idnits 2.17.1 draft-ietf-mptcp-experience-04.txt: 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: ---------------------------------------------------------------------------- No issues found here. Checking nits according to https://www.ietf.org/id-info/checklist : ---------------------------------------------------------------------------- ** The document seems to lack a Security Considerations section. ** The document seems to lack an IANA Considerations section. (See Section 2.2 of https://www.ietf.org/id-info/checklist for how to handle the case when there are no actions for IANA.) Miscellaneous warnings: ---------------------------------------------------------------------------- == The copyright year in the IETF Trust and authors Copyright Line does not match the current year -- The document date (April 01, 2016) is 2918 days in the past. Is this intentional? Checking references for intended status: Informational ---------------------------------------------------------------------------- == Outdated reference: A later version (-03) exists of draft-boucadair-mptcp-max-subflow-01 == Outdated reference: A later version (-08) exists of draft-ietf-dnsop-edns-client-subnet-07 -- Obsolete informational reference (is this intentional?): RFC 6824 (Obsoleted by RFC 8684) Summary: 2 errors (**), 0 flaws (~~), 3 warnings (==), 2 comments (--). Run idnits with the --verbose option for more detailed information about the items above. -------------------------------------------------------------------------------- 2 MPTCP Working Group O. Bonaventure 3 Internet-Draft UCLouvain 4 Intended status: Informational C. Paasch 5 Expires: October 3, 2016 Apple, Inc. 6 G. Detal 7 Tessares 8 April 01, 2016 10 Use Cases and Operational Experience with Multipath TCP 11 draft-ietf-mptcp-experience-04 13 Abstract 15 This document discusses both use cases and operational experience 16 with Multipath TCP in real world networks. It lists several 17 prominent use cases for which Multipath TCP has been considered and 18 is being used. It also gives insight to some heuristics and 19 decisions that have helped to realize these use cases. 21 Status of This Memo 23 This Internet-Draft is submitted in full conformance with the 24 provisions of BCP 78 and BCP 79. 26 Internet-Drafts are working documents of the Internet Engineering 27 Task Force (IETF). Note that other groups may also distribute 28 working documents as Internet-Drafts. The list of current Internet- 29 Drafts is at http://datatracker.ietf.org/drafts/current/. 31 Internet-Drafts are draft documents valid for a maximum of six months 32 and may be updated, replaced, or obsoleted by other documents at any 33 time. It is inappropriate to use Internet-Drafts as reference 34 material or to cite them other than as "work in progress." 36 This Internet-Draft will expire on October 3, 2016. 38 Copyright Notice 40 Copyright (c) 2016 IETF Trust and the persons identified as the 41 document authors. All rights reserved. 43 This document is subject to BCP 78 and the IETF Trust's Legal 44 Provisions Relating to IETF Documents 45 (http://trustee.ietf.org/license-info) in effect on the date of 46 publication of this document. Please review these documents 47 carefully, as they describe your rights and restrictions with respect 48 to this document. Code Components extracted from this document must 49 include Simplified BSD License text as described in Section 4.e of 50 the Trust Legal Provisions and are provided without warranty as 51 described in the Simplified BSD License. 53 Table of Contents 55 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 2 56 2. Use cases . . . . . . . . . . . . . . . . . . . . . . . . . . 4 57 2.1. Datacenters . . . . . . . . . . . . . . . . . . . . . . . 4 58 2.2. Cellular/WiFi Offload . . . . . . . . . . . . . . . . . . 5 59 2.3. Multipath TCP proxies . . . . . . . . . . . . . . . . . . 8 60 3. Operational Experience . . . . . . . . . . . . . . . . . . . 9 61 3.1. Middlebox interference . . . . . . . . . . . . . . . . . 9 62 3.2. Congestion control . . . . . . . . . . . . . . . . . . . 11 63 3.3. Subflow management . . . . . . . . . . . . . . . . . . . 12 64 3.4. Implemented subflow managers . . . . . . . . . . . . . . 12 65 3.5. Subflow destination port . . . . . . . . . . . . . . . . 14 66 3.6. Closing subflows . . . . . . . . . . . . . . . . . . . . 15 67 3.7. Packet schedulers . . . . . . . . . . . . . . . . . . . . 17 68 3.8. Segment size selection . . . . . . . . . . . . . . . . . 17 69 3.9. Interactions with the Domain Name System . . . . . . . . 18 70 3.10. Captive portals . . . . . . . . . . . . . . . . . . . . . 19 71 3.11. Stateless webservers . . . . . . . . . . . . . . . . . . 20 72 3.12. Loadbalanced serverfarms . . . . . . . . . . . . . . . . 20 73 4. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . 21 74 5. Acknowledgements . . . . . . . . . . . . . . . . . . . . . . 21 75 6. Informative References . . . . . . . . . . . . . . . . . . . 22 76 Appendix A. Changelog . . . . . . . . . . . . . . . . . . . . . 28 77 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 29 79 1. Introduction 81 Multipath TCP was standardized in [RFC6824] and five independant 82 implementations have been developed 83 [I-D.eardley-mptcp-implementations-survey]. As of September 2015, 84 Multipath TCP has been or is being implemented on the following 85 platforms : 87 o Linux kernel [MultipathTCP-Linux] 89 o Apple iOS and MacOS [Apple-MPTCP] 91 o Citrix load balancers 93 o FreeBSD [FreeBSD-MPTCP] 95 o Oracle 96 The first three implementations 97 [I-D.eardley-mptcp-implementations-survey] are known to interoperate. 98 The last two are currently being tested and improved against the 99 Linux implementation. Three of these implementations are open- 100 source. Apple's implementation is widely deployed. 102 Since the publication of [RFC6824], experience has been gathered by 103 various network researchers and users about the operational issues 104 that arise when Multipath TCP is used in today's Internet. 106 When the MPTCP working group was created, several use cases for 107 Multipath TCP were identified [RFC6182]. Since then, other use cases 108 have been proposed and some have been tested and even deployed. We 109 describe these use cases in Section 2. 111 Section 3 focuses on the operational experience with Multipath TCP. 112 Most of this experience comes from the utilisation of the Multipath 113 TCP implementation in the Linux kernel [MultipathTCP-Linux]. This 114 open-source implementation has been downloaded and is used by 115 thousands of users all over the world. Many of these users have 116 provided direct or indirect feedback by writing documents (scientific 117 articles or blog messages) or posting to the mptcp-dev mailing list 118 (see https://listes-2.sipr.ucl.ac.be/sympa/arc/mptcp-dev ). This 119 Multipath TCP implementation is actively maintained and continuously 120 improved. It is used on various types of hosts, ranging from 121 smartphones or embedded routers to high-end servers. 123 The Multipath TCP implementation in the Linux kernel is not, by far, 124 the most widespread deployment of Multipath TCP. Since September 125 2013, Multipath TCP is also supported on smartphones and tablets 126 running iOS7 [IOS7]. There are likely hundreds of millions of 127 Multipath TCP enabled devices. However, this particular Multipath 128 TCP implementation is currently only used to support a single 129 application. Unfortunately, there is no public information about the 130 lessons learned from this large scale deployment. 132 Section 3 is organized as follows. Supporting the middleboxes was 133 one of the difficult issues in designing the Multipath TCP protocol. 134 We explain in Section 3.1 which types of middleboxes the Linux Kernel 135 implementation of Multipath TCP supports and how it reacts upon 136 encountering these. Section 3.2 summarises the MPTCP specific 137 congestion controls that have been implemented. Section 3.3 and 138 Section 3.7 discuss heuristics and issues with respect to subflow 139 management as well as the scheduling across the subflows. 140 Section 3.8 explains some problems that occurred with subflows having 141 different Maximum Segment Size (MSS) values. Section 3.9 presents 142 issues with respect to content delivery networks and suggests a 143 solution to this issue. Finally, Section 3.10 documents an issue 144 with captive portals where MPTCP will behave suboptimally. 146 2. Use cases 148 Multipath TCP has been tested in several use cases. There is already 149 an abundant scientific literature on Multipath TCP [MPTCPBIB]. 150 Several of the papers published in the scientific literature have 151 identified possible improvements that are worth being discussed here. 153 2.1. Datacenters 155 A first, although initially unexpected, documented use case for 156 Multipath TCP has been in datacenters [HotNets][SIGCOMM11]. Today's 157 datacenters are designed to provide several paths between single- 158 homed servers. The multiplicity of these paths comes from the 159 utilization of Equal Cost Multipath (ECMP) and other load balancing 160 techniques inside the datacenter. Most of the deployed load 161 balancing techniques in datacenters rely on hashes computed over the 162 five tuple. Thus all packets from the same TCP connection follow the 163 same path and so are not reordered. The results in [HotNets] 164 demonstrate by simulations that Multipath TCP can achieve a better 165 utilization of the available network by using multiple subflows for 166 each Multipath TCP session. Although [RFC6182] assumes that at least 167 one of the communicating hosts has several IP addresses, [HotNets] 168 demonstrates that Multipath TCP is beneficial when both hosts are 169 single-homed. This idea is analysed in more details in [SIGCOMM11] 170 where the Multipath TCP implementation in the Linux kernel is 171 modified to be able to use several subflows from the same IP address. 172 Measurements in a public datacenter show the quantitative benefits of 173 Multipath TCP [SIGCOMM11] in this environment. 175 Although ECMP is widely used inside datacenters, this is not the only 176 environment where there are different paths between a pair of hosts. 177 ECMP and other load balancing techniques such as Link Aggregation 178 Groups (LAG) are widely used in today's networks and having multiple 179 paths between a pair of single-homed hosts is becoming the norm 180 instead of the exception. Although these multiple paths have often 181 the same cost (from an IGP metrics viewpoint), they do not 182 necessarily have the same performance. For example, [IMC13c] reports 183 the results of a long measurement study showing that load balanced 184 Internet paths between that same pair of hosts can have huge delay 185 differences. 187 2.2. Cellular/WiFi Offload 189 A second use case that has been explored by several network 190 researchers is the cellular/WiFi offload use case. Smartphones or 191 other mobile devices equipped with two wireless interfaces are a very 192 common use case for Multipath TCP. In September 2015, this is also 193 the largest deployment of Multipath-TCP enabled devices [IOS7]. It 194 has been briefly discussed during IETF88 [ietf88], but there is no 195 published paper or report that analyses this deployment. For this 196 reason, we only discuss published papers that have mainly used the 197 Multipath TCP implementation in the Linux kernel for their 198 experiments. 200 The performance of Multipath TCP in wireless networks was briefly 201 evaluated in [NSDI12]. One experiment analyzes the performance of 202 Multipath TCP on a client with two wireless interfaces. This 203 evaluation shows that when the receive window is large, Multipath TCP 204 can efficiently use the two available links. However, if the window 205 becomes smaller, then packets sent on a slow path can block the 206 transmission of packets on a faster path. In some cases, the 207 performance of Multipath TCP over two paths can become lower than the 208 performance of regular TCP over the best performing path. Two 209 heuristics, reinjection and penalization, are proposed in [NSDI12] to 210 solve this identified performance problem. These two heuristics have 211 since been used in the Multipath TCP implementation in the Linux 212 kernel. [CONEXT13] explored the problem in more detail and revealed 213 some other scenarios where Multipath TCP can have difficulties in 214 efficiently pooling the available paths. Improvements to the 215 Multipath TCP implementation in the Linux kernel are proposed in 216 [CONEXT13] to cope with some of these problems. 218 The first experimental analysis of Multipath TCP in a public wireless 219 environment was presented in [Cellnet12]. These measurements explore 220 the ability of Multipath TCP to use two wireless networks (real WiFi 221 and 3G networks). Three modes of operation are compared. The first 222 mode of operation is the simultaneous use of the two wireless 223 networks. In this mode, Multipath TCP pools the available resources 224 and uses both wireless interfaces. This mode provides fast handover 225 from WiFi to cellular or the opposite when the user moves. 226 Measurements presented in [CACM14] show that the handover from one 227 wireless network to another is not an abrupt process. When a host 228 moves, there are regions where the quality of one of the wireless 229 networks is weaker than the other, but the host considers this 230 wireless network to still be up. When a mobile host enters such 231 regions, its ability to send packets over another wireless network is 232 important to ensure a smooth handover. This is clearly illustrated 233 from the packet trace discussed in [CACM14]. 235 Many cellular networks use volume-based pricing and users often 236 prefer to use unmetered WiFi networks when available instead of 237 metered cellular networks. [Cellnet12] implements support for the 238 MP_PRIO option to explore two other modes of operation. 240 In the backup mode, Multipath TCP opens a TCP subflow over each 241 interface, but the cellular interface is configured in backup mode. 242 This implies that data only flows over only the WiFi interface when 243 both interfaces are considered to be active. If the WiFi interface 244 fails, then the traffic switches quickly to the cellular interface, 245 ensuring a smooth handover from the user's viewpoint [Cellnet12]. 246 The cost of this approach is that the WiFi and cellular interfaces 247 are likely to remain active all the time since all subflows are 248 established over the two interfaces. 250 The single-path mode is slightly different. This mode benefits from 251 the break-before-make capability of Multipath TCP. When an MPTCP 252 session is established, a subflow is created over the WiFi interface. 253 No packet is sent over the cellular interface as long as the WiFi 254 interface remains up [Cellnet12]. This implies that the cellular 255 interface can remain idle and battery capacity is preserved. When 256 the WiFi interface fails, a new subflow is established over the 257 cellular interface in order to preserve the established Multipath TCP 258 sessions. Compared to the backup mode described earlier, 259 measurements reported in [Cellnet12] indicate that this mode of 260 operation is characterised by a throughput drop while the cellular 261 interface is brought up and the subflows are reestablished. 263 From a protocol viewpoint, [Cellnet12] discusses the problem posed by 264 the unreliability of the ADD_ADDR option and proposes a small 265 protocol extension to allow hosts to reliably exchange this option. 266 It would be useful to analyze packet traces to understand whether the 267 unreliability of the REMOVE_ADDR option poses an operational problem 268 in real deployments. 270 Another study of the performance of Multipath TCP in wireless 271 networks was reported in [IMC13b]. This study uses laptops connected 272 to various cellular ISPs and WiFi hotspots. It compares various file 273 transfer scenarios. [IMC13b] observes that 4-path MPTCP outperforms 274 2-path MPTCP, especially for larger files. The comparison between 275 LIA, OLIA and Reno does not reveal a significant performance 276 difference for file sizes smaller than 4MB. 278 A different study of the performance of Multipath TCP with two 279 wireless networks is presented in [INFOCOM14]. In this study the two 280 networks had different qualities : a good network and a lossy 281 network. When using two paths with different packet loss ratios, the 282 Multipath TCP congestion control scheme moves traffic away from the 283 lossy link that is considered to be congested. However, [INFOCOM14] 284 documents an interesting scenario that is summarised in Figure 1. 286 client ----------- path1 -------- server 287 | | 288 +--------------- path2 ------------+ 290 Figure 1: Simple network topology 292 Initially, the two paths have the same quality and Multipath TCP 293 distributes the load over both of them. During the transfer, the 294 second path becomes lossy, e.g. because the client moves. Multipath 295 TCP detects the packet losses and they are retransmitted over the 296 first path. This enables the data transfer to continue over the 297 first path. However, the subflow over the second path is still up 298 and transmits one packet from time to time. Although the N packets 299 have been acknowledged over the first subflow (at the MPTCP level), 300 they have not been acknowledged at the TCP level over the second 301 subflow. To preserve the continuity of the sequence numbers over the 302 second subflow, TCP will continue to retransmit these segments until 303 either they are acknowledged or the maximum number of retransmissions 304 is reached. This behavior is clearly inefficient and may lead to 305 blocking since the second subflow will consume window space to be 306 able to retransmit these packets. [INFOCOM14] proposes a new 307 Multipath TCP option to solve this problem. In practice, a new TCP 308 option is probably not required. When the client detects that the 309 data transmitted over the second subflow has been acknowledged over 310 the first subflow, it could decide to terminate the second subflow by 311 sending a RST segment. If the interface associated to this subflow 312 is still up, a new subflow could be immediately reestablished. It 313 would then be immediately usable to send new data and would not be 314 forced to first retransmit the previously transmitted data. As of 315 this writing, this dynamic management of the subflows is not yet 316 implemented in the Multipath TCP implementation in the Linux kernel. 318 Some studies have started to analyse the performance of Multipath TCP 319 on smartphones with real applications. In contrast with the bulk 320 transfers that are used by many publications, real applications do 321 not exchange huge amounts of data and establish a large number of 322 small connections. [COMMAG2016] proposes a software testing 323 framework that allows to automate Android applications to study their 324 interactions with Multipath TCP. [PAM2016] analyses a one-month 325 packet trace of all the packets exchanged by a dozen of smartphones 326 used by regular users. This analysis reveals that short connections 327 are important on smartphones and that the main benefit of using 328 Multipath TCP on smartphones is the ability to perform seamless 329 handovers between different wireless networks. Long connections 330 benefit from these handovers. 332 2.3. Multipath TCP proxies 334 As Multipath TCP is not yet widely deployed on both clients and 335 servers, several deployments have used various forms of proxies. Two 336 families of solutions are currently being used or tested 337 [I-D.deng-mptcp-proxy]. 339 A first use case is when a Multipath TCP enabled client wants to use 340 several interfaces to reach a regular TCP server. A typical use case 341 is a smartphone that needs to use both its WiFi and its cellular 342 interface to transfer data. Several types of proxies are possible 343 for this use case. An HTTP proxy deployed on a Multipath TCP capable 344 server would enable the smartphone to use Multipath TCP to access 345 regular web servers. Obviously, this solution only works for 346 applications that rely on HTTP. Another possibility is to use a 347 proxy that can convert any Multipath TCP connection into a regular 348 TCP connection. Multipath TCP-specific proxies have been proposed 349 [I-D.wei-mptcp-proxy-mechanism] [HotMiddlebox13b] 350 [I-D.hampel-mptcp-proxies-anchors]. 352 Another possibility leverages the SOCKS protocol [RFC1928]. SOCKS is 353 often used in enterprise networks to allow clients to reach external 354 servers. For this, the client opens a TCP connection to the SOCKS 355 server that relays it to the final destination. If both the client 356 and the SOCKS server use Multipath TCP, but not the final 357 destination, then Multipath TCP can still be used on the path between 358 the client and the SOCKS server. At IETF'93, Korea Telecom announced 359 that they have deployed in June 2015 a commercial service that uses 360 Multipath TCP on smartphones. These smartphones access regular TCP 361 servers through a SOCKS proxy. This enables them to achieve 362 throughputs of up to 850 Mbps [KT]. 364 Measurements performed with Android smartphones [Mobicom15] show that 365 popular applications work correctly through a SOCKS proxy and 366 Multipath TCP enabled smartphones. Thanks to Multipath TCP, long- 367 lived connections can be spread over the two available interfaces. 368 However, for short-lived connections, most of the data is sent over 369 the initial subflow that is created over the interface corresponding 370 to the default route and the second subflow is almost not used 371 [PAM2016]. 373 A second use case is when Multipath TCP is used by middleboxes, 374 typically inside access networks. Various network operators are 375 discussing and evaluating solutions for hybrid access networks 376 [BBF-WT348]. Such networks arise when a network operator controls 377 two different access network technologies, e.g. wired and cellular, 378 and wants to combine them to improve the bandwidth offered to the 379 endusers [I-D.lhwxz-hybrid-access-network-architecture]. Several 380 solutions are currently investigated for such networks [BBF-WT348]. 381 Figure 2 shows the organisation of such a network. When a client 382 creates a normal TCP connection, it is intercepted by the Hybrid CPE 383 (HPCE) that converts it in a Multipath TCP connection so that it can 384 use the available access networks (DSL and LTE in the example). The 385 Hybrid Access Gateway (HAG) does the opposite to ensure that the 386 regular server sees a normal TCP connection. Some of the solutions 387 that are currently discussed for hybrid networks use Multipath TCP on 388 the HCPE and the HAG. Other solutions rely on tunnels between the 389 HCPE and the HAG [I-D.lhwxz-gre-notifications-hybrid-access]. 391 client --- HCPE ------ DSL ------- HAG --- internet --- server 392 | | 393 +------- LTE -----------+ 395 Figure 2: Hybrid Access Network 397 3. Operational Experience 399 3.1. Middlebox interference 401 The interference caused by various types of middleboxes has been an 402 important concern during the design of the Multipath TCP protocol. 403 Three studies on the interactions between Multipath TCP and 404 middleboxes are worth discussing. 406 The first analysis appears in [IMC11]. This paper was the main 407 motivation for Multipath TCP incorporating various techniques to cope 408 with middlebox interference. More specifically, Multipath TCP has 409 been designed to cope with middleboxes that : 411 o change source or destination addresses 413 o change source or destination port numbers 415 o change TCP sequence numbers 417 o split or coalesce segments 419 o remove TCP options 421 o modify the payload of TCP segments 422 These middlebox interferences have all been included in the MBtest 423 suite [MBTest]. This test suite is used in [HotMiddlebox13] to 424 verify the reaction of the Multipath TCP implementation in the Linux 425 kernel when faced with middlebox interference. The test environment 426 used for this evaluation is a dual-homed client connected to a 427 single-homed server. The middlebox behavior can be activated on any 428 of the paths. The main results of this analysis are : 430 o the Multipath TCP implementation in the Linux kernel is not 431 affected by a middlebox that performs NAT or modifies TCP sequence 432 numbers 434 o when a middlebox removes the MP_CAPABLE option from the initial 435 SYN segment, the Multipath TCP implementation in the Linux kernel 436 falls back correctly to regular TCP 438 o when a middlebox removes the DSS option from all data segments, 439 the Multipath TCP implementation in the Linux kernel falls back 440 correctly to regular TCP 442 o when a middlebox performs segment coalescing, the Multipath TCP 443 implementation in the Linux kernel is still able to accurately 444 extract the data corresponding to the indicated mapping 446 o when a middlebox performs segment splitting, the Multipath TCP 447 implementation in the Linux kernel correctly reassembles the data 448 corresponding to the indicated mapping. [HotMiddlebox13] shows on 449 figure 4 in section 3.3 a corner case with segment splitting that 450 may lead to a desynchronisation between the two hosts. 452 The interactions between Multipath TCP and real deployed middleboxes 453 is also analyzed in [HotMiddlebox13] and a particular scenario with 454 the FTP application level gateway running on a NAT is described. 456 Middlebox interference can also be detected by analysing packet 457 traces on Multipath TCP enabled servers. A closer look at the 458 packets received on the multipath-tcp.org server [TMA2015] shows that 459 among the 184,000 Multipath TCP connections, only 125 of them were 460 falling back to regular TCP. These connections originated from 28 461 different client IP addresses. These include 91 HTTP connections and 462 34 FTP connections. The FTP interference is expected and due to 463 Application Level Gateways running home routers. The HTTP 464 interference appeared only on the direction from server to client and 465 could have been caused by transparent proxies deployed in cellular or 466 enterprise networks. A longer trace is discussed in [COMCOM2016] and 467 similar conclusions about the middlebox interference are provided. 469 From an operational viewpoint, knowing that Multipath TCP can cope 470 with various types of middlebox interference is important. However, 471 there are situations where the network operators need to gather 472 information about where a particular middlebox interference occurs. 473 The tracebox software [tracebox] described in [IMC13a] is an 474 extension of the popular traceroute software that enables network 475 operators to check at which hop a particular field of the TCP header 476 (including options) is modified. It has been used by several network 477 operators to debug various middlebox interference problems. tracebox 478 includes a scripting language that enables its user to specify 479 precisely which packet (including IP and TCP options) is sent by the 480 source. tracebox sends packets with an increasing TTL/HopLimit and 481 compares the information returned in the ICMP messages with the 482 packet that it sent. This enables tracebox to detect any 483 interference caused by middleboxes on a given path. tracebox works 484 better when routers implement the ICMP extension defined in 485 [RFC1812]. 487 Users of the Multipath TCP implementation have reported some 488 experience with middlebox interference. The strangest scenario has 489 been a middlebox that accepts the Multipath TCP options in the SYN 490 segment but later replaces Multipath TCP options with a TCP EOL 491 option [StrangeMbox]. This causes Multipath TCP to perform a 492 fallback to regular TCP without any impact on the application. 494 3.2. Congestion control 496 Congestion control has been an important problem for Multipath TCP. 497 The standardised congestion control scheme for Multipath TCP is 498 defined in [RFC6356] and [NSDI11]. This congestion control scheme 499 has been implemented in the Linux implementation of Multipath TCP. 500 Linux uses a modular architecture to support various congestion 501 control schemes. This architecture is applicable for both regular 502 TCP and Multipath TCP. While the coupled congestion control scheme 503 defined in [RFC6356] is the default congestion control scheme in the 504 Linux implementation, other congestion control schemes have been 505 added. The second congestion control scheme is OLIA [CONEXT12]. 506 This congestion control scheme is also an adaptation of the NewReno 507 single path congestion control scheme to support multiple paths. 508 Simulations and measurements have shown that it provides some 509 performance benefits compared to the the default congestion control 510 scheme [CONEXT12]. Measurements over a wide range of parameters 511 reported in [CONEXT13] also indicate some benefits with the OLIA 512 congestion control scheme. Recently, a delay-based congestion 513 control scheme has been ported to the Multipath TCP implementation in 514 the Linux kernel. This congestion control scheme has been evaluated 515 by using simulations in [ICNP12]. The fourth congestion control 516 scheme that has been included in the Linux implementation of 517 Multipath TCP is the BALIA scheme 518 [I-D.walid-mptcp-congestion-control]. 520 These different congestion control schemes have been compared in 521 several articles. [CONEXT13] and [PaaschPhD] compare these 522 algorithms in an emulated environment. The evaluation showed that 523 the delay-based congestion control scheme is less able to efficiently 524 use the available links than the three other schemes. Reports from 525 some users indicate that they seem to favor OLIA. 527 3.3. Subflow management 529 The multipath capability of Multipath TCP comes from the utilisation 530 of one subflow per path. The Multipath TCP architecture [RFC6182] 531 and the protocol specification [RFC6824] define the basic usage of 532 the subflows and the protocol mechanisms that are required to create 533 and terminate them. However, there are no guidelines on how subflows 534 are used during the lifetime of a Multipath TCP session. Most of the 535 published experiments with Multipath TCP have been performed in 536 controlled environments. Still, based on the experience running them 537 and discussions on the mptcp-dev mailing list, interesting lessons 538 have been learned about the management of these subflows. 540 From a subflow viewpoint, the Multipath TCP protocol is completely 541 symmetrical. Both the clients and the server have the capability to 542 create subflows. However in practice the existing Multipath TCP 543 implementations [I-D.eardley-mptcp-implementations-survey] have opted 544 for a strategy where only the client creates new subflows. The main 545 motivation for this strategy is that often the client resides behind 546 a NAT or a firewall, preventing passive subflow openings on the 547 client. Although there are environments such as datacenters where 548 this problem does not occur, as of this writing, no precise 549 requirement has emerged for allowing the server to create new 550 subflows. 552 3.4. Implemented subflow managers 554 The Multipath TCP implementation in the Linux kernel includes several 555 strategies to manage the subflows that compose a Multipath TCP 556 session. The basic subflow manager is the full-mesh. As the name 557 implies, it creates a full-mesh of subflows between the communicating 558 hosts. 560 The most frequent use case for this subflow manager is a multihomed 561 client connected to a single-homed server. In this case, one subflow 562 is created for each interface on the client. The current 563 implementation of the full-mesh subflow manager is static. The 564 subflows are created immediately after the creation of the initial 565 subflow. If one subflow fails during the lifetime of the Multipath 566 TCP session (e.g. due to excessive retransmissions, or the loss of 567 the corresponding interface), it is not always reestablished. There 568 is ongoing work to enhance the full-mesh path manager to deal with 569 such events. 571 When the server is multihomed, using the full-mesh subflow manager 572 may lead to a large number of subflows being established. For 573 example, consider a dual-homed client connected to a server with 574 three interfaces. In this case, even if the subflows are only 575 created by the client, 6 subflows will be established. This may be 576 excessive in some environments, in particular when the client and/or 577 the server have a large number of interfaces. A recent draft has 578 proposed a Multipath TCP option to negotiate the maximum number of 579 subflows. However, it should be noted that there have been reports 580 on the mptcp-dev mailing indicating that users rely on Multipath TCP 581 to aggregate more than four different interfaces. Thus, there is a 582 need for supporting many interfaces efficiently. 584 Creating subflows between multihomed clients and servers may 585 sometimes lead to operational issues as observed by discussions on 586 the mptcp-dev mailing list. In some cases the network operators 587 would like to have a better control on how the subflows are created 588 by Multipath TCP [I-D.boucadair-mptcp-max-subflow]. This might 589 require the definition of policy rules to control the operation of 590 the subflow manager. The two scenarios below illustrate some of 591 these requirements. 593 host1 ---------- switch1 ----- host2 594 | | | 595 +-------------- switch2 --------+ 597 Figure 3: Simple switched network topology 599 Consider the simple network topology shown in Figure 3. From an 600 operational viewpoint, a network operator could want to create two 601 subflows between the communicating hosts. From a bandwidth 602 utilization viewpoint, the most natural paths are host1-switch1-host2 603 and host1-switch2-host2. However, a Multipath TCP implementation 604 running on these two hosts may sometimes have difficulties to obtain 605 this result. 607 To understand the difficulty, let us consider different allocation 608 strategies for the IP addresses. A first strategy is to assign two 609 subnets : subnetA (resp. subnetB) contains the IP addresses of 610 host1's interface to switch1 (resp. switch2) and host2's interface to 611 switch1 (resp. switch2). In this case, a Multipath TCP subflow 612 manager should only create one subflow per subnet. To enforce the 613 utilization of these paths, the network operator would have to 614 specify a policy that prefers the subflows in the same subnet over 615 subflows between addresses in different subnets. It should be noted 616 that the policy should probably also specify how the subflow manager 617 should react when an interface or subflow fails. 619 A second strategy is to use a single subnet for all IP addresses. In 620 this case, it becomes more difficult to specify a policy that 621 indicates which subflows should be established. 623 The second subflow manager that is currently supported by the 624 Multipath TCP implementation in the Linux kernel is the ndiffport 625 subflow manager. This manager was initially created to exploit the 626 path diversity that exists between single-homed hosts due to the 627 utilization of flow-based load balancing techniques [SIGCOMM11]. 628 This subflow manager creates N subflows between the same pair of IP 629 addresses. The N subflows are created by the client and differ only 630 in the source port selected by the client. It was not designed to be 631 used on multihomed hosts. 633 A more flexible subflow manager has been proposed, implemented and 634 evaluated in [CONEXT15]. This subflow manager exposes various kernel 635 events to a user space daemon that decides when subflows need to be 636 created and terminated based on various policies. 638 3.5. Subflow destination port 640 The Multipath TCP protocol relies on the token contained in the 641 MP_JOIN option to associate a subflow to an existing Multipath TCP 642 session. This implies that there is no restriction on the source 643 address, destination address and source or destination ports used for 644 the new subflow. The ability to use different source and destination 645 addresses is key to support multihomed servers and clients. The 646 ability to use different destination port numbers is worth discussing 647 because it has operational implications. 649 For illustration, consider a dual-homed client that creates a second 650 subflow to reach a single-homed server as illustrated in Figure 4. 652 client ------- r1 --- internet --- server 653 | | 654 +----------r2-------+ 656 Figure 4: Multihomed-client connected to single-homed server 658 When the Multipath TCP implementation in the Linux kernel creates the 659 second subflow it uses the same destination port as the initial 660 subflow. This choice is motivated by the fact that the server might 661 be protected by a firewall and only accept TCP connections (including 662 subflows) on the official port number. Using the same destination 663 port for all subflows is also useful for operators that rely on the 664 port numbers to track application usage in their network. 666 There have been suggestions from Multipath TCP users to modify the 667 implementation to allow the client to use different destination ports 668 to reach the server. This suggestion seems mainly motivated by 669 traffic shaping middleboxes that are used in some wireless networks. 670 In networks where different shaping rates are associated to different 671 destination port numbers, this could allow Multipath TCP to reach a 672 higher performance. As of this writing, we are not aware of any 673 implementation of this kind of tweaking. 675 However, from an implementation point-of-view supporting different 676 destination ports for the same Multipath TCP connection can cause 677 some issues. A legacy implementation of a TCP stack creates a 678 listening socket to react upon incoming SYN segments. The listening 679 socket is handling the SYN segments that are sent on a specific port 680 number. Demultiplexing incoming segments can thus be done solely by 681 looking at the IP addresses and the port numbers. With Multipath TCP 682 however, incoming SYN segments may have an MP_JOIN option with a 683 different destination port. This means, that all incoming segments 684 that did not match on an existing listening-socket or an already 685 established socket must be parsed for an eventual MP_JOIN option. 686 This imposes an additional cost on servers, previously not existent 687 on legacy TCP implementations. 689 3.6. Closing subflows 690 client server 691 | | 692 MPTCP: established | | MPTCP: established 693 Sub: established | | Sub: established 694 | | 695 | DATA_FIN | 696 MPTCP: close-wait | <------------------------ | close() (step 1) 697 Sub: established | DATA_ACK | 698 | ------------------------> | MPTCP: fin-wait-2 699 | | Sub: established 700 | | 701 | DATA_FIN + subflow-FIN | 702 close()/shutdown() | ------------------------> | MPTCP: time-wait 703 (step 2) | DATA_ACK | Sub: close-wait 704 MPTCP: closed | <------------------------ | 705 Sub: fin-wait-2 | | 706 | | 707 | subflow-FIN | 708 MPTCP: closed | <------------------------ | subflow-close() 709 Sub: time-wait | subflow-ACK | 710 (step 3) | ------------------------> | MPTCP: time-wait 711 | | Sub: closed 712 | | 714 Figure 5: Multipath TCP may not be able to avoid time-wait state 715 (even if enforced by the application). 717 Figure 5 shows a very particular issue within Multipath TCP. Many 718 high-performance applications try to avoid Time-Wait state by 719 deferring the closure of the connection until the peer has sent a 720 FIN. That way, the client on the left of Figure 5 does a passive 721 closure of the connection, transitioning from Close-Wait to Last-ACK 722 and finally freeing the resources after reception of the ACK of the 723 FIN. An application running on top of a Multipath TCP enabled Linux 724 kernel might also use this approach. The difference here is that the 725 close() of the connection (Step 1 in Figure 5) only triggers the 726 sending of a DATA_FIN. Nothing guarantees that the kernel is ready 727 to combine the DATA_FIN with a subflow-FIN. The reception of the 728 DATA_FIN will make the application trigger the closure of the 729 connection (step 2), trying to avoid Time-Wait state with this late 730 closure. This time, the kernel might decide to combine the DATA_FIN 731 with a subflow-FIN. This decision will be fatal, as the subflow's 732 state machine will not transition from Close-Wait to Last-Ack, but 733 rather go through Fin-Wait-2 into Time-Wait state. The Time-Wait 734 state will consume resources on the host for at least 2 MSL (Maximum 735 Segment Lifetime). Thus, a smart application that tries to avoid 736 Time-Wait state by doing late closure of the connection actually ends 737 up with one of its subflows in Time-Wait state. A high-performance 738 Multipath TCP kernel implementation should honor the desire of the 739 application to do passive closure of the connection and successfully 740 avoid Time-Wait state - even on the subflows. 742 The solution to this problem lies in an optimistic assumption that a 743 host doing active-closure of a Multipath TCP connection by sending a 744 DATA_FIN will soon also send a FIN on all its subflows. Thus, the 745 passive closer of the connection can simply wait for the peer to send 746 exactly this FIN - enforcing passive closure even on the subflows. 747 Of course, to avoid consuming resources indefinitely, a timer must 748 limit the time our implementation waits for the FIN. 750 3.7. Packet schedulers 752 In a Multipath TCP implementation, the packet scheduler is the 753 algorithm that is executed when transmitting each packet to decide on 754 which subflow it needs to be transmitted. The packet scheduler 755 itself does not have any impact on the interoperability of Multipath 756 TCP implementations. However, it may clearly impact the performance 757 of Multipath TCP sessions. The Multipath TCP implementation in the 758 Linux kernel supports a pluggable architecture for the packet 759 scheduler [PaaschPhD]. As of this writing, two schedulers have been 760 implemented: round-robin and lowest-rtt-first. The second scheduler 761 relies on the round-trip-time (rtt) measured on each TCP subflow and 762 sends first segments over the subflow having the lowest round-trip- 763 time. They are compared in [CSWS14]. The experiments and 764 measurements described in [CSWS14] show that the lowest-rtt-first 765 scheduler appears to be the best compromise from a performance 766 viewpoint. Another study of the packet schedulers is presented in 767 [PAMS2014]. This study relies on simulations with the Multipath TCP 768 implementation in the Linux kernel. They compare the lowest-rtt- 769 first with the round-robin and a random scheduler. They show some 770 situations where the lowest-rtt-first scheduler does not perform as 771 well as the other schedulers, but there are many scenarios where the 772 opposite is true. [PAMS2014] notes that "it is highly likely that 773 the optimal scheduling strategy depends on the characteristics of the 774 paths being used." 776 3.8. Segment size selection 778 When an application performs a write/send system call, the kernel 779 allocates a packet buffer (sk_buff in Linux) to store the data the 780 application wants to send. The kernel will store at most one MSS 781 (Maximum Segment Size) of data per buffer. As the MSS can differ 782 amongst subflows, an MPTCP implementation must select carefully the 783 MSS used to generate application data. The Linux kernel 784 implementation had various ways of selecting the MSS: minimum or 785 maximum amongst the different subflows. However, these heuristics of 786 MSS selection can cause significant performance issues in some 787 environment. Consider the following example. An MPTCP connection 788 has two established subflows that respectively use a MSS of 1420 and 789 1428 bytes. If MPTCP selects the maximum, then the application will 790 generate segments of 1428 bytes of data. An MPTCP implementation 791 will have to split the segment in two (a 1420-byte and 8-byte 792 segments) when pushing on the subflow with the smallest MSS. The 793 latter segment will introduce a large overhead as for a single data 794 segment 2 slots will be used in the congestion window (in packets) 795 therefore reducing by roughly twice the potential throughput (in 796 bytes/s) of this subflow. Taking the smallest MSS does not solve the 797 issue as there might be a case where the subflow with the smallest 798 MSS only sends a few packets therefore reducing the potential 799 throughput of the other subflows. 801 The Linux implementation recently took another approach [DetalMSS]. 802 Instead of selecting the minimum and maximum values, it now 803 dynamically adapts the MSS based on the contribution of all the 804 subflows to the connection's throughput. For this it computes, for 805 each subflow, the potential throughput achieved by selecting each MSS 806 value and by taking into account the lost space in the cwnd. It then 807 selects the MSS that allows to achieve the highest potential 808 throughput. 810 3.9. Interactions with the Domain Name System 812 Multihomed clients such as smartphones can send DNS queries over any 813 of their interfaces. When a single-homed client performs a DNS 814 query, it receives from its local resolver the best answer for its 815 request. If the client is multihomed, the answer returned to the DNS 816 query may vary with the interface over which it has been sent. 818 cdn1 819 | 820 client -- cellular -- internet -- cdn3 821 | | 822 +----- wifi --------+ 823 | 824 cdn2 826 Figure 6: Simple network topology 828 If the client sends a DNS query over the WiFi interface, the answer 829 will point to the cdn2 server while the same request sent over the 830 cellular interface will point to the cdn1 server. This might cause 831 problems for CDN providers that locate their servers inside ISP 832 networks and have contracts that specify that the CDN server will 833 only be accessed from within this particular ISP. Assume now that 834 both the client and the CDN servers support Multipath TCP. In this 835 case, a Multipath TCP session from cdn1 or cdn2 would potentially use 836 both the cellular network and the WiFi network. Serving the client 837 from cdn2 over the cellular interface could violate the contract 838 between the CDN provider and the network operators. A similar 839 problem occurs with regular TCP if the client caches DNS replies. 840 For example the client obtains a DNS answer over the cellular 841 interface and then stops this interface and starts to use its WiFi 842 interface. If the client retrieves data from cdn1 over its WiFi 843 interface, this may also violate the contract between the CDN and the 844 network operators. 846 A possible solution to prevent this problem would be to modify the 847 DNS resolution on the client. The client subnet EDNS extension 848 defined in [I-D.ietf-dnsop-edns-client-subnet] could be used for this 849 purpose. When the client sends a DNS query from its WiFi interface, 850 it should also send the client subnet corresponding to the cellular 851 interface in this request. This would indicate to the resolver that 852 the answer should be valid for both the WiFi and the cellular 853 interfaces (e.g., the cdn3 server). 855 3.10. Captive portals 857 Multipath TCP enables a host to use different interfaces to reach a 858 server. In theory, this should ensure connectivity when at least one 859 of the interfaces is active. In practice however, there are some 860 particular scenarios with captive portals that may cause operational 861 problems. The reference environment is shown in Figure 7. 863 client ----- network1 864 | 865 +------- internet ------------- server 867 Figure 7: Issue with captive portal 869 The client is attached to two networks : network1 that provides 870 limited connectivity and the entire Internet through the second 871 network interface. In practice, this scenario corresponds to an open 872 WiFi network with a captive portal for network1 and a cellular 873 service for the second interface. On many smartphones, the WiFi 874 interface is preferred over the cellular interface. If the 875 smartphone learns a default route via both interfaces, it will 876 typically prefer to use the WiFi interface to send its DNS request 877 and create the first subflow. This is not optimal with Multipath 878 TCP. A better approach would probably be to try a few attempts on 879 the WiFi interface and then try to use the second interface for the 880 initial subflow as well. 882 3.11. Stateless webservers 884 MPTCP has been designed to interoperate with webservers that benefit 885 from SYN-cookies to protect against SYN-flooding attacks [RFC4987]. 886 MPTCP achieves this by echoing the keys negotiated during the 887 MP_CAPABLE handshake in the third ACK of the 3-way handshake. 888 Reception of this third ACK then allows the server to reconstruct the 889 state specific to MPTCP. 891 However, one caveat to this mechanism is the non-reliable nature of 892 the third ACK. Indeed, when the third ACK gets lost, the server will 893 not be able to reconstruct the MPTCP-state. MPTCP will fallback to 894 regular TCP in this case. This is in contrast to regular TCP. When 895 the client starts sending data, the first data segment also includes 896 the SYN-cookie, which allows the server to reconstruct the TCP-state. 897 Further, this data segment will be retransmitted by the client in 898 case it gets lost and thus is resilient against loss. MPTCP does not 899 include the keys in this data segment and thus the server cannot 900 reconstruct the MPTCP state. 902 This issue might be considered as a minor one for MPTCP. Losing the 903 third ACK should only happen when packet loss is high. However, when 904 packet-loss is high MPTCP provides a lot of benefits as it can move 905 traffic away from the lossy link. It is undesirable that MPTCP has a 906 higher chance to fall back to regular TCP in those lossy 907 environments. 909 [I-D.paasch-mptcp-syncookies] discusses this issue and suggests a 910 modified handshake mechanism that ensures reliable delivery of the 911 MP_CAPABLE, following the 3-way handshake. This modification will 912 make MPTCP reliable, even in lossy environments when servers need to 913 use SYN-cookies to protect against SYN-flooding attacks. 915 3.12. Loadbalanced serverfarms 917 Large-scale serverfarms typically deploy thousands of servers behind 918 a single virtual IP (VIP). Steering traffic to these servers is done 919 through layer-4 loadbalancers that ensure that a TCP-flow will always 920 be routed to the same server [Presto08]. 922 As Multipath TCP uses multiple different TCP subflows to steer the 923 traffic across the different paths, loadbalancers need to ensure that 924 all these subflows are routed to the same server. This implies that 925 the loadbalancers need to track the MPTCP-related state, allowing 926 them to parse the token in the MP_JOIN and assign those subflows to 927 the appropriate server. However, serverfarms typically deploy 928 multiple of these loadbalancers for reliability and capacity reasons. 929 As a TCP subflow might get routed to any of these loadbalancers, they 930 would need to synchronize the MPTCP-related state - a solution that 931 is not feasible at large scale. 933 The token (carried in the MP_JOIN) contains the information 934 indicating which MPTCP-session the subflow belongs to. As the token 935 is a hash of the key, servers are not able to generate the token in 936 such a way that the token can provide the necessary information to 937 the loadbalancers which would allow them to route TCP subflows to the 938 appropriate server. [I-D.paasch-mptcp-loadbalancer] discusses this 939 issue in detail and suggests two alternative MP_CAPABLE handshakes to 940 overcome these. As of September 2015, it is not yet clear how MPTCP 941 might accomodate such use-case to enable its deployment within 942 loadbalanced serverfarms. 944 4. Conclusion 946 In this document, we have documented a few years of experience with 947 Multipath TCP. The different scientific publications that have been 948 summarised confirm that Multipath TCP works well in different use 949 cases in today's Internet. None of the cited publications has 950 identified major issues with Multipath TCP and its utilisation in the 951 current Internet. Some of these publications list directions for 952 future improvements that mainly affect the subflow managers and 953 packet schedulers. These heuristics affect the performance of 954 Multipath TCP, but not the protocol itself. It is likely that these 955 improvements will be discussed in future IETF documents. 957 Besides the published scientific literature, a number of companies 958 have deployed Multipath TCP at large. One of these deployments uses 959 Multipath TCP on the client and the server side, making it a true 960 end-to-end deployment. This deployment uses Multipath TCP to support 961 fast handover between cellular and WiFi networks. A wider deployment 962 of Multipath TCP on servers seems to be blocked by the necessity to 963 support Multipath TCP on load balancers. Given the influence that 964 middleboxes had on the design of Multipath TCP, it is interesting to 965 note that the other industrial deployments use Multipath TCP inside 966 middleboxes. These middelboxes use Multipath TCP to efficiently 967 combine several access links while still interacting with legacy TCP 968 servers. 970 5. Acknowledgements 972 This work was partially supported by the FP7-Trilogy2 project. We 973 would like to thank all the implementers and users of the Multipath 974 TCP implementation in the Linux kernel. This document has benefited 975 from the comments of John Ronan, Yoshifumi Nishida, Phil Eardley and 976 Jaehyun Hwang. 978 6. Informative References 980 [Apple-MPTCP] 981 Apple, Inc, ., "iOS - Multipath TCP Support in iOS 7", 982 n.d., . 984 [BBF-WT348] 985 Fabregas (Ed), G., "WT-348 - Hybrid Access for Broadband 986 Networks", Broadband Forum, contribution bbf2014.1139.04 , 987 June 2015. 989 [CACM14] Paasch, C. and O. Bonaventure, "Multipath TCP", 990 Communications of the ACM, 57(4):51-57 , April 2014, 991 . 993 [COMCOM2016] 994 "Observing real Multipath TCP traffic", Computer 995 Communications , April 2016, 996 . 999 [COMMAG2016] 1000 De Coninck, Q., Baerts, M., Hesmans, B., and O. 1001 Bonaventure, "Observing Real Smartphone Applications over 1002 Multipath TCP", IEEE Communications Magazine , March 2016, 1003 . 1006 [CONEXT12] 1007 Khalili, R., Gast, N., Popovic, M., Upadhyay, U., and J. 1008 Leboudec, "MPTCP is not pareto-optimal performance issues 1009 and a possible solution", Proceedings of the 8th 1010 international conference on Emerging networking 1011 experiments and technologies (CoNEXT12) , 2012. 1013 [CONEXT13] 1014 Paasch, C., Khalili, R., and O. Bonaventure, "On the 1015 Benefits of Applying Experimental Design to Improve 1016 Multipath TCP", Conference on emerging Networking 1017 EXperiments and Technologies (CoNEXT) , December 2013, 1018 . 1021 [CONEXT15] 1022 Hesmans, B., Detal, G., Barre, S., Bauduin, R., and O. 1023 Bonaventure, "SMAPP - Towards Smart Multipath TCP-enabled 1024 APPlications", Proc. Conext 2015, Heidelberg, Germany , 1025 December 2015, . 1028 [CSWS14] Paasch, C., Ferlin, S., Alay, O., and O. Bonaventure, 1029 "Experimental Evaluation of Multipath TCP Schedulers", 1030 SIGCOMM CSWS2014 workshop , August 2014. 1032 [Cellnet12] 1033 Paasch, C., Detal, G., Duchene, F., Raiciu, C., and O. 1034 Bonaventure, "Exploring Mobile/WiFi Handover with 1035 Multipath TCP", ACM SIGCOMM workshop on Cellular Networks 1036 (Cellnet12) , 2012, 1037 . 1040 [DetalMSS] 1041 Detal, G., "Adaptive MSS value", Post on the mptcp-dev 1042 mailing list , September 2014, . 1046 [FreeBSD-MPTCP] 1047 Williams, N., "Multipath TCP For FreeBSD Kernel Patch 1048 v0.5", n.d., . 1050 [HotMiddlebox13] 1051 Hesmans, B., Duchene, F., Paasch, C., Detal, G., and O. 1052 Bonaventure, "Are TCP Extensions Middlebox-proof?", CoNEXT 1053 workshop HotMiddlebox , December 2013, 1054 . 1057 [HotMiddlebox13b] 1058 Detal, G., Paasch, C., and O. Bonaventure, "Multipath in 1059 the Middle(Box)", HotMiddlebox'13 , December 2013, 1060 . 1063 [HotNets] Raiciu, C., Pluntke, C., Barre, S., Greenhalgh, A., 1064 Wischik, D., and M. Handley, "Data center networking with 1065 multipath TCP", Proceedings of the 9th ACM SIGCOMM 1066 Workshop on Hot Topics in Networks (Hotnets-IX) , 2010, 1067 . 1069 [I-D.boucadair-mptcp-max-subflow] 1070 Boucadair, M. and C. Jacquenet, "Negotiating the Maximum 1071 Number of Multipath TCP (MPTCP) Subflows", draft- 1072 boucadair-mptcp-max-subflow-01 (work in progress), 1073 December 2015. 1075 [I-D.deng-mptcp-proxy] 1076 Lingli, D., Liu, D., Sun, T., Boucadair, M., and G. 1077 Cauchie, "Use-cases and Requirements for MPTCP Proxy in 1078 ISP Networks", draft-deng-mptcp-proxy-01 (work in 1079 progress), October 2014. 1081 [I-D.eardley-mptcp-implementations-survey] 1082 Eardley, P., "Survey of MPTCP Implementations", draft- 1083 eardley-mptcp-implementations-survey-02 (work in 1084 progress), July 2013. 1086 [I-D.hampel-mptcp-proxies-anchors] 1087 Hampel, G. and T. Klein, "MPTCP Proxies and Anchors", 1088 draft-hampel-mptcp-proxies-anchors-00 (work in progress), 1089 February 2012. 1091 [I-D.ietf-dnsop-edns-client-subnet] 1092 Contavalli, C., Gaast, W., tale, t., and W. Kumari, 1093 "Client Subnet in DNS Queries", draft-ietf-dnsop-edns- 1094 client-subnet-07 (work in progress), March 2016. 1096 [I-D.lhwxz-gre-notifications-hybrid-access] 1097 Leymann, N., Heidemann, C., Wasserman, M., Xue, L., and M. 1098 Zhang, "GRE Notifications for Hybrid Access", draft-lhwxz- 1099 gre-notifications-hybrid-access-01 (work in progress), 1100 January 2015. 1102 [I-D.lhwxz-hybrid-access-network-architecture] 1103 Leymann, N., Heidemann, C., Wasserman, M., Xue, L., and M. 1104 Zhang, "Hybrid Access Network Architecture", draft-lhwxz- 1105 hybrid-access-network-architecture-02 (work in progress), 1106 January 2015. 1108 [I-D.paasch-mptcp-loadbalancer] 1109 Paasch, C., Greenway, G., and A. Ford, "Multipath TCP 1110 behind Layer-4 loadbalancers", draft-paasch-mptcp- 1111 loadbalancer-00 (work in progress), September 2015. 1113 [I-D.paasch-mptcp-syncookies] 1114 Paasch, C., Biswas, A., and D. Haas, "Making Multipath TCP 1115 robust for stateless webservers", draft-paasch-mptcp- 1116 syncookies-02 (work in progress), October 2015. 1118 [I-D.walid-mptcp-congestion-control] 1119 Walid, A., Peng, Q., Hwang, J., and S. Low, "Balanced 1120 Linked Adaptation Congestion Control Algorithm for MPTCP", 1121 draft-walid-mptcp-congestion-control-04 (work in 1122 progress), January 2016. 1124 [I-D.wei-mptcp-proxy-mechanism] 1125 Wei, X., Xiong, C., and E. Ed, "MPTCP proxy mechanisms", 1126 draft-wei-mptcp-proxy-mechanism-02 (work in progress), 1127 June 2015. 1129 [ICNP12] Cao, Y., Xu, M., and X. Fu, "Delay-based congestion 1130 control for multipath TCP", 20th IEEE International 1131 Conference on Network Protocols (ICNP) , 2012. 1133 [IMC11] Honda, M., Nishida, Y., Raiciu, C., Greenhalgh, A., 1134 Handley, M., and H. Tokuda, "Is it still possible to 1135 extend TCP?", Proceedings of the 2011 ACM SIGCOMM 1136 conference on Internet measurement conference (IMC '11) , 1137 2011, . 1139 [IMC13a] Detal, G., Hesmans, B., Bonaventure, O., Vanaubel, Y., and 1140 B. Donnet, "Revealing Middlebox Interference with 1141 Tracebox", Proceedings of the 2013 ACM SIGCOMM conference 1142 on Internet measurement conference , 2013, 1143 . 1146 [IMC13b] Chen, Y., Lim, Y., Gibbens, R., Nahum, E., Khalili, R., 1147 and D. Towsley, "A measurement-based study of MultiPath 1148 TCP performance over wireless network", Proceedings of the 1149 2013 conference on Internet measurement conference (IMC 1150 '13) , n.d., . 1152 [IMC13c] Pelsser, C., Cittadini, L., Vissicchio, S., and R. Bush, 1153 "From Paris to Tokyo on the suitability of ping to measure 1154 latency", Proceedings of the 2013 conference on Internet 1155 measurement conference (IMC '13) , 2013, 1156 . 1158 [INFOCOM14] 1159 Lim, Y., Chen, Y., Nahum, E., Towsley, D., and K. Lee, 1160 "Cross-Layer Path Management in Multi-path Transport 1161 Protocol for Mobile Devices", IEEE INFOCOM'14 , 2014. 1163 [IOS7] "Multipath TCP Support in iOS 7", January 2014, 1164 . 1166 [KT] Seo, S., "KT's GiGA LTE", July 2015, 1167 . 1170 [MBTest] Hesmans, B., "MBTest", 2013, 1171 . 1173 [MPTCPBIB] 1174 Bonaventure, O., "Multipath TCP - An annotated 1175 bibliography", Technical report , April 2015, 1176 . 1178 [Mobicom15] 1179 De Coninck, Q., Baerts, M., Hesmans, B., and O. 1180 Bonaventure, "Poster - Evaluating Android Applications 1181 with Multipath TCP", Mobicom 2015 (Poster) , September 1182 2015. 1184 [MultipathTCP-Linux] 1185 Paasch, C., Barre, S., and . et al, "Multipath TCP 1186 implementation in the Linux kernel", n.d., 1187 . 1189 [NSDI11] Wischik, D., Raiciu, C., Greenhalgh, A., and M. Handley, 1190 "Design, implementation and evaluation of congestion 1191 control for Multipath TCP", In Proceedings of the 8th 1192 USENIX conference on Networked systems design and 1193 implementation (NSDI11) , 2011. 1195 [NSDI12] Raiciu, C., Paasch, C., Barre, S., Ford, A., Honda, M., 1196 Duchene, F., Bonaventure, O., and M. Handley, "How Hard 1197 Can It Be? Designing and Implementing a Deployable 1198 Multipath TCP", USENIX Symposium of Networked Systems 1199 Design and Implementation (NSDI12) , April 2012, 1200 . 1203 [PAM2016] De Coninck, Q., Baerts, M., Hesmans, B., and O. 1204 Bonaventure, "A First Analysis of Multipath TCP on 1205 Smartphones", 17th International Passive and Active 1206 Measurements Conference (PAM2016) , March 2016, 1207 . 1210 [PAMS2014] 1211 Arzani, B., Gurney, A., Cheng, S., Guerin, R., and B. Loo, 1212 "Impact of Path Selection and Scheduling Policies on MPTCP 1213 Performance", PAMS2014 , 2014. 1215 [PaaschPhD] 1216 Paasch, C., "Improving Multipath TCP", Ph.D. Thesis , 1217 November 2014, . 1220 [Presto08] 1221 Greenberg, A., Lahiri, P., Maltz, D., Parveen, P., and S. 1222 Sengupta, "Towards a Next Generation Data Center 1223 Architecture - Scalability and Commoditization", ACM 1224 PRESTO 2008 , August 2008, 1225 . 1227 [RFC1812] Baker, F., Ed., "Requirements for IP Version 4 Routers", 1228 RFC 1812, DOI 10.17487/RFC1812, June 1995, 1229 . 1231 [RFC1928] Leech, M., Ganis, M., Lee, Y., Kuris, R., Koblas, D., and 1232 L. Jones, "SOCKS Protocol Version 5", RFC 1928, DOI 1233 10.17487/RFC1928, March 1996, 1234 . 1236 [RFC4987] Eddy, W., "TCP SYN Flooding Attacks and Common 1237 Mitigations", RFC 4987, DOI 10.17487/RFC4987, August 2007, 1238 . 1240 [RFC6182] Ford, A., Raiciu, C., Handley, M., Barre, S., and J. 1241 Iyengar, "Architectural Guidelines for Multipath TCP 1242 Development", RFC 6182, DOI 10.17487/RFC6182, March 2011, 1243 . 1245 [RFC6356] Raiciu, C., Handley, M., and D. Wischik, "Coupled 1246 Congestion Control for Multipath Transport Protocols", RFC 1247 6356, DOI 10.17487/RFC6356, October 2011, 1248 . 1250 [RFC6824] Ford, A., Raiciu, C., Handley, M., and O. Bonaventure, 1251 "TCP Extensions for Multipath Operation with Multiple 1252 Addresses", RFC 6824, DOI 10.17487/RFC6824, January 2013, 1253 . 1255 [SIGCOMM11] 1256 Raiciu, C., Barre, S., Pluntke, C., Greenhalgh, A., 1257 Wischik, D., and M. Handley, "Improving datacenter 1258 performance and robustness with multipath TCP", 1259 Proceedings of the ACM SIGCOMM 2011 conference , n.d., 1260 . 1262 [StrangeMbox] 1263 Bonaventure, O., "Multipath TCP through a strange 1264 middlebox", Blog post , January 2015, 1265 . 1268 [TMA2015] Hesmans, B., Tran Viet, H., Sadre, R., and O. Bonaventure, 1269 "A First Look at Real Multipath TCP Traffic", Traffic 1270 Monitoring and Analysis , 2015, 1271 . 1274 [ietf88] Stewart, L., "IETF'88 Meeting minutes of the MPTCP working 1275 group", n.d., . 1278 [tracebox] 1279 Detal, G. and O. Tilmans, "tracebox", 2013, 1280 . 1282 Appendix A. Changelog 1284 This section should be removed before final publication 1286 o initial version : September 16th, 2014 : Added section Section 3.8 1287 that discusses some performance problems that appeared with the 1288 Linux implementation when using subflows having different MSS 1289 values 1291 o update with a description of the middlebox that replaces an 1292 unknown TCP option with EOL [StrangeMbox] 1294 o version ietf-02 : July 2015, answer to last call comments 1296 * Reorganised text to better separate use cases and operational 1297 experience 1299 * New use case on Multipath TCP proxies in Section 2.3 1301 * Added some text on middleboxes in Section 3.1 1303 * Removed the discussion on SDN 1305 * Restructured text and improved writing in some parts 1307 o version ietf-03 : September 2015, answer to comments from Phil 1308 Eardley 1309 * Improved introduction 1311 * Added details about using SOCKS and Korea Telecom's use-case in 1312 Section 2.3. 1314 * Added issue around clients caching DNS-results in Section 3.9 1316 * Explained issue of MPTCP with stateless webservers Section 3.11 1318 * Added description of MPTCP's use behind layer-4 loadbalancers 1319 Section 3.12 1321 * Restructured text and improved writing in some parts 1323 o version ietf-04 : April 2016, answer to last comments 1325 * Updated text on measurements with smartphones 1327 * Updated conclusion 1329 Authors' Addresses 1331 Olivier Bonaventure 1332 UCLouvain 1334 Email: Olivier.Bonaventure@uclouvain.be 1336 Christoph Paasch 1337 Apple, Inc. 1339 Email: cpaasch@apple.com 1341 Gregory Detal 1342 Tessares 1344 Email: Gregory.Detal@tessares.net