idnits 2.17.1
draft-ietf-tcpm-rfc8312bis-05.txt:
-(505): Line appears to be too long, but this could be caused by non-ascii characters in UTF-8 encoding
-(507): Line appears to be too long, but this could be caused by non-ascii characters in UTF-8 encoding
-(611): Line appears to be too long, but this could be caused by non-ascii characters in UTF-8 encoding
-(695): Line appears to be too long, but this could be caused by non-ascii characters in UTF-8 encoding
Checking boilerplate required by RFC 5378 and the IETF Trust (see
https://trustee.ietf.org/license-info):
----------------------------------------------------------------------------
No issues found here.
Checking nits according to https://www.ietf.org/id-info/1id-guidelines.txt:
----------------------------------------------------------------------------
== There are 25 instances of lines with non-ascii characters in the
document.
Checking nits according to https://www.ietf.org/id-info/checklist :
----------------------------------------------------------------------------
** There are 3 instances of too long lines in the document, the longest one
being 12 characters in excess of 72.
-- The draft header indicates that this document obsoletes RFC8312, but the
abstract doesn't seem to directly say this. It does mention RFC8312
though, so this could be OK.
-- The draft header indicates that this document updates RFC5681, but the
abstract doesn't seem to directly say this. It does mention RFC5681
though, so this could be OK.
Miscellaneous warnings:
----------------------------------------------------------------------------
== The copyright year in the IETF Trust and authors Copyright Line does not
match the current year
== Line 552 has weird spacing: '... est est...'
== Line 728 has weird spacing: '... max max...'
(Using the creation date from RFC5681, updated by this document, for
RFC5378 checks: 2006-01-26)
-- The document seems to lack a disclaimer for pre-RFC5378 work, but may
have content which was first submitted before 10 November 2008. If you
have contacted all the original authors and they are all willing to grant
the BCP78 rights to the IETF Trust, then this is fine, and you can ignore
this comment. If not, you may need to add the pre-RFC5378 disclaimer.
(See the Legal Provisions document at
https://trustee.ietf.org/license-info for more information.)
-- The document date (25 October 2021) is 908 days in the past. Is this
intentional?
Checking references for intended status: Proposed Standard
----------------------------------------------------------------------------
(See RFCs 3967 and 4897 for information about using normative references
to lower-maturity documents in RFCs)
== Outdated reference: A later version (-14) exists of
draft-ietf-tcpm-hystartplusplus-03
-- Obsolete informational reference (is this intentional?): RFC 4960
(Obsoleted by RFC 9260)
-- Obsolete informational reference (is this intentional?): RFC 8312
(Obsoleted by RFC 9438)
Summary: 1 error (**), 0 flaws (~~), 5 warnings (==), 6 comments (--).
Run idnits with the --verbose option for more detailed information about
the items above.
--------------------------------------------------------------------------------
2 TCPM L. Xu
3 Internet-Draft UNL
4 Obsoletes: 8312 (if approved) S. Ha
5 Updates: 5681 (if approved) Colorado
6 Intended status: Standards Track I. Rhee
7 Expires: 28 April 2022 Bowery
8 V. Goel
9 Apple Inc.
10 L. Eggert, Ed.
11 NetApp
12 25 October 2021
14 CUBIC for Fast and Long-Distance Networks
15 draft-ietf-tcpm-rfc8312bis-05
17 Abstract
19 CUBIC is a standard TCP congestion control algorithm that uses a
20 cubic function instead of the linear window increase function on the
21 sender side to improve scalability and stability over fast and long-
22 distance networks. CUBIC has been adopted as the default TCP
23 congestion control algorithm by the Linux, Windows, and Apple stacks.
25 This document updates the specification of CUBIC to include
26 algorithmic improvements based on these implementations and recent
27 academic work. Based on the extensive deployment experience with
28 CUBIC, it also moves the specification to the Standards Track,
29 obsoleting RFC 8312. This also requires updating RFC 5681, to allow
30 for CUBIC's occasionally more aggressive sending behavior.
32 Note to Readers
34 Discussion of this draft takes place on the TCPM working group
35 mailing list (mailto:tcpm@ietf.org), which is archived at
36 https://mailarchive.ietf.org/arch/browse/tcpm/.
38 Working Group information can be found at
39 https://datatracker.ietf.org/wg/tcpm/; source code and issues list
40 for this draft can be found at https://github.com/NTAP/rfc8312bis.
42 Note to the RFC Editor
44 xml2rfc currently renders in the XML by surrounding the
45 corresponding text with underscores. This is highly distracting;
46 please manually remove the underscores when doing the final edits to
47 the text version of this document.
49 (There is an issue open against xml2rfc to stop doing this in the
50 future: https://trac.tools.ietf.org/tools/xml2rfc/trac/ticket/596)
52 Also, please manually change "Figure" to "Equation" for all artwork
53 with anchors beginning with "eq" - xml2rfc doesn't seem to be able to
54 do this.
56 Status of This Memo
58 This Internet-Draft is submitted in full conformance with the
59 provisions of BCP 78 and BCP 79.
61 Internet-Drafts are working documents of the Internet Engineering
62 Task Force (IETF). Note that other groups may also distribute
63 working documents as Internet-Drafts. The list of current Internet-
64 Drafts is at https://datatracker.ietf.org/drafts/current/.
66 Internet-Drafts are draft documents valid for a maximum of six months
67 and may be updated, replaced, or obsoleted by other documents at any
68 time. It is inappropriate to use Internet-Drafts as reference
69 material or to cite them other than as "work in progress."
71 This Internet-Draft will expire on 28 April 2022.
73 Copyright Notice
75 Copyright (c) 2021 IETF Trust and the persons identified as the
76 document authors. All rights reserved.
78 This document is subject to BCP 78 and the IETF Trust's Legal
79 Provisions Relating to IETF Documents (https://trustee.ietf.org/
80 license-info) in effect on the date of publication of this document.
81 Please review these documents carefully, as they describe your rights
82 and restrictions with respect to this document. Code Components
83 extracted from this document must include Simplified BSD License text
84 as described in Section 4.e of the Trust Legal Provisions and are
85 provided without warranty as described in the Simplified BSD License.
87 Table of Contents
89 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 4
90 2. Conventions . . . . . . . . . . . . . . . . . . . . . . . . . 5
91 3. Design Principles of CUBIC . . . . . . . . . . . . . . . . . 5
92 3.1. Principle 1 for the CUBIC Increase Function . . . . . . . 5
93 3.2. Principle 2 for Reno-Friendliness . . . . . . . . . . . . 6
94 3.3. Principle 3 for RTT Fairness . . . . . . . . . . . . . . 7
95 3.4. Principle 4 for the CUBIC Decrease Factor . . . . . . . . 7
96 4. CUBIC Congestion Control . . . . . . . . . . . . . . . . . . 8
97 4.1. Definitions . . . . . . . . . . . . . . . . . . . . . . . 8
98 4.1.1. Constants of Interest . . . . . . . . . . . . . . . . 8
99 4.1.2. Variables of Interest . . . . . . . . . . . . . . . . 8
100 4.2. Window Increase Function . . . . . . . . . . . . . . . . 9
101 4.3. Reno-Friendly Region . . . . . . . . . . . . . . . . . . 11
102 4.4. Concave Region . . . . . . . . . . . . . . . . . . . . . 13
103 4.5. Convex Region . . . . . . . . . . . . . . . . . . . . . . 13
104 4.6. Multiplicative Decrease . . . . . . . . . . . . . . . . . 14
105 4.7. Fast Convergence . . . . . . . . . . . . . . . . . . . . 15
106 4.8. Timeout . . . . . . . . . . . . . . . . . . . . . . . . . 15
107 4.9. Spurious Congestion Events . . . . . . . . . . . . . . . 16
108 4.10. Slow Start . . . . . . . . . . . . . . . . . . . . . . . 17
109 5. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . 17
110 5.1. Fairness to Reno . . . . . . . . . . . . . . . . . . . . 18
111 5.2. Using Spare Capacity . . . . . . . . . . . . . . . . . . 20
112 5.3. Difficult Environments . . . . . . . . . . . . . . . . . 21
113 5.4. Investigating a Range of Environments . . . . . . . . . . 21
114 5.5. Protection against Congestion Collapse . . . . . . . . . 22
115 5.6. Fairness within the Alternative Congestion Control
116 Algorithm . . . . . . . . . . . . . . . . . . . . . . . 22
117 5.7. Performance with Misbehaving Nodes and Outside
118 Attackers . . . . . . . . . . . . . . . . . . . . . . . 22
119 5.8. Behavior for Application-Limited Flows . . . . . . . . . 22
120 5.9. Responses to Sudden or Transient Events . . . . . . . . . 22
121 5.10. Incremental Deployment . . . . . . . . . . . . . . . . . 23
122 6. Security Considerations . . . . . . . . . . . . . . . . . . . 23
123 7. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 23
124 8. References . . . . . . . . . . . . . . . . . . . . . . . . . 23
125 8.1. Normative References . . . . . . . . . . . . . . . . . . 23
126 8.2. Informative References . . . . . . . . . . . . . . . . . 25
127 Appendix A. Acknowledgments . . . . . . . . . . . . . . . . . . 27
128 Appendix B. Evolution of CUBIC . . . . . . . . . . . . . . . . . 28
129 B.1. Since draft-ietf-tcpm-rfc8312bis-04 . . . . . . . . . . . 28
130 B.2. Since draft-ietf-tcpm-rfc8312bis-03 . . . . . . . . . . . 29
131 B.3. Since draft-ietf-tcpm-rfc8312bis-02 . . . . . . . . . . . 29
132 B.4. Since draft-ietf-tcpm-rfc8312bis-01 . . . . . . . . . . . 29
133 B.5. Since draft-ietf-tcpm-rfc8312bis-00 . . . . . . . . . . . 30
134 B.6. Since draft-eggert-tcpm-rfc8312bis-03 . . . . . . . . . . 30
135 B.7. Since draft-eggert-tcpm-rfc8312bis-02 . . . . . . . . . . 30
136 B.8. Since draft-eggert-tcpm-rfc8312bis-01 . . . . . . . . . . 30
137 B.9. Since draft-eggert-tcpm-rfc8312bis-00 . . . . . . . . . . 30
138 B.10. Since RFC8312 . . . . . . . . . . . . . . . . . . . . . . 31
139 B.11. Since the Original Paper . . . . . . . . . . . . . . . . 31
140 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 32
142 1. Introduction
144 CUBIC has been adopted as the default TCP congestion control
145 algorithm in the Linux, Windows, and Apple stacks, and has been used
146 and deployed globally. Extensive, decade-long deployment experience
147 in vastly different Internet scenarios has convincingly demonstrated
148 that CUBIC is safe for deployment on the global Internet and delivers
149 substantial benefits over classical Reno congestion control
150 [RFC5681]. It is therefore to be regarded as the currently most
151 widely deployed standard for TCP congestion control. CUBIC can also
152 be used for other transport protocols such as QUIC [RFC9000] and SCTP
153 [RFC4960] as a default congestion controller.
155 The design of CUBIC was motivated by the well-documented problem
156 classical Reno TCP has with low utilization over fast and long-
157 distance networks [K03][RFC3649]. This problem arises from a slow
158 increase of the congestion window following a congestion event in a
159 network with a large bandwidth-delay product (BDP). [HKLRX06]
160 indicates that this problem is frequently observed even in the range
161 of congestion window sizes over several hundreds of packets. This
162 problem is equally applicable to all Reno-style standards and their
163 variants, including TCP-Reno [RFC5681], TCP-NewReno
164 [RFC6582][RFC6675], SCTP [RFC4960], TFRC [RFC5348], and QUIC
165 congestion control [RFC9002], which use the same linear increase
166 function for window growth. We refer to all Reno-style standards and
167 their variants collectively as "Reno" below.
169 CUBIC, originally proposed in [HRX08], is a modification to the
170 congestion control algorithm of classical Reno to remedy this
171 problem. Specifically, CUBIC uses a cubic function instead of the
172 linear window increase function of Reno to improve scalability and
173 stability under fast and long-distance networks.
175 This document updates the specification of CUBIC to include
176 algorithmic improvements based on the Linux, Windows, and Apple
177 implementations and recent academic work. Based on the extensive
178 deployment experience with CUBIC, it also moves the specification to
179 the Standards Track, obsoleting [RFC8312]. This requires an update
180 to [RFC5681], which limits the aggressiveness of Reno TCP
181 implementations in its Section 3. Since CUBIC is occasionally more
182 aggressive than the [RFC5681] algorithms, this document updates
183 [RFC5681] to allow for CUBIC's behavior.
185 Binary Increase Congestion Control (BIC-TCP) [XHR04], a predecessor
186 of CUBIC, was selected as the default TCP congestion control
187 algorithm by Linux in the year 2005 and had been used for several
188 years by the Internet community at large.
190 CUBIC uses a similar window increase function as BIC-TCP and is
191 designed to be less aggressive and fairer to Reno in bandwidth usage
192 than BIC-TCP while maintaining the strengths of BIC-TCP such as
193 stability, window scalability, and round-trip time (RTT) fairness.
195 In the following sections, we first briefly explain the design
196 principles of CUBIC, then provide the exact specification of CUBIC,
197 and finally discuss the safety features of CUBIC following the
198 guidelines specified in [RFC5033].
200 2. Conventions
202 The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
203 "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and
204 "OPTIONAL" in this document are to be interpreted as described in
205 BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all
206 capitals, as shown here.
208 3. Design Principles of CUBIC
210 CUBIC is designed according to the following design principles:
212 Principle 1: For better network utilization and stability, CUBIC
213 uses both the concave and convex profiles of a cubic function to
214 increase the congestion window size, instead of using just a
215 convex function.
217 Principle 2: To be Reno-friendly, CUBIC is designed to behave like
218 Reno in networks with short RTTs and small bandwidth where Reno
219 performs well.
221 Principle 3: For RTT-fairness, CUBIC is designed to achieve linear
222 bandwidth sharing among flows with different RTTs.
224 Principle 4: CUBIC appropriately sets its multiplicative window
225 decrease factor in order to balance between the scalability and
226 convergence speed.
228 3.1. Principle 1 for the CUBIC Increase Function
230 For better network utilization and stability, CUBIC [HRX08] uses a
231 cubic window increase function in terms of the elapsed time from the
232 last congestion event. While most alternative congestion control
233 algorithms to Reno increase the congestion window using convex
234 functions, CUBIC uses both the concave and convex profiles of a cubic
235 function for window growth.
237 After a window reduction in response to a congestion event detected
238 by duplicate ACKs, Explicit Congestion Notification-Echo (ECN-Echo,
239 ECE) ACKs [RFC3168], TCP RACK [RFC8985] or QUIC loss detection
240 [RFC9002], CUBIC remembers the congestion window size at which it
241 received the congestion event and performs a multiplicative decrease
242 of the congestion window. When CUBIC enters into congestion
243 avoidance, it starts to increase the congestion window using the
244 concave profile of the cubic function. The cubic function is set to
245 have its plateau at the remembered congestion window size, so that
246 the concave window increase continues until then. After that, the
247 cubic function turns into a convex profile and the convex window
248 increase begins.
250 This style of window adjustment (concave and then convex) improves
251 the algorithm stability while maintaining high network utilization
252 [CEHRX07]. This is because the window size remains almost constant,
253 forming a plateau around the remembered congestion window size of the
254 last congestion event, where network utilization is deemed highest.
255 Under steady state, most window size samples of CUBIC are close to
256 that remembered congestion window size, thus promoting high network
257 utilization and stability.
259 Note that congestion control algorithms that only use convex
260 functions to increase the congestion window size have their maximum
261 increments around the remembered congestion window size of the last
262 congestion event, and thus introduce many packet bursts around the
263 saturation point of the network, likely causing frequent global loss
264 synchronizations.
266 3.2. Principle 2 for Reno-Friendliness
268 CUBIC promotes per-flow fairness to Reno. Note that Reno performs
269 well over paths with short RTTs and small bandwidths (or small BDPs).
270 There is only a scalability problem in networks with long RTTs and
271 large bandwidths (or large BDPs).
273 A congestion control algorithm designed to be friendly to Reno on a
274 per-flow basis must increase its congestion window less aggressively
275 in small BDP networks than in large BDP networks.
277 The aggressiveness of CUBIC mainly depends on the maximum window size
278 before a window reduction, which is smaller in small-BDP networks
279 than in large-BDP networks. Thus, CUBIC increases its congestion
280 window less aggressively in small-BDP networks than in large-BDP
281 networks.
283 Furthermore, in cases when the cubic function of CUBIC would increase
284 the congestion window less aggressively than Reno, CUBIC simply
285 follows the window size of Reno to ensure that CUBIC achieves at
286 least the same throughput as Reno in small-BDP networks. We call
287 this region where CUBIC behaves like Reno the "Reno-friendly region".
289 3.3. Principle 3 for RTT Fairness
291 Two CUBIC flows with different RTTs have a throughput ratio that is
292 linearly proportional to the inverse of their RTT ratio, where the
293 throughput of a flow is approximately the size of its congestion
294 window divided by its RTT.
296 Specifically, CUBIC maintains a window increase rate independent of
297 RTTs outside the Reno-friendly region, and thus flows with different
298 RTTs have similar congestion window sizes under steady state when
299 they operate outside the Reno-friendly region.
301 This notion of a linear throughput ratio is similar to that of Reno
302 under high statistical multiplexing where packet loss is independent
303 of individual flow rates. However, under low statistical
304 multiplexing, the throughput ratio of Reno flows with different RTTs
305 is quadratically proportional to the inverse of their RTT ratio
306 [XHR04].
308 CUBIC always ensures a linear throughput ratio independent of the
309 amount of statistical multiplexing. This is an improvement over
310 Reno. While there is no consensus on particular throughput ratios
311 for different RTT flows, we believe that over wired Internet paths,
312 use of a linear throughput ratio seems more reasonable than equal
313 throughputs (i.e., the same throughput for flows with different RTTs)
314 or a higher-order throughput ratio (e.g., a quadratical throughput
315 ratio of Reno under low statistical multiplexing environments).
317 3.4. Principle 4 for the CUBIC Decrease Factor
319 To balance between scalability and convergence speed, CUBIC sets the
320 multiplicative window decrease factor to 0.7, whereas Reno uses 0.5.
322 While this improves the scalability of CUBIC, a side effect of this
323 decision is slower convergence, especially under low statistical
324 multiplexing. This design choice is following the observation that
325 HighSpeed TCP (HSTCP) [RFC3649] and other approaches (e.g., [GV02])
326 made: the current Internet becomes more asynchronous with less
327 frequent loss synchronizations under high statistical multiplexing.
329 In such environments, even strict Multiplicative-Increase
330 Multiplicative-Decrease (MIMD) can converge. CUBIC flows with the
331 same RTT always converge to the same throughput independent of
332 statistical multiplexing, thus achieving intra-algorithm fairness.
333 We also find that in environments with sufficient statistical
334 multiplexing, the convergence speed of CUBIC is reasonable.
336 4. CUBIC Congestion Control
338 In this section, we discuss how the congestion window is updated
339 during the different stages of the CUBIC congestion controller.
341 4.1. Definitions
343 The unit of all window sizes in this document is segments of the
344 maximum segment size (MSS), and the unit of all times is seconds.
345 Implementations can use bytes to express window sizes, which would
346 require factoring in the maximum segment size wherever necessary and
347 replacing _segments_acked_ with the number of bytes acknowledged in
348 Figure 4.
350 4.1.1. Constants of Interest
352 β__cubic_: CUBIC multiplicative decrease factor as described in
353 Section 4.6.
355 α__cubic_: CUBIC additive increase factor used in Reno-friendly
356 region as described in Section 4.3.
358 _C_: constant that determines the aggressiveness of CUBIC in
359 competing with other congestion control algorithms in high BDP
360 networks. Please see Section 5 for more explanation on how it is
361 set. The unit for _C_ is
363 segment
364 -------
365 3
366 second
368 4.1.2. Variables of Interest
370 This section defines the variables required to implement CUBIC:
372 _RTT_: Smoothed round-trip time in seconds, calculated as described
373 in [RFC6298].
375 _cwnd_: Current congestion window in segments.
377 _ssthresh_: Current slow start threshold in segments.
379 _W_max_: Size of _cwnd_ in segments just before _cwnd_ was reduced in
380 the last congestion event when fast convergence is disabled.
381 However, if fast convergence is enabled, the size may be further
382 reduced based on the current saturation point.
384 _K_: The time period in seconds it takes to increase the congestion
385 window size at the beginning of the current congestion avoidance
386 stage to _W_max_.
388 _current_time_: Current time of the system in seconds.
390 _epoch_start_: The time in seconds at which the current congestion
391 avoidance stage started.
393 _cwnd_start_: The _cwnd_ at the beginning of the current congestion
394 avoidance stage, i.e., at time _epoch_start_.
396 W_cubic(_t_): The congestion window in segments at time _t_ in
397 seconds based on the cubic increase function, as described in
398 Section 4.2.
400 _target_: Target value of congestion window in segments after the
401 next RTT, that is, W_cubic(_t_ + _RTT_), as described in Section 4.2.
403 _W_est_: An estimate for the congestion window in segments in the
404 Reno-friendly region, that is, an estimate for the congestion window
405 of Reno.
407 _segments_acked_: Number of MSS-sized segments acked when a "new ACK"
408 is received, i.e., an ACK that cumulatively acknowledges the delivery
409 of new data. This number will be a decimal value when a new ACK
410 acknowledges an amount of data that is not MSS-sized. Specifically,
411 it can be less than 1 when a new ACK acknowledges a segment smaller
412 than the MSS.
414 4.2. Window Increase Function
416 CUBIC maintains the acknowledgment (ACK) clocking of Reno by
417 increasing the congestion window only at the reception of a new ACK.
418 It does not make any changes to the TCP Fast Recovery and Fast
419 Retransmit algorithms [RFC6582][RFC6675].
421 During congestion avoidance, after a congestion event is detected by
422 mechanisms described in Section 3.1, CUBIC changes the window
423 increase function of Reno.
425 CUBIC uses the following window increase function:
427 3
428 W (t) = C * (t - K) + W
429 cubic max
431 Figure 1
433 where _t_ is the elapsed time in seconds from the beginning of the
434 current congestion avoidance stage, that is,
436 t = current_time - epoch
437 start
439 and where _epoch_start_ is the time at which the current congestion
440 avoidance stage starts. _K_ is the time period that the above
441 function takes to increase the congestion window size at the
442 beginning of the current congestion avoidance stage to _W_max_ if
443 there are no further congestion events and is calculated using the
444 following equation:
446 ________________
447 /W - cwnd
448 3 / max start
449 K = | / ----------------
450 |/ C
452 Figure 2
454 where _cwnd_start_ is the congestion window at the beginning of the
455 current congestion avoidance stage.
457 Upon receiving a new ACK during congestion avoidance, CUBIC computes
458 the _target_ congestion window size after the next _RTT_ using
459 Figure 1 as follows, where _RTT_ is the smoothed round-trip time.
460 The lower and upper bounds below ensure that CUBIC's congestion
461 window increase rate is non-decreasing and is less than the increase
462 rate of slow start.
464 /
465 | if W (t + RTT) < cwnd
466 |cwnd cubic
467 |
468 |
469 |
470 target = < if W (t + RTT) > 1.5 * cwnd
471 |1.5 * cwnd cubic
472 |
473 |
474 |W (t + RTT)
475 | cubic otherwise
476 \
478 The elapsed time _t_ in Figure 1 MUST NOT include periods during
479 which _cwnd_ has not been updated due to an application limit (see
480 Section 5.8).
482 Depending on the value of the current congestion window size _cwnd_,
483 CUBIC runs in three different regions:
485 1. The Reno-friendly region, which ensures that CUBIC achieves at
486 least the same throughput as Reno.
488 2. The concave region, if CUBIC is not in the Reno-friendly region
489 and _cwnd_ is less than _W_max_.
491 3. The convex region, if CUBIC is not in the Reno-friendly region
492 and _cwnd_ is greater than _W_max_.
494 Below, we describe the exact actions taken by CUBIC in each region.
496 4.3. Reno-Friendly Region
498 Reno performs well in certain types of networks, for example, under
499 short RTTs and small bandwidths (or small BDPs). In these networks,
500 CUBIC remains in the Reno-friendly region to achieve at least the
501 same throughput as Reno.
503 The Reno-friendly region is designed according to the analysis in
504 [FHP00], which studies the performance of an AIMD algorithm with an
505 additive factor of α (segments per _RTT_) and a multiplicative factor
506 of β, denoted by AIMD(α, β). _p_ is the packet loss rate.
507 Specifically, the average congestion window size of AIMD(α, β) can be
508 calculated using Figure 3.
510 _______________
511 / α * (1 + β)
512 AVG_AIMD(α, β) = | / ---------------
513 |/ 2 * (1 - β) * p
515 Figure 3
517 By the same analysis, to achieve the same average window size as Reno
518 that uses AIMD(1, 0.5), α must be equal to,
520 1 - β
521 3 * -----
522 1 + β
524 Thus, CUBIC uses Figure 4 to estimate the window size _W_est_ in the
525 Reno-friendly region with
527 1 - β
528 cubic
529 α = 3 * ----------
530 cubic 1 + β
531 cubic
533 which achieves the same average window size as Reno. When receiving
534 a new ACK in congestion avoidance (where _cwnd_ could be greater than
535 or less than _W_max_), CUBIC checks whether W_cubic(_t_) is less than
536 _W_est_. If so, CUBIC is in the Reno-friendly region and _cwnd_
537 SHOULD be set to _W_est_ at each reception of a new ACK.
539 _W_est_ is set equal to _cwnd_start_ at the start of the congestion
540 avoidance stage. After that, on every new ACK, _W_est_ is updated
541 using Figure 4. Note that this equation is for a connection where
542 Appropriate Byte Counting (ABC) [RFC3465] is disabled. For a
543 connection with ABC enabled, this equation SHOULD be adjusted by
544 using the number of acknowledged bytes instead of acknowledged
545 segments. Also note that this equation works for connections with
546 enabled or disabled Delayed ACKs [RFC5681], as _segments_acked_ will
547 be different based on the segments actually acknowledged by a new
548 ACK.
550 segments_acked
551 W = W + α * --------------
552 est est cubic cwnd
554 Figure 4
556 Note that once _W_est_ reaches _W_max_, that is, _W_est_ >= _W_max_,
557 CUBIC needs to start probing to determine the new value of _W_max_.
558 At this point, α__cubic_ SHOULD be set to 1 to ensure that CUBIC can
559 achieve the same congestion window increment as Reno, which uses
560 AIMD(1, 0.5).
562 4.4. Concave Region
564 When receiving a new ACK in congestion avoidance, if CUBIC is not in
565 the Reno-friendly region and _cwnd_ is less than _W_max_, then CUBIC
566 is in the concave region. In this region, _cwnd_ MUST be incremented
567 by
569 target - cwnd
570 -------------
571 cwnd
573 for each received new ACK, where _target_ is calculated as described
574 in Section 4.2.
576 4.5. Convex Region
578 When receiving a new ACK in congestion avoidance, if CUBIC is not in
579 the Reno-friendly region and _cwnd_ is larger than or equal to
580 _W_max_, then CUBIC is in the convex region.
582 The convex region indicates that the network conditions might have
583 changed since the last congestion event, possibly implying more
584 available bandwidth after some flow departures. Since the Internet
585 is highly asynchronous, some amount of perturbation is always
586 possible without causing a major change in available bandwidth.
588 Unless it is overridden by the AIMD window increase, CUBIC is very
589 careful in this region. The convex profile aims to increase the
590 window very slowly at the beginning when _cwnd_ is around _W_max_ and
591 then gradually increases its rate of increase. We also call this
592 region the "maximum probing phase", since CUBIC is searching for a
593 new _W_max_. In this region, _cwnd_ MUST be incremented by
595 target - cwnd
596 -------------
597 cwnd
599 for each received new ACK, where _target_ is calculated as described
600 in Section 4.2.
602 4.6. Multiplicative Decrease
604 When a congestion event is detected by mechanisms described in
605 Section 3.1, CUBIC updates _W_max_ and reduces _cwnd_ and _ssthresh_
606 immediately as described below. In case of packet loss, the sender
607 MUST reduce _cwnd_ and _ssthresh_ immediately upon entering loss
608 recovery, similar to [RFC5681] (and [RFC6675]). Note that other
609 mechanisms, such as Proportional Rate Reduction [RFC6937], can be
610 used to reduce the sending rate during loss recovery more gradually.
611 The parameter β__cubic_ SHOULD be set to 0.7, which is different from
612 the multiplicative decrease factor used in [RFC5681] (and [RFC6675])
613 during fast recovery.
615 In Figure 5, _flight_size_ is the amount of outstanding data in the
616 network, as defined in [RFC5681]. Note that a rate-limited
617 application with idle periods or periods when unable to send at the
618 full rate permitted by _cwnd_ may easily encounter notable variations
619 in the volume of data sent from one RTT to another, resulting in
620 _flight_size_ that is significantly less than _cwnd_ on a congestion
621 event. This may decrease _cwnd_ to a much lower value than
622 necessary. To avoid suboptimal performance with such applications,
623 some implementations of CUBIC use _cwnd_ instead of _flight_size_ to
624 calculate the new _ssthresh_ in Figure 5. Alternatively, the
625 mechanisms described in [RFC7661] may also be adopted to mitigate
626 this issue.
628 flight_size * β // new ssthresh
629 ssthresh = cubic
631 /max(ssthresh, 2) // reduction on packet loss, cwnd is at least 2 MSS
632 |
633 cwnd = <
634 |max(ssthresh, 1) // reduction on ECE, cwnd is at least 1 MSS
635 \
637 max(ssthresh, 2) // ssthresh is at least 2 MSS
638 ssthresh =
640 Figure 5
642 A side effect of setting β__cubic_ to a value bigger than 0.5 is
643 slower convergence. We believe that while a more adaptive setting of
644 β__cubic_ could result in faster convergence, it will make the
645 analysis of CUBIC much harder.
647 Note that CUBIC will continue to reduce _cwnd_ in response to
648 congestion events due to ECN-Echo ACKs until it reaches a value of 1
649 MSS. If congestion persists, a sender with a _cwnd_ of 1 MSS needs
650 to reduce its sending rate even further. It can achieve that by
651 using a retransmission timer with exponential backoff, as described
652 in [RFC3168].
654 4.7. Fast Convergence
656 To improve convergence speed, CUBIC uses a heuristic. When a new
657 flow joins the network, existing flows need to give up some of their
658 bandwidth to allow the new flow some room for growth, if the existing
659 flows have been using all the network bandwidth. To speed up this
660 bandwidth release by existing flows, the following "Fast Convergence"
661 mechanism SHOULD be implemented.
663 With Fast Convergence, when a congestion event occurs, we update
664 _W_max_ as follows, before the window reduction as described in
665 Section 4.6.
667 /
668 | 1 + β
669 | cubic if cwnd < W and fast convergence is enabled,
670 |cwnd * ---------- max
671 | 2
672 W = <
673 max | further reduce W
674 | max
675 |
676 | otherwise, remember cwnd before reduction
677 \cwnd
679 At a congestion event, if the current _cwnd_ is less than _W_max_,
680 this indicates that the saturation point experienced by this flow is
681 getting reduced because of a change in available bandwidth. Then we
682 allow this flow to release more bandwidth by reducing _W_max_
683 further. This action effectively lengthens the time for this flow to
684 increase its congestion window, because the reduced _W_max_ forces
685 the flow to plateau earlier. This allows more time for the new flow
686 to catch up to its congestion window size.
688 Fast Convergence is designed for network environments with multiple
689 CUBIC flows. In network environments with only a single CUBIC flow
690 and without any other traffic, Fast Convergence SHOULD be disabled.
692 4.8. Timeout
694 In case of a timeout, CUBIC follows Reno to reduce _cwnd_ [RFC5681],
695 but sets _ssthresh_ using β__cubic_ (same as in Section 4.6) in a way
696 that is different from Reno TCP [RFC5681].
698 During the first congestion avoidance stage after a timeout, CUBIC
699 increases its congestion window size using Figure 1, where _t_ is the
700 elapsed time since the beginning of the current congestion avoidance,
701 _K_ is set to 0, and _W_max_ is set to the congestion window size at
702 the beginning of the current congestion avoidance stage. In
703 addition, for the Reno-friendly region, _W_est_ SHOULD be set to the
704 congestion window size at the beginning of the current congestion
705 avoidance.
707 4.9. Spurious Congestion Events
709 In cases where CUBIC reduces its congestion window in response to
710 having detected packet loss via duplicate ACKs or timeouts, there is
711 a possibility that the missing ACK would arrive after the congestion
712 window reduction and a corresponding packet retransmission. For
713 example, packet reordering could trigger this behavior. A high
714 degree of packet reordering could cause multiple congestion window
715 reduction events, where spurious losses are incorrectly interpreted
716 as congestion signals, thus degrading CUBIC's performance
717 significantly.
719 When there is a congestion event, a CUBIC implementation SHOULD save
720 the current value of the following variables before the congestion
721 window reduction.
723 prior_cwnd = cwnd
725 prior_ssthresh = ssthresh
727 prior_W = W
728 max max
730 prior_K = K
732 prior_epoch = epoch
733 start start
735 prior_W_{est} = W
736 est
738 CUBIC MAY implement an algorithm to detect spurious retransmissions,
739 such as Forward RTO-Recovery [RFC5682]. Experimental alternatives
740 include DSACK [RFC3708] and Eifel [RFC3522]. Once a spurious
741 congestion event is detected, CUBIC SHOULD restore the original
742 values of above-mentioned variables as follows if the current _cwnd_
743 is lower than _prior_cwnd_. Restoring the original values ensures
744 that CUBIC's performance is similar to what it would be without
745 spurious losses.
747 \
748 cwnd = prior_cwnd |
749 |
750 ssthresh = prior_ssthresh |
751 |
752 W = prior_W |
753 max max |
754 >if cwnd < prior_cwnd
755 K = prior_K |
756 |
757 epoch = prior_epoch |
758 start start|
759 |
760 W = prior_W |
761 est est /
763 In rare cases, when the detection happens long after a spurious loss
764 event and the current _cwnd_ is already higher than _prior_cwnd_,
765 CUBIC SHOULD continue to use the current and the most recent values
766 of these variables.
768 4.10. Slow Start
770 CUBIC MUST employ a slow-start algorithm, when _cwnd_ is no more than
771 _ssthresh_. In general, CUBIC SHOULD use the HyStart++ slow start
772 algorithm [I-D.ietf-tcpm-hystartplusplus], or MAY use the Reno TCP
773 slow start algorithm [RFC5681] in the rare cases when HyStart++ is
774 not suitable. Experimental alternatives include hybrid slow start
775 [HR08], a predecessor to HyStart++ that some CUBIC implementations
776 have used as the default for the last decade, and limited slow start
777 [RFC3742]. Whichever start-up algorithm is used, work might be
778 needed to ensure that the end of slow start and the first
779 multiplicative decrease of congestion avoidance work well together.
781 When CUBIC uses HyStart++ [I-D.ietf-tcpm-hystartplusplus], it may
782 exit the first slow start without incurring any packet loss and thus
783 _W_max_ is undefined. In this special case, CUBIC switches to
784 congestion avoidance and increases its congestion window size using
785 Figure 1, where _t_ is the elapsed time since the beginning of the
786 current congestion avoidance, _K_ is set to 0, and _W_max_ is set to
787 the congestion window size at the beginning of the current congestion
788 avoidance stage.
790 5. Discussion
792 In this section, we further discuss the safety features of CUBIC
793 following the guidelines specified in [RFC5033].
795 With a deterministic loss model where the number of packets between
796 two successive packet losses is always _1/p_, CUBIC always operates
797 with the concave window profile, which greatly simplifies the
798 performance analysis of CUBIC. The average window size of CUBIC can
799 be obtained by the following function:
801 ________________ ____
802 /C * (3 + β ) 4 / 3
803 4 / cubic |/ RTT
804 AVG_W = | / ---------------- * -------
805 cubic | / 4 * (1 - β ) __
806 |/ cubic 4 / 3
807 |/ p
809 Figure 6
811 With β__cubic_ set to 0.7, the above formula reduces to:
813 ____
814 _______ 4 / 3
815 4 /C * 3.7 |/ RTT
816 AVG_W = | / ------- * -------
817 cubic |/ 1.2 __
818 4 / 3
819 |/ p
821 Figure 7
823 We will determine the value of _C_ in the following subsection using
824 Figure 7.
826 5.1. Fairness to Reno
828 In environments where Reno is able to make reasonable use of the
829 available bandwidth, CUBIC does not significantly change this state.
831 Reno performs well in the following two types of networks:
833 1. networks with a small bandwidth-delay product (BDP)
835 2. networks with a short RTTs, but not necessarily a small BDP
837 CUBIC is designed to behave very similarly to Reno in the above two
838 types of networks. The following two tables show the average window
839 sizes of Reno TCP, HSTCP, and CUBIC TCP. The average window sizes of
840 Reno TCP and HSTCP are from [RFC3649]. The average window size of
841 CUBIC is calculated using Figure 7 and the CUBIC Reno-friendly region
842 for three different values of _C_.
844 +=============+=======+========+================+=========+========+
845 | Loss Rate P | Reno | HSTCP | CUBIC (C=0.04) | CUBIC | CUBIC |
846 | | | | | (C=0.4) | (C=4) |
847 +=============+=======+========+================+=========+========+
848 | 1.0e-02 | 12 | 12 | 12 | 12 | 12 |
849 +-------------+-------+--------+----------------+---------+--------+
850 | 1.0e-03 | 38 | 38 | 38 | 38 | 59 |
851 +-------------+-------+--------+----------------+---------+--------+
852 | 1.0e-04 | 120 | 263 | 120 | 187 | 333 |
853 +-------------+-------+--------+----------------+---------+--------+
854 | 1.0e-05 | 379 | 1795 | 593 | 1054 | 1874 |
855 +-------------+-------+--------+----------------+---------+--------+
856 | 1.0e-06 | 1200 | 12280 | 3332 | 5926 | 10538 |
857 +-------------+-------+--------+----------------+---------+--------+
858 | 1.0e-07 | 3795 | 83981 | 18740 | 33325 | 59261 |
859 +-------------+-------+--------+----------------+---------+--------+
860 | 1.0e-08 | 12000 | 574356 | 105383 | 187400 | 333250 |
861 +-------------+-------+--------+----------------+---------+--------+
863 Table 1: Reno TCP, HSTCP, and CUBIC with RTT = 0.1 seconds
865 Table 1 describes the response function of Reno TCP, HSTCP, and CUBIC
866 in networks with _RTT_ = 0.1 seconds. The average window size is in
867 MSS-sized segments.
869 +=============+=======+========+================+=========+=======+
870 | Loss Rate P | Reno | HSTCP | CUBIC (C=0.04) | CUBIC | CUBIC |
871 | | | | | (C=0.4) | (C=4) |
872 +=============+=======+========+================+=========+=======+
873 | 1.0e-02 | 12 | 12 | 12 | 12 | 12 |
874 +-------------+-------+--------+----------------+---------+-------+
875 | 1.0e-03 | 38 | 38 | 38 | 38 | 38 |
876 +-------------+-------+--------+----------------+---------+-------+
877 | 1.0e-04 | 120 | 263 | 120 | 120 | 120 |
878 +-------------+-------+--------+----------------+---------+-------+
879 | 1.0e-05 | 379 | 1795 | 379 | 379 | 379 |
880 +-------------+-------+--------+----------------+---------+-------+
881 | 1.0e-06 | 1200 | 12280 | 1200 | 1200 | 1874 |
882 +-------------+-------+--------+----------------+---------+-------+
883 | 1.0e-07 | 3795 | 83981 | 3795 | 5926 | 10538 |
884 +-------------+-------+--------+----------------+---------+-------+
885 | 1.0e-08 | 12000 | 574356 | 18740 | 33325 | 59261 |
886 +-------------+-------+--------+----------------+---------+-------+
888 Table 2: Reno TCP, HSTCP, and CUBIC with RTT = 0.01 seconds
890 Table 2 describes the response function of Reno TCP, HSTCP, and CUBIC
891 in networks with _RTT_ = 0.01 seconds. The average window size is in
892 MSS-sized segments.
894 Both tables show that CUBIC with any of these three _C_ values is
895 more friendly to Reno TCP than HSTCP, especially in networks with a
896 short _RTT_ where Reno TCP performs reasonably well. For example, in
897 a network with _RTT_ = 0.01 seconds and p=10^-6, Reno TCP has an
898 average window of 1200 packets. If the packet size is 1500 bytes,
899 then Reno TCP can achieve an average rate of 1.44 Gbps. In this
900 case, CUBIC with _C_=0.04 or _C_=0.4 achieves exactly the same rate
901 as Reno TCP, whereas HSTCP is about ten times more aggressive than
902 Reno TCP.
904 We can see that _C_ determines the aggressiveness of CUBIC in
905 competing with other congestion control algorithms for bandwidth.
906 CUBIC is more friendly to Reno TCP, if the value of _C_ is lower.
907 However, we do not recommend setting _C_ to a very low value like
908 0.04, since CUBIC with a low _C_ cannot efficiently use the bandwidth
909 in fast and long-distance networks. Based on these observations and
910 extensive deployment experience, we find _C_=0.4 gives a good balance
911 between Reno-friendliness and aggressiveness of window increase.
912 Therefore, _C_ SHOULD be set to 0.4. With _C_ set to 0.4, Figure 7
913 is reduced to:
915 ____
916 4 / 3
917 |/ RTT
918 AVG_W = 1.054 * -------
919 cubic __
920 4 / 3
921 |/ p
923 Figure 8
925 Figure 8 is then used in the next subsection to show the scalability
926 of CUBIC.
928 5.2. Using Spare Capacity
930 CUBIC uses a more aggressive window increase function than Reno for
931 fast and long-distance networks.
933 The following table shows that to achieve the 10 Gbps rate, Reno TCP
934 requires a packet loss rate of 2.0e-10, while CUBIC TCP requires a
935 packet loss rate of 2.9e-8.
937 +===================+===========+=========+=========+=========+
938 | Throughput (Mbps) | Average W | Reno P | HSTCP P | CUBIC P |
939 +===================+===========+=========+=========+=========+
940 | 1 | 8.3 | 2.0e-2 | 2.0e-2 | 2.0e-2 |
941 +-------------------+-----------+---------+---------+---------+
942 | 10 | 83.3 | 2.0e-4 | 3.9e-4 | 2.9e-4 |
943 +-------------------+-----------+---------+---------+---------+
944 | 100 | 833.3 | 2.0e-6 | 2.5e-5 | 1.4e-5 |
945 +-------------------+-----------+---------+---------+---------+
946 | 1000 | 8333.3 | 2.0e-8 | 1.5e-6 | 6.3e-7 |
947 +-------------------+-----------+---------+---------+---------+
948 | 10000 | 83333.3 | 2.0e-10 | 1.0e-7 | 2.9e-8 |
949 +-------------------+-----------+---------+---------+---------+
951 Table 3: Required packet loss rate for Reno TCP, HSTCP, and
952 CUBIC to achieve a certain throughput
954 Table 3 describes the required packet loss rate for Reno TCP, HSTCP,
955 and CUBIC to achieve a certain throughput. We use 1500-byte packets
956 and an _RTT_ of 0.1 seconds.
958 Our test results in [HKLRX06] indicate that CUBIC uses the spare
959 bandwidth left unused by existing Reno TCP flows in the same
960 bottleneck link without taking away much bandwidth from the existing
961 flows.
963 5.3. Difficult Environments
965 CUBIC is designed to remedy the poor performance of Reno in fast and
966 long-distance networks.
968 5.4. Investigating a Range of Environments
970 There is decade-long deployment experience with CUBIC on the
971 Internet. CUBIC has also been extensively studied by using both NS-2
972 simulation and testbed experiments, covering a wide range of network
973 environments. More information can be found in [HKLRX06].
975 Same as Reno, CUBIC is a loss-based congestion control algorithm.
976 Because CUBIC is designed to be more aggressive (due to a faster
977 window increase function and bigger multiplicative decrease factor)
978 than Reno in fast and long-distance networks, it can fill large drop-
979 tail buffers more quickly than Reno and increases the risk of a
980 standing queue [RFC8511]. In this case, proper queue sizing and
981 management [RFC7567] could be used to mitigate the risk to some
982 extent and reduce the packet queuing delay. Also, in large-BDP
983 networks after a congestion event, CUBIC, due its cubic window
984 increase function, recovers quickly to the highest link utilization
985 point. This means that link utilization is less sensitive to an
986 active queue management (AQM) target that is lower than the amplitude
987 of the whole sawtooth.
989 Similar to Reno, the performance of CUBIC as a loss-based congestion
990 control algorithm suffers in networks where a packet loss is not a
991 good indication of bandwidth utilization, such as wireless or mobile
992 networks [LIU16].
994 5.5. Protection against Congestion Collapse
996 With regard to the potential of causing congestion collapse, CUBIC
997 behaves like Reno, since CUBIC modifies only the window adjustment
998 algorithm of Reno. Thus, it does not modify the ACK clocking and
999 timeout behaviors of Reno.
1001 CUBIC also satisfies the "full backoff" requirement as described in
1002 [RFC5033]. After reducing the sending rate to one packet per RTT in
1003 response to congestion events due to ECN-Echo ACKs, CUBIC then
1004 exponentially increases the transmission timer for each packet
1005 retransmission while congestion persists.
1007 5.6. Fairness within the Alternative Congestion Control Algorithm
1009 CUBIC ensures convergence of competing CUBIC flows with the same RTT
1010 in the same bottleneck links to an equal throughput. When competing
1011 flows have different RTT values, their throughput ratio is linearly
1012 proportional to the inverse of their RTT ratios. This is true
1013 independently of the level of statistical multiplexing on the link.
1015 5.7. Performance with Misbehaving Nodes and Outside Attackers
1017 This is not considered in the current CUBIC design.
1019 5.8. Behavior for Application-Limited Flows
1021 CUBIC does not increase its congestion window size if a flow is
1022 currently limited by the application instead of the congestion
1023 window. Section 4.2 requires that _t_ in Figure 1 does not include
1024 application-limited periods, such as idle periods, otherwise
1025 W_cubic(_t_) might be very high after restarting from these periods.
1027 5.9. Responses to Sudden or Transient Events
1029 If there is a sudden increase in capacity, e.g., due to variable
1030 radio capacity, a routing change, or a mobility event, CUBIC is
1031 designed to utilize the newly available capacity faster than Reno.
1033 On the other hand, if there is a sudden decrease in capacity, CUBIC
1034 reduces more slowly than Reno. This remains true whether or not
1035 CUBIC is in Reno-friendly mode and whether or not fast convergence is
1036 enabled.
1038 5.10. Incremental Deployment
1040 CUBIC requires only changes to the congestion control at the sender,
1041 and it does not require any changes at receivers. That is, a CUBIC
1042 sender works correctly with Reno receivers. In addition, CUBIC does
1043 not require any changes to routers and does not require any
1044 assistance from routers.
1046 6. Security Considerations
1048 CUBIC makes no changes to the underlying security of TCP. More
1049 information about TCP security concerns can be found in [RFC5681].
1051 7. IANA Considerations
1053 This document does not require any IANA actions.
1055 8. References
1057 8.1. Normative References
1059 [I-D.ietf-tcpm-hystartplusplus]
1060 Balasubramanian, P., Huang, Y., and M. Olson, "HyStart++:
1061 Modified Slow Start for TCP", Work in Progress, Internet-
1062 Draft, draft-ietf-tcpm-hystartplusplus-03, 25 July 2021,
1063 .
1066 [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate
1067 Requirement Levels", BCP 14, RFC 2119,
1068 DOI 10.17487/RFC2119, March 1997,
1069 .
1071 [RFC3168] Ramakrishnan, K., Floyd, S., and D. Black, "The Addition
1072 of Explicit Congestion Notification (ECN) to IP",
1073 RFC 3168, DOI 10.17487/RFC3168, September 2001,
1074 .
1076 [RFC5033] Floyd, S. and M. Allman, "Specifying New Congestion
1077 Control Algorithms", BCP 133, RFC 5033,
1078 DOI 10.17487/RFC5033, August 2007,
1079 .
1081 [RFC5348] Floyd, S., Handley, M., Padhye, J., and J. Widmer, "TCP
1082 Friendly Rate Control (TFRC): Protocol Specification",
1083 RFC 5348, DOI 10.17487/RFC5348, September 2008,
1084 .
1086 [RFC5681] Allman, M., Paxson, V., and E. Blanton, "TCP Congestion
1087 Control", RFC 5681, DOI 10.17487/RFC5681, September 2009,
1088 .
1090 [RFC5682] Sarolahti, P., Kojo, M., Yamamoto, K., and M. Hata,
1091 "Forward RTO-Recovery (F-RTO): An Algorithm for Detecting
1092 Spurious Retransmission Timeouts with TCP", RFC 5682,
1093 DOI 10.17487/RFC5682, September 2009,
1094 .
1096 [RFC6298] Paxson, V., Allman, M., Chu, J., and M. Sargent,
1097 "Computing TCP's Retransmission Timer", RFC 6298,
1098 DOI 10.17487/RFC6298, June 2011,
1099 .
1101 [RFC6582] Henderson, T., Floyd, S., Gurtov, A., and Y. Nishida, "The
1102 NewReno Modification to TCP's Fast Recovery Algorithm",
1103 RFC 6582, DOI 10.17487/RFC6582, April 2012,
1104 .
1106 [RFC6675] Blanton, E., Allman, M., Wang, L., Jarvinen, I., Kojo, M.,
1107 and Y. Nishida, "A Conservative Loss Recovery Algorithm
1108 Based on Selective Acknowledgment (SACK) for TCP",
1109 RFC 6675, DOI 10.17487/RFC6675, August 2012,
1110 .
1112 [RFC7567] Baker, F., Ed. and G. Fairhurst, Ed., "IETF
1113 Recommendations Regarding Active Queue Management",
1114 BCP 197, RFC 7567, DOI 10.17487/RFC7567, July 2015,
1115 .
1117 [RFC8174] Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC
1118 2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174,
1119 May 2017, .
1121 [RFC8985] Cheng, Y., Cardwell, N., Dukkipati, N., and P. Jha, "The
1122 RACK-TLP Loss Detection Algorithm for TCP", RFC 8985,
1123 DOI 10.17487/RFC8985, February 2021,
1124 .
1126 [RFC9002] Iyengar, J., Ed. and I. Swett, Ed., "QUIC Loss Detection
1127 and Congestion Control", RFC 9002, DOI 10.17487/RFC9002,
1128 May 2021, .
1130 8.2. Informative References
1132 [CEHRX07] Cai, H., Eun, D., Ha, S., Rhee, I., and L. Xu, "Stochastic
1133 Ordering for Internet Congestion Control and its
1134 Applications", IEEE INFOCOM 2007 - 26th IEEE International
1135 Conference on Computer Communications,
1136 DOI 10.1109/infcom.2007.111, 2007,
1137 .
1139 [FHP00] Floyd, S., Handley, M., and J. Padhye, "A Comparison of
1140 Equation-Based and AIMD Congestion Control", May 2000,
1141 .
1143 [GV02] Gorinsky, S. and H. Vin, "Extended Analysis of Binary
1144 Adjustment Algorithms", Technical Report TR2002-29,
1145 Department of Computer Sciences, The University of
1146 Texas at Austin, 11 August 2002,
1147 .
1149 [HKLRX06] Ha, S., Kim, Y., Le, L., Rhee, I., and L. Xu, "A Step
1150 toward Realistic Performance Evaluation of High-Speed TCP
1151 Variants", International Workshop on Protocols for Fast
1152 Long-Distance Networks, February 2006,
1153 .
1155 [HR08] Ha, S. and I. Rhee, "Hybrid Slow Start for High-Bandwidth
1156 and Long-Distance Networks", International Workshop
1157 on Protocols for Fast Long-Distance Networks, March 2008,
1158 .
1161 [HRX08] Ha, S., Rhee, I., and L. Xu, "CUBIC: a new TCP-friendly
1162 high-speed TCP variant", ACM SIGOPS Operating Systems
1163 Review Vol. 42, pp. 64-74, DOI 10.1145/1400097.1400105,
1164 July 2008, .
1166 [K03] Kelly, T., "Scalable TCP: improving performance in
1167 highspeed wide area networks", ACM SIGCOMM Computer
1168 Communication Review Vol. 33, pp. 83-91,
1169 DOI 10.1145/956981.956989, April 2003,
1170 .
1172 [LIU16] Liu, K. and J. Lee, "On Improving TCP Performance over
1173 Mobile Data Networks", IEEE Transactions on Mobile
1174 Computing Vol. 15, pp. 2522-2536,
1175 DOI 10.1109/tmc.2015.2500227, October 2016,
1176 .
1178 [RFC3465] Allman, M., "TCP Congestion Control with Appropriate Byte
1179 Counting (ABC)", RFC 3465, DOI 10.17487/RFC3465, February
1180 2003, .
1182 [RFC3522] Ludwig, R. and M. Meyer, "The Eifel Detection Algorithm
1183 for TCP", RFC 3522, DOI 10.17487/RFC3522, April 2003,
1184 .
1186 [RFC3649] Floyd, S., "HighSpeed TCP for Large Congestion Windows",
1187 RFC 3649, DOI 10.17487/RFC3649, December 2003,
1188 .
1190 [RFC3708] Blanton, E. and M. Allman, "Using TCP Duplicate Selective
1191 Acknowledgement (DSACKs) and Stream Control Transmission
1192 Protocol (SCTP) Duplicate Transmission Sequence Numbers
1193 (TSNs) to Detect Spurious Retransmissions", RFC 3708,
1194 DOI 10.17487/RFC3708, February 2004,
1195 .
1197 [RFC3742] Floyd, S., "Limited Slow-Start for TCP with Large
1198 Congestion Windows", RFC 3742, DOI 10.17487/RFC3742, March
1199 2004, .
1201 [RFC4960] Stewart, R., Ed., "Stream Control Transmission Protocol",
1202 RFC 4960, DOI 10.17487/RFC4960, September 2007,
1203 .
1205 [RFC6937] Mathis, M., Dukkipati, N., and Y. Cheng, "Proportional
1206 Rate Reduction for TCP", RFC 6937, DOI 10.17487/RFC6937,
1207 May 2013, .
1209 [RFC7661] Fairhurst, G., Sathiaseelan, A., and R. Secchi, "Updating
1210 TCP to Support Rate-Limited Traffic", RFC 7661,
1211 DOI 10.17487/RFC7661, October 2015,
1212 .
1214 [RFC8312] Rhee, I., Xu, L., Ha, S., Zimmermann, A., Eggert, L., and
1215 R. Scheffenegger, "CUBIC for Fast Long-Distance Networks",
1216 RFC 8312, DOI 10.17487/RFC8312, February 2018,
1217 .
1219 [RFC8511] Khademi, N., Welzl, M., Armitage, G., and G. Fairhurst,
1220 "TCP Alternative Backoff with ECN (ABE)", RFC 8511,
1221 DOI 10.17487/RFC8511, December 2018,
1222 .
1224 [RFC9000] Iyengar, J., Ed. and M. Thomson, Ed., "QUIC: A UDP-Based
1225 Multiplexed and Secure Transport", RFC 9000,
1226 DOI 10.17487/RFC9000, May 2021,
1227 .
1229 [SXEZ19] Sun, W., Xu, L., Elbaum, S., and D. Zhao, "Model-Agnostic
1230 and Efficient Exploration of Numerical State Space of
1231 Real-World TCP Congestion Control Implementations", USENIX
1232 NSDI 2019, February 2019,
1233 .
1235 [XHR04] Xu, L., Harfoush, K., and I. Rhee, "Binary Increase
1236 Congestion Control (BIC) for Fast Long-Distance Networks",
1237 IEEE INFOCOM 2004, DOI 10.1109/infcom.2004.1354672, March
1238 2004, .
1240 Appendix A. Acknowledgments
1242 Richard Scheffenegger and Alexander Zimmermann originally co-authored
1243 [RFC8312].
1245 These individuals suggested improvements to this document:
1247 * Bob Briscoe
1249 * Christian Huitema
1251 * Gorry Fairhurst
1253 * Jonathan Morton
1255 * Juhamatti Kuusisaari
1257 * Junho Choi
1259 * Markku Kojo
1261 * Martin Thomson
1263 * Matt Olson
1265 * Michael Welzl
1267 * Mirja Kuehlewind
1269 * Mohit P. Tahiliani
1271 * Neal Cardwell
1272 * Praveen Balasubramanian
1274 * Richard Scheffenegger
1276 * Rod Grimes
1278 * Tom Henderson
1280 * Tom Petch
1282 * Wesley Rosenblum
1284 * Yoshifumi Nishida
1286 * Yuchung Cheng
1288 Appendix B. Evolution of CUBIC
1290 B.1. Since draft-ietf-tcpm-rfc8312bis-04
1292 * Fix incorrect math (#106 (https://github.com/NTAP/rfc8312bis/
1293 issues/106))
1295 * Update RFC5681 (#99 (https://github.com/NTAP/rfc8312bis/
1296 issues/99))
1298 * Rephrase text around algorithmic alternatives, add HyStart++ (#85
1299 (https://github.com/NTAP/rfc8312bis/issues/85), #86
1300 (https://github.com/NTAP/rfc8312bis/issues/86), #90
1301 (https://github.com/NTAP/rfc8312bis/issues/90))
1303 * Clarify what we mean by "new ACK" and use it in the text in more
1304 places. (#101 (https://github.com/NTAP/rfc8312bis/issues/101))
1306 * Rewrite the Responses to Sudden or Transient Events section (#98
1307 (https://github.com/NTAP/rfc8312bis/issues/98))
1309 * Remove confusing text about _cwnd_start_ in Section 4.2 (#100
1310 (https://github.com/NTAP/rfc8312bis/issues/100))
1312 * Change terminology from "AIMD" to "Reno" (#108
1313 (https://github.com/NTAP/rfc8312bis/issues/108))
1315 * Moved MUST NOT from app-limited section to main cubic AI section
1316 (#97 (https://github.com/NTAP/rfc8312bis/issues/97))
1318 * Clarify cwnd decrease during multiplicative decrease (#102
1319 (https://github.com/NTAP/rfc8312bis/issues/102))
1321 * Clarify text around queuing and slow adaptation of CUBIC in
1322 wireless environments (#94 (https://github.com/NTAP/rfc8312bis/
1323 issues/94))
1325 * Set lower bound of cwnd to 1 MSS and use retransmit timer
1326 thereafter (#83 (https://github.com/NTAP/rfc8312bis/issues/83))
1328 * Use FlightSize instead of cwnd to update ssthresh (#114
1329 (https://github.com/NTAP/rfc8312bis/issues/114))
1331 B.2. Since draft-ietf-tcpm-rfc8312bis-03
1333 * Remove reference from abstract (#82
1334 (https://github.com/NTAP/rfc8312bis/pull/82))
1336 B.3. Since draft-ietf-tcpm-rfc8312bis-02
1338 * Description of packet loss rate _p_ (#65
1339 (https://github.com/NTAP/rfc8312bis/issues/65))
1341 * Clarification of TCP Friendly Equation for ABC and Delayed ACK
1342 (#66 (https://github.com/NTAP/rfc8312bis/issues/66))
1344 * add applicability to QUIC and SCTP (#61
1345 (https://github.com/NTAP/rfc8312bis/issues/61))
1347 * clarity on setting alpha__aimd_ to 1 (#68
1348 (https://github.com/NTAP/rfc8312bis/issues/68))
1350 * introduce alpha__cubic_ (#64 (https://github.com/NTAP/rfc8312bis/
1351 issues/64))
1353 * clarify _cwnd_ growth in convex region (#69
1354 (https://github.com/NTAP/rfc8312bis/issues/69))
1356 * add guidance for using bytes and mention that segments count is
1357 decimal (#67 (https://github.com/NTAP/rfc8312bis/issues/67))
1359 * add loss events detected by RACK and QUIC loss detection (#62
1360 (https://github.com/NTAP/rfc8312bis/issues/62))
1362 B.4. Since draft-ietf-tcpm-rfc8312bis-01
1364 * address Michael Scharf's editorial suggestions. (#59
1365 (https://github.com/NTAP/rfc8312bis/issues/59))
1367 * add "Note to the RFC Editor" about removing underscores
1369 B.5. Since draft-ietf-tcpm-rfc8312bis-00
1371 * use updated xml2rfc with better text rendering of subscripts
1373 B.6. Since draft-eggert-tcpm-rfc8312bis-03
1375 * fix spelling nits
1377 * rename to draft-ietf
1379 * define _W_max_ more clearly
1381 B.7. Since draft-eggert-tcpm-rfc8312bis-02
1383 * add definition for segments_acked and alpha__aimd_. (#47
1384 (https://github.com/NTAP/rfc8312bis/issues/47))
1386 * fix a mistake in _W_max_ calculation in the fast convergence
1387 section. (#51 (https://github.com/NTAP/rfc8312bis/issues/51))
1389 * clarity on setting _ssthresh_ and _cwnd_start_ during
1390 multiplicative decrease. (#53 (https://github.com/NTAP/rfc8312bis/
1391 issues/53))
1393 B.8. Since draft-eggert-tcpm-rfc8312bis-01
1395 * rename TCP-Friendly to AIMD-Friendly and rename Standard TCP to
1396 AIMD TCP to avoid confusion as CUBIC has been widely used on the
1397 Internet. (#38 (https://github.com/NTAP/rfc8312bis/issues/38))
1399 * change introductory text to reflect the significant broader
1400 deployment of CUBIC on the Internet. (#39
1401 (https://github.com/NTAP/rfc8312bis/issues/39))
1403 * rephrase introduction to avoid referring to variables that have
1404 not been defined yet.
1406 B.9. Since draft-eggert-tcpm-rfc8312bis-00
1408 * acknowledge former co-authors (#15
1409 (https://github.com/NTAP/rfc8312bis/issues/15))
1411 * prevent _cwnd_ from becoming less than two (#7
1412 (https://github.com/NTAP/rfc8312bis/issues/7))
1414 * add list of variables and constants (#5
1415 (https://github.com/NTAP/rfc8312bis/issues/5), #6
1416 (https://github.com/NTAP/rfc8312bis/issues/6))
1418 * update _K_'s definition and add bounds for CUBIC _target_ _cwnd_
1419 [SXEZ19] (#1 (https://github.com/NTAP/rfc8312bis/issues/1), #14
1420 (https://github.com/NTAP/rfc8312bis/issues/14))
1422 * update _W_est_ to use AIMD approach (#20
1423 (https://github.com/NTAP/rfc8312bis/issues/20))
1425 * set alpha__aimd_ to 1 once _W_est_ reaches _W_max_ (#2
1426 (https://github.com/NTAP/rfc8312bis/issues/2))
1428 * add Vidhi as co-author (#17 (https://github.com/NTAP/rfc8312bis/
1429 issues/17))
1431 * note for Fast Recovery during _cwnd_ decrease due to congestion
1432 event (#11 (https://github.com/NTAP/rfc8312bis/issues/11))
1434 * add section for spurious congestion events (#23
1435 (https://github.com/NTAP/rfc8312bis/issues/23))
1437 * initialize _W_est_ after timeout and remove variable
1438 _W_(last_max)_ (#28 (https://github.com/NTAP/rfc8312bis/
1439 issues/28))
1441 B.10. Since RFC8312
1443 * converted to Markdown and xml2rfc v3
1445 * updated references (as part of the conversion)
1447 * updated author information
1449 * various formatting changes
1451 * move to Standards Track
1453 B.11. Since the Original Paper
1455 CUBIC has gone through a few changes since the initial release
1456 [HRX08] of its algorithm and implementation. Below we highlight the
1457 differences between its original paper and [RFC8312].
1459 * The original paper [HRX08] includes the pseudocode of CUBIC
1460 implementation using Linux's pluggable congestion control
1461 framework, which excludes system-specific optimizations. The
1462 simplified pseudocode might be a good source to start with and
1463 understand CUBIC.
1465 * [HRX08] also includes experimental results showing its performance
1466 and fairness.
1468 * The definition of beta__cubic_ constant was changed in [RFC8312].
1469 For example, beta__cubic_ in the original paper was the window
1470 decrease constant while [RFC8312] changed it to CUBIC
1471 multiplication decrease factor. With this change, the current
1472 congestion window size after a congestion event in [RFC8312] was
1473 beta__cubic_ * _W_max_ while it was (1-beta__cubic_) * _W_max_ in
1474 the original paper.
1476 * Its pseudocode used _W_(last_max)_ while [RFC8312] used _W_max_.
1478 * Its AIMD-friendly window was _W_tcp_ while [RFC8312] used _W_est_.
1480 Authors' Addresses
1482 Lisong Xu
1483 University of Nebraska-Lincoln
1484 Department of Computer Science and Engineering
1485 Lincoln, NE 68588-0115
1486 United States of America
1488 Email: xu@unl.edu
1489 URI: https://cse.unl.edu/~xu/
1491 Sangtae Ha
1492 University of Colorado at Boulder
1493 Department of Computer Science
1494 Boulder, CO 80309-0430
1495 United States of America
1497 Email: sangtae.ha@colorado.edu
1498 URI: https://netstech.org/sangtaeha/
1500 Injong Rhee
1501 Bowery Farming
1502 151 W 26TH Street, 12TH Floor
1503 New York, NY 10001
1504 United States of America
1506 Email: injongrhee@gmail.com
1507 Vidhi Goel
1508 Apple Inc.
1509 One Apple Park Way
1510 Cupertino, California 95014
1511 United States of America
1513 Email: vidhi_goel@apple.com
1515 Lars Eggert (editor)
1516 NetApp
1517 Stenbergintie 12 B
1518 FI-02700 Kauniainen
1519 Finland
1521 Email: lars@eggert.org
1522 URI: https://eggert.org/