Network Working Group C. Filsfils, Ed.
Internet-Draft D. Cai, Ed.
Intended status: Informational S. Previdi
Expires: October 3, 2016 Cisco
W. Henderickx
Alcatel-Lucent
R. Shakir
BT
D. Cooper
F. Ferguson
Level3
S. Lin
Microsoft
T. LaBerge
Cisco
B. Decraene
Orange
L. Jalil
Verizon
J. Tantsura
Ericsson
April 3, 2016
Interconnecting Millions Of Endpoints With Segment Routing
draft-filsfils-spring-large-scale-interconnect-02
Abstract
This document describes an application of Segment Routing to scale
the network to support hundreds of thousands of network nodes, and
tens of millions of physical underlay endpoints. This use-case can be
applied to the interconnection of massive-scale DC's and/or large
aggregation networks. Forwarding tables of midpoint and leaf nodes
only require a few tens of thousands of entries.
Requirements Language
Filsfils, et al. Expires October 3, 2016 [Page 1]
Internet-Draft Segment Routing April 2016
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
"SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this
document are to be interpreted as described in RFC 2119 [RFC2119].
Status of This Memo
This Internet-Draft is submitted in full conformance with the
provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering
Task Force (IETF). Note that other groups may also distribute
working documents as Internet-Drafts. The list of current Internet-
Drafts is at http://datatracker.ietf.org/drafts/current/.
Internet-Drafts are draft documents valid for a maximum of six months
and may be updated, replaced, or obsoleted by other documents at any
time. It is inappropriate to use Internet-Drafts as reference
material or to cite them other than as "work in progress."
This Internet-Draft will expire on May 1, 2016.
Copyright Notice
Copyright (c) 2014 IETF Trust and the persons identified as the
document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal
Provisions Relating to IETF Documents
(http://trustee.ietf.org/license-info) in effect on the date of
publication of this document. Please review these documents
carefully, as they describe your rights and restrictions with respect
to this document. Code Components extracted from this document must
include Simplified BSD License text as described in Section 4.e of
the Trust Legal Provisions and are provided without warranty as
described in the Simplified BSD License.
Table of Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 4
2. Reference Design . . . . . . . . . . . . . . . . . . . . . . . 4
3. Control Plane . . . . . . . . . . . . . . . . . . . . . . . . . 6
4. Illustration of the scale . . . . . . . . . .. . . . . . . . . 7
5. Optional Designs . . . . . . . . . . . . . . . . . . . . . . . 7
6. Deployment Model . . . . . . . . . . . . . . . . . . . . . . . .9
7. Benefits . . . . . . . . . . . . . . . . . . . . . . . . . . ..9
8. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 10
Filsfils, et al. Expires October 3, 2016 [Page 2]
Internet-Draft Segment Routing April 2016
9. Manageability Considerations . . . . . . . . . . . . . . . . . 10
10. Security Considerations . . . . . . . . . . . . . . . . . . . 10
11. Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . 10
12. References . . . . . . . . . . . . . . . . . . . . . . . . . .10
Appendix - Scale Example . . . . . . . . . . . . . . . . . . . . .10
Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . ..12
Filsfils, et al. Expires October 3, 2016 [Page 3]
Internet-Draft Segment Routing April 2016
1 Introduction
This document describes how SR can be used to interconnect 100s
thousands of nodes and 10's of millions of applications/humans. This
version of the document focuses on the MPLS/SR instantiation. No new
protocol extensions are required.
1.1. Terminology
Term Definition
----------- ------------------------------------------------
BGP Border Gateway Protocol
BGP-LS Border Gateway Protocol - Link State
DC Data Center
ECMP Equal Cost MultiPathing
FIB Forwarding Information Base
LDP Label Distribution Protocol
MPLS Multi-Protocol Label Switching
PCE Path Computation Element
PCEP Path Computation Element Protocol
PW Pseudowire
SR Segment Routing
SRTE Segment Routing Traffic Engineering
SRTE Policy Segment Routing Traffic Engineering Policy
TI-LFA Topology Independent Loop Free Alternate
2 Reference Design
+---------+ +---------+ +---------+
A X1a X2a C
| L1 | | C | | L2 |
B X1b X2b D
+---------+ +---------+ +---------+
A : PrefixSID 18001 is unique in L1
B : PrefixSID 18002 is unique in L1
X1a: Anycast PrefixSID 16001 is unique across all the domains
PrefixSID 16003 is unique across all the domains
X1b: Anycast PrefixSID 16001 is unique across all the domains
PrefixSID 16004 is unique across all the domains
X2a: Anycast PrefixSID 16002 is unique across all the domains
PrefixSID 16005 is unique across all the domains
X2b: Anycast PrefixSID 16002 is unique across all the domains
PrefixSID 16006 is unique across all the domains
C : PrefixSID 18001 is unique in L2
Filsfils, et al. Expires October 3, 2016 [Page 4]
Internet-Draft Segment Routing April 2016
D : PrefixSID 18002 is unique in L2
We structure the network into leaf domains (L1, L2...) interconnected
by a central core domain C. Each domain runs SR with its own
independent routing protocol (e.g.: IS-IS, OSPF, BGP).
A common SRGB of [16000-23999] is assumed (any other common block
choice is possible) across all of the domains. We further assume that
[16000-17999] is solely used to provide prefix segments in the C
domain (any other choice is possible) while [18000, 23999] is reused
to provide prefix segments in any leaf domain.
For example, we see that A and C of the leaf domain L1 and L2
respectively, receive the prefix segment 18001 while prefix segment
16003 is allocated to node X1a in the C domain and is unique across
the entire set of domains.
Each leaf domain Lk connects to the domain C with 2 or more nodes
called Xka and Xkb. Each X node runs two independent SR routing
protocols: one in the leaf domain and one in the core domain. Each X
nodes is provided with two prefix segments allocated from the domain
C: one uniquely identifies the node while the other (anycast prefix
segment) identifies the pair number k of X nodes interconnecting the
leaf domain k to the core domain.
In our reference diagram, X1a has prefix segment 16003 and anycast
prefix segment 16001 while X1b has prefix segment 16004 and anycast
prefix segment 16001.
No route is redistributed from a leaf domain to the core domain. All
the routes (and their prefix SID's) of the X nodes are redistributed
from the core domain into the leaf domains. No other route is
redistributed from the core into the leaf domains. The FIB of an
interior node within the C domain does not hold any entry for
segments in the range [18000, 23999]. A node in a leaf domain only
has FIB entries for all the segments in the local leaf domain and
prefix segments towards all the X nodes in the network. For example,
A of leaf L1 has a FIB entry for anycast segment 16002 which leads to
the pair X2a and X2b and prefix segment 16005 which leads to X2a.
2.1 Examples
We use the notation A.L1 to represent the node A of leaf domain L1.
Leveraging the above design, any leaf node can be interconnected with
any other leaf node.
Intra-leaf, shortest-path: A.L1 uses the following SID list to reach
Filsfils, et al. Expires October 3, 2016 [Page 5]
Internet-Draft Segment Routing April 2016
B.L1: {18002}
Inter-leaf, shortest-path through any X: A.L1 uses the following SID
list to reach D.L2 via any intermediate X: {16002, 18002}
Inter-leaf, shortest-path through a specific X: A.L1 uses the
following SID list to reach D.L2 via X2a: {16005, 18002}
It is out of the scope of this document to describe how the SID lists
are computed and programmed at the source nodes. As an example, a
centralized controller could be the source of the Prefix SID
allocation. The controller could continuously collect the state of
each domain (e.g. BGP-LS). Upon any new service request (e.g.: from V
to W), it could check whether W is in the same leaf domain of V. If
so, a single SID would be required (dynamically learned via IGP-SR
(IS-IS-SR, OSPF-SR) within the domain and would not be added by the
controller). Otherwise, if V and W resides on separate domains, the
SID of the X gateway to W's leaf domain would be inserted before W's
SID by the controller.
3. Control-plane
This section provides a high-level description of one example of an
implemented control-plane. The example is for L2VPN PW service with
certain SLA contract.
The service orchestration programs A with a PW to a remote next-hop C
with a given SLA contract (low-latency path, be disjoint from a
specific core plane, be disjoint from a different PW service, etc.).
A automatically detects that it does not have reachability to C. It
then automatically sends a PCEP request to an SR PCE for an SRTE
policy that provides reachability to C with the requested SLA.
The SR PCE is made of two components. A multi-domain topology and a
compute block. The multi-domain topology is continuously refreshed
from BGP-LS feeds from each domain. The compute block implements TE
algorithms designed specifically for SR path expression. Upon
receiving the PCEP request, the SR PCE computes the solution (e.g.
{16002, 18001} and provides it to A.
The SR PCE logs the request as a stateful query and hence recomputes
another solution upon any multi-domain topology changes that
invalidates the previous solution.
A receives the PCEP reply with the solution. A installs the received
SRTE policy in the dataplane. A automatically steers the PW on that
SRTE policy.
Filsfils, et al. Expires October 3, 2016 [Page 6]
Internet-Draft Segment Routing April 2016
It is out of the scope of this document to describe how the SRTE
Policies are computed and programmed at the source nodes.
4. Illustration of the scale
A review of a practical design use case can be used to determine the
true scalability of this methodology in the real world. Such an
example will demonstrate that over 50 million physical end points can
be supported in as little as 12K of FIB space. Moreover, this example
does not even suggest a maximum scale. Please refer to [Appendix] for
details.
5 Optional Designs
Section 2 describes the reference model of the design. However, there
could be multiple different design options depends on the network
scale, network node hardware capability, SLA requirement, etc.
5.1 Leaf and Core Domains Sizing
The operator might choose to not redistribute the X routes into the
leaf domains. In that case, one more segment is required in order to
compose an end-to-end path. For example, to express an "inter-leaf,
shortest-path through any X" path from A.L1 to D.L2, A.L1 uses
{16001, 16002, 18002} instead of {16002, 18002}. This model gives the
operator the ability to choose among a small number of larger leaf
domains, a large number of small leaf domains or a mix of small and
large domains.
5.2 Local Segments to Hosts/Servers
Local segments can be programmed at any leaf node in order to
identify locally-attached hosts (or VM's). For example, if D.L2 has
bound a local segment 40001 to a local host DH1, then A uses the
following SID list to reach that host: {16002, 18002, 40001}
(assuming the reference design above). Such local segment could
represent the NID (Network Interface Device) device in the context of
the SP access network, or VM in the context of the DC network.
5.3 Sub-leaf Domains
A third level of hierarchy called "Sub-Leaf" can be introduced for
further scale.
+---------+ +---------+ +---------+ +-------+
A X1a X2a Y21a E
| L1 | | C | | L2 | | SL21 |
B X1b X2b Y21b F
+---------+ +---------+ +---------+ +-------+
Filsfils, et al. Expires October 3, 2016 [Page 7]
Internet-Draft Segment Routing April 2016
In the above diagram, a sub-leaf "SL21" has been added to the leaf
domain L2. SL21 is connected to L2 via two (or more) Y nodes. The
SRGB sub-space [18000, 23999] initially allocated for the leaf is
splitted into two sub-spaces: [18000-19999] for the leaf allocation
and [20000-23999] for the sub-leaf allocation. Each Y node is
allocated with a unique anycast prefix segment and a unique prefix
segmemt within the leaf block. For example, Y21a receives anycast SID
19021 and prefix SID 19211. Each node within a subleaf domain
receives a unique prefix SID from that domain (e.g. E receives
20001).
For example, to express an "inter-leaf, shortest-path, through any X,
through any Y" path to E.L2.SL21, A.L1 uses {16002, 19021, 20001}.
Alternatively, the operator may decide not to distribute any X route
down into leaf domains, but instead, distribute Y gateways up to the
C domain. In this case, A.L1 would express the "inter-leaf, shortest-
path, through any X, through any Y" path to E.L2.SL21 with SID
list:{16001, 19021, 20001}.
5.4 Traffic Engineering
Traffic Engineering: Any leaf or core domain can use SR in order to
traffic engineer its traffic locally within the domain.
For example, a flow from A.L1 to X1a within L1 domain could be
steered via B using the SR policy {18002, 16003}. Similarly a flow
from X1a to X2a within the core domain could be steered via X2b with
the SR policy {16006, 16005}.
Similarly, a flow can be engineered across domains. For example, a
flow from A.L1 to C.L2 could be steered via B then X1a then X2b then
X2a then C using the SR policy {18002, 16003, 16006, 16005, 18001}.
The SR policy at the source can be "compressed" (in terms of number
of segments) by leveraging binding segments bound to SR policy. For
example, assuming that the local binding segment 30000 is bound by A
to the policy {18002, 16003} and that the local binding segment 30001
is bound by X1a to the policy {16006, 16005}, then the previous
inter-domain policy can also be expressed at A (or any node connected
to A) as {30000, 30001, 18001}. Using a binding segment to refer to a
remote SR policy provides other benefits such as decreasing the need
for the centralized controller in order to reflect a change from one
domain to another.
For example, let us assume that something changes within the core
domain such that the path followed by the policy 30001 at X1a
changes. The SR policy associated with 30001 is updated at X1a
without any change at A. The binding segment 30001 remains "stable"
from the viewpoint of L1 leaf domain. Updating a remote domain
becomes necessary only when the headend of the binding segment
Filsfils, et al. Expires October 3, 2016 [Page 8]
Internet-Draft Segment Routing April 2016
becomes unavailable (X1a becomes unavailable) or when the policy
attached to the binding segment is no longer achievable. An example
could be: upon a double and independent failure, a policy avoiding
some resources (e.g. another plane of the backbone) might no longer
be possible). In only these cases, the policy at A needs to be
changed. It is out of the scope of this document to describe how the
SID lists are computed in order to realize a specific traffic-
engineering objective, with or without the use of binding SID. For
example, an application could request a specific treatment via a
north-bound API to a centralized controller. The centralized
controller might collect the topology of all the domains. It might
also translate the application requirement into an end-to-end path
through the domains. Finally, it might then translate that end-to-end
path in a list of segments. It might create intermediate per-domain
policies (e.g. using PCEP provisioning) and learn their associated
binding segments (e.g. PCEP or BGP-LS) and return to the application
the resulting SID list where some of the SID's are binding segments.
6 Deployment Model
It is expected that this design be deployed as a greenfield but as
well in interworking (brownfield) with seamless-mpls design (draft-
ietf-mpls-seamless-mpls).
7 Benefit
There are some notable benefits that this proposal can bring to the
large scale network deployment:
ECMP: each policy (intra or inter-domain, with or without TE) is
expressed as a list of segments. As each segment is optimized for
ECMP, therefore the entire policy is optimized for ECMP. The ECMP
gain of anycast prefix segment should also be considered (e.g. 16001
load-shares across any gateway from L1 leaf domain to Core and 16002
load-shares across any gateway from Core to L2 leaf domain.
Sub-50msec FRR: Topology-Independent FRR using SR [draft-francois-
spring-segment-routing-ti-lfa-01] ensures sub-50msec upon any link or
node failure, in any topology.
Simple and better node redundancy: furthermore the use of anycast
segment provides for an additional high-availability mechanism (e.g.:
flows directed to 16001 can either go via X1a or X1b).
No new protocol extensions are required to support this.
Filsfils, et al. Expires October 3, 2016 [Page 9]
Internet-Draft Segment Routing April 2016
8. IANA Considerations
None
9. Manageability Considerations
TBD
10. Security Considerations
TBD
11. Acknowledgements
We would like to thank Giles Heron, Alexander Preusche and Steve
Braaten for their contribution to the content of this document.
12. References
12.1. Normative References
[RFC2119] Bradner, S., "Key words for use in RFCs to Indicate
Requirement Levels", BCP 14, RFC 2119, March 1997.
12.2. Informative References
[draft-ietf-mpls-seamless-mpls] Leymann, et al., "Seamless MPLS
Architecture", draft-ietf-mpls-seamless-mpls-07, (work in
progress), July 2015
[draft-francois-spring-segment-routing-ti-lfa-01] Pierre Francois, et
al., "Topology Independent Fast Reroute using Segment
Routing", draft-francois-spring-segment-routing-ti-lfa-01,
(work in progress), April 2015
Appendix - Scale Example
A review of a practical design use case can be used to determine the
true scalability of this methodology in the real world. This example
does not suggest a maximum scale, but it provides a simple framework
that can be deployed on emerging silicon products from a variety of
sources. In doing so, it also demonstrates huge scale potential on
the order of 50 million end-points.
The design
The practical design is based on the reference topology above, but
Filsfils, et al. Expires October 3, 2016 [Page 10]
Internet-Draft Segment Routing April 2016
differs in that each data center layer is enumerated for addressing
and the SRGB is necessarily expanded beyond the default (more on that
follows). Otherwise, the topology is the same. For the sake of
brevity, the design is described only at a high level here.
For terminology's sake, each leaf domain in the original topology is
shown below as a data center (DC). Further, the X nodes above are
referred to here as data center routers (DCR's). For example:
+---------+ +---------+ +---------+
| DCR DCR |
| DC1 | | Core | | DC2 |
| DCR DCR |
+---------+ +---------+ +---------+
The design leverages well-worn physical deployment practices (such as
Clos-based fabric principles and relatively low oversuscription) in
conjunction with emerging silicon options supporting 25G/50G/100G
interfaces.
Each data center is composed of three layers of switching: Top-of-
Rack (ToR), leaf, and spine. Sixteen ToR's and eight leaf devices
combine to create clusters supporting 1,536 physical servers each
(connected at 25G). All switching devices in the cluster are based
on 32x100G switches (leveraging 25G breakout where required).
The spine layer interconnects 272 clusters using a total of 128
288x100G modular switches. To leave each data center, 16 DCR devices
(using 256 ports of 100G) provide connectivity to 16 core devices
(also using 256 ports of 100G). From a server's perspective, north-
south oversubscription in this model is 51:1 and east-west
oversubscription is 9:1.
All told, the data center (including its portion of the core) is
composed of 6,688 networking devices:
4,352 ToR's (16 ToR x 272 clusters) 2,176 Leaf nodes (8 leaf
nodes x 272 clusters) 128 spine switches 16 DCR's 16 Core
routers
For the sake of Segment Routing scale calculations, each of these
nodes will advertise its node SID. The Leaf domain FIB size is
6,672, consisting of the spine, leaf, and ToR switching layers as
well as the DCR layer.
The core domain consists of the DCR's and the Core routers (32
Filsfils, et al. Expires October 3, 2016 [Page 11]
Internet-Draft Segment Routing April 2016
nodes). Finally, each data center site will advertise one anycast
SID into the core domain, so the per-site SID count in the core
domain is 33.
Scalability
Each data center site has the following scale characteristics:
417,792 physical server ports (at 25G per port) 6,672 SID
entries in the Leaf domain 33 SID entries in the Core domain
Each device in the Leaf domain must be aware of all other devices in
the data center, as well as all SID's advertised on the core.
Therefore, the FIB in these devices must support:
7K segments for local connectivity (from the SRGB) <300 adjacency
segments for large switches (locally installed) Up to 104 adjacency
segments on small switches (locally installed) 4K Core segments
(from the SRGB, to be covered shortly)
This represents a total FIB requirement of about 12K (rounded up).
Most of this addressing involves global addressing. This means that
the SRGB must be expanded beyond its 8K default configuration (values
16000 - 23999). For this design, the SRGB is expanded to 12K (16000
- 27999) and apportioned in the following way:
16000 - 19999 (4K) Core and inter-site connectivity (truly global)
20000 - 27999 (8K) Leaf domain (reused per DC site)
The core domain size of 4K is somewhat arbitrary and was chosen to
demonstrate huge scale without requiring a large FIB. Assuming every
data center site consumes 33 SID's in the Core domain space, a 4K
Core domain results in the following system scalability:
120 data center sites (4K / 33) 50M servers at 25G (417,792 *
120)
It should be appreciated that this example covers infrastructure
only. Further, design options are too numerous to describe
succinctly. Therefore, subjects such as an application-based
addressing architecture, multi-tenancy, service chaining, and so
forth are not accounted for. Nonetheless, deploying 50M endpoints
using a maximum FIB of 12K should generally provide super-granular
traffic steering and SDN services on low-cost hardware.
Authors' Addresses
Filsfils, et al. Expires October 3, 2016 [Page 12]
Internet-Draft Segment Routing April 2016
Clarence Filsfils (editor)
Cisco Systems, Inc.
Brussels
BE
Email: cfilsfil@cisco.com
Dennis Cai (editor)
Cisco Systems, Inc.
170, West Tasman Drive
San Jose, CA 95134
US
Email: dcai@cisco.com
Stefano Previdi
Cisco Systems, Inc.
Via Del Serafico, 200
Rome 00142
Italy
Email: sprevidi@cisco.com
Wim Henderickx
Alcatel-Lucent
Email: wim.henderickx@alcatel-lucent.com
Rob Shakir
BT
Email: rob.shakir@bt.com
Dave Cooper
Level 3
Email: Dave.Cooper@Level3.com
Francis Ferguson
Level 3
Email: Francis.Ferguson@level3.com
Tim LaBerge
Cisco
Email: tlaberge@cisco.com
Steven Lin
Microsoft
Email: slin@microsoft.com
Bruno Decraene
Orange
Email: bruno.decraene@orange.com
Filsfils, et al. Expires October 3, 2016 [Page 13]
Internet-Draft Segment Routing April 2016
Luay Jalil
Verizon
400 International Pkwy
Richardson, TX 75081
Email: luay.jalil@verizon.com
Jeff Tantsura
Ericsson
jeff.tantsura@ericsson.com
Filsfils, et al. Expires October 3, 2016 [Page 14]
