Debian Codegraph @

rfc5681.html

raw
1 <pre>Network Working Group M. Allman
2 Request for Comments: 5681 V. Paxson
3 Obsoletes: <a href="./rfc2581">2581</a> ICSI
4 Category: Standards Track E. Blanton
5 Purdue University
6 September 2009
7
8
9 <span class="h1">TCP Congestion Control</span>
10
11 Abstract
12
13 This document defines TCP's four intertwined congestion control
14 algorithms: slow start, congestion avoidance, fast retransmit, and
15 fast recovery. In addition, the document specifies how TCP should
16 begin transmission after a relatively long idle period, as well as
17 discussing various acknowledgment generation methods. This document
18 obsoletes <a href="./rfc2581">RFC 2581</a>.
19
20 Status of This Memo
21
22 This document specifies an Internet standards track protocol for the
23 Internet community, and requests discussion and suggestions for
24 improvements. Please refer to the current edition of the "Internet
25 Official Protocol Standards" (STD 1) for the standardization state
26 and status of this protocol. Distribution of this memo is unlimited.
27
28 Copyright Notice
29
30 Copyright (c) 2009 IETF Trust and the persons identified as the
31 document authors. All rights reserved.
32
33 This document is subject to <a href="https://www.rfc-editor.org/bcp/bcp78">BCP 78</a> and the IETF Trust's Legal
34 Provisions Relating to IETF Documents in effect on the date of
35 publication of this document (<a href="http://trustee.ietf.org/license-info">http://trustee.ietf.org/license-info</a>).
36 Please review these documents carefully, as they describe your rights
37 and restrictions with respect to this document.
38
39 This document may contain material from IETF Documents or IETF
40 Contributions published or made publicly available before November
41 10, 2008. The person(s) controlling the copyright in some of this
42 material may not have granted the IETF Trust the right to allow
43 modifications of such material outside the IETF Standards Process.
44 Without obtaining an adequate license from the person(s) controlling
45 the copyright in such materials, this document may not be modified
46 outside the IETF Standards Process, and derivative works of it may
47
48
49
50
51
52 <span class="grey">Allman, et al. Standards Track [Page 1]</span></pre>
53 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-2" ></span>
54 <span class="grey"><a href="./rfc5681">RFC 5681</a> TCP Congestion Control September 2009</span>
55
56
57 not be created outside the IETF Standards Process, except to format
58 it for publication as an RFC or to translate it into languages other
59 than English.
60
61 Table Of Contents
62
63 <a href="#section-1">1</a>. Introduction ....................................................<a href="#page-2">2</a>
64 <a href="#section-2">2</a>. Definitions .....................................................<a href="#page-3">3</a>
65 <a href="#section-3">3</a>. Congestion Control Algorithms ...................................<a href="#page-4">4</a>
66 <a href="#section-3.1">3.1</a>. Slow Start and Congestion Avoidance ........................<a href="#page-4">4</a>
67 <a href="#section-3.2">3.2</a>. Fast Retransmit/Fast Recovery ..............................<a href="#page-8">8</a>
68 <a href="#section-4">4</a>. Additional Considerations ......................................<a href="#page-10">10</a>
69 <a href="#section-4.1">4.1</a>. Restarting Idle Connections ...............................<a href="#page-10">10</a>
70 <a href="#section-4.2">4.2</a>. Generating Acknowledgments ................................<a href="#page-11">11</a>
71 <a href="#section-4.3">4.3</a>. Loss Recovery Mechanisms ..................................<a href="#page-12">12</a>
72 <a href="#section-5">5</a>. Security Considerations ........................................<a href="#page-13">13</a>
73 <a href="#section-6">6</a>. Changes between <a href="./rfc2001">RFC 2001</a> and <a href="./rfc2581">RFC 2581</a> ..........................<a href="#page-13">13</a>
74 <a href="#section-7">7</a>. Changes Relative to <a href="./rfc2581">RFC 2581</a> ...................................<a href="#page-14">14</a>
75 <a href="#section-8">8</a>. Acknowledgments ................................................<a href="#page-15">15</a>
76 <a href="#section-9">9</a>. References .....................................................<a href="#page-15">15</a>
77 <a href="#section-9.1">9.1</a>. Normative References ......................................<a href="#page-15">15</a>
78 <a href="#section-9.2">9.2</a>. Informative References ....................................<a href="#page-16">16</a>
79
80 <span class="h2"><a class="selflink" id="section-1" href="#section-1">1</a>. Introduction</span>
81
82 This document specifies four TCP [<a href="./rfc793" title="&quot;Transmission Control Protocol&quot;">RFC793</a>] congestion control
83 algorithms: slow start, congestion avoidance, fast retransmit and
84 fast recovery. These algorithms were devised in [<a href="#ref-Jac88" title="&quot;Congestion Avoidance and Control&quot;">Jac88</a>] and [<a href="#ref-Jac90" title="&quot;Modified TCP Congestion Avoidance Algorithm&quot;">Jac90</a>].
85 Their use with TCP is standardized in [<a href="./rfc1122" title="&quot;Requirements for Internet Hosts - Communication Layers&quot;">RFC1122</a>]. Additional early
86 work in additive-increase, multiplicative-decrease congestion control
87 is given in [<a href="#ref-CJ89" title="&quot;Analysis of the Increase/Decrease Algorithms for Congestion Avoidance in Computer Networks&quot;">CJ89</a>].
88
89 Note that [<a href="#ref-Ste94" title="&quot;TCP/IP Illustrated, Volume 1: The Protocols&quot;">Ste94</a>] provides examples of these algorithms in action and
90 [<a href="#ref-WS95" title="&quot;TCP/IP Illustrated, Volume 2: The Implementation&quot;">WS95</a>] provides an explanation of the source code for the BSD
91 implementation of these algorithms.
92
93 In addition to specifying these congestion control algorithms, this
94 document specifies what TCP connections should do after a relatively
95 long idle period, as well as specifying and clarifying some of the
96 issues pertaining to TCP ACK generation.
97
98 This document obsoletes [<a href="./rfc2581" title="&quot;TCP Congestion Control&quot;">RFC2581</a>], which in turn obsoleted [<a href="./rfc2001" title="&quot;TCP Slow Start, Congestion Avoidance, Fast Retransmit, and Fast Recovery Algorithms&quot;">RFC2001</a>].
99
100 This document is organized as follows. <a href="#section-2">Section 2</a> provides various
101 definitions that will be used throughout the document. <a href="#section-3">Section 3</a>
102 provides a specification of the congestion control algorithms.
103 <a href="#section-4">Section 4</a> outlines concerns related to the congestion control
104 algorithms and finally, <a href="#section-5">section 5</a> outlines security considerations.
105
106
107
108 <span class="grey">Allman, et al. Standards Track [Page 2]</span></pre>
109 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-3" ></span>
110 <span class="grey"><a href="./rfc5681">RFC 5681</a> TCP Congestion Control September 2009</span>
111
112
113 The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
114 "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this
115 document are to be interpreted as described in [<a href="./rfc2119" title="&quot;Key words for use in RFCs to Indicate Requirement Levels&quot;">RFC2119</a>].
116
117 <span class="h2"><a class="selflink" id="section-2" href="#section-2">2</a>. Definitions</span>
118
119 This section provides the definition of several terms that will be
120 used throughout the remainder of this document.
121
122 SEGMENT: A segment is ANY TCP/IP data or acknowledgment packet (or
123 both).
124
125 SENDER MAXIMUM SEGMENT SIZE (SMSS): The SMSS is the size of the
126 largest segment that the sender can transmit. This value can be
127 based on the maximum transmission unit of the network, the path
128 MTU discovery [RFC1191, <a href="./rfc4821">RFC4821</a>] algorithm, RMSS (see next item),
129 or other factors. The size does not include the TCP/IP headers
130 and options.
131
132 RECEIVER MAXIMUM SEGMENT SIZE (RMSS): The RMSS is the size of the
133 largest segment the receiver is willing to accept. This is the
134 value specified in the MSS option sent by the receiver during
135 connection startup. Or, if the MSS option is not used, it is 536
136 bytes [<a href="./rfc1122" title="&quot;Requirements for Internet Hosts - Communication Layers&quot;">RFC1122</a>]. The size does not include the TCP/IP headers and
137 options.
138
139 FULL-SIZED SEGMENT: A segment that contains the maximum number of
140 data bytes permitted (i.e., a segment containing SMSS bytes of
141 data).
142
143 RECEIVER WINDOW (rwnd): The most recently advertised receiver window.
144
145 CONGESTION WINDOW (cwnd): A TCP state variable that limits the amount
146 of data a TCP can send. At any given time, a TCP MUST NOT send
147 data with a sequence number higher than the sum of the highest
148 acknowledged sequence number and the minimum of cwnd and rwnd.
149
150 INITIAL WINDOW (IW): The initial window is the size of the sender's
151 congestion window after the three-way handshake is completed.
152
153 LOSS WINDOW (LW): The loss window is the size of the congestion
154 window after a TCP sender detects loss using its retransmission
155 timer.
156
157 RESTART WINDOW (RW): The restart window is the size of the congestion
158 window after a TCP restarts transmission after an idle period (if
159 the slow start algorithm is used; see <a href="#section-4.1">section 4.1</a> for more
160 discussion).
161
162
163
164 <span class="grey">Allman, et al. Standards Track [Page 3]</span></pre>
165 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-4" ></span>
166 <span class="grey"><a href="./rfc5681">RFC 5681</a> TCP Congestion Control September 2009</span>
167
168
169 FLIGHT SIZE: The amount of data that has been sent but not yet
170 cumulatively acknowledged.
171
172 DUPLICATE ACKNOWLEDGMENT: An acknowledgment is considered a
173 "duplicate" in the following algorithms when (a) the receiver of
174 the ACK has outstanding data, (b) the incoming acknowledgment
175 carries no data, (c) the SYN and FIN bits are both off, (d) the
176 acknowledgment number is equal to the greatest acknowledgment
177 received on the given connection (TCP.UNA from [<a href="./rfc793" title="&quot;Transmission Control Protocol&quot;">RFC793</a>]) and (e)
178 the advertised window in the incoming acknowledgment equals the
179 advertised window in the last incoming acknowledgment.
180
181 Alternatively, a TCP that utilizes selective acknowledgments
182 (SACKs) [RFC2018, <a href="./rfc2883">RFC2883</a>] can leverage the SACK information to
183 determine when an incoming ACK is a "duplicate" (e.g., if the ACK
184 contains previously unknown SACK information).
185
186 <span class="h2"><a class="selflink" id="section-3" href="#section-3">3</a>. Congestion Control Algorithms</span>
187
188 This section defines the four congestion control algorithms: slow
189 start, congestion avoidance, fast retransmit, and fast recovery,
190 developed in [<a href="#ref-Jac88" title="&quot;Congestion Avoidance and Control&quot;">Jac88</a>] and [<a href="#ref-Jac90" title="&quot;Modified TCP Congestion Avoidance Algorithm&quot;">Jac90</a>]. In some situations, it may be
191 beneficial for a TCP sender to be more conservative than the
192 algorithms allow; however, a TCP MUST NOT be more aggressive than the
193 following algorithms allow (that is, MUST NOT send data when the
194 value of cwnd computed by the following algorithms would not allow
195 the data to be sent).
196
197 Also, note that the algorithms specified in this document work in
198 terms of using loss as the signal of congestion. Explicit Congestion
199 Notification (ECN) could also be used as specified in [<a href="./rfc3168" title="&quot;The Addition of Explicit Congestion Notification (ECN) to IP&quot;">RFC3168</a>].
200
201 <span class="h3"><a class="selflink" id="section-3.1" href="#section-3.1">3.1</a>. Slow Start and Congestion Avoidance</span>
202
203 The slow start and congestion avoidance algorithms MUST be used by a
204 TCP sender to control the amount of outstanding data being injected
205 into the network. To implement these algorithms, two variables are
206 added to the TCP per-connection state. The congestion window (cwnd)
207 is a sender-side limit on the amount of data the sender can transmit
208 into the network before receiving an acknowledgment (ACK), while the
209 receiver's advertised window (rwnd) is a receiver-side limit on the
210 amount of outstanding data. The minimum of cwnd and rwnd governs
211 data transmission.
212
213 Another state variable, the slow start threshold (ssthresh), is used
214 to determine whether the slow start or congestion avoidance algorithm
215 is used to control data transmission, as discussed below.
216
217
218
219
220 <span class="grey">Allman, et al. Standards Track [Page 4]</span></pre>
221 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-5" ></span>
222 <span class="grey"><a href="./rfc5681">RFC 5681</a> TCP Congestion Control September 2009</span>
223
224
225 Beginning transmission into a network with unknown conditions
226 requires TCP to slowly probe the network to determine the available
227 capacity, in order to avoid congesting the network with an
228 inappropriately large burst of data. The slow start algorithm is
229 used for this purpose at the beginning of a transfer, or after
230 repairing loss detected by the retransmission timer. Slow start
231 additionally serves to start the "ACK clock" used by the TCP sender
232 to release data into the network in the slow start, congestion
233 avoidance, and loss recovery algorithms.
234
235 IW, the initial value of cwnd, MUST be set using the following
236 guidelines as an upper bound.
237
238 If SMSS &gt; 2190 bytes:
239 IW = 2 * SMSS bytes and MUST NOT be more than 2 segments
240 If (SMSS &gt; 1095 bytes) and (SMSS &lt;= 2190 bytes):
241 IW = 3 * SMSS bytes and MUST NOT be more than 3 segments
242 if SMSS &lt;= 1095 bytes:
243 IW = 4 * SMSS bytes and MUST NOT be more than 4 segments
244
245 As specified in [<a href="./rfc3390" title="&quot;Increasing TCP's Initial Window&quot;">RFC3390</a>], the SYN/ACK and the acknowledgment of the
246 SYN/ACK MUST NOT increase the size of the congestion window.
247 Further, if the SYN or SYN/ACK is lost, the initial window used by a
248 sender after a correctly transmitted SYN MUST be one segment
249 consisting of at most SMSS bytes.
250
251 A detailed rationale and discussion of the IW setting is provided in
252 [<a href="./rfc3390" title="&quot;Increasing TCP's Initial Window&quot;">RFC3390</a>].
253
254 When initial congestion windows of more than one segment are
255 implemented along with Path MTU Discovery [<a href="./rfc1191" title="&quot;Path MTU discovery&quot;">RFC1191</a>], and the MSS
256 being used is found to be too large, the congestion window cwnd
257 SHOULD be reduced to prevent large bursts of smaller segments.
258 Specifically, cwnd SHOULD be reduced by the ratio of the old segment
259 size to the new segment size.
260
261 The initial value of ssthresh SHOULD be set arbitrarily high (e.g.,
262 to the size of the largest possible advertised window), but ssthresh
263 MUST be reduced in response to congestion. Setting ssthresh as high
264 as possible allows the network conditions, rather than some arbitrary
265 host limit, to dictate the sending rate. In cases where the end
266 systems have a solid understanding of the network path, more
267 carefully setting the initial ssthresh value may have merit (e.g.,
268 such that the end host does not create congestion along the path).
269
270
271
272
273
274
275
276 <span class="grey">Allman, et al. Standards Track [Page 5]</span></pre>
277 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-6" ></span>
278 <span class="grey"><a href="./rfc5681">RFC 5681</a> TCP Congestion Control September 2009</span>
279
280
281 The slow start algorithm is used when cwnd &lt; ssthresh, while the
282 congestion avoidance algorithm is used when cwnd &gt; ssthresh. When
283 cwnd and ssthresh are equal, the sender may use either slow start or
284 congestion avoidance.
285
286 During slow start, a TCP increments cwnd by at most SMSS bytes for
287 each ACK received that cumulatively acknowledges new data. Slow
288 start ends when cwnd exceeds ssthresh (or, optionally, when it
289 reaches it, as noted above) or when congestion is observed. While
290 traditionally TCP implementations have increased cwnd by precisely
291 SMSS bytes upon receipt of an ACK covering new data, we RECOMMEND
292 that TCP implementations increase cwnd, per:
293
294 cwnd += min (N, SMSS) (2)
295
296 where N is the number of previously unacknowledged bytes acknowledged
297 in the incoming ACK. This adjustment is part of Appropriate Byte
298 Counting [<a href="./rfc3465" title="&quot;TCP Congestion Control with Appropriate Byte Counting (ABC)&quot;">RFC3465</a>] and provides robustness against misbehaving
299 receivers that may attempt to induce a sender to artificially inflate
300 cwnd using a mechanism known as "ACK Division" [<a href="#ref-SCWA99" title="&quot;TCP Congestion Control With a Misbehaving Receiver&quot;">SCWA99</a>]. ACK
301 Division consists of a receiver sending multiple ACKs for a single
302 TCP data segment, each acknowledging only a portion of its data. A
303 TCP that increments cwnd by SMSS for each such ACK will
304 inappropriately inflate the amount of data injected into the network.
305
306 During congestion avoidance, cwnd is incremented by roughly 1 full-
307 sized segment per round-trip time (RTT). Congestion avoidance
308 continues until congestion is detected. The basic guidelines for
309 incrementing cwnd during congestion avoidance are:
310
311 * MAY increment cwnd by SMSS bytes
312
313 * SHOULD increment cwnd per equation (2) once per RTT
314
315 * MUST NOT increment cwnd by more than SMSS bytes
316
317 We note that [<a href="./rfc3465" title="&quot;TCP Congestion Control with Appropriate Byte Counting (ABC)&quot;">RFC3465</a>] allows for cwnd increases of more than SMSS
318 bytes for incoming acknowledgments during slow start on an
319 experimental basis; however, such behavior is not allowed as part of
320 the standard.
321
322 The RECOMMENDED way to increase cwnd during congestion avoidance is
323 to count the number of bytes that have been acknowledged by ACKs for
324 new data. (A drawback of this implementation is that it requires
325 maintaining an additional state variable.) When the number of bytes
326 acknowledged reaches cwnd, then cwnd can be incremented by up to SMSS
327 bytes. Note that during congestion avoidance, cwnd MUST NOT be
328
329
330
331
332 <span class="grey">Allman, et al. Standards Track [Page 6]</span></pre>
333 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-7" ></span>
334 <span class="grey"><a href="./rfc5681">RFC 5681</a> TCP Congestion Control September 2009</span>
335
336
337 increased by more than SMSS bytes per RTT. This method both allows
338 TCPs to increase cwnd by one segment per RTT in the face of delayed
339 ACKs and provides robustness against ACK Division attacks.
340
341 Another common formula that a TCP MAY use to update cwnd during
342 congestion avoidance is given in equation (3):
343
344 cwnd += SMSS*SMSS/cwnd (3)
345
346 This adjustment is executed on every incoming ACK that acknowledges
347 new data. Equation (3) provides an acceptable approximation to the
348 underlying principle of increasing cwnd by 1 full-sized segment per
349 RTT. (Note that for a connection in which the receiver is
350 acknowledging every-other packet, (3) is less aggressive than allowed
351 -- roughly increasing cwnd every second RTT.)
352
353 Implementation Note: Since integer arithmetic is usually used in TCP
354 implementations, the formula given in equation (3) can fail to
355 increase cwnd when the congestion window is larger than SMSS*SMSS.
356 If the above formula yields 0, the result SHOULD be rounded up to 1
357 byte.
358
359 Implementation Note: Older implementations have an additional
360 additive constant on the right-hand side of equation (3). This is
361 incorrect and can actually lead to diminished performance [<a href="./rfc2525" title="&quot;Known TCP Implementation Problems&quot;">RFC2525</a>].
362
363 Implementation Note: Some implementations maintain cwnd in units of
364 bytes, while others in units of full-sized segments. The latter will
365 find equation (3) difficult to use, and may prefer to use the
366 counting approach discussed in the previous paragraph.
367
368 When a TCP sender detects segment loss using the retransmission timer
369 and the given segment has not yet been resent by way of the
370 retransmission timer, the value of ssthresh MUST be set to no more
371 than the value given in equation (4):
372
373 ssthresh = max (FlightSize / 2, 2*SMSS) (4)
374
375 where, as discussed above, FlightSize is the amount of outstanding
376 data in the network.
377
378 On the other hand, when a TCP sender detects segment loss using the
379 retransmission timer and the given segment has already been
380 retransmitted by way of the retransmission timer at least once, the
381 value of ssthresh is held constant.
382
383
384
385
386
387
388 <span class="grey">Allman, et al. Standards Track [Page 7]</span></pre>
389 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-8" ></span>
390 <span class="grey"><a href="./rfc5681">RFC 5681</a> TCP Congestion Control September 2009</span>
391
392
393 Implementation Note: An easy mistake to make is to simply use cwnd,
394 rather than FlightSize, which in some implementations may
395 incidentally increase well beyond rwnd.
396
397 Furthermore, upon a timeout (as specified in [<a href="./rfc2988" title="&quot;Computing TCP's Retransmission Timer&quot;">RFC2988</a>]) cwnd MUST be
398 set to no more than the loss window, LW, which equals 1 full-sized
399 segment (regardless of the value of IW). Therefore, after
400 retransmitting the dropped segment the TCP sender uses the slow start
401 algorithm to increase the window from 1 full-sized segment to the new
402 value of ssthresh, at which point congestion avoidance again takes
403 over.
404
405 As shown in [<a href="#ref-FF96" title="&quot;Simulation-based Comparisons of Tahoe, Reno and SACK TCP&quot;">FF96</a>] and [<a href="./rfc3782" title="&quot;The NewReno Modification to TCP's Fast Recovery Algorithm&quot;">RFC3782</a>], slow-start-based loss recovery
406 after a timeout can cause spurious retransmissions that trigger
407 duplicate acknowledgments. The reaction to the arrival of these
408 duplicate ACKs in TCP implementations varies widely. This document
409 does not specify how to treat such acknowledgments, but does note
410 this as an area that may benefit from additional attention,
411 experimentation and specification.
412
413 <span class="h3"><a class="selflink" id="section-3.2" href="#section-3.2">3.2</a>. Fast Retransmit/Fast Recovery</span>
414
415 A TCP receiver SHOULD send an immediate duplicate ACK when an out-
416 of-order segment arrives. The purpose of this ACK is to inform the
417 sender that a segment was received out-of-order and which sequence
418 number is expected. From the sender's perspective, duplicate ACKs
419 can be caused by a number of network problems. First, they can be
420 caused by dropped segments. In this case, all segments after the
421 dropped segment will trigger duplicate ACKs until the loss is
422 repaired. Second, duplicate ACKs can be caused by the re-ordering of
423 data segments by the network (not a rare event along some network
424 paths [<a href="#ref-Pax97" title="&quot;End-to-End Internet Packet Dynamics&quot;">Pax97</a>]). Finally, duplicate ACKs can be caused by replication
425 of ACK or data segments by the network. In addition, a TCP receiver
426 SHOULD send an immediate ACK when the incoming segment fills in all
427 or part of a gap in the sequence space. This will generate more
428 timely information for a sender recovering from a loss through a
429 retransmission timeout, a fast retransmit, or an advanced loss
430 recovery algorithm, as outlined in <a href="#section-4.3">section 4.3</a>.
431
432 The TCP sender SHOULD use the "fast retransmit" algorithm to detect
433 and repair loss, based on incoming duplicate ACKs. The fast
434 retransmit algorithm uses the arrival of 3 duplicate ACKs (as defined
435 in <a href="#section-2">section 2</a>, without any intervening ACKs which move SND.UNA) as an
436 indication that a segment has been lost. After receiving 3 duplicate
437 ACKs, TCP performs a retransmission of what appears to be the missing
438 segment, without waiting for the retransmission timer to expire.
439
440
441
442
443
444 <span class="grey">Allman, et al. Standards Track [Page 8]</span></pre>
445 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-9" ></span>
446 <span class="grey"><a href="./rfc5681">RFC 5681</a> TCP Congestion Control September 2009</span>
447
448
449 After the fast retransmit algorithm sends what appears to be the
450 missing segment, the "fast recovery" algorithm governs the
451 transmission of new data until a non-duplicate ACK arrives. The
452 reason for not performing slow start is that the receipt of the
453 duplicate ACKs not only indicates that a segment has been lost, but
454 also that segments are most likely leaving the network (although a
455 massive segment duplication by the network can invalidate this
456 conclusion). In other words, since the receiver can only generate a
457 duplicate ACK when a segment has arrived, that segment has left the
458 network and is in the receiver's buffer, so we know it is no longer
459 consuming network resources. Furthermore, since the ACK "clock"
460 [<a href="#ref-Jac88" title="&quot;Congestion Avoidance and Control&quot;">Jac88</a>] is preserved, the TCP sender can continue to transmit new
461 segments (although transmission must continue using a reduced cwnd,
462 since loss is an indication of congestion).
463
464 The fast retransmit and fast recovery algorithms are implemented
465 together as follows.
466
467 1. On the first and second duplicate ACKs received at a sender, a
468 TCP SHOULD send a segment of previously unsent data per [<a href="./rfc3042" title="&quot;Enhancing TCP's Loss Recovery Using Limited Transmit&quot;">RFC3042</a>]
469 provided that the receiver's advertised window allows, the total
470 FlightSize would remain less than or equal to cwnd plus 2*SMSS,
471 and that new data is available for transmission. Further, the
472 TCP sender MUST NOT change cwnd to reflect these two segments
473 [<a href="./rfc3042" title="&quot;Enhancing TCP's Loss Recovery Using Limited Transmit&quot;">RFC3042</a>]. Note that a sender using SACK [<a href="./rfc2018" title="&quot;TCP Selective Acknowledgment Options&quot;">RFC2018</a>] MUST NOT send
474 new data unless the incoming duplicate acknowledgment contains
475 new SACK information.
476
477 2. When the third duplicate ACK is received, a TCP MUST set ssthresh
478 to no more than the value given in equation (4). When [<a href="./rfc3042" title="&quot;Enhancing TCP's Loss Recovery Using Limited Transmit&quot;">RFC3042</a>]
479 is in use, additional data sent in limited transmit MUST NOT be
480 included in this calculation.
481
482 3. The lost segment starting at SND.UNA MUST be retransmitted and
483 cwnd set to ssthresh plus 3*SMSS. This artificially "inflates"
484 the congestion window by the number of segments (three) that have
485 left the network and which the receiver has buffered.
486
487 4. For each additional duplicate ACK received (after the third),
488 cwnd MUST be incremented by SMSS. This artificially inflates the
489 congestion window in order to reflect the additional segment that
490 has left the network.
491
492 Note: [<a href="#ref-SCWA99" title="&quot;TCP Congestion Control With a Misbehaving Receiver&quot;">SCWA99</a>] discusses a receiver-based attack whereby many
493 bogus duplicate ACKs are sent to the data sender in order to
494 artificially inflate cwnd and cause a higher than appropriate
495
496
497
498
499
500 <span class="grey">Allman, et al. Standards Track [Page 9]</span></pre>
501 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-10" ></span>
502 <span class="grey"><a href="./rfc5681">RFC 5681</a> TCP Congestion Control September 2009</span>
503
504
505 sending rate to be used. A TCP MAY therefore limit the number of
506 times cwnd is artificially inflated during loss recovery to the
507 number of outstanding segments (or, an approximation thereof).
508
509 Note: When an advanced loss recovery mechanism (such as outlined
510 in <a href="#section-4.3">section 4.3</a>) is not in use, this increase in FlightSize can
511 cause equation (4) to slightly inflate cwnd and ssthresh, as some
512 of the segments between SND.UNA and SND.NXT are assumed to have
513 left the network but are still reflected in FlightSize.
514
515 5. When previously unsent data is available and the new value of
516 cwnd and the receiver's advertised window allow, a TCP SHOULD
517 send 1*SMSS bytes of previously unsent data.
518
519 6. When the next ACK arrives that acknowledges previously
520 unacknowledged data, a TCP MUST set cwnd to ssthresh (the value
521 set in step 2). This is termed "deflating" the window.
522
523 This ACK should be the acknowledgment elicited by the
524 retransmission from step 3, one RTT after the retransmission
525 (though it may arrive sooner in the presence of significant out-
526 of-order delivery of data segments at the receiver).
527 Additionally, this ACK should acknowledge all the intermediate
528 segments sent between the lost segment and the receipt of the
529 third duplicate ACK, if none of these were lost.
530
531 Note: This algorithm is known to generally not recover efficiently
532 from multiple losses in a single flight of packets [<a href="#ref-FF96" title="&quot;Simulation-based Comparisons of Tahoe, Reno and SACK TCP&quot;">FF96</a>]. <a href="#section-4.3">Section</a>
533 <a href="#section-4.3">4.3</a> below addresses such cases.
534
535 <span class="h2"><a class="selflink" id="section-4" href="#section-4">4</a>. Additional Considerations</span>
536
537 <span class="h3"><a class="selflink" id="section-4.1" href="#section-4.1">4.1</a>. Restarting Idle Connections</span>
538
539 A known problem with the TCP congestion control algorithms described
540 above is that they allow a potentially inappropriate burst of traffic
541 to be transmitted after TCP has been idle for a relatively long
542 period of time. After an idle period, TCP cannot use the ACK clock
543 to strobe new segments into the network, as all the ACKs have drained
544 from the network. Therefore, as specified above, TCP can potentially
545 send a cwnd-size line-rate burst into the network after an idle
546 period. In addition, changing network conditions may have rendered
547 TCP's notion of the available end-to-end network capacity between two
548 endpoints, as estimated by cwnd, inaccurate during the course of a
549 long idle period.
550
551
552
553
554
555
556 <span class="grey">Allman, et al. Standards Track [Page 10]</span></pre>
557 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-11" ></span>
558 <span class="grey"><a href="./rfc5681">RFC 5681</a> TCP Congestion Control September 2009</span>
559
560
561 [<a id="ref-Jac88">Jac88</a>] recommends that a TCP use slow start to restart transmission
562 after a relatively long idle period. Slow start serves to restart
563 the ACK clock, just as it does at the beginning of a transfer. This
564 mechanism has been widely deployed in the following manner. When TCP
565 has not received a segment for more than one retransmission timeout,
566 cwnd is reduced to the value of the restart window (RW) before
567 transmission begins.
568
569 For the purposes of this standard, we define RW = min(IW,cwnd).
570
571 Using the last time a segment was received to determine whether or
572 not to decrease cwnd can fail to deflate cwnd in the common case of
573 persistent HTTP connections [<a href="#ref-HTH98" title="&quot;Issues in TCP Slow-Start Restart After Idle&quot;">HTH98</a>]. In this case, a Web server
574 receives a request before transmitting data to the Web client. The
575 reception of the request makes the test for an idle connection fail,
576 and allows the TCP to begin transmission with a possibly
577 inappropriately large cwnd.
578
579 Therefore, a TCP SHOULD set cwnd to no more than RW before beginning
580 transmission if the TCP has not sent data in an interval exceeding
581 the retransmission timeout.
582
583 <span class="h3"><a class="selflink" id="section-4.2" href="#section-4.2">4.2</a>. Generating Acknowledgments</span>
584
585 The delayed ACK algorithm specified in [<a href="./rfc1122" title="&quot;Requirements for Internet Hosts - Communication Layers&quot;">RFC1122</a>] SHOULD be used by a
586 TCP receiver. When using delayed ACKs, a TCP receiver MUST NOT
587 excessively delay acknowledgments. Specifically, an ACK SHOULD be
588 generated for at least every second full-sized segment, and MUST be
589 generated within 500 ms of the arrival of the first unacknowledged
590 packet.
591
592 The requirement that an ACK "SHOULD" be generated for at least every
593 second full-sized segment is listed in [<a href="./rfc1122" title="&quot;Requirements for Internet Hosts - Communication Layers&quot;">RFC1122</a>] in one place as a
594 SHOULD and another as a MUST. Here we unambiguously state it is a
595 SHOULD. We also emphasize that this is a SHOULD, meaning that an
596 implementor should indeed only deviate from this requirement after
597 careful consideration of the implications. See the discussion of
598 "Stretch ACK violation" in [<a href="./rfc2525" title="&quot;Known TCP Implementation Problems&quot;">RFC2525</a>] and the references therein for a
599 discussion of the possible performance problems with generating ACKs
600 less frequently than every second full-sized segment.
601
602 In some cases, the sender and receiver may not agree on what
603 constitutes a full-sized segment. An implementation is deemed to
604 comply with this requirement if it sends at least one acknowledgment
605 every time it receives 2*RMSS bytes of new data from the sender,
606 where RMSS is the Maximum Segment Size specified by the receiver to
607 the sender (or the default value of 536 bytes, per [<a href="./rfc1122" title="&quot;Requirements for Internet Hosts - Communication Layers&quot;">RFC1122</a>], if the
608 receiver does not specify an MSS option during connection
609
610
611
612 <span class="grey">Allman, et al. Standards Track [Page 11]</span></pre>
613 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-12" ></span>
614 <span class="grey"><a href="./rfc5681">RFC 5681</a> TCP Congestion Control September 2009</span>
615
616
617 establishment). The sender may be forced to use a segment size less
618 than RMSS due to the maximum transmission unit (MTU), the path MTU
619 discovery algorithm or other factors. For instance, consider the
620 case when the receiver announces an RMSS of X bytes but the sender
621 ends up using a segment size of Y bytes (Y &lt; X) due to path MTU
622 discovery (or the sender's MTU size). The receiver will generate
623 stretch ACKs if it waits for 2*X bytes to arrive before an ACK is
624 sent. Clearly this will take more than 2 segments of size Y bytes.
625 Therefore, while a specific algorithm is not defined, it is desirable
626 for receivers to attempt to prevent this situation, for example, by
627 acknowledging at least every second segment, regardless of size.
628 Finally, we repeat that an ACK MUST NOT be delayed for more than 500
629 ms waiting on a second full-sized segment to arrive.
630
631 Out-of-order data segments SHOULD be acknowledged immediately, in
632 order to accelerate loss recovery. To trigger the fast retransmit
633 algorithm, the receiver SHOULD send an immediate duplicate ACK when
634 it receives a data segment above a gap in the sequence space. To
635 provide feedback to senders recovering from losses, the receiver
636 SHOULD send an immediate ACK when it receives a data segment that
637 fills in all or part of a gap in the sequence space.
638
639 A TCP receiver MUST NOT generate more than one ACK for every incoming
640 segment, other than to update the offered window as the receiving
641 application consumes new data (see [<a href="./rfc813" title="&quot;Window and Acknowledgement Strategy in TCP&quot;">RFC813</a>] and page 42 of [<a href="./rfc793" title="&quot;Transmission Control Protocol&quot;">RFC793</a>]).
642
643 <span class="h3"><a class="selflink" id="section-4.3" href="#section-4.3">4.3</a>. Loss Recovery Mechanisms</span>
644
645 A number of loss recovery algorithms that augment fast retransmit and
646 fast recovery have been suggested by TCP researchers and specified in
647 the RFC series. While some of these algorithms are based on the TCP
648 selective acknowledgment (SACK) option [<a href="./rfc2018" title="&quot;TCP Selective Acknowledgment Options&quot;">RFC2018</a>], such as [<a href="#ref-FF96" title="&quot;Simulation-based Comparisons of Tahoe, Reno and SACK TCP&quot;">FF96</a>],
649 [<a href="#ref-MM96a" title="&quot;Forward Acknowledgment: Refining TCP Congestion Control&quot;">MM96a</a>], [<a href="#ref-MM96b" title="&quot;TCP Rate-Halving with Bounding Parameters&quot;">MM96b</a>], and [<a href="./rfc3517" title="&quot;A Conservative Selective Acknowledgment (SACK)-based Loss Recovery Algorithm for TCP&quot;">RFC3517</a>], others do not require SACKs, such as
650 [<a href="#ref-Hoe96" title="&quot;Improving the Start-up Behavior of a Congestion Control Scheme for TCP&quot;">Hoe96</a>], [<a href="#ref-FF96" title="&quot;Simulation-based Comparisons of Tahoe, Reno and SACK TCP&quot;">FF96</a>], and [<a href="./rfc3782" title="&quot;The NewReno Modification to TCP's Fast Recovery Algorithm&quot;">RFC3782</a>]. The non-SACK algorithms use "partial
651 acknowledgments" (ACKs that cover previously unacknowledged data, but
652 not all the data outstanding when loss was detected) to trigger
653 retransmissions. While this document does not standardize any of the
654 specific algorithms that may improve fast retransmit/fast recovery,
655 these enhanced algorithms are implicitly allowed, as long as they
656 follow the general principles of the basic four algorithms outlined
657 above.
658
659 That is, when the first loss in a window of data is detected,
660 ssthresh MUST be set to no more than the value given by equation (4).
661 Second, until all lost segments in the window of data in question are
662 repaired, the number of segments transmitted in each RTT MUST be no
663 more than half the number of outstanding segments when the loss was
664 detected. Finally, after all loss in the given window of segments
665
666
667
668 <span class="grey">Allman, et al. Standards Track [Page 12]</span></pre>
669 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-13" ></span>
670 <span class="grey"><a href="./rfc5681">RFC 5681</a> TCP Congestion Control September 2009</span>
671
672
673 has been successfully retransmitted, cwnd MUST be set to no more than
674 ssthresh and congestion avoidance MUST be used to further increase
675 cwnd. Loss in two successive windows of data, or the loss of a
676 retransmission, should be taken as two indications of congestion and,
677 therefore, cwnd (and ssthresh) MUST be lowered twice in this case.
678
679 We RECOMMEND that TCP implementors employ some form of advanced loss
680 recovery that can cope with multiple losses in a window of data. The
681 algorithms detailed in [<a href="./rfc3782" title="&quot;The NewReno Modification to TCP's Fast Recovery Algorithm&quot;">RFC3782</a>] and [<a href="./rfc3517" title="&quot;A Conservative Selective Acknowledgment (SACK)-based Loss Recovery Algorithm for TCP&quot;">RFC3517</a>] conform to the general
682 principles outlined above. We note that while these are not the only
683 two algorithms that conform to the above general principles these two
684 algorithms have been vetted by the community and are currently on the
685 Standards Track.
686
687 <span class="h2"><a class="selflink" id="section-5" href="#section-5">5</a>. Security Considerations</span>
688
689 This document requires a TCP to diminish its sending rate in the
690 presence of retransmission timeouts and the arrival of duplicate
691 acknowledgments. An attacker can therefore impair the performance of
692 a TCP connection by either causing data packets or their
693 acknowledgments to be lost, or by forging excessive duplicate
694 acknowledgments.
695
696 In response to the ACK division attack outlined in [<a href="#ref-SCWA99" title="&quot;TCP Congestion Control With a Misbehaving Receiver&quot;">SCWA99</a>], this
697 document RECOMMENDS increasing the congestion window based on the
698 number of bytes newly acknowledged in each arriving ACK rather than
699 by a particular constant on each arriving ACK (as outlined in <a href="#section-3.1">section</a>
700 <a href="#section-3.1">3.1</a>).
701
702 The Internet, to a considerable degree, relies on the correct
703 implementation of these algorithms in order to preserve network
704 stability and avoid congestion collapse. An attacker could cause TCP
705 endpoints to respond more aggressively in the face of congestion by
706 forging excessive duplicate acknowledgments or excessive
707 acknowledgments for new data. Conceivably, such an attack could
708 drive a portion of the network into congestion collapse.
709
710 <span class="h2"><a class="selflink" id="section-6" href="#section-6">6</a>. Changes between <a href="./rfc2001">RFC 2001</a> and <a href="./rfc2581">RFC 2581</a></span>
711
712 [<a id="ref-RFC2001">RFC2001</a>] was extensively rewritten editorially and it is not
713 feasible to itemize the list of changes between [<a href="./rfc2001" title="&quot;TCP Slow Start, Congestion Avoidance, Fast Retransmit, and Fast Recovery Algorithms&quot;">RFC2001</a>] and
714 [<a href="./rfc2581" title="&quot;TCP Congestion Control&quot;">RFC2581</a>]. The intention of [<a href="./rfc2581" title="&quot;TCP Congestion Control&quot;">RFC2581</a>] was to not change any of the
715 recommendations given in [<a href="./rfc2001" title="&quot;TCP Slow Start, Congestion Avoidance, Fast Retransmit, and Fast Recovery Algorithms&quot;">RFC2001</a>], but to further clarify cases that
716 were not discussed in detail in [<a href="./rfc2001" title="&quot;TCP Slow Start, Congestion Avoidance, Fast Retransmit, and Fast Recovery Algorithms&quot;">RFC2001</a>]. Specifically, [<a href="./rfc2581" title="&quot;TCP Congestion Control&quot;">RFC2581</a>]
717 suggested what TCP connections should do after a relatively long idle
718 period, as well as specified and clarified some of the issues
719
720
721
722
723
724 <span class="grey">Allman, et al. Standards Track [Page 13]</span></pre>
725 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-14" ></span>
726 <span class="grey"><a href="./rfc5681">RFC 5681</a> TCP Congestion Control September 2009</span>
727
728
729 pertaining to TCP ACK generation. Finally, the allowable upper bound
730 for the initial congestion window was raised from one to two
731 segments.
732
733 <span class="h2"><a class="selflink" id="section-7" href="#section-7">7</a>. Changes Relative to <a href="./rfc2581">RFC 2581</a></span>
734
735 A specific definition for "duplicate acknowledgment" has been added,
736 based on the definition used by BSD TCP.
737
738 The document now notes that what to do with duplicate ACKs after the
739 retransmission timer has fired is future work and explicitly
740 unspecified in this document.
741
742 The initial window requirements were changed to allow Larger Initial
743 Windows as standardized in [<a href="./rfc3390" title="&quot;Increasing TCP's Initial Window&quot;">RFC3390</a>]. Additionally, the steps to
744 take when an initial window is discovered to be too large due to Path
745 MTU Discovery [<a href="./rfc1191" title="&quot;Path MTU discovery&quot;">RFC1191</a>] are detailed.
746
747 The recommended initial value for ssthresh has been changed to say
748 that it SHOULD be arbitrarily high, where it was previously MAY.
749 This is to provide additional guidance to implementors on the matter.
750
751 During slow start, the usage of Appropriate Byte Counting [<a href="./rfc3465" title="&quot;TCP Congestion Control with Appropriate Byte Counting (ABC)&quot;">RFC3465</a>]
752 with L=1*SMSS is explicitly recommended. The method of increasing
753 cwnd given in [<a href="./rfc2581" title="&quot;TCP Congestion Control&quot;">RFC2581</a>] is still explicitly allowed. Byte counting
754 during congestion avoidance is also recommended, while the method
755 from [<a href="./rfc2581" title="&quot;TCP Congestion Control&quot;">RFC2581</a>] and other safe methods are still allowed.
756
757 The treatment of ssthresh on retransmission timeout was clarified.
758 In particular, ssthresh must be set to half the FlightSize on the
759 first retransmission of a given segment and then is held constant on
760 subsequent retransmissions of the same segment.
761
762 The description of fast retransmit and fast recovery has been
763 clarified, and the use of Limited Transmit [<a href="./rfc3042" title="&quot;Enhancing TCP's Loss Recovery Using Limited Transmit&quot;">RFC3042</a>] is now
764 recommended.
765
766 TCPs now MAY limit the number of duplicate ACKs that artificially
767 inflate cwnd during loss recovery to the number of segments
768 outstanding to avoid the duplicate ACK spoofing attack described in
769 [<a href="#ref-SCWA99" title="&quot;TCP Congestion Control With a Misbehaving Receiver&quot;">SCWA99</a>].
770
771 The restart window has been changed to min(IW,cwnd) from IW. This
772 behavior was described as "experimental" in [<a href="./rfc2581" title="&quot;TCP Congestion Control&quot;">RFC2581</a>].
773
774 It is now recommended that TCP implementors implement an advanced
775 loss recovery algorithm conforming to the principles outlined in this
776 document.
777
778
779
780 <span class="grey">Allman, et al. Standards Track [Page 14]</span></pre>
781 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-15" ></span>
782 <span class="grey"><a href="./rfc5681">RFC 5681</a> TCP Congestion Control September 2009</span>
783
784
785 The security considerations have been updated to discuss ACK division
786 and recommend byte counting as a counter to this attack.
787
788 <span class="h2"><a class="selflink" id="section-8" href="#section-8">8</a>. Acknowledgments</span>
789
790 The core algorithms we describe were developed by Van Jacobson
791 [<a href="#ref-Jac88" title="&quot;Congestion Avoidance and Control&quot;">Jac88</a>, <a href="#ref-Jac90" title="&quot;Modified TCP Congestion Avoidance Algorithm&quot;">Jac90</a>]. In addition, Limited Transmit [<a href="./rfc3042" title="&quot;Enhancing TCP's Loss Recovery Using Limited Transmit&quot;">RFC3042</a>] was
792 developed in conjunction with Hari Balakrishnan and Sally Floyd. The
793 initial congestion window size specified in this document is a result
794 of work with Sally Floyd and Craig Partridge [RFC2414, <a href="./rfc3390">RFC3390</a>].
795
796 W. Richard ("Rich") Stevens wrote the first version of this document
797 [<a href="./rfc2001" title="&quot;TCP Slow Start, Congestion Avoidance, Fast Retransmit, and Fast Recovery Algorithms&quot;">RFC2001</a>] and co-authored the second version [<a href="./rfc2581" title="&quot;TCP Congestion Control&quot;">RFC2581</a>]. This present
798 version much benefits from his clarity and thoughtfulness of
799 description, and we are grateful for Rich's contributions in
800 elucidating TCP congestion control, as well as in more broadly
801 helping us understand numerous issues relating to networking.
802
803 We wish to emphasize that the shortcomings and mistakes of this
804 document are solely the responsibility of the current authors.
805
806 Some of the text from this document is taken from "TCP/IP
807 Illustrated, Volume 1: The Protocols" by W. Richard Stevens
808 (Addison-Wesley, 1994) and "TCP/IP Illustrated, Volume 2: The
809 Implementation" by Gary R. Wright and W. Richard Stevens (Addison-
810 Wesley, 1995). This material is used with the permission of
811 Addison-Wesley.
812
813 Anil Agarwal, Steve Arden, Neal Cardwell, Noritoshi Demizu, Gorry
814 Fairhurst, Kevin Fall, John Heffner, Alfred Hoenes, Sally Floyd,
815 Reiner Ludwig, Matt Mathis, Craig Partridge, and Joe Touch
816 contributed a number of helpful suggestions.
817
818 <span class="h2"><a class="selflink" id="section-9" href="#section-9">9</a>. References</span>
819
820 <span class="h3"><a class="selflink" id="section-9.1" href="#section-9.1">9.1</a>. Normative References</span>
821
822 [<a id="ref-RFC793">RFC793</a>] Postel, J., "Transmission Control Protocol", STD 7, <a href="./rfc793">RFC</a>
823 <a href="./rfc793">793</a>, September 1981.
824
825 [<a id="ref-RFC1122">RFC1122</a>] Braden, R., Ed., "Requirements for Internet Hosts -
826 Communication Layers", STD 3, <a href="./rfc1122">RFC 1122</a>, October 1989.
827
828 [<a id="ref-RFC1191">RFC1191</a>] Mogul, J. and S. Deering, "Path MTU discovery", <a href="./rfc1191">RFC 1191</a>,
829 November 1990.
830
831 [<a id="ref-RFC2119">RFC2119</a>] Bradner, S., "Key words for use in RFCs to Indicate
832 Requirement Levels", <a href="https://www.rfc-editor.org/bcp/bcp14">BCP 14</a>, <a href="./rfc2119">RFC 2119</a>, March 1997.
833
834
835
836 <span class="grey">Allman, et al. Standards Track [Page 15]</span></pre>
837 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-16" ></span>
838 <span class="grey"><a href="./rfc5681">RFC 5681</a> TCP Congestion Control September 2009</span>
839
840
841 <span class="h3"><a class="selflink" id="section-9.2" href="#section-9.2">9.2</a>. Informative References</span>
842
843 [<a id="ref-CJ89">CJ89</a>] Chiu, D. and R. Jain, "Analysis of the Increase/Decrease
844 Algorithms for Congestion Avoidance in Computer Networks",
845 Journal of Computer Networks and ISDN Systems, vol. 17, no.
846 1, pp. 1-14, June 1989.
847
848 [<a id="ref-FF96">FF96</a>] Fall, K. and S. Floyd, "Simulation-based Comparisons of
849 Tahoe, Reno and SACK TCP", Computer Communication Review,
850 July 1996, <a href="ftp://ftp.ee.lbl.gov/papers/sacks.ps.Z">ftp://ftp.ee.lbl.gov/papers/sacks.ps.Z</a>.
851
852 [<a id="ref-Hoe96">Hoe96</a>] Hoe, J., "Improving the Start-up Behavior of a Congestion
853 Control Scheme for TCP", In ACM SIGCOMM, August 1996.
854
855 [<a id="ref-HTH98">HTH98</a>] Hughes, A., Touch, J., and J. Heidemann, "Issues in TCP
856 Slow-Start Restart After Idle", Work in Progress, March
857 1998.
858
859 [<a id="ref-Jac88">Jac88</a>] Jacobson, V., "Congestion Avoidance and Control", Computer
860 Communication Review, vol. 18, no. 4, pp. 314-329, Aug.
861 1988. <a href="ftp://ftp.ee.lbl.gov/papers/congavoid.ps.Z">ftp://ftp.ee.lbl.gov/papers/congavoid.ps.Z</a>.
862
863 [<a id="ref-Jac90">Jac90</a>] Jacobson, V., "Modified TCP Congestion Avoidance
864 Algorithm", end2end-interest mailing list, April 30, 1990.
865 <a href="ftp://ftp.isi.edu/end2end/end2end-interest-1990.mail">ftp://ftp.isi.edu/end2end/end2end-interest-1990.mail</a>.
866
867 [<a id="ref-MM96a">MM96a</a>] Mathis, M. and J. Mahdavi, "Forward Acknowledgment:
868 Refining TCP Congestion Control", Proceedings of
869 SIGCOMM'96, August, 1996, Stanford, CA. Available from
870 <a href="http://www.psc.edu/networking/papers/papers.html">http://www.psc.edu/networking/papers/papers.html</a>
871
872 [<a id="ref-MM96b">MM96b</a>] Mathis, M. and J. Mahdavi, "TCP Rate-Halving with Bounding
873 Parameters", Technical report. Available from
874 <a href="http://www.psc.edu/networking/papers/FACKnotes/current">http://www.psc.edu/networking/papers/FACKnotes/current</a>.
875
876 [<a id="ref-Pax97">Pax97</a>] Paxson, V., "End-to-End Internet Packet Dynamics",
877 Proceedings of SIGCOMM '97, Cannes, France, Sep. 1997.
878
879 [<a id="ref-RFC813">RFC813</a>] Clark, D., "Window and Acknowledgement Strategy in TCP",
880 <a href="./rfc813">RFC 813</a>, July 1982.
881
882 [<a id="ref-RFC2001">RFC2001</a>] Stevens, W., "TCP Slow Start, Congestion Avoidance, Fast
883 Retransmit, and Fast Recovery Algorithms", <a href="./rfc2001">RFC 2001</a>,
884 January 1997.
885
886 [<a id="ref-RFC2018">RFC2018</a>] Mathis, M., Mahdavi, J., Floyd, S., and A. Romanow, "TCP
887 Selective Acknowledgment Options", <a href="./rfc2018">RFC 2018</a>, October 1996.
888
889
890
891
892 <span class="grey">Allman, et al. Standards Track [Page 16]</span></pre>
893 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-17" ></span>
894 <span class="grey"><a href="./rfc5681">RFC 5681</a> TCP Congestion Control September 2009</span>
895
896
897 [<a id="ref-RFC2414">RFC2414</a>] Allman, M., Floyd, S., and C. Partridge, "Increasing TCP's
898 Initial Window", <a href="./rfc2414">RFC 2414</a>, September 1998.
899
900 [<a id="ref-RFC2525">RFC2525</a>] Paxson, V., Allman, M., Dawson, S., Fenner, W., Griner, J.,
901 Heavens, I., Lahey, K., Semke, J., and B. Volz, "Known TCP
902 Implementation Problems", <a href="./rfc2525">RFC 2525</a>, March 1999.
903
904 [<a id="ref-RFC2581">RFC2581</a>] Allman, M., Paxson, V., and W. Stevens, "TCP Congestion
905 Control", <a href="./rfc2581">RFC 2581</a>, April 1999.
906
907 [<a id="ref-RFC2883">RFC2883</a>] Floyd, S., Mahdavi, J., Mathis, M., and M. Podolsky, "An
908 Extension to the Selective Acknowledgement (SACK) Option
909 for TCP", <a href="./rfc2883">RFC 2883</a>, July 2000.
910
911 [<a id="ref-RFC2988">RFC2988</a>] Paxson, V. and M. Allman, "Computing TCP's Retransmission
912 Timer", <a href="./rfc2988">RFC 2988</a>, November 2000.
913
914 [<a id="ref-RFC3042">RFC3042</a>] Allman, M., Balakrishnan, H., and S. Floyd, "Enhancing
915 TCP's Loss Recovery Using Limited Transmit", <a href="./rfc3042">RFC 3042</a>,
916 January 2001.
917
918 [<a id="ref-RFC3168">RFC3168</a>] Ramakrishnan, K., Floyd, S., and D. Black, "The Addition of
919 Explicit Congestion Notification (ECN) to IP", <a href="./rfc3168">RFC 3168</a>,
920 September 2001.
921
922 [<a id="ref-RFC3390">RFC3390</a>] Allman, M., Floyd, S., and C. Partridge, "Increasing TCP's
923 Initial Window", <a href="./rfc3390">RFC 3390</a>, October 2002.
924
925 [<a id="ref-RFC3465">RFC3465</a>] Allman, M., "TCP Congestion Control with Appropriate Byte
926 Counting (ABC)", <a href="./rfc3465">RFC 3465</a>, February 2003.
927
928 [<a id="ref-RFC3517">RFC3517</a>] Blanton, E., Allman, M., Fall, K., and L. Wang, "A
929 Conservative Selective Acknowledgment (SACK)-based Loss
930 Recovery Algorithm for TCP", <a href="./rfc3517">RFC 3517</a>, April 2003.
931
932 [<a id="ref-RFC3782">RFC3782</a>] Floyd, S., Henderson, T., and A. Gurtov, "The NewReno
933 Modification to TCP's Fast Recovery Algorithm", <a href="./rfc3782">RFC 3782</a>,
934 April 2004.
935
936 [<a id="ref-RFC4821">RFC4821</a>] Mathis, M. and J. Heffner, "Packetization Layer Path MTU
937 Discovery", <a href="./rfc4821">RFC 4821</a>, March 2007.
938
939 [<a id="ref-SCWA99">SCWA99</a>] Savage, S., Cardwell, N., Wetherall, D., and T. Anderson,
940 "TCP Congestion Control With a Misbehaving Receiver", ACM
941 Computer Communication Review, 29(5), October 1999.
942
943 [<a id="ref-Ste94">Ste94</a>] Stevens, W., "TCP/IP Illustrated, Volume 1: The Protocols",
944 Addison-Wesley, 1994.
945
946
947
948 <span class="grey">Allman, et al. Standards Track [Page 17]</span></pre>
949 <hr class='noprint'/><!--NewPage--><pre class='newpage'><span id="page-18" ></span>
950 <span class="grey"><a href="./rfc5681">RFC 5681</a> TCP Congestion Control September 2009</span>
951
952
953 [<a id="ref-WS95">WS95</a>] Wright, G. and W. Stevens, "TCP/IP Illustrated, Volume 2:
954 The Implementation", Addison-Wesley, 1995.
955
956 Authors' Addresses
957
958 Mark Allman
959 International Computer Science Institute (ICSI)
960 1947 Center Street
961 Suite 600
962 Berkeley, CA 94704-1198
963 Phone: +1 440 235 1792
964 EMail: [email protected]
965 <a href="http://www.icir.org/mallman/">http://www.icir.org/mallman/</a>
966
967
968 Vern Paxson
969 International Computer Science Institute (ICSI)
970 1947 Center Street
971 Suite 600
972 Berkeley, CA 94704-1198
973 Phone: +1 510/642-4274 x302
974 EMail: [email protected]
975 <a href="http://www.icir.org/vern/">http://www.icir.org/vern/</a>
976
977
978 Ethan Blanton
979 Purdue University Computer Sciences
980 305 North University Street
981 West Lafayette, IN 47907
982 EMail: [email protected]
983 <a href="http://www.cs.purdue.edu/homes/eblanton/">http://www.cs.purdue.edu/homes/eblanton/</a>
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004 Allman, et al. Standards Track [Page 18]
1005 </pre>