idnits 2.17.00 (12 Aug 2021)
/tmp/idnits49725/draft-ietf-tcpm-rfc8312bis-07.txt:
-(512): Line appears to be too long, but this could be caused by non-ascii characters in UTF-8 encoding
-(514): Line appears to be too long, but this could be caused by non-ascii characters in UTF-8 encoding
-(618): Line appears to be too long, but this could be caused by non-ascii characters in UTF-8 encoding
-(707): 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:
----------------------------------------------------------------------------
== Line 559 has weird spacing: '... est est...'
== Line 772 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 (4 March 2022) is 71 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)
No issues found here.
Summary: 1 error (**), 0 flaws (~~), 3 warnings (==), 4 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: 5 September 2022 Bowery
8 V. Goel
9 Apple Inc.
10 L. Eggert, Ed.
11 NetApp
12 4 March 2022
14 CUBIC for Fast and Long-Distance Networks
15 draft-ietf-tcpm-rfc8312bis-07
17 Abstract
19 CUBIC is a standard TCP congestion control algorithm that uses a
20 cubic function instead of a linear congestion window increase
21 function 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 About This Document
34 This note is to be removed before publishing as an RFC.
36 Status information for this document may be found at
37 https://datatracker.ietf.org/doc/draft-ietf-tcpm-rfc8312bis/.
39 Discussion of this document takes place on the TCPM Working Group
40 mailing list (mailto:tcpm@ietf.org), which is archived at
41 https://mailarchive.ietf.org/arch/browse/tcpm/.
43 Source for this draft and an issue tracker can be found at
44 https://github.com/NTAP/rfc8312bis.
46 Note to the RFC Editor
47 xml2rfc currently renders in the XML by surrounding the
48 corresponding text with underscores. This is highly distracting;
49 please manually remove the underscores when doing the final edits to
50 the text version of this document.
52 (There is an issue open against xml2rfc to stop doing this in the
53 future: https://trac.tools.ietf.org/tools/xml2rfc/trac/ticket/596)
55 Also, please manually change "Figure" to "Equation" for all artwork
56 with anchors beginning with "eq" - xml2rfc doesn't seem to be able to
57 do this.
59 Status of This Memo
61 This Internet-Draft is submitted in full conformance with the
62 provisions of BCP 78 and BCP 79.
64 Internet-Drafts are working documents of the Internet Engineering
65 Task Force (IETF). Note that other groups may also distribute
66 working documents as Internet-Drafts. The list of current Internet-
67 Drafts is at https://datatracker.ietf.org/drafts/current/.
69 Internet-Drafts are draft documents valid for a maximum of six months
70 and may be updated, replaced, or obsoleted by other documents at any
71 time. It is inappropriate to use Internet-Drafts as reference
72 material or to cite them other than as "work in progress."
74 This Internet-Draft will expire on 5 September 2022.
76 Copyright Notice
78 Copyright (c) 2022 IETF Trust and the persons identified as the
79 document authors. All rights reserved.
81 This document is subject to BCP 78 and the IETF Trust's Legal
82 Provisions Relating to IETF Documents (https://trustee.ietf.org/
83 license-info) in effect on the date of publication of this document.
84 Please review these documents carefully, as they describe your rights
85 and restrictions with respect to this document. Code Components
86 extracted from this document must include Revised BSD License text as
87 described in Section 4.e of the Trust Legal Provisions and are
88 provided without warranty as described in the Revised BSD License.
90 Table of Contents
92 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 4
93 2. Conventions . . . . . . . . . . . . . . . . . . . . . . . . . 5
94 3. Design Principles of CUBIC . . . . . . . . . . . . . . . . . 5
95 3.1. Principle 1 for the CUBIC Increase Function . . . . . . . 6
96 3.2. Principle 2 for Reno-Friendliness . . . . . . . . . . . . 6
97 3.3. Principle 3 for RTT Fairness . . . . . . . . . . . . . . 7
98 3.4. Principle 4 for the CUBIC Decrease Factor . . . . . . . . 7
99 4. CUBIC Congestion Control . . . . . . . . . . . . . . . . . . 8
100 4.1. Definitions . . . . . . . . . . . . . . . . . . . . . . . 8
101 4.1.1. Constants of Interest . . . . . . . . . . . . . . . . 8
102 4.1.2. Variables of Interest . . . . . . . . . . . . . . . . 8
103 4.2. Window Increase Function . . . . . . . . . . . . . . . . 9
104 4.3. Reno-Friendly Region . . . . . . . . . . . . . . . . . . 11
105 4.4. Concave Region . . . . . . . . . . . . . . . . . . . . . 13
106 4.5. Convex Region . . . . . . . . . . . . . . . . . . . . . . 13
107 4.6. Multiplicative Decrease . . . . . . . . . . . . . . . . . 14
108 4.7. Fast Convergence . . . . . . . . . . . . . . . . . . . . 15
109 4.8. Timeout . . . . . . . . . . . . . . . . . . . . . . . . . 16
110 4.9. Spurious Congestion Events . . . . . . . . . . . . . . . 16
111 4.9.1. Spurious timeout . . . . . . . . . . . . . . . . . . 16
112 4.9.2. Spurious loss detected by acknowledgements . . . . . 17
113 4.10. Slow Start . . . . . . . . . . . . . . . . . . . . . . . 18
114 5. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . 18
115 5.1. Fairness to Reno . . . . . . . . . . . . . . . . . . . . 19
116 5.2. Using Spare Capacity . . . . . . . . . . . . . . . . . . 21
117 5.3. Difficult Environments . . . . . . . . . . . . . . . . . 22
118 5.4. Investigating a Range of Environments . . . . . . . . . . 22
119 5.5. Protection against Congestion Collapse . . . . . . . . . 23
120 5.6. Fairness within the Alternative Congestion Control
121 Algorithm . . . . . . . . . . . . . . . . . . . . . . . 23
122 5.7. Performance with Misbehaving Nodes and Outside
123 Attackers . . . . . . . . . . . . . . . . . . . . . . . 23
124 5.8. Behavior for Application-Limited Flows . . . . . . . . . 23
125 5.9. Responses to Sudden or Transient Events . . . . . . . . . 24
126 5.10. Incremental Deployment . . . . . . . . . . . . . . . . . 24
127 6. Security Considerations . . . . . . . . . . . . . . . . . . . 24
128 7. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 24
129 8. References . . . . . . . . . . . . . . . . . . . . . . . . . 24
130 8.1. Normative References . . . . . . . . . . . . . . . . . . 24
131 8.2. Informative References . . . . . . . . . . . . . . . . . 26
132 Appendix A. Acknowledgments . . . . . . . . . . . . . . . . . . 29
133 Appendix B. Evolution of CUBIC . . . . . . . . . . . . . . . . . 30
134 B.1. Since draft-ietf-tcpm-rfc8312bis-06 . . . . . . . . . . . 30
135 B.2. Since draft-ietf-tcpm-rfc8312bis-05 . . . . . . . . . . . 30
136 B.3. Since draft-ietf-tcpm-rfc8312bis-04 . . . . . . . . . . . 30
137 B.4. Since draft-ietf-tcpm-rfc8312bis-03 . . . . . . . . . . . 31
138 B.5. Since draft-ietf-tcpm-rfc8312bis-02 . . . . . . . . . . . 31
139 B.6. Since draft-ietf-tcpm-rfc8312bis-01 . . . . . . . . . . . 32
140 B.7. Since draft-ietf-tcpm-rfc8312bis-00 . . . . . . . . . . . 32
141 B.8. Since draft-eggert-tcpm-rfc8312bis-03 . . . . . . . . . . 32
142 B.9. Since draft-eggert-tcpm-rfc8312bis-02 . . . . . . . . . . 32
143 B.10. Since draft-eggert-tcpm-rfc8312bis-01 . . . . . . . . . . 32
144 B.11. Since draft-eggert-tcpm-rfc8312bis-00 . . . . . . . . . . 33
145 B.12. Since RFC8312 . . . . . . . . . . . . . . . . . . . . . . 33
146 B.13. Since the Original Paper . . . . . . . . . . . . . . . . 34
147 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 34
149 1. Introduction
151 CUBIC has been adopted as the default TCP congestion control
152 algorithm in the Linux, Windows, and Apple stacks, and has been used
153 and deployed globally. Extensive, decade-long deployment experience
154 in vastly different Internet scenarios has convincingly demonstrated
155 that CUBIC is safe for deployment on the global Internet and delivers
156 substantial benefits over classical Reno congestion control
157 [RFC5681]. It is therefore to be regarded as the currently most
158 widely deployed standard for TCP congestion control. CUBIC can also
159 be used for other transport protocols such as QUIC [RFC9000] and SCTP
160 [RFC4960] as a default congestion controller.
162 The design of CUBIC was motivated by the well-documented problem
163 classical Reno TCP has with low utilization over fast and long-
164 distance networks [K03][RFC3649]. This problem arises from a slow
165 increase of the congestion window following a congestion event in a
166 network with a large bandwidth-delay product (BDP). [HLRX07]
167 indicates that this problem is frequently observed even in the range
168 of congestion window sizes over several hundreds of packets. This
169 problem is equally applicable to all Reno-style standards and their
170 variants, including TCP-Reno [RFC5681], TCP-NewReno
171 [RFC6582][RFC6675], SCTP [RFC4960], TFRC [RFC5348], and QUIC
172 congestion control [RFC9002], which use the same linear increase
173 function for window growth. We refer to all Reno-style standards and
174 their variants collectively as "Reno" below.
176 CUBIC, originally proposed in [HRX08], is a modification to the
177 congestion control algorithm of classical Reno to remedy this
178 problem. Specifically, CUBIC uses a cubic function instead of the
179 linear window increase function of Reno to improve scalability and
180 stability under fast and long-distance networks.
182 This document updates the specification of CUBIC to include
183 algorithmic improvements based on the Linux, Windows, and Apple
184 implementations and recent academic work. Based on the extensive
185 deployment experience with CUBIC, it also moves the specification to
186 the Standards Track, obsoleting [RFC8312]. This requires an update
187 to [RFC5681], which limits the aggressiveness of Reno TCP
188 implementations in its Section 3. Since CUBIC is occasionally more
189 aggressive than the [RFC5681] algorithms, this document updates
190 [RFC5681] to allow for CUBIC's behavior.
192 Binary Increase Congestion Control (BIC-TCP) [XHR04], a predecessor
193 of CUBIC, was selected as the default TCP congestion control
194 algorithm by Linux in the year 2005 and had been used for several
195 years by the Internet community at large.
197 CUBIC uses a similar window increase function as BIC-TCP and is
198 designed to be less aggressive and fairer to Reno in bandwidth usage
199 than BIC-TCP while maintaining the strengths of BIC-TCP such as
200 stability, window scalability, and round-trip time (RTT) fairness.
202 In the following sections, we first briefly explain the design
203 principles of CUBIC, then provide the exact specification of CUBIC,
204 and finally discuss the safety features of CUBIC following the
205 guidelines specified in [RFC5033].
207 2. Conventions
209 The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
210 "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and
211 "OPTIONAL" in this document are to be interpreted as described in
212 BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all
213 capitals, as shown here.
215 3. Design Principles of CUBIC
217 CUBIC is designed according to the following design principles:
219 Principle 1: For better network utilization and stability, CUBIC
220 uses both the concave and convex profiles of a cubic function to
221 increase the congestion window size, instead of using just a
222 convex function.
224 Principle 2: To be Reno-friendly, CUBIC is designed to behave like
225 Reno in networks with short RTTs and small bandwidth where Reno
226 performs well.
228 Principle 3: For RTT-fairness, CUBIC is designed to achieve linear
229 bandwidth sharing among flows with different RTTs.
231 Principle 4: CUBIC appropriately sets its multiplicative window
232 decrease factor in order to balance between the scalability and
233 convergence speed.
235 3.1. Principle 1 for the CUBIC Increase Function
237 For better network utilization and stability, CUBIC [HRX08] uses a
238 cubic window increase function in terms of the elapsed time from the
239 last congestion event. While most alternative congestion control
240 algorithms to Reno increase the congestion window using convex
241 functions, CUBIC uses both the concave and convex profiles of a cubic
242 function for window growth.
244 After a window reduction in response to a congestion event detected
245 by duplicate ACKs, Explicit Congestion Notification-Echo (ECN-Echo,
246 ECE) ACKs [RFC3168], TCP RACK [RFC8985] or QUIC loss detection
247 [RFC9002], CUBIC remembers the congestion window size at which it
248 received the congestion event and performs a multiplicative decrease
249 of the congestion window. When CUBIC enters into congestion
250 avoidance, it starts to increase the congestion window using the
251 concave profile of the cubic function. The cubic function is set to
252 have its plateau at the remembered congestion window size, so that
253 the concave window increase continues until then. After that, the
254 cubic function turns into a convex profile and the convex window
255 increase begins.
257 This style of window adjustment (concave and then convex) improves
258 the algorithm stability while maintaining high network utilization
259 [CEHRX09]. This is because the window size remains almost constant,
260 forming a plateau around the remembered congestion window size of the
261 last congestion event, where network utilization is deemed highest.
262 Under steady state, most window size samples of CUBIC are close to
263 that remembered congestion window size, thus promoting high network
264 utilization and stability.
266 Note that congestion control algorithms that only use convex
267 functions to increase the congestion window size have their maximum
268 increments around the remembered congestion window size of the last
269 congestion event, and thus introduce many packet bursts around the
270 saturation point of the network, likely causing frequent global loss
271 synchronizations.
273 3.2. Principle 2 for Reno-Friendliness
275 CUBIC promotes per-flow fairness to Reno. Note that Reno performs
276 well over paths with short RTTs and small bandwidths (or small BDPs).
277 There is only a scalability problem in networks with long RTTs and
278 large bandwidths (or large BDPs).
280 A congestion control algorithm designed to be friendly to Reno on a
281 per-flow basis must increase its congestion window less aggressively
282 in small BDP networks than in large BDP networks.
284 The aggressiveness of CUBIC mainly depends on the maximum window size
285 before a window reduction, which is smaller in small-BDP networks
286 than in large-BDP networks. Thus, CUBIC increases its congestion
287 window less aggressively in small-BDP networks than in large-BDP
288 networks.
290 Furthermore, in cases when the cubic function of CUBIC would increase
291 the congestion window less aggressively than Reno, CUBIC simply
292 follows the window size of Reno to ensure that CUBIC achieves at
293 least the same throughput as Reno in small-BDP networks. We call
294 this region where CUBIC behaves like Reno the "Reno-friendly region".
296 3.3. Principle 3 for RTT Fairness
298 Two CUBIC flows with different RTTs have a throughput ratio that is
299 linearly proportional to the inverse of their RTT ratio, where the
300 throughput of a flow is approximately the size of its congestion
301 window divided by its RTT.
303 Specifically, CUBIC maintains a window increase rate independent of
304 RTTs outside the Reno-friendly region, and thus flows with different
305 RTTs have similar congestion window sizes under steady state when
306 they operate outside the Reno-friendly region.
308 This notion of a linear throughput ratio is similar to that of Reno
309 under high statistical multiplexing where packet loss is independent
310 of individual flow rates. However, under low statistical
311 multiplexing, the throughput ratio of Reno flows with different RTTs
312 is quadratically proportional to the inverse of their RTT ratio
313 [XHR04].
315 CUBIC always ensures a linear throughput ratio independent of the
316 amount of statistical multiplexing. This is an improvement over
317 Reno. While there is no consensus on particular throughput ratios
318 for different RTT flows, we believe that over wired Internet paths,
319 use of a linear throughput ratio seems more reasonable than equal
320 throughputs (i.e., the same throughput for flows with different RTTs)
321 or a higher-order throughput ratio (e.g., a quadratical throughput
322 ratio of Reno under low statistical multiplexing environments).
324 3.4. Principle 4 for the CUBIC Decrease Factor
326 To balance between scalability and convergence speed, CUBIC sets the
327 multiplicative window decrease factor to 0.7, whereas Reno uses 0.5.
329 While this improves the scalability of CUBIC, a side effect of this
330 decision is slower convergence, especially under low statistical
331 multiplexing. This design choice is following the observation that
332 HighSpeed TCP (HSTCP) [RFC3649] and other approaches (e.g., [GV02])
333 made: the current Internet becomes more asynchronous with less
334 frequent loss synchronizations under high statistical multiplexing.
336 In such environments, even strict Multiplicative-Increase
337 Multiplicative-Decrease (MIMD) can converge. CUBIC flows with the
338 same RTT always converge to the same throughput independent of
339 statistical multiplexing, thus achieving intra-algorithm fairness.
340 We also find that in environments with sufficient statistical
341 multiplexing, the convergence speed of CUBIC is reasonable.
343 4. CUBIC Congestion Control
345 In this section, we discuss how the congestion window is updated
346 during the different stages of the CUBIC congestion controller.
348 4.1. Definitions
350 The unit of all window sizes in this document is segments of the
351 maximum segment size (MSS), and the unit of all times is seconds.
352 Implementations can use bytes to express window sizes, which would
353 require factoring in the maximum segment size wherever necessary and
354 replacing _segments_acked_ with the number of bytes acknowledged in
355 Figure 4.
357 4.1.1. Constants of Interest
359 β__cubic_: CUBIC multiplicative decrease factor as described in
360 Section 4.6.
362 α__cubic_: CUBIC additive increase factor used in Reno-friendly
363 region as described in Section 4.3.
365 _C_: constant that determines the aggressiveness of CUBIC in
366 competing with other congestion control algorithms in high BDP
367 networks. Please see Section 5 for more explanation on how it is
368 set. The unit for _C_ is
370 segment
371 -------
372 3
373 second
375 4.1.2. Variables of Interest
377 This section defines the variables required to implement CUBIC:
379 _RTT_: Smoothed round-trip time in seconds, calculated as described
380 in [RFC6298].
382 _cwnd_: Current congestion window in segments.
384 _ssthresh_: Current slow start threshold in segments.
386 _W_max_: Size of _cwnd_ in segments just before _cwnd_ was reduced in
387 the last congestion event when fast convergence is disabled.
388 However, if fast convergence is enabled, the size may be further
389 reduced based on the current saturation point.
391 _K_: The time period in seconds it takes to increase the congestion
392 window size at the beginning of the current congestion avoidance
393 stage to _W_max_.
395 _current_time_: Current time of the system in seconds.
397 _epoch_start_: The time in seconds at which the current congestion
398 avoidance stage started.
400 _cwnd_start_: The _cwnd_ at the beginning of the current congestion
401 avoidance stage, i.e., at time _epoch_start_.
403 W_cubic(_t_): The congestion window in segments at time _t_ in
404 seconds based on the cubic increase function, as described in
405 Section 4.2.
407 _target_: Target value of congestion window in segments after the
408 next RTT, that is, W_cubic(_t_ + _RTT_), as described in Section 4.2.
410 _W_est_: An estimate for the congestion window in segments in the
411 Reno-friendly region, that is, an estimate for the congestion window
412 of Reno.
414 _segments_acked_: Number of MSS-sized segments acked when a "new ACK"
415 is received, i.e., an ACK that cumulatively acknowledges the delivery
416 of new data. This number will be a decimal value when a new ACK
417 acknowledges an amount of data that is not MSS-sized. Specifically,
418 it can be less than 1 when a new ACK acknowledges a segment smaller
419 than the MSS.
421 4.2. Window Increase Function
423 CUBIC maintains the acknowledgment (ACK) clocking of Reno by
424 increasing the congestion window only at the reception of a new ACK.
425 It does not make any changes to the TCP Fast Recovery and Fast
426 Retransmit algorithms [RFC6582][RFC6675].
428 During congestion avoidance, after a congestion event is detected by
429 mechanisms described in Section 3.1, CUBIC uses a window increase
430 function different from Reno.
432 CUBIC uses the following window increase function:
434 3
435 W (t) = C * (t - K) + W
436 cubic max
438 Figure 1
440 where _t_ is the elapsed time in seconds from the beginning of the
441 current congestion avoidance stage, that is,
443 t = current_time - epoch
444 start
446 and where _epoch_start_ is the time at which the current congestion
447 avoidance stage starts. _K_ is the time period that the above
448 function takes to increase the congestion window size at the
449 beginning of the current congestion avoidance stage to _W_max_ if
450 there are no further congestion events and is calculated using the
451 following equation:
453 ________________
454 /W - cwnd
455 3 / max start
456 K = | / ----------------
457 |/ C
459 Figure 2
461 where _cwnd_start_ is the congestion window at the beginning of the
462 current congestion avoidance stage.
464 Upon receiving a new ACK during congestion avoidance, CUBIC computes
465 the _target_ congestion window size after the next _RTT_ using
466 Figure 1 as follows, where _RTT_ is the smoothed round-trip time.
467 The lower and upper bounds below ensure that CUBIC's congestion
468 window increase rate is non-decreasing and is less than the increase
469 rate of slow start [SXEZ19].
471 /
472 | if W (t + RTT) < cwnd
473 |cwnd cubic
474 |
475 |
476 |
477 target = < if W (t + RTT) > 1.5 * cwnd
478 |1.5 * cwnd cubic
479 |
480 |
481 |W (t + RTT)
482 | cubic otherwise
483 \
485 The elapsed time _t_ in Figure 1 MUST NOT include periods during
486 which _cwnd_ has not been updated due to application-limited behavior
487 (see Section 5.8).
489 Depending on the value of the current congestion window size _cwnd_,
490 CUBIC runs in three different regions:
492 1. The Reno-friendly region, which ensures that CUBIC achieves at
493 least the same throughput as Reno.
495 2. The concave region, if CUBIC is not in the Reno-friendly region
496 and _cwnd_ is less than _W_max_.
498 3. The convex region, if CUBIC is not in the Reno-friendly region
499 and _cwnd_ is greater than _W_max_.
501 Below, we describe the exact actions taken by CUBIC in each region.
503 4.3. Reno-Friendly Region
505 Reno performs well in certain types of networks, for example, under
506 short RTTs and small bandwidths (or small BDPs). In these networks,
507 CUBIC remains in the Reno-friendly region to achieve at least the
508 same throughput as Reno.
510 The Reno-friendly region is designed according to the analysis in
511 [FHP00], which studies the performance of an AIMD algorithm with an
512 additive factor of α (segments per _RTT_) and a multiplicative factor
513 of β, denoted by AIMD(α, β). _p_ is the packet loss rate.
514 Specifically, the average congestion window size of AIMD(α, β) can be
515 calculated using Figure 3.
517 _______________
518 / α * (1 + β)
519 AVG_AIMD(α, β) = | / ---------------
520 |/ 2 * (1 - β) * p
522 Figure 3
524 By the same analysis, to achieve the same average window size as Reno
525 that uses AIMD(1, 0.5), α must be equal to,
527 1 - β
528 3 * -----
529 1 + β
531 Thus, CUBIC uses Figure 4 to estimate the window size _W_est_ in the
532 Reno-friendly region with
534 1 - β
535 cubic
536 α = 3 * ----------
537 cubic 1 + β
538 cubic
540 which achieves the same average window size as Reno. When receiving
541 a new ACK in congestion avoidance (where _cwnd_ could be greater than
542 or less than _W_max_), CUBIC checks whether W_cubic(_t_) is less than
543 _W_est_. If so, CUBIC is in the Reno-friendly region and _cwnd_
544 SHOULD be set to _W_est_ at each reception of a new ACK.
546 _W_est_ is set equal to _cwnd_start_ at the start of the congestion
547 avoidance stage. After that, on every new ACK, _W_est_ is updated
548 using Figure 4. Note that this equation is for a connection where
549 Appropriate Byte Counting (ABC) [RFC3465] is disabled. For a
550 connection with ABC enabled, this equation SHOULD be adjusted by
551 using the number of acknowledged bytes instead of acknowledged
552 segments. Also note that this equation works for connections with
553 enabled or disabled Delayed ACKs [RFC5681], as _segments_acked_ will
554 be different based on the segments actually acknowledged by a new
555 ACK.
557 segments_acked
558 W = W + α * --------------
559 est est cubic cwnd
561 Figure 4
563 Note that once _W_est_ reaches _W_max_, that is, _W_est_ >= _W_max_,
564 CUBIC needs to start probing to determine the new value of _W_max_.
565 At this point, α__cubic_ SHOULD be set to 1 to ensure that CUBIC can
566 achieve the same congestion window increment as Reno, which uses
567 AIMD(1, 0.5).
569 4.4. Concave Region
571 When receiving a new ACK in congestion avoidance, if CUBIC is not in
572 the Reno-friendly region and _cwnd_ is less than _W_max_, then CUBIC
573 is in the concave region. In this region, _cwnd_ MUST be incremented
574 by
576 target - cwnd
577 -------------
578 cwnd
580 for each received new ACK, where _target_ is calculated as described
581 in Section 4.2.
583 4.5. Convex Region
585 When receiving a new ACK in congestion avoidance, if CUBIC is not in
586 the Reno-friendly region and _cwnd_ is larger than or equal to
587 _W_max_, then CUBIC is in the convex region.
589 The convex region indicates that the network conditions might have
590 changed since the last congestion event, possibly implying more
591 available bandwidth after some flow departures. Since the Internet
592 is highly asynchronous, some amount of perturbation is always
593 possible without causing a major change in available bandwidth.
595 Unless it is overridden by the AIMD window increase, CUBIC is very
596 careful in this region. The convex profile aims to increase the
597 window very slowly at the beginning when _cwnd_ is around _W_max_ and
598 then gradually increases its rate of increase. We also call this
599 region the "maximum probing phase", since CUBIC is searching for a
600 new _W_max_. In this region, _cwnd_ MUST be incremented by
602 target - cwnd
603 -------------
604 cwnd
606 for each received new ACK, where _target_ is calculated as described
607 in Section 4.2.
609 4.6. Multiplicative Decrease
611 When a congestion event is detected by mechanisms described in
612 Section 3.1, CUBIC updates _W_max_ and reduces _cwnd_ and _ssthresh_
613 immediately as described below. In case of packet loss, the sender
614 MUST reduce _cwnd_ and _ssthresh_ immediately upon entering loss
615 recovery, similar to [RFC5681] (and [RFC6675]). Note that other
616 mechanisms, such as Proportional Rate Reduction [RFC6937], can be
617 used to reduce the sending rate during loss recovery more gradually.
618 The parameter β__cubic_ SHOULD be set to 0.7, which is different from
619 the multiplicative decrease factor used in [RFC5681] (and [RFC6675])
620 during fast recovery.
622 In Figure 5, _flight_size_ is the amount of outstanding data in the
623 network, as defined in [RFC5681]. Note that a rate-limited
624 application with idle periods or periods when unable to send at the
625 full rate permitted by _cwnd_ may easily encounter notable variations
626 in the volume of data sent from one RTT to another, resulting in
627 _flight_size_ that is significantly less than _cwnd_ on a congestion
628 event. This may decrease _cwnd_ to a much lower value than
629 necessary. To avoid suboptimal performance with such applications,
630 the mechanisms described in [RFC7661] can be used to mitigate this
631 issue as it would allow using a value between _cwnd_ and
632 _flight_size_ to calculate the new _ssthresh_ in Figure 5. The
633 congestion window growth mechanism defined in [RFC7661] is safe to
634 use even when _cwnd_ is greater than the receive window as it
635 validates _cwnd_ based on the amount of data acknowledged by the
636 network in an RTT which implicitly accounts for the allowed receive
637 window. Some implementations of CUBIC currently use _cwnd_ instead
638 of _flight_size_ when calculating a new _ssthresh_ using Figure 5.
640 flight_size * β // new ssthresh
641 ssthresh = cubic
643 /max(ssthresh, 2) // reduction on packet loss, cwnd is at least 2 MSS
644 |
645 cwnd = <
646 |max(ssthresh, 1) // reduction on ECE, cwnd is at least 1 MSS
647 \
649 max(ssthresh, 2) // ssthresh is at least 2 MSS
650 ssthresh =
652 Figure 5
654 A side effect of setting β__cubic_ to a value bigger than 0.5 is
655 slower convergence. We believe that while a more adaptive setting of
656 β__cubic_ could result in faster convergence, it will make the
657 analysis of CUBIC much harder.
659 Note that CUBIC MUST continue to reduce _cwnd_ in response to
660 congestion events due to ECN-Echo ACKs until it reaches a value of 1
661 MSS. If congestion events indicated by ECN-Echo ACKs persist, a
662 sender with a _cwnd_ of 1 MSS MUST reduce its sending rate even
663 further. It can achieve that by using a retransmission timer with
664 exponential backoff, as described in [RFC3168].
666 4.7. Fast Convergence
668 To improve convergence speed, CUBIC uses a heuristic. When a new
669 flow joins the network, existing flows need to give up some of their
670 bandwidth to allow the new flow some room for growth, if the existing
671 flows have been using all the network bandwidth. To speed up this
672 bandwidth release by existing flows, the following "Fast Convergence"
673 mechanism SHOULD be implemented.
675 With Fast Convergence, when a congestion event occurs, we update
676 _W_max_ as follows, before the window reduction as described in
677 Section 4.6.
679 /
680 | 1 + β
681 | cubic if cwnd < W and fast convergence is enabled,
682 |cwnd * ---------- max
683 | 2
684 W = <
685 max | further reduce W
686 | max
687 |
688 | otherwise, remember cwnd before reduction
689 \cwnd
691 At a congestion event, if the current _cwnd_ is less than _W_max_,
692 this indicates that the saturation point experienced by this flow is
693 getting reduced because of a change in available bandwidth. Then we
694 allow this flow to release more bandwidth by reducing _W_max_
695 further. This action effectively lengthens the time for this flow to
696 increase its congestion window, because the reduced _W_max_ forces
697 the flow to plateau earlier. This allows more time for the new flow
698 to catch up to its congestion window size.
700 Fast Convergence is designed for network environments with multiple
701 CUBIC flows. In network environments with only a single CUBIC flow
702 and without any other traffic, Fast Convergence SHOULD be disabled.
704 4.8. Timeout
706 In case of a timeout, CUBIC follows Reno to reduce _cwnd_ [RFC5681],
707 but sets _ssthresh_ using β__cubic_ (same as in Section 4.6) in a way
708 that is different from Reno TCP [RFC5681].
710 During the first congestion avoidance stage after a timeout, CUBIC
711 increases its congestion window size using Figure 1, where _t_ is the
712 elapsed time since the beginning of the current congestion avoidance,
713 _K_ is set to 0, and _W_max_ is set to the congestion window size at
714 the beginning of the current congestion avoidance stage. In
715 addition, for the Reno-friendly region, _W_est_ SHOULD be set to the
716 congestion window size at the beginning of the current congestion
717 avoidance.
719 4.9. Spurious Congestion Events
721 In cases where CUBIC reduces its congestion window in response to
722 having detected packet loss via duplicate ACKs or timeouts, there is
723 a possibility that the missing ACK would arrive after the congestion
724 window reduction and a corresponding packet retransmission. For
725 example, packet reordering could trigger this behavior. A high
726 degree of packet reordering could cause multiple congestion window
727 reduction events, where spurious losses are incorrectly interpreted
728 as congestion signals, thus degrading CUBIC's performance
729 significantly.
731 For TCP, there are two types of spurious events - spurious timeouts
732 and spurious fast retransmits. In case of QUIC, there are no
733 spurious timeouts as the loss is only detected after receiving an
734 ACK.
736 4.9.1. Spurious timeout
738 An implementation MAY detect spurious timeouts based on the
739 mechanisms described in Forward RTO-Recovery [RFC5682]. Experimental
740 alternatives include Eifel [RFC3522]. When a spurious timeout is
741 detected, a TCP implementation MAY follow the response algorithm
742 described in [RFC4015] to restore the congestion control state and
743 adapt the retransmission timer to avoid further spurious timeouts.
745 4.9.2. Spurious loss detected by acknowledgements
747 Upon receiving an ACK, a TCP implementation MAY detect spurious
748 losses either using TCP Timestamps or via D-SACK[RFC2883].
749 Experimental alternatives include Eifel detection algorithm [RFC3522]
750 which uses TCP Timestamps and DSACK based detection [RFC3708] which
751 uses DSACK information. A QUIC implementation can easily determine a
752 spurious loss if a QUIC packet is acknowledged after it has been
753 marked as lost and the original data has been retransmitted with a
754 new QUIC packet.
756 In this section, we specify a simple response algorithm when a
757 spurious loss is detected by acknowledgements. Implementations would
758 need to carefully evaluate the impact of using this algorithm in
759 different environments that may experience sudden change in available
760 capacity (e.g., due to variable radio capacity, a routing change, or
761 a mobility event).
763 When a packet loss is detected via acknowledgements, a CUBIC
764 implementation MAY save the current value of the following variables
765 before the congestion window is reduced.
767 prior_cwnd = cwnd
769 prior_ssthresh = ssthresh
771 prior_W = W
772 max max
774 prior_K = K
776 prior_epoch = epoch
777 start start
779 prior_W_{est} = W
780 est
782 Once the previously declared packet loss is confirmed to be spurious,
783 CUBIC MAY restore the original values of the above-mentioned
784 variables as follows if the current _cwnd_ is lower than
785 _prior_cwnd_. Restoring the original values ensures that CUBIC's
786 performance is similar to what it would be without spurious losses.
788 \
789 cwnd = prior_cwnd |
790 |
791 ssthresh = prior_ssthresh |
792 |
793 W = prior_W |
794 max max |
795 >if cwnd < prior_cwnd
796 K = prior_K |
797 |
798 epoch = prior_epoch |
799 start start|
800 |
801 W = prior_W |
802 est est /
804 In rare cases, when the detection happens long after a spurious loss
805 event and the current _cwnd_ is already higher than _prior_cwnd_,
806 CUBIC SHOULD continue to use the current and the most recent values
807 of these variables.
809 4.10. Slow Start
811 CUBIC MUST employ a slow-start algorithm, when _cwnd_ is no more than
812 _ssthresh_. In general, CUBIC SHOULD use the HyStart++ slow start
813 algorithm [I-D.ietf-tcpm-hystartplusplus], or MAY use the Reno TCP
814 slow start algorithm [RFC5681] in the rare cases when HyStart++ is
815 not suitable. Experimental alternatives include hybrid slow start
816 [HR11], a predecessor to HyStart++ that some CUBIC implementations
817 have used as the default for the last decade, and limited slow start
818 [RFC3742]. Whichever start-up algorithm is used, work might be
819 needed to ensure that the end of slow start and the first
820 multiplicative decrease of congestion avoidance work well together.
822 When CUBIC uses HyStart++ [I-D.ietf-tcpm-hystartplusplus], it may
823 exit the first slow start without incurring any packet loss and thus
824 _W_max_ is undefined. In this special case, CUBIC switches to
825 congestion avoidance and increases its congestion window size using
826 Figure 1, where _t_ is the elapsed time since the beginning of the
827 current congestion avoidance, _K_ is set to 0, and _W_max_ is set to
828 the congestion window size at the beginning of the current congestion
829 avoidance stage.
831 5. Discussion
833 In this section, we further discuss the safety features of CUBIC
834 following the guidelines specified in [RFC5033].
836 With a deterministic loss model where the number of packets between
837 two successive packet losses is always _1/p_, CUBIC always operates
838 with the concave window profile, which greatly simplifies the
839 performance analysis of CUBIC. The average window size of CUBIC can
840 be obtained by the following function:
842 ________________ ____
843 /C * (3 + β ) 4 / 3
844 4 / cubic |/ RTT
845 AVG_W = | / ---------------- * -------
846 cubic | / 4 * (1 - β ) __
847 |/ cubic 4 / 3
848 |/ p
850 Figure 6
852 With β__cubic_ set to 0.7, the above formula reduces to:
854 ____
855 _______ 4 / 3
856 4 /C * 3.7 |/ RTT
857 AVG_W = | / ------- * -------
858 cubic |/ 1.2 __
859 4 / 3
860 |/ p
862 Figure 7
864 We will determine the value of _C_ in the following subsection using
865 Figure 7.
867 5.1. Fairness to Reno
869 In environments where Reno is able to make reasonable use of the
870 available bandwidth, CUBIC does not significantly change this state.
872 Reno performs well in the following two types of networks:
874 1. networks with a small bandwidth-delay product (BDP)
876 2. networks with a short RTTs, but not necessarily a small BDP
878 CUBIC is designed to behave very similarly to Reno in the above two
879 types of networks. The following two tables show the average window
880 sizes of Reno TCP, HSTCP, and CUBIC TCP. The average window sizes of
881 Reno TCP and HSTCP are from [RFC3649]. The average window size of
882 CUBIC is calculated using Figure 7 and the CUBIC Reno-friendly region
883 for three different values of _C_.
885 +=============+=======+========+================+=========+========+
886 | Loss Rate P | Reno | HSTCP | CUBIC (C=0.04) | CUBIC | CUBIC |
887 | | | | | (C=0.4) | (C=4) |
888 +=============+=======+========+================+=========+========+
889 | 1.0e-02 | 12 | 12 | 12 | 12 | 12 |
890 +-------------+-------+--------+----------------+---------+--------+
891 | 1.0e-03 | 38 | 38 | 38 | 38 | 59 |
892 +-------------+-------+--------+----------------+---------+--------+
893 | 1.0e-04 | 120 | 263 | 120 | 187 | 333 |
894 +-------------+-------+--------+----------------+---------+--------+
895 | 1.0e-05 | 379 | 1795 | 593 | 1054 | 1874 |
896 +-------------+-------+--------+----------------+---------+--------+
897 | 1.0e-06 | 1200 | 12280 | 3332 | 5926 | 10538 |
898 +-------------+-------+--------+----------------+---------+--------+
899 | 1.0e-07 | 3795 | 83981 | 18740 | 33325 | 59261 |
900 +-------------+-------+--------+----------------+---------+--------+
901 | 1.0e-08 | 12000 | 574356 | 105383 | 187400 | 333250 |
902 +-------------+-------+--------+----------------+---------+--------+
904 Table 1: Reno TCP, HSTCP, and CUBIC with RTT = 0.1 seconds
906 Table 1 describes the response function of Reno TCP, HSTCP, and CUBIC
907 in networks with _RTT_ = 0.1 seconds. The average window size is in
908 MSS-sized segments.
910 +=============+=======+========+================+=========+=======+
911 | Loss Rate P | Reno | HSTCP | CUBIC (C=0.04) | CUBIC | CUBIC |
912 | | | | | (C=0.4) | (C=4) |
913 +=============+=======+========+================+=========+=======+
914 | 1.0e-02 | 12 | 12 | 12 | 12 | 12 |
915 +-------------+-------+--------+----------------+---------+-------+
916 | 1.0e-03 | 38 | 38 | 38 | 38 | 38 |
917 +-------------+-------+--------+----------------+---------+-------+
918 | 1.0e-04 | 120 | 263 | 120 | 120 | 120 |
919 +-------------+-------+--------+----------------+---------+-------+
920 | 1.0e-05 | 379 | 1795 | 379 | 379 | 379 |
921 +-------------+-------+--------+----------------+---------+-------+
922 | 1.0e-06 | 1200 | 12280 | 1200 | 1200 | 1874 |
923 +-------------+-------+--------+----------------+---------+-------+
924 | 1.0e-07 | 3795 | 83981 | 3795 | 5926 | 10538 |
925 +-------------+-------+--------+----------------+---------+-------+
926 | 1.0e-08 | 12000 | 574356 | 18740 | 33325 | 59261 |
927 +-------------+-------+--------+----------------+---------+-------+
929 Table 2: Reno TCP, HSTCP, and CUBIC with RTT = 0.01 seconds
931 Table 2 describes the response function of Reno TCP, HSTCP, and CUBIC
932 in networks with _RTT_ = 0.01 seconds. The average window size is in
933 MSS-sized segments.
935 Both tables show that CUBIC with any of these three _C_ values is
936 more friendly to Reno TCP than HSTCP, especially in networks with a
937 short _RTT_ where Reno TCP performs reasonably well. For example, in
938 a network with _RTT_ = 0.01 seconds and p=10^-6, Reno TCP has an
939 average window of 1200 packets. If the packet size is 1500 bytes,
940 then Reno TCP can achieve an average rate of 1.44 Gbps. In this
941 case, CUBIC with _C_=0.04 or _C_=0.4 achieves exactly the same rate
942 as Reno TCP, whereas HSTCP is about ten times more aggressive than
943 Reno TCP.
945 We can see that _C_ determines the aggressiveness of CUBIC in
946 competing with other congestion control algorithms for bandwidth.
947 CUBIC is more friendly to Reno TCP, if the value of _C_ is lower.
948 However, we do not recommend setting _C_ to a very low value like
949 0.04, since CUBIC with a low _C_ cannot efficiently use the bandwidth
950 in fast and long-distance networks. Based on these observations and
951 extensive deployment experience, we find _C_=0.4 gives a good balance
952 between Reno-friendliness and aggressiveness of window increase.
953 Therefore, _C_ SHOULD be set to 0.4. With _C_ set to 0.4, Figure 7
954 is reduced to:
956 ____
957 4 / 3
958 |/ RTT
959 AVG_W = 1.054 * -------
960 cubic __
961 4 / 3
962 |/ p
964 Figure 8
966 Figure 8 is then used in the next subsection to show the scalability
967 of CUBIC.
969 5.2. Using Spare Capacity
971 CUBIC uses a more aggressive window increase function than Reno for
972 fast and long-distance networks.
974 The following table shows that to achieve the 10 Gbps rate, Reno TCP
975 requires a packet loss rate of 2.0e-10, while CUBIC TCP requires a
976 packet loss rate of 2.9e-8.
978 +===================+===========+=========+=========+=========+
979 | Throughput (Mbps) | Average W | Reno P | HSTCP P | CUBIC P |
980 +===================+===========+=========+=========+=========+
981 | 1 | 8.3 | 2.0e-2 | 2.0e-2 | 2.0e-2 |
982 +-------------------+-----------+---------+---------+---------+
983 | 10 | 83.3 | 2.0e-4 | 3.9e-4 | 2.9e-4 |
984 +-------------------+-----------+---------+---------+---------+
985 | 100 | 833.3 | 2.0e-6 | 2.5e-5 | 1.4e-5 |
986 +-------------------+-----------+---------+---------+---------+
987 | 1000 | 8333.3 | 2.0e-8 | 1.5e-6 | 6.3e-7 |
988 +-------------------+-----------+---------+---------+---------+
989 | 10000 | 83333.3 | 2.0e-10 | 1.0e-7 | 2.9e-8 |
990 +-------------------+-----------+---------+---------+---------+
992 Table 3: Required packet loss rate for Reno TCP, HSTCP, and
993 CUBIC to achieve a certain throughput
995 Table 3 describes the required packet loss rate for Reno TCP, HSTCP,
996 and CUBIC to achieve a certain throughput. We use 1500-byte packets
997 and an _RTT_ of 0.1 seconds.
999 Our test results in [HLRX07] indicate that CUBIC uses the spare
1000 bandwidth left unused by existing Reno TCP flows in the same
1001 bottleneck link without taking away much bandwidth from the existing
1002 flows.
1004 5.3. Difficult Environments
1006 CUBIC is designed to remedy the poor performance of Reno in fast and
1007 long-distance networks.
1009 5.4. Investigating a Range of Environments
1011 CUBIC has been extensively studied using simulations, testbed
1012 emulations, Internet experiments, and Internet measurements, covering
1013 a wide range of network environments
1014 [HLRX07][H16][CEHRX09][HR11][BSCLU13][LBEWK16]. They have
1015 convincingly demonstrated that CUBIC delivers substantial benefits
1016 over classical Reno congestion control [RFC5681].
1018 Same as Reno, CUBIC is a loss-based congestion control algorithm.
1019 Because CUBIC is designed to be more aggressive (due to a faster
1020 window increase function and bigger multiplicative decrease factor)
1021 than Reno in fast and long-distance networks, it can fill large drop-
1022 tail buffers more quickly than Reno and increases the risk of a
1023 standing queue [RFC8511]. In this case, proper queue sizing and
1024 management [RFC7567] could be used to mitigate the risk to some
1025 extent and reduce the packet queuing delay. Also, in large-BDP
1026 networks after a congestion event, CUBIC, due its cubic window
1027 increase function, recovers quickly to the highest link utilization
1028 point. This means that link utilization is less sensitive to an
1029 active queue management (AQM) target that is lower than the amplitude
1030 of the whole sawtooth.
1032 Similar to Reno, the performance of CUBIC as a loss-based congestion
1033 control algorithm suffers in networks where a packet loss is not a
1034 good indication of bandwidth utilization, such as wireless or mobile
1035 networks [LIU16].
1037 5.5. Protection against Congestion Collapse
1039 With regard to the potential of causing congestion collapse, CUBIC
1040 behaves like Reno, since CUBIC modifies only the window adjustment
1041 algorithm of Reno. Thus, it does not modify the ACK clocking and
1042 timeout behaviors of Reno.
1044 CUBIC also satisfies the "full backoff" requirement as described in
1045 [RFC5033]. After reducing the sending rate to one packet per RTT in
1046 response to congestion events due to ECN-Echo ACKs, CUBIC then
1047 exponentially increases the transmission timer for each packet
1048 retransmission while congestion persists.
1050 5.6. Fairness within the Alternative Congestion Control Algorithm
1052 CUBIC ensures convergence of competing CUBIC flows with the same RTT
1053 in the same bottleneck links to an equal throughput. When competing
1054 flows have different RTT values, their throughput ratio is linearly
1055 proportional to the inverse of their RTT ratios. This is true
1056 independently of the level of statistical multiplexing on the link.
1057 The convergence time depends on the network environments (e.g.,
1058 bandwidth, RTT) and the level of statistical multiplexing, as
1059 mentioned in Section 3.4.
1061 5.7. Performance with Misbehaving Nodes and Outside Attackers
1063 This is not considered in the current CUBIC design.
1065 5.8. Behavior for Application-Limited Flows
1067 A flow is application-limited if it is currently sending less than
1068 what is allowed by the congestion window. This can happen if the
1069 flow is limited by either the sender application or the receiver
1070 application (via the receiver advertised window) and thus sends less
1071 data than what is allowed by the sender's congestion window.
1073 CUBIC does not increase its congestion window if a flow is
1074 application-limited. Section 4.2 requires that _t_ in Figure 1 does
1075 not include application-limited periods, such as idle periods,
1076 otherwise W_cubic(_t_) might be very high after restarting from these
1077 periods.
1079 5.9. Responses to Sudden or Transient Events
1081 If there is a sudden increase in capacity, e.g., due to variable
1082 radio capacity, a routing change, or a mobility event, CUBIC is
1083 designed to utilize the newly available capacity faster than Reno.
1085 On the other hand, if there is a sudden decrease in capacity, CUBIC
1086 reduces more slowly than Reno. This remains true whether or not
1087 CUBIC is in Reno-friendly mode and whether or not fast convergence is
1088 enabled.
1090 5.10. Incremental Deployment
1092 CUBIC requires only changes to the congestion control at the sender,
1093 and it does not require any changes at receivers. That is, a CUBIC
1094 sender works correctly with Reno receivers. In addition, CUBIC does
1095 not require any changes to routers and does not require any
1096 assistance from routers.
1098 6. Security Considerations
1100 CUBIC makes no changes to the underlying security of TCP. More
1101 information about TCP security concerns can be found in [RFC5681].
1103 7. IANA Considerations
1105 This document does not require any IANA actions.
1107 8. References
1109 8.1. Normative References
1111 [I-D.ietf-tcpm-hystartplusplus]
1112 Balasubramanian, P., Huang, Y., and M. Olson, "HyStart++:
1113 Modified Slow Start for TCP", Work in Progress, Internet-
1114 Draft, draft-ietf-tcpm-hystartplusplus-04, 23 January
1115 2022, .
1118 [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate
1119 Requirement Levels", BCP 14, RFC 2119,
1120 DOI 10.17487/RFC2119, March 1997,
1121 .
1123 [RFC2883] Floyd, S., Mahdavi, J., Mathis, M., and M. Podolsky, "An
1124 Extension to the Selective Acknowledgement (SACK) Option
1125 for TCP", RFC 2883, DOI 10.17487/RFC2883, July 2000,
1126 .
1128 [RFC3168] Ramakrishnan, K., Floyd, S., and D. Black, "The Addition
1129 of Explicit Congestion Notification (ECN) to IP",
1130 RFC 3168, DOI 10.17487/RFC3168, September 2001,
1131 .
1133 [RFC4015] Ludwig, R. and A. Gurtov, "The Eifel Response Algorithm
1134 for TCP", RFC 4015, DOI 10.17487/RFC4015, February 2005,
1135 .
1137 [RFC5033] Floyd, S. and M. Allman, "Specifying New Congestion
1138 Control Algorithms", BCP 133, RFC 5033,
1139 DOI 10.17487/RFC5033, August 2007,
1140 .
1142 [RFC5348] Floyd, S., Handley, M., Padhye, J., and J. Widmer, "TCP
1143 Friendly Rate Control (TFRC): Protocol Specification",
1144 RFC 5348, DOI 10.17487/RFC5348, September 2008,
1145 .
1147 [RFC5681] Allman, M., Paxson, V., and E. Blanton, "TCP Congestion
1148 Control", RFC 5681, DOI 10.17487/RFC5681, September 2009,
1149 .
1151 [RFC5682] Sarolahti, P., Kojo, M., Yamamoto, K., and M. Hata,
1152 "Forward RTO-Recovery (F-RTO): An Algorithm for Detecting
1153 Spurious Retransmission Timeouts with TCP", RFC 5682,
1154 DOI 10.17487/RFC5682, September 2009,
1155 .
1157 [RFC6298] Paxson, V., Allman, M., Chu, J., and M. Sargent,
1158 "Computing TCP's Retransmission Timer", RFC 6298,
1159 DOI 10.17487/RFC6298, June 2011,
1160 .
1162 [RFC6582] Henderson, T., Floyd, S., Gurtov, A., and Y. Nishida, "The
1163 NewReno Modification to TCP's Fast Recovery Algorithm",
1164 RFC 6582, DOI 10.17487/RFC6582, April 2012,
1165 .
1167 [RFC6675] Blanton, E., Allman, M., Wang, L., Jarvinen, I., Kojo, M.,
1168 and Y. Nishida, "A Conservative Loss Recovery Algorithm
1169 Based on Selective Acknowledgment (SACK) for TCP",
1170 RFC 6675, DOI 10.17487/RFC6675, August 2012,
1171 .
1173 [RFC7567] Baker, F., Ed. and G. Fairhurst, Ed., "IETF
1174 Recommendations Regarding Active Queue Management",
1175 BCP 197, RFC 7567, DOI 10.17487/RFC7567, July 2015,
1176 .
1178 [RFC8174] Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC
1179 2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174,
1180 May 2017, .
1182 [RFC8985] Cheng, Y., Cardwell, N., Dukkipati, N., and P. Jha, "The
1183 RACK-TLP Loss Detection Algorithm for TCP", RFC 8985,
1184 DOI 10.17487/RFC8985, February 2021,
1185 .
1187 [RFC9002] Iyengar, J., Ed. and I. Swett, Ed., "QUIC Loss Detection
1188 and Congestion Control", RFC 9002, DOI 10.17487/RFC9002,
1189 May 2021, .
1191 8.2. Informative References
1193 [BSCLU13] Belhareth, S., Sassatelli, L., Collange, D., Lopez-
1194 Pacheco, D., and G. Urvoy-Keller, "Understanding TCP cubic
1195 performance in the cloud: A mean-field approach", 2013
1196 IEEE 2nd International Conference on Cloud
1197 Networking (CloudNet), DOI 10.1109/cloudnet.2013.6710576,
1198 November 2013,
1199 .
1201 [CEHRX09] Cai, H., Eun, D., Ha, S., Rhee, I., and L. Xu, "Stochastic
1202 convex ordering for multiplicative decrease internet
1203 congestion control", Computer Networks Vol. 53, pp.
1204 365-381, DOI 10.1016/j.comnet.2008.10.012, February 2009,
1205 .
1207 [FHP00] Floyd, S., Handley, M., and J. Padhye, "A Comparison of
1208 Equation-Based and AIMD Congestion Control", May 2000,
1209 .
1211 [GV02] Gorinsky, S. and H. Vin, "Extended Analysis of Binary
1212 Adjustment Algorithms", Technical Report TR2002-29,
1213 Department of Computer Sciences, The University of
1214 Texas at Austin, 11 August 2002,
1215 .
1217 [H16] Sangtae Ha, "Simulation, Testbed, and Deployment Testing
1218 Results of CUBIC", 3 November 2016,
1219 .
1222 [HLRX07] Ha, S., Le, L., Rhee, I., and L. Xu, "Impact of background
1223 traffic on performance of high-speed TCP variant
1224 protocols", Computer Networks Vol. 51, pp. 1748-1762,
1225 DOI 10.1016/j.comnet.2006.11.005, May 2007,
1226 .
1228 [HR11] Ha, S. and I. Rhee, "Taming the elephants: New TCP slow
1229 start", Computer Networks Vol. 55, pp. 2092-2110,
1230 DOI 10.1016/j.comnet.2011.01.014, June 2011,
1231 .
1233 [HRX08] Ha, S., Rhee, I., and L. Xu, "CUBIC: a new TCP-friendly
1234 high-speed TCP variant", ACM SIGOPS Operating Systems
1235 Review Vol. 42, pp. 64-74, DOI 10.1145/1400097.1400105,
1236 July 2008, .
1238 [K03] Kelly, T., "Scalable TCP: improving performance in
1239 highspeed wide area networks", ACM SIGCOMM Computer
1240 Communication Review Vol. 33, pp. 83-91,
1241 DOI 10.1145/956981.956989, April 2003,
1242 .
1244 [LBEWK16] Lukaseder, T., Bradatsch, L., Erb, B., Van Der Heijden,
1245 R., and F. Kargl, "A Comparison of TCP Congestion Control
1246 Algorithms in 10G Networks", 2016 IEEE 41st Conference on
1247 Local Computer Networks (LCN), DOI 10.1109/lcn.2016.121,
1248 November 2016, .
1250 [LIU16] Liu, K. and J. Lee, "On Improving TCP Performance over
1251 Mobile Data Networks", IEEE Transactions on Mobile
1252 Computing Vol. 15, pp. 2522-2536,
1253 DOI 10.1109/tmc.2015.2500227, October 2016,
1254 .
1256 [RFC3465] Allman, M., "TCP Congestion Control with Appropriate Byte
1257 Counting (ABC)", RFC 3465, DOI 10.17487/RFC3465, February
1258 2003, .
1260 [RFC3522] Ludwig, R. and M. Meyer, "The Eifel Detection Algorithm
1261 for TCP", RFC 3522, DOI 10.17487/RFC3522, April 2003,
1262 .
1264 [RFC3649] Floyd, S., "HighSpeed TCP for Large Congestion Windows",
1265 RFC 3649, DOI 10.17487/RFC3649, December 2003,
1266 .
1268 [RFC3708] Blanton, E. and M. Allman, "Using TCP Duplicate Selective
1269 Acknowledgement (DSACKs) and Stream Control Transmission
1270 Protocol (SCTP) Duplicate Transmission Sequence Numbers
1271 (TSNs) to Detect Spurious Retransmissions", RFC 3708,
1272 DOI 10.17487/RFC3708, February 2004,
1273 .
1275 [RFC3742] Floyd, S., "Limited Slow-Start for TCP with Large
1276 Congestion Windows", RFC 3742, DOI 10.17487/RFC3742, March
1277 2004, .
1279 [RFC4960] Stewart, R., Ed., "Stream Control Transmission Protocol",
1280 RFC 4960, DOI 10.17487/RFC4960, September 2007,
1281 .
1283 [RFC6937] Mathis, M., Dukkipati, N., and Y. Cheng, "Proportional
1284 Rate Reduction for TCP", RFC 6937, DOI 10.17487/RFC6937,
1285 May 2013, .
1287 [RFC7661] Fairhurst, G., Sathiaseelan, A., and R. Secchi, "Updating
1288 TCP to Support Rate-Limited Traffic", RFC 7661,
1289 DOI 10.17487/RFC7661, October 2015,
1290 .
1292 [RFC8312] Rhee, I., Xu, L., Ha, S., Zimmermann, A., Eggert, L., and
1293 R. Scheffenegger, "CUBIC for Fast Long-Distance Networks",
1294 RFC 8312, DOI 10.17487/RFC8312, February 2018,
1295 .
1297 [RFC8511] Khademi, N., Welzl, M., Armitage, G., and G. Fairhurst,
1298 "TCP Alternative Backoff with ECN (ABE)", RFC 8511,
1299 DOI 10.17487/RFC8511, December 2018,
1300 .
1302 [RFC9000] Iyengar, J., Ed. and M. Thomson, Ed., "QUIC: A UDP-Based
1303 Multiplexed and Secure Transport", RFC 9000,
1304 DOI 10.17487/RFC9000, May 2021,
1305 .
1307 [SXEZ19] Sun, W., Xu, L., Elbaum, S., and D. Zhao, "Model-Agnostic
1308 and Efficient Exploration of Numerical Congestion Control
1309 State Space of Real-World TCP Implementations", IEEE/ACM
1310 Transactions on Networking Vol. 29, pp. 1990-2004,
1311 DOI 10.1109/tnet.2021.3078161, October 2021,
1312 .
1314 [XHR04] Xu, L., Harfoush, K., and I. Rhee, "Binary increase
1315 congestion control (BIC) for fast long-distance networks",
1316 IEEE INFOCOM 2004, DOI 10.1109/infcom.2004.1354672, n.d.,
1317 .
1319 Appendix A. Acknowledgments
1321 Richard Scheffenegger and Alexander Zimmermann originally co-authored
1322 [RFC8312].
1324 These individuals suggested improvements to this document:
1326 * Bob Briscoe
1328 * Christian Huitema
1330 * Gorry Fairhurst
1332 * Jonathan Morton
1334 * Juhamatti Kuusisaari
1336 * Junho Choi
1338 * Markku Kojo
1340 * Martin Thomson
1342 * Matt Mathis
1344 * Matt Olson
1346 * Michael Welzl
1348 * Mirja Kuehlewind
1350 * Mohit P. Tahiliani
1352 * Neal Cardwell
1354 * Praveen Balasubramanian
1355 * Randall Stewart
1357 * Richard Scheffenegger
1359 * Rod Grimes
1361 * Tom Henderson
1363 * Tom Petch
1365 * Wesley Rosenblum
1367 * Yoshifumi Nishida
1369 * Yuchung Cheng
1371 Appendix B. Evolution of CUBIC
1373 B.1. Since draft-ietf-tcpm-rfc8312bis-06
1375 * RFC7661 is safe even when cwnd grows beyond rwnd (#143
1376 (https://github.com/NTAP/rfc8312bis/issues/143))
1378 B.2. Since draft-ietf-tcpm-rfc8312bis-05
1380 * Clarify meaning of "application-limited" in Section 5.8 (#137
1381 (https://github.com/NTAP/rfc8312bis/issues/137))
1383 * Create new subsections for spurious timeouts and spurious loss via
1384 ACK (#90 (https://github.com/NTAP/rfc8312bis/issues/90))
1386 * Brief discussion of convergence in Section 5.6 (#96
1387 (https://github.com/NTAP/rfc8312bis/issues/96))
1389 * Add more test results to Section 5 and update some references (#91
1390 (https://github.com/NTAP/rfc8312bis/issues/91))
1392 * Change wording around setting ssthresh (#131
1393 (https://github.com/NTAP/rfc8312bis/issues/131))
1395 B.3. Since draft-ietf-tcpm-rfc8312bis-04
1397 * Fix incorrect math (#106 (https://github.com/NTAP/rfc8312bis/
1398 issues/106))
1400 * Update RFC5681 (#99 (https://github.com/NTAP/rfc8312bis/
1401 issues/99))
1403 * Rephrase text around algorithmic alternatives, add HyStart++ (#85
1404 (https://github.com/NTAP/rfc8312bis/issues/85), #86
1405 (https://github.com/NTAP/rfc8312bis/issues/86), #90
1406 (https://github.com/NTAP/rfc8312bis/issues/90))
1408 * Clarify what we mean by "new ACK" and use it in the text in more
1409 places. (#101 (https://github.com/NTAP/rfc8312bis/issues/101))
1411 * Rewrite the Responses to Sudden or Transient Events section (#98
1412 (https://github.com/NTAP/rfc8312bis/issues/98))
1414 * Remove confusing text about _cwnd_start_ in Section 4.2 (#100
1415 (https://github.com/NTAP/rfc8312bis/issues/100))
1417 * Change terminology from "AIMD" to "Reno" (#108
1418 (https://github.com/NTAP/rfc8312bis/issues/108))
1420 * Moved MUST NOT from app-limited section to main cubic AI section
1421 (#97 (https://github.com/NTAP/rfc8312bis/issues/97))
1423 * Clarify cwnd decrease during multiplicative decrease (#102
1424 (https://github.com/NTAP/rfc8312bis/issues/102))
1426 * Clarify text around queuing and slow adaptation of CUBIC in
1427 wireless environments (#94 (https://github.com/NTAP/rfc8312bis/
1428 issues/94))
1430 * Set lower bound of cwnd to 1 MSS and use retransmit timer
1431 thereafter (#83 (https://github.com/NTAP/rfc8312bis/issues/83))
1433 * Use FlightSize instead of cwnd to update ssthresh (#114
1434 (https://github.com/NTAP/rfc8312bis/issues/114))
1436 B.4. Since draft-ietf-tcpm-rfc8312bis-03
1438 * Remove reference from abstract (#82
1439 (https://github.com/NTAP/rfc8312bis/pull/82))
1441 B.5. Since draft-ietf-tcpm-rfc8312bis-02
1443 * Description of packet loss rate _p_ (#65
1444 (https://github.com/NTAP/rfc8312bis/issues/65))
1446 * Clarification of TCP Friendly Equation for ABC and Delayed ACK
1447 (#66 (https://github.com/NTAP/rfc8312bis/issues/66))
1449 * add applicability to QUIC and SCTP (#61
1450 (https://github.com/NTAP/rfc8312bis/issues/61))
1452 * clarity on setting alpha__aimd_ to 1 (#68
1453 (https://github.com/NTAP/rfc8312bis/issues/68))
1455 * introduce alpha__cubic_ (#64 (https://github.com/NTAP/rfc8312bis/
1456 issues/64))
1458 * clarify _cwnd_ growth in convex region (#69
1459 (https://github.com/NTAP/rfc8312bis/issues/69))
1461 * add guidance for using bytes and mention that segments count is
1462 decimal (#67 (https://github.com/NTAP/rfc8312bis/issues/67))
1464 * add loss events detected by RACK and QUIC loss detection (#62
1465 (https://github.com/NTAP/rfc8312bis/issues/62))
1467 B.6. Since draft-ietf-tcpm-rfc8312bis-01
1469 * address Michael Scharf's editorial suggestions. (#59
1470 (https://github.com/NTAP/rfc8312bis/issues/59))
1472 * add "Note to the RFC Editor" about removing underscores
1474 B.7. Since draft-ietf-tcpm-rfc8312bis-00
1476 * use updated xml2rfc with better text rendering of subscripts
1478 B.8. Since draft-eggert-tcpm-rfc8312bis-03
1480 * fix spelling nits
1482 * rename to draft-ietf
1484 * define _W_max_ more clearly
1486 B.9. Since draft-eggert-tcpm-rfc8312bis-02
1488 * add definition for segments_acked and alpha__aimd_. (#47
1489 (https://github.com/NTAP/rfc8312bis/issues/47))
1491 * fix a mistake in _W_max_ calculation in the fast convergence
1492 section. (#51 (https://github.com/NTAP/rfc8312bis/issues/51))
1494 * clarity on setting _ssthresh_ and _cwnd_start_ during
1495 multiplicative decrease. (#53 (https://github.com/NTAP/rfc8312bis/
1496 issues/53))
1498 B.10. Since draft-eggert-tcpm-rfc8312bis-01
1499 * rename TCP-Friendly to AIMD-Friendly and rename Standard TCP to
1500 AIMD TCP to avoid confusion as CUBIC has been widely used on the
1501 Internet. (#38 (https://github.com/NTAP/rfc8312bis/issues/38))
1503 * change introductory text to reflect the significant broader
1504 deployment of CUBIC on the Internet. (#39
1505 (https://github.com/NTAP/rfc8312bis/issues/39))
1507 * rephrase introduction to avoid referring to variables that have
1508 not been defined yet.
1510 B.11. Since draft-eggert-tcpm-rfc8312bis-00
1512 * acknowledge former co-authors (#15
1513 (https://github.com/NTAP/rfc8312bis/issues/15))
1515 * prevent _cwnd_ from becoming less than two (#7
1516 (https://github.com/NTAP/rfc8312bis/issues/7))
1518 * add list of variables and constants (#5
1519 (https://github.com/NTAP/rfc8312bis/issues/5), #6
1520 (https://github.com/NTAP/rfc8312bis/issues/6))
1522 * update _K_'s definition and add bounds for CUBIC _target_ _cwnd_
1523 [SXEZ19] (#1 (https://github.com/NTAP/rfc8312bis/issues/1), #14
1524 (https://github.com/NTAP/rfc8312bis/issues/14))
1526 * update _W_est_ to use AIMD approach (#20
1527 (https://github.com/NTAP/rfc8312bis/issues/20))
1529 * set alpha__aimd_ to 1 once _W_est_ reaches _W_max_ (#2
1530 (https://github.com/NTAP/rfc8312bis/issues/2))
1532 * add Vidhi as co-author (#17 (https://github.com/NTAP/rfc8312bis/
1533 issues/17))
1535 * note for Fast Recovery during _cwnd_ decrease due to congestion
1536 event (#11 (https://github.com/NTAP/rfc8312bis/issues/11))
1538 * add section for spurious congestion events (#23
1539 (https://github.com/NTAP/rfc8312bis/issues/23))
1541 * initialize _W_est_ after timeout and remove variable
1542 _W_(last_max)_ (#28 (https://github.com/NTAP/rfc8312bis/
1543 issues/28))
1545 B.12. Since RFC8312
1546 * converted to Markdown and xml2rfc v3
1548 * updated references (as part of the conversion)
1550 * updated author information
1552 * various formatting changes
1554 * move to Standards Track
1556 B.13. Since the Original Paper
1558 CUBIC has gone through a few changes since the initial release
1559 [HRX08] of its algorithm and implementation. Below we highlight the
1560 differences between its original paper and [RFC8312].
1562 * The original paper [HRX08] includes the pseudocode of CUBIC
1563 implementation using Linux's pluggable congestion control
1564 framework, which excludes system-specific optimizations. The
1565 simplified pseudocode might be a good source to start with and
1566 understand CUBIC.
1568 * [HRX08] also includes experimental results showing its performance
1569 and fairness.
1571 * The definition of beta__cubic_ constant was changed in [RFC8312].
1572 For example, beta__cubic_ in the original paper was the window
1573 decrease constant while [RFC8312] changed it to CUBIC
1574 multiplication decrease factor. With this change, the current
1575 congestion window size after a congestion event in [RFC8312] was
1576 beta__cubic_ * _W_max_ while it was (1-beta__cubic_) * _W_max_ in
1577 the original paper.
1579 * Its pseudocode used _W_(last_max)_ while [RFC8312] used _W_max_.
1581 * Its AIMD-friendly window was _W_tcp_ while [RFC8312] used _W_est_.
1583 Authors' Addresses
1585 Lisong Xu
1586 University of Nebraska-Lincoln
1587 Department of Computer Science and Engineering
1588 Lincoln, NE 68588-0115
1589 United States of America
1590 Email: xu@unl.edu
1591 URI: https://cse.unl.edu/~xu/
1592 Sangtae Ha
1593 University of Colorado at Boulder
1594 Department of Computer Science
1595 Boulder, CO 80309-0430
1596 United States of America
1597 Email: sangtae.ha@colorado.edu
1598 URI: https://netstech.org/sangtaeha/
1600 Injong Rhee
1601 Bowery Farming
1602 151 W 26TH Street, 12TH Floor
1603 New York, NY 10001
1604 United States of America
1605 Email: injongrhee@gmail.com
1607 Vidhi Goel
1608 Apple Inc.
1609 One Apple Park Way
1610 Cupertino, California 95014
1611 United States of America
1612 Email: vidhi_goel@apple.com
1614 Lars Eggert (editor)
1615 NetApp
1616 Stenbergintie 12 B
1617 FI-02700 Kauniainen
1618 Finland
1619 Email: lars@eggert.org
1620 URI: https://eggert.org/