ECN CE marks be at an appropriate point and progressive. If the network bottleneck does not introduce significant queuing delay, modify the procedure described in Section 6.3 to start the scan at a window equal to or slightly smaller than the test_window. Use the procedure in Section 6.3 to sweep the window across the onset of queuing and the onset of loss. The tests below all assume that the scan emulates standard additive increase and delayed ACK by incrementing the window by one packet for every 2*target_window_size packets delivered. A scan can typically be divided into three regions: below the onset of queuing, a standing queue, and at or beyond the onset of loss. Below the onset of queuing, the RTT is typically fairly constant, and the data rate varies in proportion to the window size. Once the data rate reaches the subpath IP rate, the data rate becomes fairly constant, and the RTT increases in proportion to the increase in window size. The precise transition across the start of queuing can be identified by the maximum network power, defined to be the ratio
data rate over the RTT. The network power can be computed at each window size, and the window with the maximum is taken as the start of the queuing region. If there is random background loss (e.g., bit errors), precise determination of the onset of queue-induced packet loss may require multiple scans. At window sizes large enough to cause loss in queues, all transport protocols are expected to experience periodic losses determined by the interaction between the congestion control and AQM algorithms. For standard congestion control algorithms, the periodic losses are likely to be relatively widely spaced, and the details are typically dominated by the behavior of the transport protocol itself. For the case of stiffened transport protocols (with non-standard, aggressive congestion control algorithms), the details of periodic losses will be dominated by how the window increase function responds to loss.8.2.1. Congestion Avoidance
A subpath passes the congestion avoidance standing queue test if more than target_run_length packets are delivered between the onset of queuing (as determined by the window with the maximum network power as described above) and the first loss or ECN CE mark. If this test is implemented using a standard congestion control algorithm with a clamp, it can be performed in situ in the production internet as a capacity test. For an example of such a test, see [Pathdiag]. For technologies that do not have conventional queues, use the test_window in place of the onset of queuing. That is, a subpath passes the congestion avoidance standing queue test if more than target_run_length packets are delivered between the start of the scan at test_window and the first loss or ECN CE mark.8.2.2. Bufferbloat
This test confirms that there is some mechanism to limit buffer occupancy (e.g., that prevents bufferbloat). Note that this is not strictly a requirement for single-stream bulk transport capacity; however, if there is no mechanism to limit buffer queue occupancy, then a single stream with sufficient data to deliver is likely to cause the problems described in [RFC7567] and [wikiBloat]. This may cause only minor symptoms for the dominant flow but has the potential to make the subpath unusable for other flows and applications. The test will pass if the onset of loss occurs before a standing queue has introduced delay greater than twice the target_RTT or another well-defined and specified limit. Note that there is not yet a model for how much standing queue is acceptable. The factor of two
chosen here reflects a rule of thumb. In conjunction with the previous test, this test implies that the first loss should occur at a queuing delay that is between one and two times the target_RTT. Specified RTT limits that are larger than twice the target_RTT must be fully justified in the FSTIDS.8.2.3. Non-excessive Loss
This test confirms that the onset of loss is not excessive. The test will pass if losses are equal to or less than the increase in the cross traffic plus the test stream window increase since the previous RTT. This could be restated as non-decreasing total throughput of the subpath at the onset of loss. (Note that when there is a transient drop in subpath throughput and there is not already a standing queue, a subpath that passes other queue tests in this document will have sufficient queue space to hold one full RTT worth of data). Note that token bucket policers will not pass this test, which is as intended. TCP often stumbles badly if more than a small fraction of the packets are dropped in one RTT. Many TCP implementations will require a timeout and slowstart to recover their self-clock. Even if they can recover from the massive losses, the sudden change in available capacity at the bottleneck wastes serving and front-path capacity until TCP can adapt to the new rate [Policing].8.2.4. Duplex Self-Interference
This engineering test confirms a bound on the interactions between the forward data path and the ACK return path when they share a half- duplex link. Some historical half-duplex technologies had the property that each direction held the channel until it completely drained its queue. When a self-clocked transport protocol, such as TCP, has data and ACKs passing in opposite directions through such a link, the behavior often reverts to stop-and-wait. Each additional packet added to the window raises the observed RTT by two packet times, once as the additional packet passes through the data path and once for the additional delay incurred by the ACK waiting on the return path. The Duplex Self-Interference Test fails if the RTT rises by more than a fixed bound above the expected queuing time computed from the excess window divided by the subpath IP capacity. This bound must be smaller than target_RTT/2 to avoid reverting to stop-and-wait behavior (e.g., data packets and ACKs both have to be released at least twice per RTT).
8.3. Slowstart Tests
These tests mimic slowstart: data is sent at twice the effective bottleneck rate to exercise the queue at the dominant bottleneck.8.3.1. Full Window Slowstart Test
This capacity test confirms that slowstart is not likely to exit prematurely. To perform this test, send slowstart bursts that are target_window_size total packets and accumulate packet transfer statistics as described in Section 7.2 to score the outcome. The test will pass if it is statistically significant that the observed number of good packets delivered between losses or ECN CE marks is larger than the target_run_length. The test will fail if it is statistically significant that the observed interval between losses or ECN CE marks is smaller than the target_run_length. The test is deemed inconclusive if the elapsed time to send the data burst is not less than half of the time to receive the ACKs. (That is, it is acceptable to send data too fast, but sending it slower than twice the actual bottleneck rate as indicated by the ACKs is deemed inconclusive). The headway for the slowstart bursts should be the target_RTT. Note that these are the same parameters that are used for the Sustained Full-Rate Bursts Test, except the burst rate is at slowstart rate rather than sender interface rate.8.3.2. Slowstart AQM Test
To perform this test, do a continuous slowstart (send data continuously at twice the implied IP bottleneck capacity) until the first loss; stop and allow the network to drain and repeat; gather statistics on how many packets were delivered before the loss, the pattern of losses, maximum observed RTT, and window size; and justify the results. There is not currently sufficient theory to justify requiring any particular result; however, design decisions that affect the outcome of this tests also affect how the network balances between long and short flows (the "mice vs. elephants" problem). The queue sojourn time for the first packet delivered after the first loss should be at least one half of the target_RTT. This engineering test should be performed on a quiescent network or testbed, since cross traffic has the potential to change the results in ill-defined ways.
8.4. Sender Rate Burst Tests
These tests determine how well the network can deliver bursts sent at the sender's interface rate. Note that this test most heavily exercises the front path and is likely to include infrastructure that may be out of scope for an access ISP, even though the bursts might be caused by ACK compression, thinning, or channel arbitration in the access ISP. See Appendix B. Also, there are a several details about sender interface rate bursts that are not fully defined here. These details, such as the assumed sender interface rate, should be explicitly stated in an FSTIDS. Current standards permit TCP to send full window bursts following an application pause. (Congestion Window Validation [RFC2861] and updates to support Rate-Limited Traffic [RFC7661] are not required). Since full window bursts are consistent with standard behavior, it is desirable that the network be able to deliver such bursts; otherwise, application pauses will cause unwarranted losses. Note that the AIMD sawtooth requires a peak window that is twice target_window_size, so the worst-case burst may be 2*target_window_size. It is also understood in the application and serving community that interface rate bursts have a cost to the network that has to be balanced against other costs in the servers themselves. For example, TCP Segmentation Offload (TSO) reduces server CPU in exchange for larger network bursts, which increase the stress on network buffer memory. Some newer TCP implementations can pace traffic at scale [TSO_pacing] [TSO_fq_pacing]. It remains to be determined if and how quickly these changes will be deployed. There is not yet theory to unify these costs or to provide a framework for trying to optimize global efficiency. We do not yet have a model for how many server rate bursts should be tolerated by the network. Some bursts must be tolerated by the network, but it is probably unreasonable to expect the network to be able to efficiently deliver all data as a series of bursts. For this reason, this is the only test for which we encourage derating. A TIDS could include a table containing pairs of derating parameters: burst sizes and how much each burst size is permitted to reduce the run length, relative to the target_run_length.
8.5. Combined and Implicit Tests
Combined tests efficiently confirm multiple network properties in a single test, possibly as a side effect of normal content delivery. They require less measurement traffic than other testing strategies at the cost of conflating diagnostic signatures when they fail. These are by far the most efficient for monitoring networks that are nominally expected to pass all tests.8.5.1. Sustained Full-Rate Bursts Test
The Sustained Full-Rate Bursts Test implements a combined worst-case version of all of the capacity tests above. To perform this test, send target_window_size bursts of packets at server interface rate with target_RTT burst headway (burst start to next burst start), and verify that the observed packet transfer statistics meets the target_run_length. Key observations: o The subpath under test is expected to go idle for some fraction of the time, determined by the difference between the time to drain the queue at the subpath_IP_capacity and the target_RTT. If the queue does not drain completely, it may be an indication that the subpath has insufficient IP capacity or that there is some other problem with the test (e.g., it is inconclusive). o The burst sensitivity can be derated by sending smaller bursts more frequently (e.g., by sending target_window_size*derate packet bursts every target_RTT*derate, where "derate" is less than one). o When not derated, this test is the most strenuous capacity test. o A subpath that passes this test is likely to be able to sustain higher rates (close to subpath_IP_capacity) for paths with RTTs significantly smaller than the target_RTT. o This test can be implemented with instrumented TCP [RFC4898], using a specialized measurement application at one end (e.g., [MBMSource]) and a minimal service at the other end (e.g., [RFC863] and [RFC864]). o This test is efficient to implement, since it does not require per-packet timers, and can make use of TSO in modern network interfaces.
o If a subpath is known to pass the standing queue engineering tests
(particularly that it has a progressive onset of loss at an
appropriate queue depth), then the Sustained Full-Rate Bursts Test
is sufficient to assure that the subpath under test will not
impair Bulk Transport Capacity at the target performance under all
conditions. See Section 8.2 for a discussion of the standing
queue tests.
Note that this test is clearly independent of the subpath RTT or
other details of the measurement infrastructure, as long as the
measurement infrastructure can accurately and reliably deliver the
required bursts to the subpath under test.
8.5.2. Passive Measurements
Any non-throughput-maximizing application, such as fixed-rate
streaming media, can be used to implement passive or hybrid (defined
in [RFC7799]) versions of Model-Based Metrics with some additional
instrumentation and possibly a traffic shaper or other controls in
the servers. The essential requirement is that the data transmission
be constrained such that even with arbitrary application pauses and
bursts, the data rate and burst sizes stay within the envelope
defined by the individual tests described above.
If the application's serving data rate can be constrained to be less
than or equal to the target_data_rate and the serving_RTT (the RTT
between the sender and client) is less than the target_RTT, this
constraint is most easily implemented by clamping the transport
window size to serving_window_clamp (which is set to the test_window
and computed for the actual serving path).
Under the above constraints, the serving_window_clamp will limit both
the serving data rate and burst sizes to be no larger than the
parameters specified by the procedures in Section 8.1.2, 8.4, or
8.5.1. Since the serving RTT is smaller than the target_RTT, the
worst-case bursts that might be generated under these conditions will
be smaller than called for by Section 8.4, and the sender rate burst
sizes are implicitly derated by the serving_window_clamp divided by
the target_window_size at the very least. (Depending on the
application behavior, the data might be significantly smoother than
specified by any of the burst tests.)
In an alternative implementation, the data rate and bursts might be
explicitly controlled by a programmable traffic shaper or by pacing
at the sender. This would provide better control over transmissions
but is more complicated to implement, although the required
technology is available [TSO_pacing] [TSO_fq_pacing].
Note that these techniques can be applied to any content delivery that can be operated at a constrained data rate to inhibit TCP equilibrium behavior. Furthermore, note that Dynamic Adaptive Streaming over HTTP (DASH) is generally in conflict with passive Model-Based Metrics measurement, because it is a rate-maximizing protocol. It can still meet the requirement here if the rate can be capped, for example, by knowing a priori the maximum rate needed to deliver a particular piece of content.9. Example
In this section, we illustrate a TIDS designed to confirm that an access ISP can reliably deliver HD video from multiple content providers to all of its customers. With modern codecs, minimal HD video (720p) generally fits in 2.5 Mb/s. Due to the ISP's geographical size, network topology, and modem characteristics, the ISP determines that most content is within a 50 ms RTT of its users. (This example RTT is sufficient to cover the propagation delay to continental Europe or to either coast of the United States with low- delay modems; it is sufficient to cover somewhat smaller geographical regions if the modems require additional delay to implement advanced compression and error recovery.) +----------------------+-------+---------+ | End-to-End Parameter | value | units | +----------------------+-------+---------+ | target_rate | 2.5 | Mb/s | | target_RTT | 50 | ms | | target_MTU | 1500 | bytes | | header_overhead | 64 | bytes | | | | | | target_window_size | 11 | packets | | target_run_length | 363 | packets | +----------------------+-------+---------+ Table 1: 2.5 Mb/s over a 50 ms Path Table 1 shows the default TCP model with no derating and, as such, is quite conservative. The simplest TIDS would be to use the Sustained Full-Rate Bursts Test, described in Section 8.5.1. Such a test would send 11 packet bursts every 50 ms and confirm that there was no more than 1 packet loss per 33 bursts (363 total packets in 1.650 seconds).
Since this number represents the entire end-to-end loss budget, independent subpath tests could be implemented by apportioning the packet loss ratio across subpaths. For example, 50% of the losses might be allocated to the access or last mile link to the user, 40% to the network interconnections with other ISPs, and 1% to each internal hop (assuming no more than 10 internal hops). Then, all of the subpaths can be tested independently, and the spatial composition of passing subpaths would be expected to be within the end-to-end loss budget.9.1. Observations about Applicability
Guidance on deploying and using MBM belong in a future document. However, the example above illustrates some of the issues that may need to be considered. Note that another ISP, with different geographical coverage, topology, or modem technology may need to assume a different target_RTT and, as a consequence, a different target_window_size and target_run_length, even for the same target_data rate. One of the implications of this is that infrastructure shared by multiple ISPs, such as Internet Exchange Points (IXPs) and other interconnects may need to be evaluated on the basis of the most stringent target_window_size and target_run_length of any participating ISP. One way to do this might be to choose target parameters for evaluating such shared infrastructure on the basis of a hypothetical reference path that does not necessarily match any actual paths. Testing interconnects has generally been problematic: conventional performance tests run between measurement points adjacent to either side of the interconnect are not generally useful. Unconstrained TCP tests, such as iPerf [iPerf], are usually overly aggressive due to the small RTT (often less than 1 ms). With a short RTT, these tools are likely to report inflated data rates because on a short RTT, these tools can tolerate very high packet loss ratios and can push other cross traffic off of the network. As a consequence, these measurements are useless for predicting actual user performance over longer paths and may themselves be quite disruptive. Model-Based Metrics solves this problem. The interconnect can be evaluated with the same TIDS as other subpaths. Continuing our example, if the interconnect is apportioned 40% of the losses, 11 packet bursts sent every 50 ms should have fewer than one loss per 82 bursts (902 packets).
10. Validation
Since some aspects of the models are likely to be too conservative, Section 5.2 permits alternate protocol models, and Section 5.3 permits test parameter derating. If either of these techniques is used, we require demonstrations that such a TIDS can robustly detect subpaths that will prevent authentic applications using state-of-the- art protocol implementations from meeting the specified Target Transport Performance. This correctness criteria is potentially difficult to prove, because it implicitly requires validating a TIDS against all possible paths and subpaths. The procedures described here are still experimental. We suggest two approaches, both of which should be applied. First, publish a fully open description of the TIDS, including what assumptions were used and how it was derived, such that the research community can evaluate the design decisions, test them, and comment on their applicability. Second, demonstrate that applications do meet the Target Transport Performance when running over a network testbed that has the tightest possible constraints that still allow the tests in the TIDS to pass. This procedure resembles an epsilon-delta proof in calculus. Construct a test network such that all of the individual tests of the TIDS pass by only small (infinitesimal) margins, and demonstrate that a variety of authentic applications running over real TCP implementations (or other protocols as appropriate) meets the Target Transport Performance over such a network. The workloads should include multiple types of streaming media and transaction-oriented short flows (e.g., synthetic web traffic). For example, for the HD streaming video TIDS described in Section 9, the IP capacity should be exactly the header_overhead above 2.5 Mb/s, the per packet random background loss ratio should be 1/363 (for a run length of 363 packets), the bottleneck queue should be 11 packets, and the front path should have just enough buffering to withstand 11 packet interface rate bursts. We want every one of the TIDS tests to fail if we slightly increase the relevant test parameter, so, for example, sending a 12-packet burst should cause excess (possibly deterministic) packet drops at the dominant queue at the bottleneck. This network has the tightest possible constraints that can be expected to pass the TIDS, yet it should be possible for a real application using a stock TCP implementation in the vendor's default configuration to attain 2.5 Mb/s over a 50 ms path. The most difficult part of setting up such a testbed is arranging for it to have the tightest possible constraints that still allow it to pass the individual tests. Two approaches are suggested:
o constraining (configuring) the network devices not to use all
available resources (e.g., by limiting available buffer space or
data rate)
o pre-loading subpaths with cross traffic
Note that it is important that a single tightly constrained
environment just barely passes all tests; otherwise, there is a
chance that TCP can exploit extra latitude in some parameters (such
as data rate) to partially compensate for constraints in other
parameters (e.g., queue space). This effect is potentially
bidirectional: extra latitude in the queue space tests has the
potential to enable TCP to compensate for insufficient data-rate
headroom.
To the extent that a TIDS is used to inform public dialog, it should
be fully documented publicly, including the details of the tests,
what assumptions were used, and how it was derived. All of the
details of the validation experiment should also be published with
sufficient detail for the experiments to be replicated by other
researchers. All components should be either open source or fully
described proprietary implementations that are available to the
research community.
11. Security Considerations
Measurement is often used to inform business and policy decisions
and, as a consequence, is potentially subject to manipulation.
Model-Based Metrics are expected to be a huge step forward because
equivalent measurements can be performed from multiple vantage
points, such that performance claims can be independently validated
by multiple parties.
Much of the acrimony in the Net Neutrality debate is due to the
historical lack of any effective vantage-independent tools to
characterize network performance. Traditional methods for measuring
Bulk Transport Capacity are sensitive to RTT and as a consequence
often yield very different results when run local to an ISP or
interconnect and when run over a customer's complete path. Neither
the ISP nor customer can repeat the other's measurements, leading to
high levels of distrust and acrimony. Model-Based Metrics are
expected to greatly improve this situation.
Note that in situ measurements sometimes require sending synthetic
measurement traffic between arbitrary locations in the network and,
as such, are potentially attractive platforms for launching DDoS
attacks. All active measurement tools and protocols must be designed to minimize the opportunities for these misuses. See the discussion in Section 7 of [RFC7594]. Some of the tests described in this document are not intended for frequent network monitoring since they have the potential to cause high network loads and might adversely affect other traffic. This document only describes a framework for designing a Fully Specified Targeted IP Diagnostic Suite. Each FSTIDS must include its own security section.12. IANA Considerations
This document has no IANA actions.13. Informative References
[RFC863] Postel, J., "Discard Protocol", STD 21, RFC 863, DOI 10.17487/RFC0863, May 1983, <https://www.rfc-editor.org/info/rfc863>. [RFC864] Postel, J., "Character Generator Protocol", STD 22, RFC 864, DOI 10.17487/RFC0864, May 1983, <https://www.rfc-editor.org/info/rfc864>. [RFC2330] Paxson, V., Almes, G., Mahdavi, J., and M. Mathis, "Framework for IP Performance Metrics", RFC 2330, DOI 10.17487/RFC2330, May 1998, <https://www.rfc-editor.org/info/rfc2330>. [RFC2861] Handley, M., Padhye, J., and S. Floyd, "TCP Congestion Window Validation", RFC 2861, DOI 10.17487/RFC2861, June 2000, <https://www.rfc-editor.org/info/rfc2861>. [RFC3148] Mathis, M. and M. Allman, "A Framework for Defining Empirical Bulk Transfer Capacity Metrics", RFC 3148, DOI 10.17487/RFC3148, July 2001, <https://www.rfc-editor.org/info/rfc3148>. [RFC3168] Ramakrishnan, K., Floyd, S., and D. Black, "The Addition of Explicit Congestion Notification (ECN) to IP", RFC 3168, DOI 10.17487/RFC3168, September 2001, <https://www.rfc-editor.org/info/rfc3168>. [RFC3465] Allman, M., "TCP Congestion Control with Appropriate Byte Counting (ABC)", RFC 3465, DOI 10.17487/RFC3465, February 2003, <https://www.rfc-editor.org/info/rfc3465>.
[RFC4737] Morton, A., Ciavattone, L., Ramachandran, G., Shalunov, S., and J. Perser, "Packet Reordering Metrics", RFC 4737, DOI 10.17487/RFC4737, November 2006, <https://www.rfc-editor.org/info/rfc4737>. [RFC4898] Mathis, M., Heffner, J., and R. Raghunarayan, "TCP Extended Statistics MIB", RFC 4898, DOI 10.17487/RFC4898, May 2007, <https://www.rfc-editor.org/info/rfc4898>. [RFC5136] Chimento, P. and J. Ishac, "Defining Network Capacity", RFC 5136, DOI 10.17487/RFC5136, February 2008, <https://www.rfc-editor.org/info/rfc5136>. [RFC5681] Allman, M., Paxson, V., and E. Blanton, "TCP Congestion Control", RFC 5681, DOI 10.17487/RFC5681, September 2009, <https://www.rfc-editor.org/info/rfc5681>. [RFC5827] Allman, M., Avrachenkov, K., Ayesta, U., Blanton, J., and P. Hurtig, "Early Retransmit for TCP and Stream Control Transmission Protocol (SCTP)", RFC 5827, DOI 10.17487/RFC5827, May 2010, <https://www.rfc-editor.org/info/rfc5827>. [RFC5835] Morton, A., Ed. and S. Van den Berghe, Ed., "Framework for Metric Composition", RFC 5835, DOI 10.17487/RFC5835, April 2010, <https://www.rfc-editor.org/info/rfc5835>. [RFC6049] Morton, A. and E. Stephan, "Spatial Composition of Metrics", RFC 6049, DOI 10.17487/RFC6049, January 2011, <https://www.rfc-editor.org/info/rfc6049>. [RFC6576] Geib, R., Ed., Morton, A., Fardid, R., and A. Steinmitz, "IP Performance Metrics (IPPM) Standard Advancement Testing", BCP 176, RFC 6576, DOI 10.17487/RFC6576, March 2012, <https://www.rfc-editor.org/info/rfc6576>. [RFC6673] Morton, A., "Round-Trip Packet Loss Metrics", RFC 6673, DOI 10.17487/RFC6673, August 2012, <https://www.rfc-editor.org/info/rfc6673>. [RFC6928] Chu, J., Dukkipati, N., Cheng, Y., and M. Mathis, "Increasing TCP's Initial Window", RFC 6928, DOI 10.17487/RFC6928, April 2013, <https://www.rfc-editor.org/info/rfc6928>.
[RFC7312] Fabini, J. and A. Morton, "Advanced Stream and Sampling Framework for IP Performance Metrics (IPPM)", RFC 7312, DOI 10.17487/RFC7312, August 2014, <https://www.rfc-editor.org/info/rfc7312>. [RFC7398] Bagnulo, M., Burbridge, T., Crawford, S., Eardley, P., and A. Morton, "A Reference Path and Measurement Points for Large-Scale Measurement of Broadband Performance", RFC 7398, DOI 10.17487/RFC7398, February 2015, <https://www.rfc-editor.org/info/rfc7398>. [RFC7567] Baker, F., Ed. and G. Fairhurst, Ed., "IETF Recommendations Regarding Active Queue Management", BCP 197, RFC 7567, DOI 10.17487/RFC7567, July 2015, <https://www.rfc-editor.org/info/rfc7567>. [RFC7594] Eardley, P., Morton, A., Bagnulo, M., Burbridge, T., Aitken, P., and A. Akhter, "A Framework for Large-Scale Measurement of Broadband Performance (LMAP)", RFC 7594, DOI 10.17487/RFC7594, September 2015, <https://www.rfc-editor.org/info/rfc7594>. [RFC7661] Fairhurst, G., Sathiaseelan, A., and R. Secchi, "Updating TCP to Support Rate-Limited Traffic", RFC 7661, DOI 10.17487/RFC7661, October 2015, <https://www.rfc-editor.org/info/rfc7661>. [RFC7680] Almes, G., Kalidindi, S., Zekauskas, M., and A. Morton, Ed., "A One-Way Loss Metric for IP Performance Metrics (IPPM)", STD 82, RFC 7680, DOI 10.17487/RFC7680, January 2016, <https://www.rfc-editor.org/info/rfc7680>. [RFC7799] Morton, A., "Active and Passive Metrics and Methods (with Hybrid Types In-Between)", RFC 7799, DOI 10.17487/RFC7799, May 2016, <https://www.rfc-editor.org/info/rfc7799>. [AFD] Pan, R., Breslau, L., Prabhakar, B., and S. Shenker, "Approximate fairness through differential dropping", ACM SIGCOMM Computer Communication Review, Volume 33, Issue 2, DOI 10.1145/956981.956985, April 2003. [CCscaling] Paganini, F., Doyle, J., and S. Low, "Scalable laws for stable network congestion control", Proceedings of IEEE Conference on Decision and Control,, DOI 10.1109/CDC.2001.980095, December 2001.
[CVST] Krueger, T. and M. Braun, "R package: Fast Cross- Validation via Sequential Testing", version 0.1, 11 2012. [iPerf] Wikipedia, "iPerf", November 2017, <https://en.wikipedia.org/w/ index.php?title=Iperf&oldid=810583885>. [MBMSource] "mbm", July 2016, <https://github.com/m-lab/MBM>. [Montgomery90] Montgomery, D., "Introduction to Statistical Quality Control", 2nd Edition, ISBN 0-471-51988-X, 1990. [mpingSource] "mping", July 2016, <https://github.com/m-lab/mping>. [MSMO97] Mathis, M., Semke, J., Mahdavi, J., and T. Ott, "The Macroscopic Behavior of the TCP Congestion Avoidance Algorithm", Computer Communications Review, Volume 27, Issue 3, DOI 10.1145/263932.264023, July 1997. [Pathdiag] Mathis, M., Heffner, J., O'Neil, P., and P. Siemsen, "Pathdiag: Automated TCP Diagnosis", Passive and Active Network Measurement, Lecture Notes in Computer Science, Volume 4979, DOI 10.1007/978-3-540-79232-1_16, 2008. [Policing] Flach, T., Papageorge, P., Terzis, A., Pedrosa, L., Cheng, Y., Karim, T., Katz-Bassett, E., and R. Govindan, "An Internet-Wide Analysis of Traffic Policing", Proceedings of ACM SIGCOMM, DOI 10.1145/2934872.2934873, August 2016. [RACK] Cheng, Y., Cardwell, N., Dukkipati, N., and P. Jha, "RACK: a time-based fast loss detection algorithm for TCP", Work in Progress, draft-ietf-tcpm-rack-03, March 2018. [Rtool] R Development Core Team, "R: A language and environment for statistical computing", R Foundation for Statistical Computing, Vienna, Austria, ISBN 3-900051-07-0, 2011, <http://www.R-project.org/>. [TSO_fq_pacing] Dumazet, E. and Y. Chen, "TSO, fair queuing, pacing: three's a charm", Proceedings of IETF 88, TCPM WG, November 2013, <https://www.ietf.org/proceedings/88/slides/ slides-88-tcpm-9.pdf>.
[TSO_pacing] Corbet, J., "TSO sizing and the FQ scheduler", August 2013, <https://lwn.net/Articles/564978/>. [Wald45] Wald, A., "Sequential Tests of Statistical Hypotheses", The Annals of Mathematical Statistics, Volume 16, Number 2, pp. 117-186, June 1945, <http://www.jstor.org/stable/2235829>. [wikiBloat] Wikipedia, "Bufferbloat", January 2018, <https://en.wikipedia.org/w/ index.php?title=Bufferbloat&oldid=819293377>. [WPING] Mathis, M., "Windowed Ping: An IP Level Performance Diagnostic", Computer Networks and ISDN Systems, Volume 27, Issue 3, DOI 10.1016/0169-7552(94)90119-8, June 1994.
Appendix A. Model Derivations
The reference target_run_length described in Section 5.2 is based on very conservative assumptions: that all excess data in flight (i.e., the window size) above the target_window_size contributes to a standing queue that raises the RTT and that classic Reno congestion control with delayed ACKs is in effect. In this section we provide two alternative calculations using different assumptions. It may seem out of place to allow such latitude in a measurement method, but this section provides offsetting requirements. The estimates provided by these models make the most sense if network performance is viewed logarithmically. In the operational Internet, data rates span more than eight orders of magnitude, RTT spans more than three orders of magnitude, and packet loss ratio spans at least eight orders of magnitude if not more. When viewed logarithmically (as in decibels), these correspond to 80 dB of dynamic range. On an 80 dB scale, a 3 dB error is less than 4% of the scale, even though it represents a factor of 2 in untransformed parameter. This document gives a lot of latitude for calculating target_run_length; however, people designing a TIDS should consider the effect of their choices on the ongoing tussle about the relevance of "TCP friendliness" as an appropriate model for Internet capacity allocation. Choosing a target_run_length that is substantially smaller than the reference target_run_length specified in Section 5.2 strengthens the argument that it may be appropriate to abandon "TCP friendliness" as the Internet fairness model. This gives developers incentive and permission to develop even more aggressive applications and protocols, for example, by increasing the number of connections that they open concurrently.A.1. Queueless Reno
In Section 5.2, models were derived based on the assumption that the subpath IP rate matches the target rate plus overhead, such that the excess window needed for the AIMD sawtooth causes a fluctuating queue at the bottleneck. An alternate situation would be a bottleneck where there is no significant queue and losses are caused by some mechanism that does not involve extra delay, for example, by the use of a virtual queue as done in Approximate Fair Dropping [AFD]. A flow controlled by such a bottleneck would have a constant RTT and a data rate that fluctuates in a sawtooth due to AIMD congestion control. Assume the
losses are being controlled to make the average data rate meet some goal that is equal to or greater than the target_rate. The necessary run length to meet the target_rate can be computed as follows: For some value of Wmin, the window will sweep from Wmin packets to 2*Wmin packets in 2*Wmin RTT (due to delayed ACK). Unlike the queuing case where Wmin = target_window_size, we want the average of Wmin and 2*Wmin to be the target_window_size, so the average data rate is the target rate. Thus, we want Wmin = (2/3)*target_window_size. Between losses, each sawtooth delivers (1/2)(Wmin+2*Wmin)(2Wmin) packets in 2*Wmin RTTs. Substituting these together, we get: target_run_length = (4/3)(target_window_size^2) Note that this is 44% of the reference_run_length computed earlier. This makes sense because under the assumptions in Section 5.2, the AMID sawtooth caused a queue at the bottleneck, which raised the effective RTT by 50%.Appendix B. The Effects of ACK Scheduling
For many network technologies, simple queuing models don't apply: the network schedules, thins, or otherwise alters the timing of ACKs and data, generally to raise the efficiency of the channel allocation algorithms when confronted with relatively widely spaced small ACKs. These efficiency strategies are ubiquitous for half-duplex, wireless, and broadcast media. Altering the ACK stream by holding or thinning ACKs typically has two consequences: it raises the implied bottleneck IP capacity, making the fine-grained slowstart bursts either faster or larger, and it raises the effective RTT by the average time that the ACKs and data are delayed. The first effect can be partially mitigated by re-clocking ACKs once they are beyond the bottleneck on the return path to the sender; however, this further raises the effective RTT. The most extreme example of this sort of behavior would be a half- duplex channel that is not released as long as the endpoint currently holding the channel has more traffic (data or ACKs) to send. Such environments cause self-clocked protocols under full load to revert to extremely inefficient stop-and-wait behavior. The channel constrains the protocol to send an entire window of data as a single
contiguous burst on the forward path, followed by the entire window of ACKs on the return path. (A channel with this behavior would fail the Duplex Self-Interference Test described in Section 8.2.4). If a particular return path contains a subpath or device that alters the timing of the ACK stream, then the entire front path from the sender up to the bottleneck must be tested at the burst parameters implied by the ACK scheduling algorithm. The most important parameter is the implied bottleneck IP capacity, which is the average rate at which the ACKs advance snd.una. Note that thinning the ACK stream (relying on the cumulative nature of seg.ack to permit discarding some ACKs) causes most TCP implementations to send interface rate bursts to offset the longer times between ACKs in order to maintain the average data rate. Note that due to ubiquitous self-clocking in Internet protocols, ill-conceived channel allocation mechanisms are likely to increases the queuing stress on the front path because they cause larger full sender rate data bursts. Holding data or ACKs for channel allocation or other reasons (such as forward error correction) always raises the effective RTT relative to the minimum delay for the path. Therefore, it may be necessary to replace target_RTT in the calculation in Section 5.2 by an effective_RTT, which includes the target_RTT plus a term to account for the extra delays introduced by these mechanisms.
Acknowledgments
Ganga Maguluri suggested the statistical test for measuring loss probability in the target run length. Alex Gilgur and Merry Mou helped with the statistics. Meredith Whittaker improved the clarity of the communications. Ruediger Geib provided feedback that greatly improved the document. This work was inspired by Measurement Lab: open tools running on an open platform, using open tools to collect open data. See <http://www.measurementlab.net/>.Authors' Addresses
Matt Mathis Google, Inc 1600 Amphitheatre Parkway Mountain View, CA 94043 United States of America Email: mattmathis@google.com Al Morton AT&T Labs 200 Laurel Avenue South Middletown, NJ 07748 United States of America Phone: +1 732 420 1571 Email: acmorton@att.com