idnits 2.17.1 draft-filsfils-spring-segment-routing-msdc-03.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 : ---------------------------------------------------------------------------- No issues found here. Miscellaneous warnings: ---------------------------------------------------------------------------- == The copyright year in the IETF Trust and authors Copyright Line does not match the current year -- The document date (July 20, 2015) is 3204 days in the past. Is this intentional? Checking references for intended status: Informational ---------------------------------------------------------------------------- -- Looks like a reference, but probably isn't: '16000' on line 284 -- Looks like a reference, but probably isn't: '23999' on line 284 -- Looks like a reference, but probably isn't: '1000' on line 873 -- Looks like a reference, but probably isn't: '1999' on line 873 -- Looks like a reference, but probably isn't: '2000' on line 873 -- Looks like a reference, but probably isn't: '2999' on line 873 ** Obsolete normative reference: RFC 3107 (Obsoleted by RFC 8277) == Outdated reference: A later version (-05) exists of draft-filsfils-spring-segment-routing-central-epe-04 == Outdated reference: A later version (-11) exists of draft-ietf-rtgwg-bgp-routing-large-dc-03 == Outdated reference: A later version (-05) exists of draft-keyupate-idr-bgp-prefix-sid-04 == Outdated reference: A later version (-08) exists of draft-previdi-6man-segment-routing-header-06 Summary: 1 error (**), 0 flaws (~~), 5 warnings (==), 7 comments (--). Run idnits with the --verbose option for more detailed information about the items above. -------------------------------------------------------------------------------- 2 Network Working Group C. Filsfils, Ed. 3 Internet-Draft S. Previdi, Ed. 4 Intended status: Informational Cisco Systems, Inc. 5 Expires: January 21, 2016 J. Mitchell 6 Unaffiliated 7 E. Aries 8 P. Lapukhov 9 G. Nagarajan 10 Facebook 11 D. Afanasiev 12 Yandex 13 T. Laberge 14 E. Nkposong 15 M. Nanduri 16 Microsoft 17 J. Uttaro 18 ATT 19 S. Ray 20 Unaffiliated 21 July 20, 2015 23 BGP-Prefix Segment in large-scale data centers 24 draft-filsfils-spring-segment-routing-msdc-03 26 Abstract 28 This document describes the motivation and benefits for applying 29 segment routing in the data-center. It describes the design to 30 deploy segment routing in the data-center, for both the MPLS and IPv6 31 dataplanes. 33 Requirements Language 35 The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", 36 "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this 37 document are to be interpreted as described in RFC 2119 [RFC2119]. 39 Status of This Memo 41 This Internet-Draft is submitted in full conformance with the 42 provisions of BCP 78 and BCP 79. 44 Internet-Drafts are working documents of the Internet Engineering 45 Task Force (IETF). Note that other groups may also distribute 46 working documents as Internet-Drafts. The list of current Internet- 47 Drafts is at http://datatracker.ietf.org/drafts/current/. 49 Internet-Drafts are draft documents valid for a maximum of six months 50 and may be updated, replaced, or obsoleted by other documents at any 51 time. It is inappropriate to use Internet-Drafts as reference 52 material or to cite them other than as "work in progress." 54 This Internet-Draft will expire on January 21, 2016. 56 Copyright Notice 58 Copyright (c) 2015 IETF Trust and the persons identified as the 59 document authors. All rights reserved. 61 This document is subject to BCP 78 and the IETF Trust's Legal 62 Provisions Relating to IETF Documents 63 (http://trustee.ietf.org/license-info) in effect on the date of 64 publication of this document. Please review these documents 65 carefully, as they describe your rights and restrictions with respect 66 to this document. Code Components extracted from this document must 67 include Simplified BSD License text as described in Section 4.e of 68 the Trust Legal Provisions and are provided without warranty as 69 described in the Simplified BSD License. 71 Table of Contents 73 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 3 74 2. Large Scale Data Center Network Design Summary . . . . . . . 3 75 2.1. Reference design . . . . . . . . . . . . . . . . . . . . 4 76 3. Some open problems in large data-center networks . . . . . . 5 77 4. Applying Segment Routing in the DC with MPLS dataplane . . . 6 78 4.1. BGP Prefix Segment . . . . . . . . . . . . . . . . . . . 6 79 4.2. eBGP Labeled Unicast (RFC3107) . . . . . . . . . . . . . 7 80 4.2.1. Control Plane . . . . . . . . . . . . . . . . . . . . 8 81 4.2.2. Data Plane . . . . . . . . . . . . . . . . . . . . . 9 82 4.2.3. Network Design Variation . . . . . . . . . . . . . . 10 83 4.2.4. Global BGP Prefix Segment through the fabric . . . . 10 84 4.2.5. Incremental Deployments . . . . . . . . . . . . . . . 11 85 4.3. iBGP Labeled Unicast (RFC3107) . . . . . . . . . . . . . 12 86 5. Applying Segment Routing in the DC with IPv6 dataplane . . . 12 87 6. Communicating path information to the host . . . . . . . . . 13 88 7. Addressing the open problems . . . . . . . . . . . . . . . . 14 89 7.1. Per-packet and flowlet switching . . . . . . . . . . . . 14 90 7.2. Performance-aware routing . . . . . . . . . . . . . . . . 15 91 7.3. Non-oblivious routing . . . . . . . . . . . . . . . . . . 16 92 7.4. Deterministic network probing . . . . . . . . . . . . . . 16 93 8. Additional Benefits . . . . . . . . . . . . . . . . . . . . . 16 94 8.1. MPLS Dataplane with operational simplicity . . . . . . . 16 95 8.2. Minimizing the FIB table . . . . . . . . . . . . . . . . 17 96 8.3. Egress Peer Engineering . . . . . . . . . . . . . . . . . 17 97 8.4. Incremental Deployments . . . . . . . . . . . . . . . . . 18 98 8.5. Anycast . . . . . . . . . . . . . . . . . . . . . . . . . 18 99 9. Preferred SRGB Allocation . . . . . . . . . . . . . . . . . . 18 100 10. Alternative Options . . . . . . . . . . . . . . . . . . . . . 19 101 11. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 20 102 12. Manageability Considerations . . . . . . . . . . . . . . . . 20 103 13. Security Considerations . . . . . . . . . . . . . . . . . . . 20 104 14. Contributors . . . . . . . . . . . . . . . . . . . . . . . . 20 105 15. Acknowledgements . . . . . . . . . . . . . . . . . . . . . . 20 106 16. References . . . . . . . . . . . . . . . . . . . . . . . . . 20 107 16.1. Normative References . . . . . . . . . . . . . . . . . . 20 108 16.2. Informative References . . . . . . . . . . . . . . . . . 20 109 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 21 111 1. Introduction 113 Segment Routing (SR), as described in 114 [I-D.filsfils-spring-segment-routing] leverages the source routing 115 paradigm. A node steers a packet through an ordered list of 116 instructions, called segments. A segment can represent any 117 instruction, topological or service-based. A segment can have a 118 local semantic to an SR node or global within an SR domain. SR 119 allows to enforce a flow through any topological path and service 120 chain while maintaining per-flow state only at the ingress node to 121 the SR domain. Segment Routing can be applied to the MPLS and IPv6 122 data-planes. 124 The use-case use-cases described in this document should be 125 considered in the context of the BGP-based large-scale data-center 126 (DC) design described in[I-D.ietf-rtgwg-bgp-routing-large-dc]We 127 extend it by applying SR both with IPv6 and MPLS dataplane. 129 2. Large Scale Data Center Network Design Summary 131 This section provides a brief summary of the informational document 132 [I-D.ietf-rtgwg-bgp-routing-large-dc] that outlines a practical 133 network design suitable for data-centers of various scales: 135 o Data-center networks have highly symmetric topologies with 136 multiple parallel paths between two server attachment points. The 137 well-known Clos topology is most popular among the operators. In 138 a Clos topology, the minimum number of parallel paths between two 139 elements is determined by the "width" of the middle stage. See 140 Figure 1 below for an illustration of the concept. 142 o Large-scale data-centers commonly use a routing protocol, such as 143 BGP4 [RFC4271] in order to provide endpoint connectivity. 144 Recovery after a network failure is therefore driven either by 145 local knowledge of directly available backup paths or by 146 distributed signaling between the network devices. 148 o Within data-center networks, traffic is load-shared using the 149 Equal Cost Multipath (ECMP) mechanism. With ECMP, every network 150 device implements a pseudo-random decision, mapping packets to one 151 of the parallel paths by means of a hash function calculated over 152 certain parts of the packet, typically a combination of various 153 packet header fields. 155 The following is a schematic of a five-stage Clos topology, with four 156 devices in the middle stage. Notice that number of paths between 157 Node1 and Node12 equals to four: the paths have to cross all of 158 Tier-1 devices. At the same time, the number of paths between Node1 159 and Node2 equals two, and the paths only cross Tier-2 devices. Other 160 topologies are possible, but for simplicity we'll only look into the 161 topologies that have a single path from Tier-1 to Tier-3. The rest 162 could be treated similarly, with a few modifications to the logic. 164 2.1. Reference design 166 Tier-1 167 +-----+ 168 |NODE | 169 +->| 5 |--+ 170 | +-----+ | 171 Tier-2 | | Tier-2 172 +-----+ | +-----+ | +-----+ 173 +------------>|NODE |--+->|NODE |--+--|NODE |-------------+ 174 | +-----| 3 |--+ | 6 | +--| 9 |-----+ | 175 | | +-----+ +-----+ +-----+ | | 176 | | | | 177 | | +-----+ +-----+ +-----+ | | 178 | +-----+---->|NODE |--+ |NODE | +--|NODE |-----+-----+ | 179 | | | +---| 4 |--+->| 7 |--+--| 10 |--+ | | | 180 | | | | +-----+ | +-----+ | +-----+ | | | | 181 | | | | | | | | | | 182 +-----+ +-----+ | +-----+ | +-----+ +-----+ 183 |NODE | |NODE | Tier-3 +->|NODE |--+ Tier-3 |NODE | |NODE | 184 | 1 | | 2 | | 8 | | 11 | | 12 | 185 +-----+ +-----+ +-----+ +-----+ +-----+ 186 | | | | | | | | 187 A O B O <- Servers -> Z O O O 189 Figure 1: 5-stage Clos topology 191 In the reference topology illustrated in Figure 1, we assume: 193 o Each node is its own AS (Node X has AS X) 195 * For simple and efficient route propagation filtering, Nodes 5, 196 6, 7 and 8 share the same AS, Nodes 3 and 4 share the same AS, 197 Nodes 9 and 10 share the same AS. 199 * For efficient usage of the scarce 2-byte Private Use AS pool, 200 different Tier-3 nodes might share the same AS. 202 * Without loss of generality, we will simplify these details in 203 this document and assume that each node has its own AS. 205 o Each node peers with its neighbors via BGP session 207 * If not specified, eBGP is assumed. In a specific use-case, 208 iBGP will be used but this will be called out explicitly in 209 that case. 211 o Each node originates the IPv4 address of it's loopback interface 212 into BGP and announces it to its neighbors. 214 * The loopback of Node X is 192.0.2.x/32. 216 In this document, we also refer to the Tier-1, Tier-2 and Tier-3 217 switches respectively as Spine, Leaf and ToR (top of rack) switches. 218 When a ToR switch acts as a gateway to the "outside world", we call 219 it a border switch. 221 3. Some open problems in large data-center networks 223 The data-center network design summarized above provides means for 224 moving traffic between hosts with reasonable efficiency. There are 225 few open performance and reliability problems that arise in such 226 design: 228 o ECMP routing is most commonly realized per-flow. This means that 229 large, long-lived "elephant" flows may affect performance of 230 smaller, short-lived "mouse" flows and reduce efficiency of per- 231 flow load-sharing. In other words, per-flow ECMP that does not 232 perform efficiently when flow life-time distribution is heavy- 233 tailed. Furthermore, due to hash-function inefficiencies it is 234 possible to have frequent flow collisions, where more flows get 235 placed on one path over the others 237 o Shortest-path routing with ECMP implements oblivious routing 238 model, which is not aware of the network imbalances. If the 239 network symmetry is broken, for example due to link failures, 240 utilization hotspots may appear. For example, if a link fails 241 between Tier-1 and Tier-2 devices (e.g. "Node5" and "Node9"), 242 Tier-3 devices "Node1" and "Node2" will not be aware of that, 243 since there are other paths available from perspective of "Node3". 244 They will continue sending roughly equal traffic to Node3 and 245 Node4 as if the failure didn't exist which may cause a traffic 246 hotspot. 248 o Absence of path visibility leaves transport protocols, such as 249 TCP, with a "blackbox" view of the network. Some TCP metrics, 250 such as SRTT, MSS, CWND and few others could be inferred and 251 cached based on past history, but those apply to destinations, 252 regardless of the path that has been chosen to get there. Thus, 253 for instance, TCP is not capable of remembering "bad" paths, such 254 as those that exhibited poor performance in the past. This means 255 that every new connection will be established obliviously (memory- 256 less) with regards to the paths chosen before, or chosen by other 257 nodes. 259 o Isolating faults in the network with multiple parallel paths and 260 ECMP-based routing is non-trivial due to lack of determinism. 261 Specifically, the connections from HostA to HostB may take a 262 different path every time a new connection is formed, thus making 263 consistent reproduction of a failure much more difficult. This 264 complexity scales linearly with the number of parallel paths in 265 the network, and stems from the random nature of path selection by 266 the network devices. 268 Further in this document, we are going to demonstrate how these 269 problems could be addressed within the framework of Segment Routing. 271 First, we will explain how to apply SR in the DC, for MPLS and IPv6 272 data-planes. 274 4. Applying Segment Routing in the DC with MPLS dataplane 276 4.1. BGP Prefix Segment 278 A BGP-Prefix Segment is a segment associated with a BGP prefix. A 279 BGP-Prefix Segment is a network-wide instruction to forward the 280 packet along the ECMP-aware best path to the related prefix 281 ([I-D.keyupate-idr-bgp-prefix-sid]). 283 In this document, we make the network design decision to assume that 284 all the nodes are allocated the same SRGB, e.g. [16000, 23999]. This 285 is important to fulfill the recommendation for operational 286 simplification as explained in [I-D.filsfils-spring-segment-routing]. 288 Note well that the use of a common SRGB in all nodes is not a 289 requirement, one could use a different SRGB at every node. However, 290 this would make the operation of the DC fabric more complex as the 291 label allocated to the loopback of a remote switch is then different 292 at every node. This also may increase the complexity of the 293 centralized controller. 295 For illustration purpose, when considering an MPLS data-plane, we 296 assume that the segment index allocated to prefix 192.0.2.x/32 is X. 297 As a result, a local label 1600x is allocated for prefix 192.0.2.x/32 298 by each node throughout the DC fabric. 300 When IPv6 data-plane is considered, we assume that Node X is 301 allocated IPv6 address (segment) 2001:DB8::X. 303 4.2. eBGP Labeled Unicast (RFC3107) 305 Referring to Figure 1 and [[I-D.ietf-rtgwg-bgp-routing-large-dc], the 306 following design modifications are introduced: 308 o Each node peers with its neighbors via eBGP3107 session 310 o The forwarding plane at Tier-2 and Tier-1 is MPLS. 312 o The forwarding plane at Tier-3 is either IP2MPLS (if the host 313 sends IP traffic) or MPLS2MPLS (if the host sends MPLS- 314 encapsulated traffic). 316 Figure 2 zooms on a path from server A to server Z within the 317 topology of Figure 1. 319 +-----+ +-----+ +-----+ 320 +---------->|NODE | |NODE | |NODE | 321 | | 4 |--+->| 7 |--+--| 10 |---+ 322 | +-----+ +-----+ +-----+ | 323 | | 324 +-----+ +-----+ 325 |NODE | |NODE | 326 | 1 | | 11 | 327 +-----+ +-----+ 328 | | 329 A <- Servers -> Z 331 Figure 2: Path from A to Z via nodes 1, 4, 7, 10 and 11 333 Referring to Figure 1 and Figure 2 and assuming the IP address, AS 334 and index allocation previously described, the following sections 335 detail the control plane operation and the data plane states for the 336 prefix 192.0.2.11/32 (loopback of Node11) 338 4.2.1. Control Plane 340 Node11 originates 192.0.2.11/32 in BGP and allocates to it the BGP- 341 Prefix Segment attribute (index11). 343 Node11 sends the following eBGP3107 update to Node10: 345 . NLRI: 192.0.2.11/32 346 . Label: Implicit-Null 347 . Next-hop: Node11's interface address on the link to Node10 348 . AS Path: {11} 349 . BGP-Prefix Attribute: Index 11 351 Node10 receives the above update. As it is SR capable, Node10 is 352 able to interpret the BGP-Prefix Attribute and hence understands that 353 it should allocate the label LOCAL-SRGB (16000) + "index" 11 (hence 354 16011) to the NLRI instead of allocating an nondeterministic label 355 out of a dynamically allocated portion of the local label space. The 356 implicit-null label in the NLRI tells Node10 that it is the 357 penultimate hop and MUST pop the top label on the stack before 358 forwarding traffic for this prefix to Node11. 360 Then, Node10 sends the following eBGP3107 update to Node7: 362 . NLRI: 192.0.2.11/32 363 . Label: 16011 364 . Next-hop: Node10's interface address on the link to Node7 365 . AS Path: {10, 11} 366 . BGP-Prefix Attribute: Index 11 368 Node7 receives the above update. As it is SR capable, Node7 is able 369 to interpret the BGP-Prefix Attribute and hence allocates the local 370 (incoming) label 16011 (16000 + 11) to the NLRI (instead of 371 allocating a "dynamic" local label from its label manager). Node7 372 uses the label in the received eBGP3107 NLRI as the outgoing label 373 (the index is only used to derive the local/incoming label). 375 Node7 sends the following eBGP3107 update to Node4: 377 . NLRI: 192.0.2.11/32 378 . Label: 16011 379 . Next-hop: Node7's interface address on the link to Node4 380 . AS Path: {7, 10, 11} 381 . BGP-Prefix Attribute: Index 11 382 Node4 receives the above update. As it is SR capable, Node4 is able 383 to interpret the BGP-Prefix Attribute and hence allocates the local 384 (incoming) label 16011 to the NLRI (instead of allocating a "dynamic" 385 local label from its label manager). Node4 uses the label in the 386 received eBGP3107 NLRI as outgoing label (the index is only used to 387 derive the local/incoming label). 389 Node4 sends the following eBGP3107 update to Node1: 391 . NLRI: 192.0.2.11/32 392 . Label: 16011 393 . Next-hop: Node4's interface address on the link to Node1 394 . AS Path: {4, 7, 10, 11} 395 . BGP-Prefix Attribute: Index 11 397 Node1 receives the above update. As it is SR capable, Node1 is able 398 to interpret the BGP-Prefix Attribute and hence allocates the local 399 (incoming) label 16011 to the NLRI (instead of allocating a "dynamic" 400 local label from its label manager). Node1 uses the label in the 401 received eBGP3107 NLRI as outgoing label (the index is only used to 402 derive the local/incoming label). 404 4.2.2. Data Plane 406 Referring to Figure 1Referring to Figure 1, and assuming all nodes 407 apply the same advertisement rules described above and all nodes have 408 the same SRGB (16000-23999), here are the IP/MPLS forwarding tables 409 for prefix 192.0.2.11/32 at Nodes 1, 4, 7 and 10. 411 ----------------------------------------------- 412 Incoming label | outgoing label | Outgoing 413 or IP destination | | Interface 414 ------------------+----------------+----------- 415 16011 | 16011 | ECMP{3, 4} 416 192.0.2.11/32 | 16011 | ECMP{3, 4} 417 ------------------+----------------+----------- 419 Figure 3: Node1 Forwarding Table 421 ----------------------------------------------- 422 Incoming label | outgoing label | Outgoing 423 or IP destination | | Interface 424 ------------------+----------------+----------- 425 16011 | 16011 | ECMP{7, 8} 426 192.0.2.11/32 | 16011 | ECMP{7, 8} 427 ------------------+----------------+----------- 429 Figure 4: Node4 Forwarding Table 431 ----------------------------------------------- 432 Incoming label | outgoing label | Outgoing 433 or IP destination | | Interface 434 ------------------+----------------+----------- 435 16011 | 16011 | 10 436 192.0.2.11/32 | 16011 | 10 437 ------------------+----------------+----------- 439 Figure 5: Node7 Forwarding Table 441 ----------------------------------------------- 442 Incoming label | outgoing label | Outgoing 443 or IP destination | | Interface 444 ------------------+----------------+----------- 445 16011 | POP | 11 446 192.0.2.11/32 | N/A | 11 447 ------------------+----------------+----------- 449 Node10 Forwarding Table 451 4.2.3. Network Design Variation 453 A network design choice could consist of switching all the traffic 454 through Tier-1 and Tier-2 as MPLS traffic. In this case, one could 455 filter away the IP entries at Nodes 4, 7 and 10. This might be 456 beneficial in order to optimize the forwarding table size. 458 A network design choice could consist in allowing the hosts to send 459 MPLS-encapsulated traffic (based on EPE use-case, 460 [I-D.filsfils-spring-segment-routing-central-epe]). For example, 461 applications at HostA would send their Z-destined traffic to Node1 462 with an MPLS label stack where the top label is 16011 and the next 463 label is an EPE peer segment at Node11 directing the traffic to Z. 465 4.2.4. Global BGP Prefix Segment through the fabric 467 When the previous design is deployed, the operator enjoys global BGP 468 prefix segment (label) allocation throughout the DC fabric. 470 A few examples follow: 472 o Normal forwarding to Node11: a packet with top label 16011 473 received by any switch in the fabric will be forwarded along the 474 ECMP-aware BGP best-path towards Node11 and the label 16011 is 475 penultimate-popped at Node10. 477 o Traffic-engineered path to Node11: an application on a host behind 478 Node1 might want to restrict its traffic to paths via the Spine 479 switch Node5. The application achieves this by sending its 480 packets with a label stack of {16005, 16011}. BGP Prefix segment 481 16005 directs the packet up to Node5 along the path (Node1, Node3, 482 Node5). BGP Prefix Segment 16011 then directs the packet down to 483 Node11 along the path (Node5, Node9, Node11). 485 4.2.5. Incremental Deployments 487 The design previously described can be deployed incrementally. Let 488 us assume that Node7 does not support the BGP-Prefix Segment 489 attribute and let us show how the fabric connectivity is preserved. 491 From a signaling viewpoint, nothing would change: if Node7 does not 492 understand the BGP-Prefix Segment attribute, it does propagate the 493 attribute unmodified to its neighbors. 495 From a label allocation viewpoint, the only difference is that Node7 496 would allocate a dynamic (random) label to the prefix 192.0.2.11/32 497 (e.g. 123456) instead of the "hinted" label as instructed by the BGP 498 Prefix Segment attribute. The neighbors of Node7 adapt automatically 499 as they always use the label in the BGP3107 NLRI as outgoing label. 501 Node4 does understand the BGP-Prefix Segment attribute and hence 502 allocates the indexed label in the SRGB (16011) for 192.0.2.11/32. 504 As a result, all the data-plane entries across the network would be 505 unchanged except the entries at Node7 and its neighbor Node4 as shown 506 in the figures below. 508 The key point is that the end-to-end LSP is preserved because the 509 outgoing label is always derived from the received label within the 510 BGP3107 NLRI. The index in the BGP Prefix SID is only used as a hint 511 on how to allocate the local label (the incoming label) but never for 512 the outgoing label. 514 ------------------------------------------ 515 Incoming label | outgoing | Outgoing 516 or IP destination | label | Interface 517 -------------------+---------------------- 518 12345 | 16011 | 10 520 Figure 7: Node7 Forwarding Table 522 ------------------------------------------ 523 Incoming label | outgoing | Outgoing 524 or IP destination | label | Interface 525 -------------------+---------------------- 526 16011 | 12345 | 7 528 Figure 8: Node4 Forwarding Table 530 The BGP-Prefix Segment functionality can thus be deployed 531 incrementally one node at a time. 533 When deployed together with a homogeneous SRGB (same SRGB across the 534 fabric), the operator incrementally enjoys the global prefix segment 535 benefits as the deployment progresses through the fabric. 537 4.3. iBGP Labeled Unicast (RFC3107) 539 The same exact design as eBGP3107 is used with the following 540 modifications: 542 All switches share the same AS 544 iBGP3107 reflection with nhop-self is used instead of eBGP3107 546 For simple and efficient route propagation filtering, Nodes 5, 6, 547 7 and 8 share the same Cluster ID, Nodes 3 and 4 share the same 548 Cluster ID, Nodes 9 and 10 share the same Cluster ID. 550 AIGP metric ([RFC7311]) is likely applied to the BGP prefix 551 segments as part of a large-scale multi-domain design such as 552 Seamless MPLS [I-D.ietf-mpls-seamless-mpls]. 554 The control-plane behavior is mostly the same as described in the 555 previous section: the only difference is that the eBGP3107 path 556 propagation is simply replaced by an iBGP3107 path reflection with 557 next-hop changed to self. 559 The data-plane tables are exactly the same. 561 5. Applying Segment Routing in the DC with IPv6 dataplane 563 The design described in I-D.ietf-rtgwg-bgp-routing-large-dc 564 [I-D.ietf-rtgwg-bgp-routing-large-dc] is reused with one single 565 modification. We highlight it using the example of the reachability 566 to Node11 via spine switch Node5. 568 Spine5 originates 2001:DB8::5/128 with the attached BGP Prefix 569 Attribute adverting the support of the Segment Routing extension 570 header (SRH, [I-D.previdi-6man-segment-routing-header]) for IPv6 571 packets destined to segment 2001:DB8::5. 573 Tor11 originates 2001:DB8::11/128 with the attached BGP Prefix 574 Attribute adverting the support of the Segment Routing extension 575 header (SRH, [I-D.previdi-6man-segment-routing-header]) for IPv6 576 packets destined to segment 2001:DB8::11. 578 The control-plane and data-plane processing of all the other nodes in 579 the fabric is unchanged. Specifically, the routes to 2001:DB8::5 and 580 2001:DB8::11 are installed in the FIB along the eBGP best-path to 581 Node5 (spine node) and Node11 (ToR node) respectively. 583 An application on HostA which needs to send traffic to HostZ via only 584 Node5 (spine node) can do so by sending IPv6 packets with a SR 585 extension header. The destination address and active segment is set 586 to 2001:DB8::5. The next and last segment is set to 2001:DB8::11. 588 The application must only use IPv6 addresses that have been 589 advertised as capable for SRv6 segment processing (e.g. for which the 590 BGP prefix segment capability has been advertised). How applications 591 learn this (e.g.: centralized controller and orchestration) is 592 outside the scope of this document. 594 6. Communicating path information to the host 596 There are two general methods for communicating path information to 597 the end-hosts: "proactive" and "reactive", aka "push" and "pull" 598 models. There are multiple ways to implement either of these 599 methods. Here, we note that one way could be using a centralized 600 controller: the controller either tells the hosts of the prefix-to- 601 path mappings beforehand and updates them as needed (network event 602 driven push), or responds to the hosts making request for a path to 603 specific destination (host event driven pull). It is also possible 604 to use a hybrid model, i.e., pushing some state from the controller 605 in response to particular network events, while the host pulls other 606 state on demand. 608 We note, that when disseminating network-related data to the end- 609 hosts a trade-off is made to balance the amount of information vs the 610 level of visibility in the network state. This applies both to push 611 and pull models. In the extreme case, the host would request path 612 information on every flow, and keep no local state at all. On the 613 other end of the spectrum, information for every prefix in the 614 network along with available paths could be pushed and continuously 615 updated on all hosts. 617 7. Addressing the open problems 619 This section demonstrates how the problems describe above could be 620 solved using the segment routing concept. It is worth noting that 621 segment routing signaling and data-plane are only parts of the 622 solution. Additional enhancements, e.g. such as the centralized 623 controller mentioned previously, and host networking stack support 624 are required to implement the proposed solutions. 626 7.1. Per-packet and flowlet switching 628 With the ability to choose paths on the host, one may go from per- 629 flow load-sharing in the network to per-packet or per-flowlet (see 630 [KANDULA04] for information on flowlets). The host may select 631 different segment routing instructions either per packet, or per 632 flowlet, and route them over different paths. This allows for 633 solving the "elephant flow" problem in the data-center and avoiding 634 link imbalances. 636 Note that traditional ECMP routing could be easily simulated with on- 637 host path selection, using method proposed in VL2 (see 638 [GREENBERG09]). The hosts would randomly pick a Tier-2 or Tier-1 639 device to "bounce" the packet off of, depending on whether the 640 destination is under the same Tier-2 switches, or has to be reached 641 across Tier-1. The host would use a hash function that operates on 642 per-flow invariants, to simulate per-flow load-sharing in the 643 network. 645 Using Figure 1 as reference, let's illustrate this assuming that 646 HostA has an elephant flow to Z called Flow-f. 648 Normally, a flow is hashed on to a single path. Let's assume HostA 649 sends its packets associated with Flow-f with top label 16011 (the 650 label for the remote ToR, Node11, where HostZ is connected) and Node1 651 would hash all the packets of Flow-F via the same nhop (e.g. Node3). 652 Similarly, let's assume that leaf Node3 would hash all the packets of 653 Flow-F via the same next-hop (e.g.: spine switch Node1). This normal 654 operation would restrict the elephant flow on a small subset of the 655 ECMP paths to HostZ and potentially create imbalance and congestion 656 in the fabric. 658 Leveraging the flowlet proposal, assuming A is made aware of 4 659 disjoint paths via intermediate segment 16005, 16006, 16007 and 16008 660 (the BGP prefix SID's of the 4 spine switches) and also made aware of 661 the prefix segment of the remote ToR connected to the destination 662 (16011), then the application can break the elephant flow F into 663 flowlets F1, F2, F3, F4 and associate each flowlet with one of the 664 following 4 label stacks: {16005, 16011}, {16006, 16011}, {16007, 665 16011} and {16008, 16011}. This would spread the load of the elephant 666 flow through all the ECMP paths available in the fabric and rebalance 667 the load. 669 7.2. Performance-aware routing 671 Knowing the path associated with flows/packets, the end host may 672 deduce certain characteristics of the path on its own, and 673 additionally use the information supplied with path information 674 pushed from the controller or received via pull request. The host 675 may further share its path observations with the centralized agent, 676 so that the latter may keep up-to-date network health map to assist 677 other hosts with this information. 679 For example, an application A.1 at HostA may pin a TCP flow destined 680 to HostZ via Spine switch Node5 using label stack {16005, 16011}. The 681 application A.1 may collect information on packet loss, deduced from 682 TCP retransmissions and other signals (e.g. RTT increases). A.1 may 683 additionally publish this information to a centralized agent, e.g. 684 after a flow completes, or periodically for longer lived flows. 685 Next, using both local and/or global performance data, application 686 A.1 as well as other applications sharing the same resources in the 687 DC fabric may pick up the best path for the new flow, or update an 688 existing path (e.g.: when informed of congestion on an existing 689 path). 691 One particularly interesting instance of performance-aware routing is 692 dynamic fault-avoidance. If some links or devices in the network 693 start discarding packets due to a fault, the end-hosts could detect 694 the path(s) being affected and steer their flows away from the 695 problem spot. Similar logic applies to failure cases where packets 696 get completely black-holed, e.g. when a link goes down. 698 For example, an application A.1 informed about 5 paths to Z {16005, 699 16011}, {16006, 16011}, {16007, 16011}, {16008, 16011} and {16011} 700 might use the latter one by default (for simplicity). When 701 performance is degrading, A.1 might then start to pin TCP flows to 702 each of the 4 other paths (each via a distinct spine) and monitor the 703 performance. It would then detect the faulty path and assign a 704 negative preference to the faulty path to avoid further flows using 705 it. Gradually, over time, it may re-assign flows on the faulty path 706 to eventually detect the resolution of the trouble and start reusing 707 the path. 709 7.3. Non-oblivious routing 711 By leveraging Segment Routing, one avoids issues associated with 712 oblivious ECMP hashing. For example, if in the topology depicted on 713 Figure 1 a link between spine switch Node5 and leaf node Node9 fails, 714 HostA may exclude the segment corresponding to Node5 from the prefix 715 matching the servers under Tier-2 devices Node9. In the push path 716 discovery model, the affected path mappings may be explicitly pushed 717 to all the servers for the duration of the failure. The new mapping 718 would instruct them to avoid the particular Tier-1 switch until the 719 link has recovered. Alternatively, in pull path, the centralized 720 controller may start steering new flows immediately after it 721 discovers the issue. Until then, the existing flows may recover 722 using local detection of the path issues, as described in 723 Section 7.2. 725 7.4. Deterministic network probing 727 Active probing is a well-known technique for monitoring network 728 elements health, constituting of sending continuous packet streams 729 simulating network traffic to the hosts in the data-center. Segment 730 routing makes possible to prescribe the exact paths that each probe 731 or series of probes would be taking toward their destination. This 732 allows for fast correlation and detection of failed paths, by 733 processing information from multiple actively probing agents. This 734 complements the data collected from the hosts routing stacks as 735 described inSection 7.2. 737 For example, imagine a probe agent sending packets to all machines in 738 the data-center. For every host, it may send packets over each of 739 the possible paths, knowing exactly which links and devices these 740 packets will be crossing. Correlating results for multiple 741 destinations with the topological data, it may automatically isolate 742 possible problem to a link or device in the network. 744 8. Additional Benefits 746 8.1. MPLS Dataplane with operational simplicity 748 As required by [I-D.ietf-rtgwg-bgp-routing-large-dc], no new 749 signaling protocol is introduced. The Prefix Segment is a 750 lightweight extension to BGP Labelled Unicast (RFC3107 [RFC3107]). 751 It applies either to eBGP or iBGP based designs. 753 Specifically, LDP and RSVP-TE are not used. These protocols would 754 drastically impact the operational complexity of the Data Center and 755 would not scale. This is in line with the requirements expressed in 756 [I-D.ietf-rtgwg-bgp-routing-large-dc] 757 A key element of the operational simplicity is the deployment of the 758 design with a single and consistent SRGB across the DC fabric. 760 At every node in the fabric, the same label is associated to a given 761 BGP prefix segment and hence a notion of global prefix segment 762 arises. 764 When a controller programs HostA to send traffic to HostZ via the 765 normally available BGP ECMP paths, the controller uses label 16011 766 associated with the ToR switch connected to the HostZ. The 767 controller does not need to pick the label based on the ToR that the 768 source host is connected to. 770 In a classic BGP Labelled Unicast design applied to the DC fabric 771 illustrated in Figure 1, the ToR Node1 connected to HostA would most 772 likely allocate a different label for 192.0.2.11/32 than the one 773 allocated by ToR Node2. As a consequence, the controller would need 774 to adapt the SR policy to each host, based on the ToR switch that 775 they are connected to. This adds state maintenance and 776 synchronization problems. All of this unnecessary complexity is 777 eliminated if a single consistent SRGB is utilized across the fabric. 779 8.2. Minimizing the FIB table 781 The designer may decide to switch all the traffic at Tier-1 and Tier- 782 2's based on MPLS, hence drastically decreasing the IP table size at 783 these nodes. 785 This is easily accomplished by encapsulating the traffic either 786 directly at the host or at the source ToR switch by pushing the BGP- 787 Prefix Segment of the destination ToR for intra-DC traffic or border 788 switch for inter-DC or DC-to-outside-world traffic. 790 8.3. Egress Peer Engineering 792 It is straightforward to combine the design illustrated in this 793 document with the Egress Peer Engineering (EPE) use-case described in 794 [I-D.filsfils-spring-segment-routing-central-epe]. 796 In such case, the operator is able to engineer its outbound traffic 797 on a per host-flow basis, without incurring any additional state at 798 intermediate points in the DC fabric. 800 For example, the controller only needs to inject a per-flow state on 801 the HostA to force it to send its traffic destined to a specific 802 Internet destination D via a selected border switch (say Node12 in 803 Figure 1 instead of another border switch Node11) and a specific 804 egress peer of Node12 (say peer AS 9999 of local PeerNode segment 805 9999 at Node12 instead of any other peer which provides a path to the 806 destination D). Any packet matching this state at host A would be 807 encapsulated with SR segment list (label stack) {16012, 9999}. 16012 808 would steer the flow through the DC fabric, leveraging any ECMP, 809 along the best path to border switch Node12. Once the flow gets to 810 border switch Node12, the active segment is 9999 (thanks to PHP on 811 the upstream neighbor of Node12). This EPE PeerNode segment forces 812 border switch Node12 to forward the packet to peer AS 9999, without 813 any IP lookup at the border switch. There is no per-flow state for 814 this engineered flow in the DC fabric. A benefit of segment routing 815 is the per-flow state is only required at the source. 817 As well as allowing full traffic engineering control such a design 818 also offers FIB table minimization benefits as the Internet- scale 819 FIB at border switch Node12 is not required if all FIB lookups are 820 avoided there by using EPE. 822 8.4. Incremental Deployments 824 As explained in Section 4.2.5, this design can be deployed 825 incrementally. 827 8.5. Anycast 829 The design presented in this document preserves the availability and 830 load-balancing properties of the base design presented in 831 [I-D.filsfils-spring-segment-routing]. 833 For example, one could assign an anycast loopback 192.0.2.20/32 and 834 associate segment index 20 to it on the border switches 11 and 12 (in 835 addition to their node-specific loopbacks). Doing so, the EPE 836 controller could express a default "go-to-the- Internet via any 837 border switch" policy as segment list {16020}. Indeed, from any host 838 in the DC fabric or from any ToR switch, 16020 steers the packet 839 towards the border switches 11 or 12 leveraging ECMP where available 840 along the best paths to these switches. 842 9. Preferred SRGB Allocation 844 In the MPLS case, we do not recommend to use different SRGBs at each 845 node. 847 Different SRGBs in each node likely increase the complexity of the 848 solution both from an operation viewpoint and from a controller 849 viewpoint. 851 From an operation viewpoint, it is much simpler to have the same 852 global label at every node for the same destination (the MPLS 853 troubleshooting is then similar to the IPv6 troubleshooting where 854 this global property is a given). 856 From a controller viewpoint, this allows to construct simple policies 857 applicable across the fabric. 859 Let us consider two applications A and B respectively connected to 860 ToR1 and ToR2. A has two flows FA1 and FA2 destined to Z. B has two 861 flows FB1 and FB2 destined to Z. The controller wants FA1 and FB1 to 862 be load-shared across the fabric while FA2 and FB2 must be 863 respectively steered via Spine5 and spine 8. 865 Assuming a consistent unique SRGB across the fabric as described in 866 the document, the controller can simply do it by instructing A and B 867 to use {16011} respectively for FA1 and FB1 and by instructing A and 868 B to use {16005 16011} and {16008 16011} respectively for FA2 and 869 FB2. 871 Let us assume a design where the SRGB is different at every node: 872 SRGB of Node K starts at value K*1000 and the SRGB length is 1000 873 (e.g. ToR1's SRGB is [1000, 1999], ToR2's SRGB is [2000, 2999]...). 875 In this case, not only the controller would need to collect and store 876 all of these different SRGB's, furthermore it would need to adapt the 877 policy for each host. Indeed, the controller would instruct A to use 878 {1011} for FA1 while it would have to instruct B to use {2011} for 879 FB1 (while with the same SRGB, both policies are the same {16011}). 881 Even worse, the controller would instruct A to use {1005, 5011} for 882 FA1 while it would instruct B to use {2011, 8011} for FB1 (while with 883 the same SRGB, the second segment is the same across both policies: 884 16011). When combining segments to create a policy, one need to 885 carefully update the label of each segment. This is obviously more 886 error-prone, more complex and more difficult to troubleshoot. 888 10. Alternative Options 890 In order to support all the requirements and get consensus, the BGP 891 Prefix SID attribute has been extended to allow this design. 893 Specifically, the ORIGINATOR_SRGB TLV in the BGP Prefix SID signals 894 the SRGB of the switch that originated the BGP Prefix Segment. 896 This allows to determine the local label allocated by any switch for 897 any BGP Prefix Segment, despite the lack of a consistent unique SRGB 898 in the domain. 900 11. IANA Considerations 902 TBD 904 12. Manageability Considerations 906 TBD 908 13. Security Considerations 910 TBD 912 14. Contributors 914 Benjamin Black, Arjun Arjun Sreekantiah and Keyur Patel have 915 contributed to the content of this document. 917 15. Acknowledgements 919 The authors would like to thank Acee Lindem for his review. 921 16. References 923 16.1. Normative References 925 [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate 926 Requirement Levels", BCP 14, RFC 2119, March 1997. 928 [RFC3107] Rekhter, Y. and E. Rosen, "Carrying Label Information in 929 BGP-4", RFC 3107, May 2001. 931 [RFC4271] Rekhter, Y., Li, T., and S. Hares, "A Border Gateway 932 Protocol 4 (BGP-4)", RFC 4271, January 2006. 934 [RFC7311] Mohapatra, P., Fernando, R., Rosen, E., and J. Uttaro, 935 "The Accumulated IGP Metric Attribute for BGP", RFC 7311, 936 August 2014. 938 16.2. Informative References 940 [GREENBERG09] 941 Greenberg, A., Hamilton, J., Jain, N., Kadula, S., Kim, 942 C., Lahiri, P., Maltz, D., Patel, P., and S. Sengupta, 943 "VL2: A Scalable and Flexible Data Center Network", 2009. 945 [I-D.filsfils-spring-segment-routing] 946 Filsfils, C., Previdi, S., Bashandy, A., Decraene, B., 947 Litkowski, S., Horneffer, M., Milojevic, I., Shakir, R., 948 Ytti, S., Henderickx, W., Tantsura, J., and E. Crabbe, 949 "Segment Routing Architecture", draft-filsfils-spring- 950 segment-routing-04 (work in progress), July 2014. 952 [I-D.filsfils-spring-segment-routing-central-epe] 953 Filsfils, C., Previdi, S., Patel, K., Aries, E., 954 shaw@fb.com, s., Ginsburg, D., and D. Afanasiev, "Segment 955 Routing Centralized Egress Peer Engineering", draft- 956 filsfils-spring-segment-routing-central-epe-04 (work in 957 progress), July 2015. 959 [I-D.ietf-mpls-seamless-mpls] 960 Leymann, N., Decraene, B., Filsfils, C., Konstantynowicz, 961 M., and D. Steinberg, "Seamless MPLS Architecture", draft- 962 ietf-mpls-seamless-mpls-07 (work in progress), June 2014. 964 [I-D.ietf-rtgwg-bgp-routing-large-dc] 965 Lapukhov, P., Premji, A., and J. Mitchell, "Use of BGP for 966 routing in large-scale data centers", draft-ietf-rtgwg- 967 bgp-routing-large-dc-03 (work in progress), June 2015. 969 [I-D.keyupate-idr-bgp-prefix-sid] 970 Patel, K., Previdi, S., Filsfils, C., Sreekantiah, A., 971 Ray, S., and H. Gredler, "Segment Routing Prefix SID 972 extensions for BGP", draft-keyupate-idr-bgp-prefix-sid-04 973 (work in progress), July 2015. 975 [I-D.previdi-6man-segment-routing-header] 976 Previdi, S., Filsfils, C., Field, B., and I. Leung, "IPv6 977 Segment Routing Header (SRH)", draft-previdi-6man-segment- 978 routing-header-06 (work in progress), May 2015. 980 [KANDULA04] 981 Sinha, S., Kandula, S., and D. Katabi, "Harnessing TCP's 982 Burstiness with Flowlet Switching", 2004. 984 Authors' Addresses 986 Clarence Filsfils (editor) 987 Cisco Systems, Inc. 988 Brussels 989 BE 991 Email: cfilsfil@cisco.com 992 Stefano Previdi (editor) 993 Cisco Systems, Inc. 994 Via Del Serafico, 200 995 Rome 00142 996 Italy 998 Email: sprevidi@cisco.com 1000 Jon Mitchell 1001 Unaffiliated 1003 Email: jrmitche@puck.nether.net 1005 Ebben Aries 1006 Facebook 1007 US 1009 Email: exa@fb.com 1011 P. Lapukhov 1012 Facebook 1013 US 1015 Email: petr@fb.com 1017 G. Nagarajan 1018 Facebook 1019 US 1021 Email: gaya@fb.com 1023 Dmitry Afanasiev 1024 Yandex 1025 RU 1027 Email: fl0w@yandex-team.ru 1029 Tim Laberge 1030 Microsoft 1032 Email: tim.laberge@microsoft.com 1033 Edet Nkposong 1034 Microsoft 1036 Email: edetn@microsoft.com 1038 Mohan Nanduri 1039 Microsoft 1041 Email: mnanduri@microsoft.com 1043 James Uttaro 1044 ATT 1046 Email: ju1738@att.com 1048 Saikat Ray 1049 Unaffiliated 1051 Email: raysaikat@gmail.com