idnits 2.17.00 (12 Aug 2021) /tmp/idnits11406/draft-ietf-spring-segment-routing-msdc-09.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 (May 29, 2018) is 1452 days in the past. Is this intentional? Checking references for intended status: Informational ---------------------------------------------------------------------------- -- Looks like a reference, but probably isn't: '16000' on line 259 -- Looks like a reference, but probably isn't: '23999' on line 259 -- Looks like a reference, but probably isn't: '1000' on line 876 -- Looks like a reference, but probably isn't: '1999' on line 876 -- Looks like a reference, but probably isn't: '2000' on line 877 -- Looks like a reference, but probably isn't: '2999' on line 877 == Unused Reference: 'RFC2119' is defined on line 1024, but no explicit reference was found in the text == Outdated reference: draft-ietf-idr-bgp-prefix-sid has been published as RFC 8669 == Outdated reference: draft-ietf-spring-segment-routing has been published as RFC 8402 == Outdated reference: draft-ietf-spring-segment-routing-central-epe has been published as RFC 9087 == Outdated reference: draft-ietf-6man-segment-routing-header has been published as RFC 8754 Summary: 0 errors (**), 0 flaws (~~), 6 warnings (==), 7 comments (--).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 4 Intended status: Informational Cisco Systems, Inc. 5 Expires: November 30, 2018 G. Dawra 6 LinkedIn 7 E. Aries 8 Juniper Networks 9 P. Lapukhov 10 Facebook 11 May 29, 2018 13 BGP-Prefix Segment in large-scale data centers 14 draft-ietf-spring-segment-routing-msdc-09 16 Abstract 18 This document describes the motivation and benefits for applying 19 segment routing in BGP-based large-scale data-centers. It describes 20 the design to deploy segment routing in those data-centers, for both 21 the MPLS and IPv6 dataplanes. 23 Status of This Memo 25 This Internet-Draft is submitted in full conformance with the 26 provisions of BCP 78 and BCP 79. 28 Internet-Drafts are working documents of the Internet Engineering 29 Task Force (IETF). Note that other groups may also distribute 30 working documents as Internet-Drafts. The list of current Internet- 31 Drafts is at https://datatracker.ietf.org/drafts/current/. 33 Internet-Drafts are draft documents valid for a maximum of six months 34 and may be updated, replaced, or obsoleted by other documents at any 35 time. It is inappropriate to use Internet-Drafts as reference 36 material or to cite them other than as "work in progress." 38 This Internet-Draft will expire on November 30, 2018. 40 Copyright Notice 42 Copyright (c) 2018 IETF Trust and the persons identified as the 43 document authors. All rights reserved. 45 This document is subject to BCP 78 and the IETF Trust's Legal 46 Provisions Relating to IETF Documents 47 (https://trustee.ietf.org/license-info) in effect on the date of 48 publication of this document. Please review these documents 49 carefully, as they describe your rights and restrictions with respect 50 to this document. Code Components extracted from this document must 51 include Simplified BSD License text as described in Section 4.e of 52 the Trust Legal Provisions and are provided without warranty as 53 described in the Simplified BSD License. 55 Table of Contents 57 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 2 58 2. Large Scale Data Center Network Design Summary . . . . . . . 3 59 2.1. Reference design . . . . . . . . . . . . . . . . . . . . 4 60 3. Some open problems in large data-center networks . . . . . . 5 61 4. Applying Segment Routing in the DC with MPLS dataplane . . . 6 62 4.1. BGP Prefix Segment (BGP-Prefix-SID) . . . . . . . . . . . 6 63 4.2. eBGP Labeled Unicast (RFC8277) . . . . . . . . . . . . . 6 64 4.2.1. Control Plane . . . . . . . . . . . . . . . . . . . . 7 65 4.2.2. Data Plane . . . . . . . . . . . . . . . . . . . . . 8 66 4.2.3. Network Design Variation . . . . . . . . . . . . . . 9 67 4.2.4. Global BGP Prefix Segment through the fabric . . . . 10 68 4.2.5. Incremental Deployments . . . . . . . . . . . . . . . 10 69 4.3. iBGP Labeled Unicast (RFC8277) . . . . . . . . . . . . . 11 70 5. Applying Segment Routing in the DC with IPv6 dataplane . . . 13 71 6. Communicating path information to the host . . . . . . . . . 13 72 7. Addressing the open problems . . . . . . . . . . . . . . . . 14 73 7.1. Per-packet and flowlet switching . . . . . . . . . . . . 14 74 7.2. Performance-aware routing . . . . . . . . . . . . . . . . 15 75 7.3. Deterministic network probing . . . . . . . . . . . . . . 16 76 8. Additional Benefits . . . . . . . . . . . . . . . . . . . . . 17 77 8.1. MPLS Dataplane with operational simplicity . . . . . . . 17 78 8.2. Minimizing the FIB table . . . . . . . . . . . . . . . . 17 79 8.3. Egress Peer Engineering . . . . . . . . . . . . . . . . . 17 80 8.4. Anycast . . . . . . . . . . . . . . . . . . . . . . . . . 18 81 9. Preferred SRGB Allocation . . . . . . . . . . . . . . . . . . 18 82 10. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 19 83 11. Manageability Considerations . . . . . . . . . . . . . . . . 19 84 12. Security Considerations . . . . . . . . . . . . . . . . . . . 20 85 13. Acknowledgements . . . . . . . . . . . . . . . . . . . . . . 20 86 14. Contributors . . . . . . . . . . . . . . . . . . . . . . . . 20 87 15. References . . . . . . . . . . . . . . . . . . . . . . . . . 22 88 15.1. Normative References . . . . . . . . . . . . . . . . . . 22 89 15.2. Informative References . . . . . . . . . . . . . . . . . 23 90 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 23 92 1. Introduction 94 Segment Routing (SR), as described in 95 [I-D.ietf-spring-segment-routing] leverages the source routing 96 paradigm. A node steers a packet through an ordered list of 97 instructions, called segments. A segment can represent any 98 instruction, topological or service-based. A segment can have a 99 local semantic to an SR node or global within an SR domain. SR 100 allows to enforce a flow through any topological path while 101 maintaining per-flow state only at the ingress node to the SR domain. 102 Segment Routing can be applied to the MPLS and IPv6 data-planes. 104 The use-cases described in this document should be considered in the 105 context of the BGP-based large-scale data-center (DC) design 106 described in [RFC7938]. This document extends it by applying SR both 107 with IPv6 and MPLS dataplane. 109 2. Large Scale Data Center Network Design Summary 111 This section provides a brief summary of the informational document 112 [RFC7938] that outlines a practical network design suitable for data- 113 centers of various scales: 115 o Data-center networks have highly symmetric topologies with 116 multiple parallel paths between two server attachment points. The 117 well-known Clos topology is most popular among the operators (as 118 described in [RFC7938]). In a Clos topology, the minimum number 119 of parallel paths between two elements is determined by the 120 "width" of the "Tier-1" stage. See Figure 1 below for an 121 illustration of the concept. 123 o Large-scale data-centers commonly use a routing protocol, such as 124 BGP-4 [RFC4271] in order to provide endpoint connectivity. 125 Recovery after a network failure is therefore driven either by 126 local knowledge of directly available backup paths or by 127 distributed signaling between the network devices. 129 o Within data-center networks, traffic is load-shared using the 130 Equal Cost Multipath (ECMP) mechanism. With ECMP, every network 131 device implements a pseudo-random decision, mapping packets to one 132 of the parallel paths by means of a hash function calculated over 133 certain parts of the packet, typically a combination of various 134 packet header fields. 136 The following is a schematic of a five-stage Clos topology, with four 137 devices in the "Tier-1" stage. Notice that number of paths between 138 Node1 and Node12 equals to four: the paths have to cross all of 139 Tier-1 devices. At the same time, the number of paths between Node1 140 and Node2 equals two, and the paths only cross Tier-2 devices. Other 141 topologies are possible, but for simplicity only the topologies that 142 have a single path from Tier-1 to Tier-3 are considered below. The 143 rest could be treated similarly, with a few modifications to the 144 logic. 146 2.1. Reference design 148 Tier-1 149 +-----+ 150 |NODE | 151 +->| 5 |--+ 152 | +-----+ | 153 Tier-2 | | Tier-2 154 +-----+ | +-----+ | +-----+ 155 +------------>|NODE |--+->|NODE |--+--|NODE |-------------+ 156 | +-----| 3 |--+ | 6 | +--| 9 |-----+ | 157 | | +-----+ +-----+ +-----+ | | 158 | | | | 159 | | +-----+ +-----+ +-----+ | | 160 | +-----+---->|NODE |--+ |NODE | +--|NODE |-----+-----+ | 161 | | | +---| 4 |--+->| 7 |--+--| 10 |---+ | | | 162 | | | | +-----+ | +-----+ | +-----+ | | | | 163 | | | | | | | | | | 164 +-----+ +-----+ | +-----+ | +-----+ +-----+ 165 |NODE | |NODE | Tier-3 +->|NODE |--+ Tier-3 |NODE | |NODE | 166 | 1 | | 2 | | 8 | | 11 | | 12 | 167 +-----+ +-----+ +-----+ +-----+ +-----+ 168 | | | | | | | | 169 A O B O <- Servers -> Z O O O 171 Figure 1: 5-stage Clos topology 173 In the reference topology illustrated in Figure 1, It is assumed: 175 o Each node is its own AS (Node X has AS X). 4-byte AS numbers are 176 recommended ([RFC6793]). 178 * For simple and efficient route propagation filtering, Node5, 179 Node6, Node7 and Node8 use the same AS, Node3 and Node4 use the 180 same AS, Node9 and Node10 use the same AS. 182 * In case of 2-byte autonomous system numbers are used and for 183 efficient usage of the scarce 2-byte Private Use AS pool, 184 different Tier-3 nodes might use the same AS. 186 * Without loss of generality, these details will be simplified in 187 this document and assume that each node has its own AS. 189 o Each node peers with its neighbors with a BGP session. If not 190 specified, eBGP is assumed. In a specific use-case, iBGP will be 191 used but this will be called out explicitly in that case. 193 o Each node originates the IPv4 address of its loopback interface 194 into BGP and announces it to its neighbors. 196 * The loopback of Node X is 192.0.2.x/32. 198 In this document, the Tier-1, Tier-2 and Tier-3 nodes are referred to 199 respectively as Spine, Leaf and ToR (top of rack) nodes. When a ToR 200 node acts as a gateway to the "outside world", it is referred to as a 201 border node. 203 3. Some open problems in large data-center networks 205 The data-center network design summarized above provides means for 206 moving traffic between hosts with reasonable efficiency. There are 207 few open performance and reliability problems that arise in such 208 design: 210 o ECMP routing is most commonly realized per-flow. This means that 211 large, long-lived "elephant" flows may affect performance of 212 smaller, short-lived "mouse" flows and reduce efficiency of per- 213 flow load-sharing. In other words, per-flow ECMP does not perform 214 efficiently when flow lifetime distribution is heavy-tailed. 215 Furthermore, due to hash-function inefficiencies it is possible to 216 have frequent flow collisions, where more flows get placed on one 217 path over the others. 219 o Shortest-path routing with ECMP implements an oblivious routing 220 model, which is not aware of the network imbalances. If the 221 network symmetry is broken, for example due to link failures, 222 utilization hotspots may appear. For example, if a link fails 223 between Tier-1 and Tier-2 devices (e.g. Node5 and Node9), Tier-3 224 devices Node1 and Node2 will not be aware of that, since there are 225 other paths available from perspective of Node3. They will 226 continue sending roughly equal traffic to Node3 and Node4 as if 227 the failure didn't exist which may cause a traffic hotspot. 229 o Isolating faults in the network with multiple parallel paths and 230 ECMP-based routing is non-trivial due to lack of determinism. 231 Specifically, the connections from HostA to HostB may take a 232 different path every time a new connection is formed, thus making 233 consistent reproduction of a failure much more difficult. This 234 complexity scales linearly with the number of parallel paths in 235 the network, and stems from the random nature of path selection by 236 the network devices. 238 Further in this document (Section 7), it is demonstrated how these 239 problems could be addressed within the framework of Segment Routing. 241 First, it will be explained how to apply SR in the DC, for MPLS and 242 IPv6 data-planes. 244 4. Applying Segment Routing in the DC with MPLS dataplane 246 4.1. BGP Prefix Segment (BGP-Prefix-SID) 248 A BGP Prefix Segment is a segment associated with a BGP prefix. A 249 BGP Prefix Segment is a network-wide instruction to forward the 250 packet along the ECMP-aware best path to the related prefix. 252 The BGP Prefix Segment is defined as the BGP-Prefix-SID Attribute in 253 [I-D.ietf-idr-bgp-prefix-sid] which contains an index. Throughout 254 this document the BGP Prefix Segment Attribute is referred as the 255 BGP-Prefix-SID and the encoded index as the label-index. 257 In this document, the network design decision has been made to assume 258 that all the nodes are allocated the same SRGB (Segment Routing 259 Global Block), e.g. [16000, 23999]. This provides operational 260 simplification as explained in Section 9, but this is not a 261 requirement. 263 For illustration purpose, when considering an MPLS data-plane, it is 264 assumed that the label-index allocated to prefix 192.0.2.x/32 is X. 265 As a result, a local label (16000+x) is allocated for prefix 266 192.0.2.x/32 by each node throughout the DC fabric. 268 When IPv6 data-plane is considered, it is assumed that Node X is 269 allocated IPv6 address (segment) 2001:DB8::X. 271 4.2. eBGP Labeled Unicast (RFC8277) 273 Referring to Figure 1 and [RFC7938], the following design 274 modifications are introduced: 276 o Each node peers with its neighbors via a eBGP session with 277 extensions defined in [RFC8277] (named "eBGP8277" throughout this 278 document) and with the BGP-Prefix-SID attribute extension as 279 defined in [I-D.ietf-idr-bgp-prefix-sid]. 281 o The forwarding plane at Tier-2 and Tier-1 is MPLS. 283 o The forwarding plane at Tier-3 is either IP2MPLS (if the host 284 sends IP traffic) or MPLS2MPLS (if the host sends MPLS- 285 encapsulated traffic). 287 Figure 2 zooms into a path from server A to server Z within the 288 topology of Figure 1. 290 +-----+ +-----+ +-----+ 291 +---------->|NODE | |NODE | |NODE | 292 | | 4 |--+->| 7 |--+--| 10 |---+ 293 | +-----+ +-----+ +-----+ | 294 | | 295 +-----+ +-----+ 296 |NODE | |NODE | 297 | 1 | | 11 | 298 +-----+ +-----+ 299 | | 300 A <- Servers -> Z 302 Figure 2: Path from A to Z via nodes 1, 4, 7, 10 and 11 304 Referring to Figure 1 and Figure 2 and assuming the IP address with 305 the AS and label-index allocation previously described, the following 306 sections detail the control plane operation and the data plane states 307 for the prefix 192.0.2.11/32 (loopback of Node11) 309 4.2.1. Control Plane 311 Node11 originates 192.0.2.11/32 in BGP and allocates to it a BGP- 312 Prefix-SID with label-index: index11 [I-D.ietf-idr-bgp-prefix-sid]. 314 Node11 sends the following eBGP8277 update to Node10: 316 . IP Prefix: 192.0.2.11/32 317 . Label: Implicit-Null 318 . Next-hop: Node11's interface address on the link to Node10 319 . AS Path: {11} 320 . BGP-Prefix-SID: Label-Index 11 322 Node10 receives the above update. As it is SR capable, Node10 is 323 able to interpret the BGP-Prefix-SID and hence understands that it 324 should allocate the label from its own SRGB block, offset by the 325 Label-Index received in the BGP-Prefix-SID (16000+11 hence 16011) to 326 the NLRI instead of allocating a non-deterministic label out of a 327 dynamically allocated portion of the local label space. The 328 implicit-null label in the NLRI tells Node10 that it is the 329 penultimate hop and must pop the top label on the stack before 330 forwarding traffic for this prefix to Node11. 332 Then, Node10 sends the following eBGP8277 update to Node7: 334 . IP Prefix: 192.0.2.11/32 335 . Label: 16011 336 . Next-hop: Node10's interface address on the link to Node7 337 . AS Path: {10, 11} 338 . BGP-Prefix-SID: Label-Index 11 340 Node7 receives the above update. As it is SR capable, Node7 is able 341 to interpret the BGP-Prefix-SID and hence allocates the local 342 (incoming) label 16011 (16000 + 11) to the NLRI (instead of 343 allocating a "dynamic" local label from its label manager). Node7 344 uses the label in the received eBGP8277 NLRI as the outgoing label 345 (the index is only used to derive the local/incoming label). 347 Node7 sends the following eBGP8277 update to Node4: 349 . IP Prefix: 192.0.2.11/32 350 . Label: 16011 351 . Next-hop: Node7's interface address on the link to Node4 352 . AS Path: {7, 10, 11} 353 . BGP-Prefix-SID: Label-Index 11 355 Node4 receives the above update. As it is SR capable, Node4 is able 356 to interpret the BGP-Prefix-SID and hence allocates the local 357 (incoming) label 16011 to the NLRI (instead of allocating a "dynamic" 358 local label from its label manager). Node4 uses the label in the 359 received eBGP8277 NLRI as outgoing label (the index is only used to 360 derive the local/incoming label). 362 Node4 sends the following eBGP8277 update to Node1: 364 . IP Prefix: 192.0.2.11/32 365 . Label: 16011 366 . Next-hop: Node4's interface address on the link to Node1 367 . AS Path: {4, 7, 10, 11} 368 . BGP-Prefix-SID: Label-Index 11 370 Node1 receives the above update. As it is SR capable, Node1 is able 371 to interpret the BGP-Prefix-SID and hence allocates the local 372 (incoming) label 16011 to the NLRI (instead of allocating a "dynamic" 373 local label from its label manager). Node1 uses the label in the 374 received eBGP8277 NLRI as outgoing label (the index is only used to 375 derive the local/incoming label). 377 4.2.2. Data Plane 379 Referring to Figure 1, and assuming all nodes apply the same 380 advertisement rules described above and all nodes have the same SRGB 381 (16000-23999), here are the IP/MPLS forwarding tables for prefix 382 192.0.2.11/32 at Node1, Node4, Node7 and Node10. 384 ----------------------------------------------- 385 Incoming label | outgoing label | Outgoing 386 or IP destination | | Interface 387 ------------------+----------------+----------- 388 16011 | 16011 | ECMP{3, 4} 389 192.0.2.11/32 | 16011 | ECMP{3, 4} 390 ------------------+----------------+----------- 392 Figure 3: Node1 Forwarding Table 394 ----------------------------------------------- 395 Incoming label | outgoing label | Outgoing 396 or IP destination | | Interface 397 ------------------+----------------+----------- 398 16011 | 16011 | ECMP{7, 8} 399 192.0.2.11/32 | 16011 | ECMP{7, 8} 400 ------------------+----------------+----------- 402 Figure 4: Node4 Forwarding Table 404 ----------------------------------------------- 405 Incoming label | outgoing label | Outgoing 406 or IP destination | | Interface 407 ------------------+----------------+----------- 408 16011 | 16011 | 10 409 192.0.2.11/32 | 16011 | 10 410 ------------------+----------------+----------- 412 Figure 5: Node7 Forwarding Table 414 ----------------------------------------------- 415 Incoming label | outgoing label | Outgoing 416 or IP destination | | Interface 417 ------------------+----------------+----------- 418 16011 | POP | 11 419 192.0.2.11/32 | N/A | 11 420 ------------------+----------------+----------- 422 Node10 Forwarding Table 424 4.2.3. Network Design Variation 426 A network design choice could consist of switching all the traffic 427 through Tier-1 and Tier-2 as MPLS traffic. In this case, one could 428 filter away the IP entries at Node4, Node7 and Node10. This might be 429 beneficial in order to optimize the forwarding table size. 431 A network design choice could consist in allowing the hosts to send 432 MPLS-encapsulated traffic based on the Egress Peer Engineering (EPE) 433 use-case as defined in [I-D.ietf-spring-segment-routing-central-epe]. 434 For example, applications at HostA would send their Z-destined 435 traffic to Node1 with an MPLS label stack where the top label is 436 16011 and the next label is an EPE peer segment 437 ([I-D.ietf-spring-segment-routing-central-epe]) at Node11 directing 438 the traffic to Z. 440 4.2.4. Global BGP Prefix Segment through the fabric 442 When the previous design is deployed, the operator enjoys global BGP- 443 Prefix-SID and label allocation throughout the DC fabric. 445 A few examples follow: 447 o Normal forwarding to Node11: a packet with top label 16011 448 received by any node in the fabric will be forwarded along the 449 ECMP-aware BGP best-path towards Node11 and the label 16011 is 450 penultimate-popped at Node10 (or at Node 9). 452 o Traffic-engineered path to Node11: an application on a host behind 453 Node1 might want to restrict its traffic to paths via the Spine 454 node Node5. The application achieves this by sending its packets 455 with a label stack of {16005, 16011}. BGP Prefix SID 16005 directs 456 the packet up to Node5 along the path (Node1, Node3, Node5). BGP- 457 Prefix-SID 16011 then directs the packet down to Node11 along the 458 path (Node5, Node9, Node11). 460 4.2.5. Incremental Deployments 462 The design previously described can be deployed incrementally. Let 463 us assume that Node7 does not support the BGP-Prefix-SID and let us 464 show how the fabric connectivity is preserved. 466 From a signaling viewpoint, nothing would change: even though Node7 467 does not support the BGP-Prefix-SID, it does propagate the attribute 468 unmodified to its neighbors. 470 From a label allocation viewpoint, the only difference is that Node7 471 would allocate a dynamic (random) label to the prefix 192.0.2.11/32 472 (e.g. 123456) instead of the "hinted" label as instructed by the BGP- 473 Prefix-SID. The neighbors of Node7 adapt automatically as they 474 always use the label in the BGP8277 NLRI as outgoing label. 476 Node4 does understand the BGP-Prefix-SID and hence allocates the 477 indexed label in the SRGB (16011) for 192.0.2.11/32. 479 As a result, all the data-plane entries across the network would be 480 unchanged except the entries at Node7 and its neighbor Node4 as shown 481 in the figures below. 483 The key point is that the end-to-end Label Switched Path (LSP) is 484 preserved because the outgoing label is always derived from the 485 received label within the BGP8277 NLRI. The index in the BGP-Prefix- 486 SID is only used as a hint on how to allocate the local label (the 487 incoming label) but never for the outgoing label. 489 ------------------------------------------ 490 Incoming label | outgoing | Outgoing 491 or IP destination | label | Interface 492 -------------------+---------------------- 493 12345 | 16011 | 10 495 Figure 7: Node7 Forwarding Table 497 ------------------------------------------ 498 Incoming label | outgoing | Outgoing 499 or IP destination | label | Interface 500 -------------------+---------------------- 501 16011 | 12345 | 7 503 Figure 8: Node4 Forwarding Table 505 The BGP-Prefix-SID can thus be deployed incrementally one node at a 506 time. 508 When deployed together with a homogeneous SRGB (same SRGB across the 509 fabric), the operator incrementally enjoys the global prefix segment 510 benefits as the deployment progresses through the fabric. 512 4.3. iBGP Labeled Unicast (RFC8277) 514 The same exact design as eBGP8277 is used with the following 515 modifications: 517 All nodes use the same AS number. 519 Each node peers with its neighbors via an internal BGP session 520 (iBGP) with extensions defined in [RFC8277] (named "iBGP8277" 521 throughout this document). 523 Each node acts as a route-reflector for each of its neighbors and 524 with the next-hop-self option. Next-hop-self is a well known 525 operational feature which consists of rewriting the next-hop of a 526 BGP update prior to send it to the neighbor. Usually, it's a 527 common practice to apply next-hop-self behavior towards iBGP peers 528 for eBGP learned routes. In the case outlined in this section it 529 is proposed to use the next-hop-self mechanism also to iBGP 530 learned routes. 532 Cluster-1 533 +-----------+ 534 | Tier-1 | 535 | +-----+ | 536 | |NODE | | 537 | | 5 | | 538 Cluster-2 | +-----+ | Cluster-3 539 +---------+ | | +---------+ 540 | Tier-2 | | | | Tier-2 | 541 | +-----+ | | +-----+ | | +-----+ | 542 | |NODE | | | |NODE | | | |NODE | | 543 | | 3 | | | | 6 | | | | 9 | | 544 | +-----+ | | +-----+ | | +-----+ | 545 | | | | | | 546 | | | | | | 547 | +-----+ | | +-----+ | | +-----+ | 548 | |NODE | | | |NODE | | | |NODE | | 549 | | 4 | | | | 7 | | | | 10 | | 550 | +-----+ | | +-----+ | | +-----+ | 551 +---------+ | | +---------+ 552 | | 553 | +-----+ | 554 | |NODE | | 555 Tier-3 | | 8 | | Tier-3 556 +-----+ +-----+ | +-----+ | +-----+ +-----+ 557 |NODE | |NODE | +-----------+ |NODE | |NODE | 558 | 1 | | 2 | | 11 | | 12 | 559 +-----+ +-----+ +-----+ +-----+ 561 Figure 9: iBGP Sessions with Reflection and Next-Hop-Self 563 For simple and efficient route propagation filtering and as 564 illustrated in Figure 9: 566 Node5, Node6, Node7 and Node8 use the same Cluster ID (Cluster- 567 1) 568 Node3 and Node4 use the same Cluster ID (Cluster-2) 570 Node9 and Node10 use the same Cluster ID (Cluster-3) 572 The control-plane behavior is mostly the same as described in the 573 previous section: the only difference is that the eBGP8277 path 574 propagation is simply replaced by an iBGP8277 path reflection with 575 next-hop changed to self. 577 The data-plane tables are exactly the same. 579 5. Applying Segment Routing in the DC with IPv6 dataplane 581 The design described in [RFC7938] is reused with one single 582 modification. It is highlighted using the example of the 583 reachability to Node11 via spine node Node5. 585 Node5 originates 2001:DB8::5/128 with the attached BGP-Prefix-SID for 586 IPv6 packets destined to segment 2001:DB8::5 587 ([I-D.ietf-idr-bgp-prefix-sid]). 589 Node11 originates 2001:DB8::11/128 with the attached BGP-Prefix-SID 590 advertising the support of the SRH for IPv6 packets destined to 591 segment 2001:DB8::11. 593 The control-plane and data-plane processing of all the other nodes in 594 the fabric is unchanged. Specifically, the routes to 2001:DB8::5 and 595 2001:DB8::11 are installed in the FIB along the eBGP best-path to 596 Node5 (spine node) and Node11 (ToR node) respectively. 598 An application on HostA which needs to send traffic to HostZ via only 599 Node5 (spine node) can do so by sending IPv6 packets with a Segment 600 Routing header (SRH, [I-D.ietf-6man-segment-routing-header]). The 601 destination address and active segment is set to 2001:DB8::5. The 602 next and last segment is set to 2001:DB8::11. 604 The application must only use IPv6 addresses that have been 605 advertised as capable for SRv6 segment processing (e.g. for which the 606 BGP prefix segment capability has been advertised). How applications 607 learn this (e.g.: centralized controller and orchestration) is 608 outside the scope of this document. 610 6. Communicating path information to the host 612 There are two general methods for communicating path information to 613 the end-hosts: "proactive" and "reactive", aka "push" and "pull" 614 models. There are multiple ways to implement either of these 615 methods. Here, it is noted that one way could be using a centralized 616 controller: the controller either tells the hosts of the prefix-to- 617 path mappings beforehand and updates them as needed (network event 618 driven push), or responds to the hosts making request for a path to 619 specific destination (host event driven pull). It is also possible 620 to use a hybrid model, i.e., pushing some state from the controller 621 in response to particular network events, while the host pulls other 622 state on demand. 624 It is also noted, that when disseminating network-related data to the 625 end-hosts a trade-off is made to balance the amount of information 626 Vs. the level of visibility in the network state. This applies both 627 to push and pull models. In the extreme case, the host would request 628 path information on every flow, and keep no local state at all. On 629 the other end of the spectrum, information for every prefix in the 630 network along with available paths could be pushed and continuously 631 updated on all hosts. 633 7. Addressing the open problems 635 This section demonstrates how the problems described above (in 636 section 3) could be solved using the segment routing concept. It is 637 worth noting that segment routing signaling and data-plane are only 638 parts of the solution. Additional enhancements, e.g., such as the 639 centralized controller mentioned previously, and host networking 640 stack support are required to implement the proposed solutions. Also 641 the applicability of the solutions described below are not restricted 642 to the data-center alone, the same could be re-used in context of 643 other domains as well 645 7.1. Per-packet and flowlet switching 647 A flowlet is defined as a burst of packets from the same flow 648 followed by an idle interval. 650 With some ability to choose paths on the host, one may go from per- 651 flow load-sharing in the network to per-packet or per-flowlet. The 652 host may select different segment routing instructions either per 653 packet, or per flowlet, and route them over different paths. This 654 allows for solving the "elephant flow" problem in the data-center and 655 avoiding link imbalances. 657 Note that traditional ECMP routing could be easily simulated with on- 658 host path selection, using method proposed in [GREENBERG09]. The 659 hosts would randomly pick a Tier-2 or Tier-1 device to "bounce" the 660 packet off of, depending on whether the destination is under the same 661 Tier-2 nodes, or has to be reached across Tier-1. The host would use 662 a hash function that operates on per-flow invariants, to simulate 663 per-flow load-sharing in the network. 665 Using Figure 1 as reference, let us illustrate this concept assuming 666 that HostA has an elephant flow to HostZ called Flow-f. 668 Normally, a flow is hashed on to a single path. Let's assume HostA 669 sends its packets associated with Flow-f with top label 16011 (the 670 label for the remote ToR, Node11, where HostZ is connected) and Node1 671 would hash all the packets of Flow-F via the same next-hop (e.g. 672 Node3). Similarly, let's assume that leaf Node3 would hash all the 673 packets of Flow-F via the same next-hop (e.g.: spine node Node5). 674 This normal operation would restrict the elephant flow on a small 675 subset of the ECMP paths to HostZ and potentially create imbalance 676 and congestion in the fabric. 678 Leveraging the flowlet proposal, assuming HostA is made aware of 4 679 disjoint paths via intermediate segment 16005, 16006, 16007 and 16008 680 (the BGP prefix SID's of the 4 spine nodes) and also made aware of 681 the prefix segment of the remote ToR connected to the destination 682 (16011), then the application can break the elephant flow F into 683 flowlets F1, F2, F3, F4 and associate each flowlet with one of the 684 following 4 label stacks: {16005, 16011}, {16006, 16011}, {16007, 685 16011} and {16008, 16011}. This would spread the load of the elephant 686 flow through all the ECMP paths available in the fabric and re- 687 balance the load. 689 7.2. Performance-aware routing 691 Knowing the path associated with flows/packets, the end host may 692 deduce certain characteristics of the path on its own, and 693 additionally use the information supplied with path information 694 pushed from the controller or received via pull request. The host 695 may further share its path observations with the centralized agent, 696 so that the latter may keep up-to-date network health map to assist 697 other hosts with this information. 699 For example, an application A.1 at HostA may pin a flow destined to 700 HostZ via Spine node Node5 using label stack {16005, 16011}. The 701 application A.1 may collect information on packet loss or other 702 metrics. A.1 may additionally publish this information to a 703 centralized agent, e.g. after a flow completes, or periodically for 704 longer lived flows. Next, using both local and/or global performance 705 data, application A.1 as well as other applications sharing the same 706 resources in the DC fabric may pick up the best path for the new 707 flow, or update an existing path (e.g.: when informed of congestion 708 on an existing path). The mechanisms for collecting the flow 709 metrics, their publishing to a centralized agent and the decision 710 process at the centralized agent and the application/host to pick a 711 path through the network based on this collected information is 712 outside the scope of this document. 714 One particularly interesting instance of performance-aware routing is 715 dynamic fault-avoidance. If some links or devices in the network 716 start discarding packets due to a fault, the end-hosts could probe 717 and detect the path(s) that are affected and hence steer the affected 718 flows away from the problem spot. Similar logic applies to failure 719 cases where packets get completely black-holed, e.g., when a link 720 goes down and the failure is detected by the host while probing the 721 path. 723 For example, an application A.1 informed about 5 paths to Z {16005, 724 16011}, {16006, 16011}, {16007, 16011}, {16008, 16011} and {16011} 725 might use the last one by default (for simplicity). When performance 726 is degrading, A.1 might then start to pin flows to each of the 4 727 other paths (each via a distinct spine) and monitor the performance. 728 It would then detect the faulty path and assign a negative preference 729 to the faulty path to avoid further flows using it. Gradually, over 730 time, it may re-assign flows on the faulty path to eventually detect 731 the resolution of the trouble and start reusing the path. The 732 mechanisms for monitoring performance for a specific flow and for the 733 various paths and the deduction of optimal paths to improve the same 734 for the flow are outside the scope of this document. 736 By leveraging Segment Routing, one avoids issues associated with 737 oblivious ECMP hashing. For example, if in the topology depicted on 738 Figure 1 a link between spine node Node5 and leaf node Node9 fails, 739 HostA may exclude the segment corresponding to Node5 from the prefix 740 matching the servers under Tier-2 devices Node9. In the push path 741 discovery model, the affected path mappings may be explicitly pushed 742 to all the servers for the duration of the failure. The new mapping 743 would instruct them to avoid the particular Tier-1 node until the 744 link has recovered. Alternatively, in pull path, the centralized 745 controller may start steering new flows immediately after it 746 discovers the issue. Until then, the existing flows may recover 747 using local detection of the path issues. 749 7.3. Deterministic network probing 751 Active probing is a well-known technique for monitoring network 752 elements' health, constituting of sending continuous packet streams 753 simulating network traffic to the hosts in the data-center. Segment 754 routing makes possible to prescribe the exact paths that each probe 755 or series of probes would be taking toward their destination. This 756 allows for fast correlation and detection of failed paths, by 757 processing information from multiple actively probing agents. This 758 complements the data collected from the hosts routing stacks as 759 described in Section 7.2. 761 For example, imagine a probe agent sending packets to all machines in 762 the data-center. For every host, it may send packets over each of 763 the possible paths, knowing exactly which links and devices these 764 packets will be crossing. Correlating results for multiple 765 destinations with the topological data, it may automatically isolate 766 possible problem to a link or device in the network. 768 8. Additional Benefits 770 8.1. MPLS Dataplane with operational simplicity 772 As required by [RFC7938], no new signaling protocol is introduced. 773 The BGP-Prefix-SID is a lightweight extension to BGP Labeled Unicast 774 [RFC8277]. It applies either to eBGP or iBGP based designs. 776 Specifically, LDP and RSVP-TE are not used. These protocols would 777 drastically impact the operational complexity of the Data Center and 778 would not scale. This is in line with the requirements expressed in 779 [RFC7938]. 781 Provided the same SRGB is configured on all nodes, all nodes use the 782 same MPLS label for a given IP prefix. This is simpler from an 783 operation standpoint, as discussed in Section 9 785 8.2. Minimizing the FIB table 787 The designer may decide to switch all the traffic at Tier-1 and Tier- 788 2's based on MPLS, hence drastically decreasing the IP table size at 789 these nodes. 791 This is easily accomplished by encapsulating the traffic either 792 directly at the host or the source ToR node by pushing the BGP- 793 Prefix-SID of the destination ToR for intra-DC traffic, or the BGP- 794 Prefix-SID for the the border node for inter-DC or DC-to-outside- 795 world traffic. 797 8.3. Egress Peer Engineering 799 It is straightforward to combine the design illustrated in this 800 document with the Egress Peer Engineering (EPE) use-case described in 801 [I-D.ietf-spring-segment-routing-central-epe]. 803 In such case, the operator is able to engineer its outbound traffic 804 on a per host-flow basis, without incurring any additional state at 805 intermediate points in the DC fabric. 807 For example, the controller only needs to inject a per-flow state on 808 the HostA to force it to send its traffic destined to a specific 809 Internet destination D via a selected border node (say Node12 in 810 Figure 1 instead of another border node, Node11) and a specific 811 egress peer of Node12 (say peer AS 9999 of local PeerNode segment 812 9999 at Node12 instead of any other peer which provides a path to the 813 destination D). Any packet matching this state at host A would be 814 encapsulated with SR segment list (label stack) {16012, 9999}. 16012 815 would steer the flow through the DC fabric, leveraging any ECMP, 816 along the best path to border node Node12. Once the flow gets to 817 border node Node12, the active segment is 9999 (because of PHP on the 818 upstream neighbor of Node12). This EPE PeerNode segment forces 819 border node Node12 to forward the packet to peer AS 9999, without any 820 IP lookup at the border node. There is no per-flow state for this 821 engineered flow in the DC fabric. A benefit of segment routing is 822 the per-flow state is only required at the source. 824 As well as allowing full traffic engineering control such a design 825 also offers FIB table minimization benefits as the Internet-scale FIB 826 at border node Node12 is not required if all FIB lookups are avoided 827 there by using EPE. 829 8.4. Anycast 831 The design presented in this document preserves the availability and 832 load-balancing properties of the base design presented in 833 [I-D.ietf-spring-segment-routing]. 835 For example, one could assign an anycast loopback 192.0.2.20/32 and 836 associate segment index 20 to it on the border Node11 and Node12 (in 837 addition to their node-specific loopbacks). Doing so, the EPE 838 controller could express a default "go-to-the-Internet via any border 839 node" policy as segment list {16020}. Indeed, from any host in the DC 840 fabric or from any ToR node, 16020 steers the packet towards the 841 border Node11 or Node12 leveraging ECMP where available along the 842 best paths to these nodes. 844 9. Preferred SRGB Allocation 846 In the MPLS case, it is recommend to use same SRGBs at each node. 848 Different SRGBs in each node likely increase the complexity of the 849 solution both from an operational viewpoint and from a controller 850 viewpoint. 852 From an operation viewpoint, it is much simpler to have the same 853 global label at every node for the same destination (the MPLS 854 troubleshooting is then similar to the IPv6 troubleshooting where 855 this global property is a given). 857 From a controller viewpoint, this allows us to construct simple 858 policies applicable across the fabric. 860 Let us consider two applications A and B respectively connected to 861 Node1 and Node2 (ToR nodes). A has two flows FA1 and FA2 destined to 862 Z. B has two flows FB1 and FB2 destined to Z. The controller wants 863 FA1 and FB1 to be load-shared across the fabric while FA2 and FB2 864 must be respectively steered via Node5 and Node8. 866 Assuming a consistent unique SRGB across the fabric as described in 867 the document, the controller can simply do it by instructing A and B 868 to use {16011} respectively for FA1 and FB1 and by instructing A and 869 B to use {16005 16011} and {16008 16011} respectively for FA2 and 870 FB2. 872 Let us assume a design where the SRGB is different at every node and 873 where the SRGB of each node is advertised using the Originator SRGB 874 TLV of the BGP-Prefix-SID as defined in 875 [I-D.ietf-idr-bgp-prefix-sid]: SRGB of Node K starts at value K*1000 876 and the SRGB length is 1000 (e.g. Node1's SRGB is [1000, 1999], 877 Node2's SRGB is [2000, 2999], ...). 879 In this case, not only the controller would need to collect and store 880 all of these different SRGB's (e.g., through the Originator SRGB TLV 881 of the BGP-Prefix-SID), furthermore it would need to adapt the policy 882 for each host. Indeed, the controller would instruct A to use {1011} 883 for FA1 while it would have to instruct B to use {2011} for FB1 884 (while with the same SRGB, both policies are the same {16011}). 886 Even worse, the controller would instruct A to use {1005, 5011} for 887 FA1 while it would instruct B to use {2011, 8011} for FB1 (while with 888 the same SRGB, the second segment is the same across both policies: 889 16011). When combining segments to create a policy, one need to 890 carefully update the label of each segment. This is obviously more 891 error-prone, more complex and more difficult to troubleshoot. 893 10. IANA Considerations 895 This document does not make any IANA request. 897 11. Manageability Considerations 899 The design and deployment guidelines described in this document are 900 based on the network design described in [RFC7938]. 902 The deployment model assumed in this document is based on a single 903 domain where the interconnected DCs are part of the same 904 administrative domain (which, of course, is split into different 905 autonomous systems). The operator has full control of the whole 906 domain and the usual operational and management mechanisms and 907 procedures are used in order to prevent any information related to 908 internal prefixes and topology to be leaked outside the domain. 910 As recommended in [I-D.ietf-spring-segment-routing], the same SRGB 911 should be allocated in all nodes in order to facilitate the design, 912 deployment and operations of the domain. 914 When EPE ([I-D.ietf-spring-segment-routing-central-epe]) is used (as 915 explained in Section 8.3, the same operational model is assumed. EPE 916 information is originated and propagated throughout the domain 917 towards an internal server and unless explicitly configured by the 918 operator, no EPE information is leaked outside the domain boundaries. 920 12. Security Considerations 922 This document proposes to apply Segment Routing to a well known 923 scalability requirement expressed in [RFC7938] using the BGP-Prefix- 924 SID as defined in [I-D.ietf-idr-bgp-prefix-sid]. 926 It has to be noted, as described in Section 11 that the design 927 illustrated in [RFC7938] and in this document, refer to a deployment 928 model where all nodes are under the same administration. In this 929 context, it is assumed that the operator doesn't want to leak outside 930 of the domain any information related to internal prefixes and 931 topology. The internal information includes prefix-sid and EPE 932 information. In order to prevent such leaking, the standard BGP 933 mechanisms (filters) are applied on the boundary of the domain. 935 Therefore, the solution proposed in this document does not introduce 936 any additional security concerns from what expressed in [RFC7938] and 937 [I-D.ietf-idr-bgp-prefix-sid]. It is assumed that the security and 938 confidentiality of the prefix and topology information is preserved 939 by outbound filters at each peering point of the domain as described 940 in Section 11. 942 13. Acknowledgements 944 The authors would like to thank Benjamin Black, Arjun Sreekantiah, 945 Keyur Patel, Acee Lindem and Anoop Ghanwani for their comments and 946 review of this document. 948 14. Contributors 949 Gaya Nagarajan 950 Facebook 951 US 953 Email: gaya@fb.com 955 Gaurav Dawra 956 Cisco Systems 957 US 959 Email: gdawra.ietf@gmail.com 961 Dmitry Afanasiev 962 Yandex 963 RU 965 Email: fl0w@yandex-team.ru 967 Tim Laberge 968 Cisco 969 US 971 Email: tlaberge@cisco.com 973 Edet Nkposong 974 Salesforce.com Inc. 975 US 977 Email: enkposong@salesforce.com 979 Mohan Nanduri 980 Microsoft 981 US 983 Email: mnanduri@microsoft.com 985 James Uttaro 986 ATT 987 US 989 Email: ju1738@att.com 991 Saikat Ray 992 Unaffiliated 993 US 995 Email: raysaikat@gmail.com 996 Jon Mitchell 997 Unaffiliated 998 US 1000 Email: jrmitche@puck.nether.net 1002 15. References 1004 15.1. Normative References 1006 [I-D.ietf-idr-bgp-prefix-sid] 1007 Previdi, S., Filsfils, C., Lindem, A., Sreekantiah, A., 1008 and H. Gredler, "Segment Routing Prefix SID extensions for 1009 BGP", draft-ietf-idr-bgp-prefix-sid-21 (work in progress), 1010 May 2018. 1012 [I-D.ietf-spring-segment-routing] 1013 Filsfils, C., Previdi, S., Ginsberg, L., Decraene, B., 1014 Litkowski, S., and R. Shakir, "Segment Routing 1015 Architecture", draft-ietf-spring-segment-routing-15 (work 1016 in progress), January 2018. 1018 [I-D.ietf-spring-segment-routing-central-epe] 1019 Filsfils, C., Previdi, S., Dawra, G., Aries, E., and D. 1020 Afanasiev, "Segment Routing Centralized BGP Egress Peer 1021 Engineering", draft-ietf-spring-segment-routing-central- 1022 epe-10 (work in progress), December 2017. 1024 [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate 1025 Requirement Levels", BCP 14, RFC 2119, 1026 DOI 10.17487/RFC2119, March 1997, 1027 . 1029 [RFC4271] Rekhter, Y., Ed., Li, T., Ed., and S. Hares, Ed., "A 1030 Border Gateway Protocol 4 (BGP-4)", RFC 4271, 1031 DOI 10.17487/RFC4271, January 2006, 1032 . 1034 [RFC7938] Lapukhov, P., Premji, A., and J. Mitchell, Ed., "Use of 1035 BGP for Routing in Large-Scale Data Centers", RFC 7938, 1036 DOI 10.17487/RFC7938, August 2016, 1037 . 1039 [RFC8277] Rosen, E., "Using BGP to Bind MPLS Labels to Address 1040 Prefixes", RFC 8277, DOI 10.17487/RFC8277, October 2017, 1041 . 1043 15.2. Informative References 1045 [GREENBERG09] 1046 Greenberg, A., Hamilton, J., Jain, N., Kadula, S., Kim, 1047 C., Lahiri, P., Maltz, D., Patel, P., and S. Sengupta, 1048 "VL2: A Scalable and Flexible Data Center Network", 2009. 1050 [I-D.ietf-6man-segment-routing-header] 1051 Previdi, S., Filsfils, C., Leddy, J., Matsushima, S., and 1052 d. daniel.voyer@bell.ca, "IPv6 Segment Routing Header 1053 (SRH)", draft-ietf-6man-segment-routing-header-13 (work in 1054 progress), May 2018. 1056 [RFC6793] Vohra, Q. and E. Chen, "BGP Support for Four-Octet 1057 Autonomous System (AS) Number Space", RFC 6793, 1058 DOI 10.17487/RFC6793, December 2012, 1059 . 1061 Authors' Addresses 1063 Clarence Filsfils (editor) 1064 Cisco Systems, Inc. 1065 Brussels 1066 BE 1068 Email: cfilsfil@cisco.com 1070 Stefano Previdi 1071 Cisco Systems, Inc. 1072 Italy 1074 Email: stefano@previdi.net 1076 Gaurav Dawra 1077 LinkedIn 1078 USA 1080 Email: gdawra.ietf@gmail.com 1081 Ebben Aries 1082 Juniper Networks 1083 1133 Innovation Way 1084 Sunnyvale CA 94089 1085 US 1087 Email: exa@juniper.net 1089 Petr Lapukhov 1090 Facebook 1091 US 1093 Email: petr@fb.com