idnits 2.17.1 draft-ietf-mptcp-experience-00.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 (September 16, 2014) is 3509 days in the past. Is this intentional? Checking references for intended status: Informational ---------------------------------------------------------------------------- -- Obsolete informational reference (is this intentional?): RFC 6824 (Obsoleted by RFC 8684) Summary: 2 errors (**), 0 flaws (~~), 1 warning (==), 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 C. Paasch 4 Intended status: Informational G. Detal 5 Expires: March 20, 2015 UCLouvain 6 September 16, 2014 8 Experience with Multipath TCP 9 draft-ietf-mptcp-experience-00 11 Abstract 13 This document discusses operational experiences of using Multipath 14 TCP in real world networks. It lists several prominent use cases for 15 which Multipath TCP has been considered and is being used. It also 16 gives insight in some heuristics and decisions that have helped to 17 realize these use cases. Further, it presents several open issues 18 that are yet unclear on how they can be solved. 20 Status of This Memo 22 This Internet-Draft is submitted in full conformance with the 23 provisions of BCP 78 and BCP 79. 25 Internet-Drafts are working documents of the Internet Engineering 26 Task Force (IETF). Note that other groups may also distribute 27 working documents as Internet-Drafts. The list of current Internet- 28 Drafts is at http://datatracker.ietf.org/drafts/current/. 30 Internet-Drafts are draft documents valid for a maximum of six months 31 and may be updated, replaced, or obsoleted by other documents at any 32 time. It is inappropriate to use Internet-Drafts as reference 33 material or to cite them other than as "work in progress." 35 This Internet-Draft will expire on March 20, 2015. 37 Copyright Notice 39 Copyright (c) 2014 IETF Trust and the persons identified as the 40 document authors. All rights reserved. 42 This document is subject to BCP 78 and the IETF Trust's Legal 43 Provisions Relating to IETF Documents 44 (http://trustee.ietf.org/license-info) in effect on the date of 45 publication of this document. Please review these documents 46 carefully, as they describe your rights and restrictions with respect 47 to this document. Code Components extracted from this document must 48 include Simplified BSD License text as described in Section 4.e of 49 the Trust Legal Provisions and are provided without warranty as 50 described in the Simplified BSD License. 52 Table of Contents 54 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 2 55 2. Middlebox interference . . . . . . . . . . . . . . . . . . . 3 56 3. Use cases . . . . . . . . . . . . . . . . . . . . . . . . . . 4 57 4. Congestion control . . . . . . . . . . . . . . . . . . . . . 8 58 5. Subflow management . . . . . . . . . . . . . . . . . . . . . 9 59 5.1. Implemented subflow managers . . . . . . . . . . . . . . 9 60 5.2. Subflow destination port . . . . . . . . . . . . . . . . 11 61 5.3. Closing subflows . . . . . . . . . . . . . . . . . . . . 12 62 6. Packet schedulers . . . . . . . . . . . . . . . . . . . . . . 13 63 7. Segment size selection . . . . . . . . . . . . . . . . . . . 14 64 8. Interactions with the Domain Name System . . . . . . . . . . 15 65 9. Captive portals . . . . . . . . . . . . . . . . . . . . . . . 15 66 10. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . 16 67 11. Acknowledgements . . . . . . . . . . . . . . . . . . . . . . 16 68 12. Changelog . . . . . . . . . . . . . . . . . . . . . . . . . . 16 69 13. Informative References . . . . . . . . . . . . . . . . . . . 16 70 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 20 72 1. Introduction 74 Multipath TCP was standardized in [RFC6824] and four implementations 75 have been developed [I-D.eardley-mptcp-implementations-survey]. 76 Since the publication of [RFC6824], some experience has been gathered 77 by various network researchers and users about the issues that arise 78 when Multipath TCP is used in the Internet. 80 Most of the experience reported in this document comes from the 81 utilization of the Multipath TCP implementation in the Linux kernel 82 [MultipathTCP-Linux]. It has been downloaded and is used by 83 thousands of users all over the world. Many of these users have 84 provided direct or indirect feedback by writing documents (scientific 85 articles or blog messages) or posting to the mptcp-dev mailing list ( 86 https://listes-2.sipr.ucl.ac.be/sympa/arc/mptcp-dev ) . This 87 Multipath TCP implementation is actively maintained and continuously 88 improved. It is used on various types of hosts, ranging from 89 smartphones or embedded systems to high-end servers. 91 This is not, by far, the most widespread deployment of Multipath TCP. 92 Since September 2013, Multipath TCP is also supported on smartphones 93 and tablets running iOS7 [IOS7]. There are likely hundreds of 94 millions of Multipath TCP enabled devices. However, this particular 95 Multipath TCP implementation is currently only used to support a 96 single application. Unfortunately, there is no public information 97 about the lessons learned from this large scale deployment. 99 This document is organized as follows. We explain in 100 Section Section 2 which types of middleboxes the Linux Kernel 101 implementation of Multipath TCP supports and how it reacts upon 102 encountering these. Next, we list several use cases of Multipath TCP 103 in Section {{usecases}. Section {{congestion} summarises the MPTCP 104 specific congestion controls that have been implemented. Sections 105 Section 5 and Section 6 discuss heuristics and issues with respect to 106 subflow management as well as the scheduling across the subflows. 107 Section Section 7 explains some problems that occurred with subflows 108 having different MSS values. Section Section 8 presents issues with 109 respect to content delivery networks and suggests a solution to this 110 issue. Finally, Section Section 9 shows an issue with captive 111 portals where MPTCP will behave suboptimal. 113 2. Middlebox interference 115 The interference caused by various types of middleboxes has been an 116 important concern during the design of the Multipath TCP protocol. 117 Three studies on the interactions between Multipath TCP and 118 middleboxes are worth being discussed. 120 The first analysis was described in [IMC11]. This paper was the main 121 motivation for including inside Multipath TCP various techniques to 122 cope with middlebox interference. More specifically, Multipath TCP 123 has been designed to cope with middleboxes that : - change source or 124 destination addresses - change source or destination port numbers - 125 change TCP sequence numbers - split or coalesce segments - remove TCP 126 options - modify the payload of TCP segments 128 These middlebox interferences have all been included in the MBtest 129 suite [MBTest]. This test suite has been used [HotMiddlebox13] to 130 verify the reaction of the Multipath TCP implementation in the Linux 131 kernel when faced with middlebox interference. The test environment 132 used for this evaluation is a dual-homed client connected to a 133 single-homed server. The middlebox behavior can be activated on any 134 of the paths. The main results of this analysis are : 136 o the Multipath TCP implementation in the Linux kernel is not 137 affected by a middlebox that performs NAT or modifies TCP sequence 138 numbers 140 o when a middlebox removes the MP_CAPABLE option from the initial 141 SYN segment, the Multipath TCP implementation in the Linux kernel 142 falls back correctly to regular TCP 144 o when a middlebox removes the DSS option from all data segments, 145 the Multipath TCP implementation in the Linux kernel falls back 146 correctly to regular TCP 148 o when a middlebox performs segment coalescing, the Multipath TCP 149 implementation in the Linux kernel is still able to accurately 150 extract the data corresponding to the indicated mapping 152 o when a middlebox performs segment splitting, the Multipath TCP 153 implementation in the Linux kernel correctly reassembles the data 154 corresponding to the indicated mapping. [HotMiddlebox13] 155 documents a corner case with segment splitting that may lead to 156 desynchronisation between the two hosts. 158 The interactions between Multipath TCP and real deployed middleboxes 159 is also analyzed in [HotMiddlebox13] and a particular scenario with 160 the FTP application level gateway running on a NAT is described. 162 From an operational viewpoint, knowing that Multipath TCP can cope 163 with various types of middlebox interference is important. However, 164 there are situations where the network operators need to gather 165 information about where a particular middlebox interference occurs. 166 The tracebox software [tracebox] described in [IMC13a] is an 167 extension of the popular traceroute software that enables network 168 operators to check at which hop a particular field of the TCP header 169 (including options) is modified. It has been used by several network 170 operators to debug various middlebox interference problems. tracebox 171 includes a scripting language that enables its user to specify 172 precisely which packet is sent by the source. tracebox sends packets 173 with an increasing TTL/HopLimit and compares the information returned 174 in the ICMP messages with the packet that it sends. This enables 175 tracebox to detect any interference caused by middleboxes on a given 176 path. tracebox works better when routers implement the ICMP extension 177 defined in [RFC1812]. 179 3. Use cases 181 Multipath TCP has been tested in several use cases. Several of the 182 papers published in the scientific litterature have identified 183 possible improvements that are worth being discussed here. 185 A first, although initially unexpected, documented use case for 186 Multipath TCP has been the datacenters [HotNets][SIGCOMM11]. Today's 187 datacenters are designed to provide several paths between single- 188 homed servers. The multiplicity of these paths comes from the 189 utilization of Equal Cost Multipath (ECMP) and other load balancing 190 techniques inside the datacenter. Most of the deployed load 191 balancing techniques in these datacenters rely on hashes computed or 192 the five tuple to ensure that all packets from the same TCP 193 connection will follow the same path to prevent packet reordering. 194 The results presented in [HotNets] demonstrate by simulations that 195 Multipath TCP can achieve a better utilization of the available 196 network by using multiple subflows for each Multipath TCP session. 197 Although [RFC6182] assumes that at least one of the communicating 198 hosts has several IP addresses, [HotNets] demonstrates that there are 199 also benefits when both hosts are single-homed. This idea was 200 pursued further in [SIGCOMM11] where the Multipath TCP implementation 201 in the Linux kernel was modified to be able to use several subflows 202 from the same IP address. Measurements performed in a public 203 datacenter showed performance improvements with Multipath TCP. 205 Although ECMP is widely used inside datacenters, this is not the only 206 environment where there are different paths between a pair of hosts. 207 ECMP and other load balancing techniques such as LAG are widely used 208 in today's network and having multiple paths between a pair of 209 single-homed hosts is becoming the norm instead of the exception. 210 Although these multiple paths have often the same cost (from an IGP 211 metrics viewpoint), they do not necessarily have the same 212 performance. For example, [IMC13c] reports the results of a long 213 measurement study showing that load balanced Internet paths between 214 that same pair of hosts can have huge delay differences. 216 A second use case that has been explored by several network 217 researchers is the cellular/WiFi offload use case. Smartphones or 218 other mobile devices equipped with two wireless interfaces are a very 219 common use case for Multipath TCP. As of this writing, this is also 220 the largest deployment of Multipath-TCP enabled devices [IOS7]. 221 Unfortunately, as there are no public measurements about this 222 deployment, we can only rely on published papers that have mainly 223 used the Multipath TCP implementation in the Linux kernel for their 224 experiment. 226 The performance of Multipath TCP in wireless networks was briefly 227 evaluated in [NSDI12]. One experiment analyzes the performance of 228 Multipath TCP on a client with two wireless interfaces. This 229 evaluation shows that when the receive window is large, Multipath TCP 230 can efficiently use the two available links. However, if the window 231 becomes smaller, then packets sent on a slow path can block the 232 transmission of packets on a faster path. In some cases, the 233 performance of Multipath TCP over two paths can become lower than the 234 performance of regular TCP over the best performing path. Two 235 heuristics, reinjection and penalization, are proposed in [NSDI12] to 236 solve this identified performance problem. These two heuristics have 237 since been used in the Multipath TCP implementation in the Linux 238 kernel. [CONEXT13] explored the problem in more details and revealed 239 some other scenarios where Multipath TCP can have difficulties in 240 efficiently pooling the available paths. Improvements to the 241 Multipath TCP implementation in the Linux kernel are proposed in 242 [CONEXT13] to cope with some of these problems. 244 The first experimental analysis of Multipath TCP in a public wireless 245 environment was presented in [Cellnet12]. These measurements explore 246 the ability of Multipath TCP to use two wireless networks (real WiFi 247 and 3G networks). Three modes of operation are compared. The first 248 mode of operation is the simultaneous use of the two wireless 249 networks. In this mode, Multipath TCP pools the available resources 250 and uses both wireless interfaces. This mode provides fast handover 251 from WiFi to cellular or the opposite when the user moves. 252 Measurements presented in [CACM14] show that the handover from one 253 wireless network to another is not an abrupt process. When a host 254 moves, it does not experience either excellent connectivity or no 255 connectivity at all. Instead, there are regions where the quality of 256 one of the wireless networks is weaker than the other, but the host 257 considers this wireless network to still be up. When a mobile host 258 enters such regions, its ability to send packets over another 259 wireless network is important to ensure a smooth handover. This is 260 clearly illustrated from the packet trace discussed in [CACM14]. 262 Many cellular networks use volume-based pricing and users often 263 prefer to use unmetered WiFi networks when available instead of 264 metered cellular networks. [Cellnet12] implements the support for 265 the MP_PRIO option to explore two other modes of operation. 267 In the backup mode, Multipath TCP opens a TCP subflow over each 268 interface, but the cellular interface is configured in backup mode. 269 This implies that data only flows over the WiFi interface when both 270 interfaces are considered to be active. If the WiFi interface fails, 271 then the traffic switches quickly to the cellular interface, ensuring 272 a smooth handover from the user's viewpoint [Cellnet12]. The cost of 273 this approach is that the WiFi and cellular interfaces likely remain 274 active all the time since all subflows are established over the two 275 interfaces. 277 The single-path mode is slightly different. This mode benefits from 278 the break-before-make capability of Multipath TCP. When an MPTCP 279 session is established, a subflow is created over the WiFi interface. 280 No packet is sent over the cellular interface as long as the WiFi 281 interface remains up [Cellnet12]. This implies that the cellular 282 interface can remain idle and battery capacity is preserved. When 283 the WiFi interface fails, new subflows are established over the 284 cellular interface in order to preserve the established Multipath TCP 285 sessions. Compared to the backup mode described earlier, this mode 286 of operation is characterized by a throughput drop while the cellular 287 interface is brought up and the subflows are reestablished. During 288 this time, no data packet is transmitted. 290 From a protocol viewpoint, [Cellnet12] discusses the problem posed by 291 the unreliability of the ADD_ADDR option and proposes a small 292 protocol extension to allow hosts to reliably exchange this option. 293 It would be useful to analyze packet traces to understand whether the 294 unreliability of the REMOVE_ADDR option poses an operational problem 295 in real deployments. 297 Another study of the performance of Multipath TCP in wireless 298 networks was reported in [IMC13b]. This study uses laptops connected 299 to various cellular ISPs and WiFi hotspots. It compares various file 300 transfer scenarios and concludes based on measurements with the 301 Multipath TCP implementation in the Linux kernel that "MPTCP provides 302 a robust data transport and reduces variations in download 303 latencies". 305 A different study of the performance of Multipath TCP with two 306 wireless networks is presented in [INFOCOM14]. In this study the two 307 networks had different qualities : a good network and a lossy 308 network. When using two paths with different packet loss ratios, the 309 Multipath TCP congestion control scheme moves traffic away from the 310 lossy link that is considered to be congested. However, [INFOCOM14] 311 documents an interesting scenario that is summarised in the figure 312 below. 314 client ----------- path1 -------- server 315 | | 316 +--------------- path2 ------------+ 318 Figure 1: Simple network topology 320 Initially, the two paths have the same quality and Multipath TCP 321 distributes the load over both of them. During the transfer, the 322 second path becomes lossy, e.g. because the client moves. Multipath 323 TCP detects the packet losses and they are retransmitted over the 324 first path. This enables the data transfer to continue over the 325 first path. However, the subflow over the second path is still up 326 and transmits one packet from time to time. Although the N packets 327 have been acknowledged over the first subflow (at the MPTCP level), 328 they have not been acknowledged at the TCP level over the second 329 subflow. To preserve the continuity of the sequence numbers over the 330 second subflow, TCP will continue to retransmit these segments until 331 either they are acknowledged or the maximum number of retransmissions 332 is reached. This behavior is clearly inefficient and may lead to 333 blocking since the second subflow will consume window space to be 334 able to retransmit these packets. [INFOCOM14] proposes a new 335 Multipath TCP option to solve this problem. In practice, a new TCP 336 option is probably not required. When the client detects that the 337 data transmitted over the second subflow has been acknowledged over 338 the first subflow, it could decide to terminate the second subflow by 339 sending a RST segment. If the interface associated to this subflow 340 is still up, a new subflow could be immediately reestablished. It 341 would then be immediately usable to send new data and would not be 342 forced to first retransmit the previously transmitted data. As of 343 this writing, this dynamic management of the subflows is not yet 344 implemented in the Multipath TCP implementation in the Linux kernel. 346 A third use case has been the coupling between software defined 347 networking techniques such as Openflow and Multipath TCP. Openflow 348 can be used to configure different paths inside a network. Using an 349 international network, [TNC13] demonstrates that Multipath TCP can 350 achieve high throughput in the wide area. An interesting point to 351 note about the measurements reported in [TNC13] is that the 352 measurement setup used four paths through the WAN. Only two of these 353 paths were disjoint. When Multipath TCP was used, the congestion 354 control scheme ensured that only two of these paths were actually 355 used. 357 4. Congestion control 359 Congestion control has been an important problem for Multipath TCP. 360 The standardised congestion control scheme for Multipath TCP is 361 defined in [RFC6356] and [NSDI11]. This congestion control scheme 362 has been implemented in the Linux implementation of Multipath TCP. 363 Linux uses a modular architecture to support various congestion 364 control schemes. This architecture is applicable for both regular 365 TCP and Multipath TCP. While the coupled congestion control scheme 366 defined in [RFC6356] is the default congestion control scheme in the 367 Linux implementation, other congestion control schemes have been 368 added. The second congestion control scheme is OLIA [CONEXT12]. 369 This congestion control scheme is also an adaptation of the NewReno 370 single path congestion control scheme to support multiple paths. 371 Simulations and measurements have shown that it provides some 372 performance benefits compared to the the default congestion control 373 scheme [CONEXT12]. Measurement over a wide range of parameters 374 reported in [CONEXT13] also indicate some benefits with the OLIA 375 congestion control scheme. Recently, a delay-based congestion 376 control scheme has been ported to the Multipath TCP implementation in 377 the Linux kernel. This congestion control scheme has been evaluated 378 by using simulations in [ICNP12]. As of this writing, it has not yet 379 been evaluated by performing large measurement campaigns. 381 5. Subflow management 383 The multipath capability of Multipath TCP comes from the utilization 384 of one subflow per path. The Multipath TCP architecture [RFC6182] 385 and the protocol specification [RFC6824] define the basic usage of 386 the subflows and the protocol mechanisms that are required to create 387 and terminate them. However, there are no guidelines on how subflows 388 are used during the lifetime of a Multipath TCP session. Most of the 389 experiments with Multipath TCP have been performed in controlled 390 environments. Still, based on the experience running them and 391 discussions on the mptcp-dev mailing list, interesting lessons have 392 been learned about the management of these subflows. 394 From a subflow viewpoint, the Multipath TCP protocol is completely 395 symmetrical. Both the clients and the server have the capability to 396 create subflows. However in practice the existing Multipath TCP 397 implementations [I-D.eardley-mptcp-implementations-survey] have opted 398 for a strategy where only the client creates new subflows. The main 399 motivation for this strategy is that often the client resides behind 400 a NAT or a firewall, preventing passive subflow openings on the 401 client. Although there are environments such as datacenters where 402 this problem does not occur, as of this writing, no precise 403 requirement has emerged for allowing the server to create new 404 subflows. 406 5.1. Implemented subflow managers 408 The Multipath TCP implementation in the Linux kernel includes several 409 strategies to manage the subflows that compose a Multipath TCP 410 session. The basic subflow manager is the full-mesh. As the name 411 implies, it creates a full-mesh of subflows between the communicating 412 hosts. 414 The most frequent use case for this subflow manager is a multihomed 415 client connected to a single-homed server. In this case, one subflow 416 is created for each interface on the client. The current 417 implementation of the full-mesh subflow manager is static. The 418 subflows are created immediately after the creation of the initial 419 subflow. If one subflow fails during the lifetime of the Multipath 420 TCP session (e.g. due to excessive retransmissions, or the loss of 421 the corresponding interface), it is not always reestablished. There 422 is ongoing work to enhance the full-mesh path manager to deal with 423 such events. 425 When the server is multihomed, using the full-mesh subflow manager 426 may lead to a large number of subflows being established. For 427 example, consider a dual-homed client connected to a server with 428 three interfaces. In this case, even if the subflows are only 429 created by the client, 6 subflows will be established. This may be 430 excessive in some environments, in particular when the client and/or 431 the server have a large number of interfaces. It should be noted 432 that there have been reports on the mptcp-dev mailing indicating that 433 users rely on Multipath TCP to aggregate more than four different 434 interfaces. Thus, there is a need for supporting many interfaces 435 efficiently. 437 It should be noted that creating subflows between multihomed clients 438 and servers may sometimes lead to operational issues as observed by 439 discussions on the mptcp-dev mailing list. In some cases the network 440 operators would like to have a better control on how the subflows are 441 created by Multipath TCP. This might require the definition of 442 policy rules to control the operation of the subflow manager. The 443 two scenarios below illustrate some of these requirements. 445 host1 ---------- switch1 ----- host2 446 | | | 447 +-------------- switch2 --------+ 449 Figure 2: Simple switched network topology 451 Consider the simple network topology shown in Figure 2. From an 452 operational viewpoint, a network operator could want to create two 453 subflows between the communicating hosts. From a bandwidth 454 utilization viewpoint, the most natural paths are host1-switch1-host2 455 and host1-switch2-host2. However, a Multipath TCP implementation 456 running on these two hosts may sometimes have difficulties to achieve 457 this result. 459 To understand the difficulty, let us consider different allocation 460 strategies for the IP addresses. A first strategy is to assign two 461 subnets : subnetA (resp. subnetB) contains the IP addresses of 462 host1's interface to switch1 (resp. switch2) and host2's interface to 463 switch1 (resp. switch2). In this case, a Multipath TCP subflow 464 manager should only create one subflow per subnet. To enforce the 465 utilization of these paths, the network operator would have to 466 specify a policy that prefers the subflows in the same subnet over 467 subflows between addresses in different subnets. It should be noted 468 that the policy should probably also specify how the subflow manager 469 should react when an interface or subflow fails. 471 A second strategy is to use a single subnet for all IP addresses. In 472 this case, it becomes more difficult to specify a policy that 473 indicates which subflows should be established. 475 The second subflow manager that is currently supported by the 476 Multipath TCP implementation in the Linux kernel is the ndiffport 477 subflow manager. This manager was initially created to exploit the 478 path diversity that exists between single-homed hosts due to the 479 utilization of flow-based load balancing techniques. This subflow 480 manager creates N subflows between the same pair of IP addresses. 481 The N subflows are created by the client and differ only in the 482 source port selected by the client. 484 5.2. Subflow destination port 486 The Multipath TCP protocol relies on the token contained in the 487 MP_JOIN option to associate a subflow to an existing Multipath TCP 488 session. This implies that there is no restriction on the source 489 address, destination address and source or destination ports used for 490 the new subflow. The ability to use different source and destination 491 addresses is key to support multihomed servers and clients. The 492 ability to use different destination port numbers is worth being 493 discussed because it has operational implications. 495 For illustration, consider a dual-homed client that creates a second 496 subflow to reach a single-homed server as illustrated in the 497 Figure 3. 499 client ------- r1 --- internet --- server 500 | | 501 +----------r2-------+ 503 Figure 3: Multihomed-client connected to single-homed server 505 When the Multipath TCP implementation in the Linux kernel creates the 506 second subflow it uses the same destination port as the initial 507 subflow. This choice is motivated by the fact that the server might 508 be protected by a firewall and only accept TCP connections (including 509 subflows) on the official port number. Using the same destination 510 port for all subflows is also useful for operators that rely on the 511 port numbers to track application usage in their network. 513 There have been suggestions from Multipath TCP users to modify the 514 implementation to allow the client to use different destination ports 515 to reach the server. This suggestion seems mainly motivated by 516 traffic shaping middleboxes that are used in some wireless networks. 517 In networks where different shaping rates are associated to different 518 destination port numbers, this could allow Multipath TCP to reach a 519 higher performance. As of this writing, we are not aware of any 520 implementation of this kind of tweaking. 522 However, from an implementation point-of-view supporting different 523 destination ports for the same Multipath TCP connection introduces a 524 new performance issue. A legacy implementation of a TCP stack 525 creates a listening socket to react upon incoming SYN segments. The 526 listening socket is handling the SYN segments that are sent on a 527 specific port number. Demultiplexing incoming segments can thus be 528 done solely by looking at the IP addresses and the port numbers. 529 With Multipath TCP however, incoming SYN segments may have an MP_JOIN 530 option with a different destination port. This means, that all 531 incoming segments that did not match on an existing listening-socket 532 or an already established socket must be parsed for an eventual 533 MP_JOIN option. This imposes an additional cost on servers, 534 previously not existent on legacy TCP implementations. 536 5.3. Closing subflows 538 client server 539 | | 540 MPTCP: established | | MPTCP: established 541 Sub: established | | Sub: established 542 | | 543 | DATA_FIN | 544 MPTCP: close-wait | <------------------------ | close() (step 1) 545 Sub: established | DATA_ACK | 546 | ------------------------> | MPTCP: fin-wait-2 547 | | Sub: established 548 | | 549 | DATA_FIN + subflow-FIN | 550 close()/shutdown() | ------------------------> | MPTCP: time-wait 551 (step 2) | DATA_ACK | Sub: close-wait 552 MPTCP: closed | <------------------------ | 553 Sub: fin-wait-2 | | 554 | | 555 | subflow-FIN | 556 MPTCP: closed | <------------------------ | subflow-close() 557 Sub: time-wait | subflow-ACK | 558 (step 3) | ------------------------> | MPTCP: time-wait 559 | | Sub: closed 560 | | 562 Figure 4: Multipath TCP may not be able to avoid time-wait state 563 (even if enforced by the application). 565 Figure 4 shows a very particular issue within Multipath TCP. Many 566 high-performance applications try to avoid Time-Wait state by 567 deferring the closure of the connection until the peer has sent a 568 FIN. That way, the client on the left of Figure 4 does a passive 569 closure of the connection, transitioning from Close-Wait to Last-ACK 570 and finally freeing the resources after reception of the ACK of the 571 FIN. An application running on top of a Multipath TCP enabled Linux 572 kernel might also use this approach. The difference here is that the 573 close() of the connection (Step 1 in Figure 4) only triggers the 574 sending of a DATA_FIN. Nothing guarantees that the kernel is ready 575 to combine the DATA_FIN with a subflow-FIN. The reception of the 576 DATA_FIN will make the application trigger the closure of the 577 connection (step 2), trying to avoid Time-Wait state with this late 578 closure. This time, the kernel might decide to combine the DATA_FIN 579 with a subflow-FIN. This decision will be fatal, as the subflow's 580 state machine will not transition from Close-Wait to Last-Ack, but 581 rather go through Fin-Wait-2 into Time-Wait state. The Time-Wait 582 state will consume resources on the host for at least 2 MSL (Maximum 583 Segment Lifetime). Thus, a smart application, that tries to avoid 584 Time-Wait state by doing late closure of the connection actually ends 585 up with one of its subflows in Time-Wait state. A high-performance 586 Multipath TCP kernel implementation should honor the desire of the 587 application to do passive closure of the connection and successfully 588 avoid Time-Wait state - even on the subflows. 590 The solution to this problem lies in an optimistic assumption that a 591 host doing active-closure of a Multipath TCP connection by sending a 592 DATA_FIN will soon also send a FIN on all its in subflows. Thus, the 593 passive closer of the connection can simply wait for the peer to send 594 exactly this FIN - enforcing passive closure even on the subflows. 595 Of course, to avoid consuming resources indefinitely, a timer must 596 limit the time our implementation waits for the FIN. 598 6. Packet schedulers 600 In a Multipath TCP implementation, the packet scheduler is the 601 algorithm that is executed when transmitting each packet to decide on 602 which subflow it needs to be transmitted. The packet scheduler 603 itself does not have any impact on the interoperability of Multipath 604 TCP implementations. However, it may clearly impact the performance 605 of Multipath TCP sessions. It is important to note that the problem 606 of scheduling Multipath TCP packets among subflows is different from 607 the problem of scheduling SCTP messages. SCTP implementations also 608 include schedulers, but these are used to schedule the different 609 streams. Multipath TCP uses a single data stream. 611 Various researchers have explored theoretically and by simulations 612 the problem of scheduling packets among Multipath TCP subflows 613 [ICC14]. Unfortunately, none of the proposed techniques have been 614 implemented and used in real deployment. A detailed analysis of the 615 impact of the packet scheduler will appear in [CSWS14]. This article 616 proposes a pluggable architecture for the scheduler used by the 617 Multipath TCP implementation in the Linux kernel. This architecture 618 allows researchers to experiment with different types of schedulers. 619 Two schedulers are compared in [CSWS14] : round-robin and lowest-rtt- 620 first. The experiments and measurements described in [CSWS14] show 621 that the lowest-rtt-first scheduler appears to be the best compromise 622 from a performance viewpoint. 624 Another study of the packet schedulers is presented in [PAMS2014]. 625 This study relies on simulations with the Multipath TCP 626 implementation in the Linux kernel. The simulation scenarios 627 discussed in [PAMS2014] confirm the impact of the packet scheduler on 628 the performance of Multipath TCP. 630 7. Segment size selection 632 When an application performs a write/send system call, the kernel 633 allocates a packet buffer (sk_buff in Linux) to store the data the 634 application wants to send. The kernel will store at most one MSS 635 (Maximum Segment Size) of data per buffer. As MSS can differ amongst 636 subflows, an MPTCP implementation must select carefully the MSS used 637 to generate application data. The Linux kernel implementation had 638 various ways of selecting the MSS: minimum or maximum amongst the 639 different subflows. However, these heuristics of MSS selection can 640 cause significant performances issues in some environment. Consider 641 the following example. An MPTCP connection has two established 642 subflows that respectively use a MSS of 1420 and 1428 bytes. If 643 MPTCP selects the maximum, then the application will generate 644 segments of 1428 bytes of data. An MPTCP implementation will have to 645 split the segment in two (a 1420-byte and 8-byte segments) when 646 pushing on the subflow with the smallest MSS. The latter segment 647 will introduce a large overhead as for a single data segment 2 slots 648 will be used in the congestion window (in packets) therefore reducing 649 by ~2 the potential throughput (in bytes/s) of this subflow. Taking 650 the smallest MSS does not solve the issue as there might be a case 651 where the sublow with the smallest MSS will only participate 652 marginally to the overall performance therefore reducing the 653 potential throughput of the other subflows. 655 The Linux implementation recently took another approach [DetalMSS]. 656 Instead of selecting the minimum and maximum values, it now 657 dynamically adapts the MSS based on the contribution of all the 658 subflows to the connection's throughput. For this it computes, for 659 each subflow, the potential throughput achieved by selecting each MSS 660 value and by taking into account the lost space in the cwnd. It then 661 selects the MSS that allows to achieve the highest potential 662 throughput. 664 8. Interactions with the Domain Name System 666 Multihomed clients such as smartphones could lead to operational 667 problems when interacting with the Domain Name System. When a 668 single-homed client performs a DNS query, it receives from its local 669 resolver the best answer for its request. If the client is 670 multihomed, the answer returned to the DNS query may vary with the 671 interface over which it has been sent. 673 cdn1 674 | 675 client -- cellular -- internet -- cdn3 676 | | 677 +----- wifi --------+ 678 | 679 cdn2 681 Figure 5: Simple network topology 683 If the client sends a DNS query over the WiFi interface, the answer 684 will point to the cdn2 server while the same request sent over the 685 cellular interface will point to the cdn1 server. This might cause 686 problems for CDN providers that locate their servers inside ISP 687 networks and have contracts that specify that the CDN server will 688 only be accessed from within this particular ISP. Assume now that 689 both the client and the CDN servers support Multipath TCP. In this 690 case, a Multipath TCP session from cdn1 or cdn2 would potentially use 691 both the cellular network and the WiFi network. This would violate 692 the contract between the CDN provider and the network operators. A 693 possible solution to prevent this problem would be to modify the DNS 694 resolution on the client. The client subnet EDNS extension defined 695 in [I-D.vandergaast-edns-client-subnet] could be used for this 696 purpose. When the client sends a DNS query from its WiFi interface, 697 it should also send the client subnet corresponding to the cellular 698 interface in this request. This would indicate to the resolver that 699 the answer should be valid for both the WiFi and the cellular 700 interfaces (e.g., the cdn3 server). 702 9. Captive portals 704 Multipath TCP enables a host to use different interfaces to reach a 705 server. In theory, this should ensure connectivity when at least one 706 of the interfaces is active. In practice however, there are some 707 particular scenarios with captive portals that may cause operational 708 problems. The reference environment is the following : 710 client ----- network1 711 | 712 +------- internet ------------- server 714 Figure 6: Issue with captive portal 716 The client is attached to two networks : network1 that provides 717 limited connectivity and the entire Internet through the second 718 network interface. In practice, this scenario corresponds to an open 719 WiFi network with a captive portal for network1 and a cellular 720 service for the second interface. On many smartphones, the WiFi 721 interface is preferred over the cellular interface. If the 722 smartphone learns a default route via both interfaces, it will 723 typically prefer to use the WiFi interface to send its DNS request 724 and create the first subflow. This is not optimal with Multipath 725 TCP. A better approach would probably be to try a few attempts on 726 the WiFi interface and then try to use the second interface for the 727 initial subflow as well. 729 10. Conclusion 731 In this document, we have documented a few years of experience with 732 Multipath TCP. The information presented in this document was 733 gathered from scientific publications and discussions with various 734 users of the Multipath TCP implementation in the Linux kernel. 736 11. Acknowledgements 738 This work was partially supported by the FP7-Trilogy2 project. We 739 would like to thank all the implementers and users of the Multipath 740 TCP implementation in the Linux kernel. 742 12. Changelog 744 o initial version : 746 13. Informative References 748 [CACM14] Paasch, C. and O. Bonaventure, "Multipath TCP", 749 Communications of the ACM, 57(4):51-57 , April 2014, 750 . 752 [CONEXT12] 753 Khalili, R., Gast, N., Popovic, M., Upadhyay, U., and J. 754 Leboudec, "MPTCP is not pareto-optimal performance issues 755 and a possible solution", Proceedings of the 8th 756 international conference on Emerging networking 757 experiments and technologies (CoNEXT12) , 2012. 759 [CONEXT13] 760 Paasch, C., Khalili, R., and O. Bonaventure, "On the 761 Benefits of Applying Experimental Design to Improve 762 Multipath TCP", Conference on emerging Networking 763 EXperiments and Technologies (CoNEXT) , December 2013, 764 . 767 [CSWS14] Paasch, C., Ferlin, S., Alay, O., and O. Bonaventure, 768 "Experimental Evaluation of Multipath TCP Schedulers", 769 SIGCOMM CSWS2014 workshop , August 2014. 771 [Cellnet12] 772 Paasch, C., Detal, G., Duchene, F., Raiciu, C., and O. 773 Bonaventure, "Exploring Mobile/WiFi Handover with 774 Multipath TCP", ACM SIGCOMM workshop on Cellular Networks 775 (Cellnet12) , 2012, 776 . 779 [DetalMSS] 780 Detal, G., "Adaptive MSS value", Post on the mptcp-dev 781 mailing list , September 2014, . 785 [HotMiddlebox13] 786 Hesmans, B., Duchene, F., Paasch, C., Detal, G., and O. 787 Bonaventure, "Are TCP Extensions Middlebox-proof?", CoNEXT 788 workshop HotMiddlebox , December 2013, 789 . 792 [HotNets] Raiciu, C., Pluntke, C., Barre, S., Greenhalgh, A., 793 Wischik, D., and M. Handley, "Data center networking with 794 multipath TCP", Proceedings of the 9th ACM SIGCOMM 795 Workshop on Hot Topics in Networks (Hotnets-IX) , 2010, 796 . 798 [I-D.eardley-mptcp-implementations-survey] 799 Eardley, P., "Survey of MPTCP Implementations", draft- 800 eardley-mptcp-implementations-survey-02 (work in 801 progress), July 2013. 803 [I-D.vandergaast-edns-client-subnet] 804 Contavalli, C., Gaast, W., Leach, S., and E. Lewis, 805 "Client Subnet in DNS Requests", draft-vandergaast-edns- 806 client-subnet-02 (work in progress), July 2013. 808 [ICC14] Kuhn, N., Lochin, E., Mifdaoui, A., Sarwar, G., Mehani, 809 O., and R. Boreli, "DAPS Intelligent Delay-Aware Packet 810 Scheduling For Multipath Transport", IEEE ICC 2014 , 2014. 812 [ICNP12] Cao, Y., Xu, M., and X. Fu, "Delay-based congestion 813 control for multipath TCP", 20th IEEE International 814 Conference on Network Protocols (ICNP) , 2012. 816 [IMC11] Honda, M., Nishida, Y., Raiciu, C., Greenhalgh, A., 817 Handley, M., and H. Tokuda, "Is it still possible to 818 extend TCP?", Proceedings of the 2011 ACM SIGCOMM 819 conference on Internet measurement conference (IMC '11) , 820 2011, . 822 [IMC13a] Detal, G., Hesmans, B., Bonaventure, O., Vanaubel, Y., and 823 B. Donnet, "Revealing Middlebox Interference with 824 Tracebox", Proceedings of the 2013 ACM SIGCOMM conference 825 on Internet measurement conference , 2013, 826 . 829 [IMC13b] Chen, Y., Lim, Y., Gibbens, R., Nahum, E., Khalili, R., 830 and D. Towsley, "A measurement-based study of MultiPath 831 TCP performance over wireless network", Proceedings of the 832 2013 conference on Internet measurement conference (IMC 833 '13) , n.d., . 835 [IMC13c] Pelsser, C., Cittadini, L., Vissicchio, S., and R. Bush, 836 "From Paris to Tokyo on the suitability of ping to measure 837 latency", Proceedings of the 2013 conference on Internet 838 measurement conference (IMC '13) , 2013, 839 . 841 [INFOCOM14] 842 Lim, Y., Chen, Y., Nahum, E., Towsley, D., and K. Lee, 843 "Cross-Layer Path Management in Multi-path Transport 844 Protocol for Mobile Devices", IEEE INFOCOM'14 , 2014. 846 [IOS7] "Multipath TCP Support in iOS 7", January 2014, 847 . 849 [MBTest] Hesmans, B., "MBTest", 2013, 850 . 852 [MultipathTCP-Linux] 853 Paasch, C., Barre, S., and . et al, "Multipath TCP 854 implementation in the Linux kernel", n.d., 855 . 857 [NSDI11] Wischik, D., Raiciu, C., Greenhalgh, A., and M. Handley, 858 "Design, implementation and evaluation of congestion 859 control for Multipath TCP", In Proceedings of the 8th 860 USENIX conference on Networked systems design and 861 implementation (NSDI11) , 2011. 863 [NSDI12] Raiciu, C., Paasch, C., Barre, S., Ford, A., Honda, M., 864 Duchene, F., Bonaventure, O., and M. Handley, "How Hard 865 Can It Be? Designing and Implementing a Deployable 866 Multipath TCP", USENIX Symposium of Networked Systems 867 Design and Implementation (NSDI12) , April 2012, 868 . 871 [PAMS2014] 872 Arzani, B., Gurney, A., Cheng, S., Guerin, R., and B. Loo, 873 "Impact of Path Selection and Scheduling Policies on MPTCP 874 Performance", PAMS2014 , 2014. 876 [RFC1812] Baker, F., "Requirements for IP Version 4 Routers", RFC 877 1812, June 1995. 879 [RFC6182] Ford, A., Raiciu, C., Handley, M., Barre, S., and J. 880 Iyengar, "Architectural Guidelines for Multipath TCP 881 Development", RFC 6182, March 2011. 883 [RFC6356] Raiciu, C., Handley, M., and D. Wischik, "Coupled 884 Congestion Control for Multipath Transport Protocols", RFC 885 6356, October 2011. 887 [RFC6824] Ford, A., Raiciu, C., Handley, M., and O. Bonaventure, 888 "TCP Extensions for Multipath Operation with Multiple 889 Addresses", RFC 6824, January 2013. 891 [SIGCOMM11] 892 Raiciu, C., Barre, S., Pluntke, C., Greenhalgh, A., 893 Wischik, D., and M. Handley, "Improving datacenter 894 performance and robustness with multipath TCP", 895 Proceedings of the ACM SIGCOMM 2011 conference , n.d., 896 . 898 [TNC13] van der Pol, R., Bredel, M., and A. Barczyk, "Experiences 899 with MPTCP in an intercontinental multipathed OpenFlow 900 network", TNC2013 , 2013. 902 [tracebox] 903 Detal, G., "tracebox", 2013, . 905 Authors' Addresses 907 Olivier Bonaventure 908 UCLouvain 910 Email: Olivier.Bonaventure@uclouvain.be 912 Christoph Paasch 913 UCLouvain 915 Email: Christoph.Paasch@uclouvain.be 917 Gregory Detal 918 UCLouvain 920 Email: Gregory.Detal@uclouvain.be