INTERNET-DRAFT Luyuan Fang
Intended Status: Informational Deepak Bansal
Expires: January 7, 2016 Microsoft
Fabio Chiussi
Cisco Systems
Chandra Ramachandran
Juniper Networks
Ebben Aries
Facebook
Shahram Davari
Broadcom
Barak Gafni
Mellanox
Daniel Voyer
Bell Canada
July 6, 2015
MPLS-Based Hierarchical SDN for Hyper-Scale DC/Cloud
draft-fang-mpls-hsdn-for-hsdc-03
Abstract
This document describes Hierarchical SDN (HSDN), an architectural
solution to scale the Data Center (DC) and Data Center Interconnect
(DCI) networks to support tens of millions of physical underlay
endpoints, while efficiently handling both Equal Cost Multi Path
(ECMP) load-balanced traffic and any-to-any end-to-end Traffic
Engineered (TE) traffic. HSDN achieves massive scale using
surprisingly small forwarding tables in the network nodes. HSDN
introduces a new paradigm for both forwarding and control planes, in
that all paths in the network are pre-established in the forwarding
tables and the labels can identify entire paths rather than simply
destinations. The HSDN forwarding architecture is based on four main
concepts: 1. Dividing the DC and DCI in a hierarchically-partitioned
structure; 2. Assigning groups of Underlay Border Nodes in charge of
forwarding within each partition; 3. Constructing HSDN MPLS label
stacks to identify endpoints and paths according to the HSDN
structure; and 4. Forwarding using the HSDN MPLS labels.
Status of this Memo
This Internet-Draft is submitted to IETF in full conformance with the
provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering
Task Force (IETF), its areas, and its working groups. Note that
other groups may also distribute working documents as
Fang et al. Expires <January 7, 2016> [Page 1]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
Internet-Drafts.
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."
The list of current Internet-Drafts can be accessed at
http://www.ietf.org/1id-abstracts.html
The list of Internet-Draft Shadow Directories can be accessed at
http://www.ietf.org/shadow.html
Copyright Notice
Copyright (c) 2015 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
1.1. Terminology . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2. DC and DCI Reference Model . . . . . . . . . . . . . . . . 8
2. Requirements . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1. MPLS-Based HSDN Design Requirements . . . . . . . . . . . . 10
2.2. Hardware Requirements . . . . . . . . . . . . . . . . . . . 11
3. HSDN Architecture - Forwarding Plane . . . . . . . . . . . . . 11
3.1. Hierarchical Underlay Partitioning . . . . . . . . . . . . 12
3.2. Underlay Partition Border Nodes . . . . . . . . . . . . . . 14
3.2.1. UPBN and UPBG Naming Convention . . . . . . . . . . . . 17
3.2.2. HSDN Label Stack . . . . . . . . . . . . . . . . . . . 17
3.2.3. HSDN Design Example . . . . . . . . . . . . . . . . . . 18
3.3. MPLS-Based HSDN Forwarding . . . . . . . . . . . . . . . . 20
3.3.1 Non-TE Traffic . . . . . . . . . . . . . . . . . . . . . 21
3.3.2 TE Traffic . . . . . . . . . . . . . . . . . . . . . . . 23
4. Scalability Analysis . . . . . . . . . . . . . . . . . . . . . 24
Fang et al. Expires <January 7, 2016> [Page 2]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
4.1. LFIB Sizing - ECMP . . . . . . . . . . . . . . . . . . . . 24
4.2. LFIB Sizing - TE . . . . . . . . . . . . . . . . . . . . . 25
5. HSDN Label Stack Assignment Scheme . . . . . . . . . . . . . . 26
6. HSDN Architecture - Control Plane . . . . . . . . . . . . . . . 28
6.1. The SDN Approach . . . . . . . . . . . . . . . . . . . . . 28
6.2. HSDN Distributed Control Plane . . . . . . . . . . . . . . 29
7. Security Considerations . . . . . . . . . . . . . . . . . . . 29
8. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 29
9. Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . 30
10. Contributors . . . . . . . . . . . . . . . . . . . . . . . . 30
11. References . . . . . . . . . . . . . . . . . . . . . . . . . 31
11.1 Normative References . . . . . . . . . . . . . . . . . . . 31
11.2 Informative References . . . . . . . . . . . . . . . . . . 31
Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . . 32
Fang et al. Expires <January 7, 2016> [Page 3]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
1. Introduction
With the growth in the demand for cloud services, the end-to-end
cloud network, which includes Data Center (DC) and Data Center
Interconnect (DCI) networks, has to scale to support millions to tens
of millions of underlay network endpoints. These endpoints can be
bare-metal servers, virtualized servers, or physical and virtualized
network functions and appliances.
The scalability challenge is twofold: 1. Being able to scale using
low-cost network nodes while achieving high resource utilization in
the network; and 2. Being able to scale at low operational and
computational complexity while supporting both Equal-Cost Multi-Path
(ECMP) load-balanced traffic and any-to-any Traffic Engineering (TE)
traffic.
Being able to scale at low cost requires to avoid the potential
explosion of the routing tables in the network nodes as the number of
underlay network endpoints increases. Current commodity switches have
relatively small routing and forwarding tables. For example, the
typical Forwarding Information Base (FIBs) and Label Forwarding
Information Base (LFIBs) tables in current low-cost network nodes
contain 16K or 32K entries. These small sizes are clearly
insufficient to support entries for all the endpoints in the hyper-
scale cloud. Address aggregation is used to ameliorate the problem,
but the scalability challenges remain, since the dynamic and elastic
environment in the DC/cloud often brings the need to handle finely
granular prefixes in the network in order to support Virtual Machine
(VM) and Virtualized Network Function (VNF) mobility.
Other factors contribute to the FIB/LFIB explosion. For example, in a
typical DC using a fat Clos topology, even the support of ECMP load
balancing may become an issue if the individual outgoing paths
belonging to an ECMP group carry different outgoing labels, since a
single destination may contribute multiple entries in the tables.
Another key scalability issue to resolve is the complexity of certain
desired functions that should be supported in the network, the most
prominent one being TE. Currently, any-to-any server-to-server TE in
the DC/DCI is simply unfeasible, as path computation and bandwidth
allocation at scale, an NP-complete problem, becomes rapidly
unmanageable. Furthermore, the forwarding state needed in the network
nodes for TE tunnels contributes in a major way to the explosion of
the LFIBs, since each TE tunnel corresponds to an entry in the
tables.
Other major scalability issues are related to the efficient creation,
management, and use of tunnels, for example the configuration of
Fang et al. Expires <January 7, 2016> [Page 4]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
protection paths for fast restoration.
Many additional scalability issues in terms of operational and
computational complexity need to be resolved in order to scale the
control plane and the network state. In particular, the controller-
centric approach of Software Defined Networks (SDNs), which is
increasingly accepted as "the way to build the next generation
clouds," still needs to be demonstrated to be scalable to the levels
required in the hyper-scale DC and cloud.
Finally, the underlay network architecture should offer certain
capabilities to facilitate the support of the demands of the overlay
network.
In this document, we present Hierarchical SDN (HSDN), a set of
solutions for all these scalability challenges in the underlay
network, both in the forwarding and in the control plane.
Although HSDN can be used with any forwarding technology, including
IPv4 and IPv6, it has been designed to leverage Multi Protocol Label
Switching (MPLS)-based forwarding [RFC3031], using label stacks
[RFC3032] constructed according to the HSDN structure. This document
therefore describes MPLS-based HSDN. Here, we describe end-to-end
(host-to-host) MPLS-based HSDN, where the entire HSDN label stacks
from source to destination are imposed at the server's Network
Interface Cards (NICs), and thus all the IP lookups are confined to
the network edges. However, MPLS-based HSDN does not need to be end-
to-end, since label imposition could happen instead at the network
nodes (e.g., at the Top-of-Rack (ToR) switches), or intermediate
lookups in the network could be introduced, or even a combination of
MPLS and IP forwarding could be deployed as part of the HSDN network.
The HSDN underlay network is suited to support any Layer 2 or Layer 3
virtualized overlay network technology. In this document, we assume a
MPLS-based overlay technology using a Virtual Network (VN) Label,
which is encapsulated in the HSDN label stack. However the
description can be easily generalized to any overlay technology, such
as BGP/MPLS IP VPNs [RFC4364], EVPN [RFC7432], VXLAN [RFC7348], NVGRE
[I-D.sridharan-virtualization-nvgre], Geneve
[I-D.draft-gross-geneve], and other technologies.
HSDN achieves massive scale using surprisingly small LFIBs in the
network nodes, while supporting both ECMP load-balanced traffic and
any-to-any end-to-end TE traffic [HSDNSOSR15]. HSDN also brings
important simplifications in the control plane and in the
architecture of the SDN controller.
The HSDN architecture and operation is characterized by two
Fang et al. Expires <January 7, 2016> [Page 5]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
fundamental properties. First, all paths in the network are pre-
established in the forwarding tables. Second, the HSDN labels can
identify entire paths or groups of paths rather than simply
destinations.
These two properties radically simplify establishing and handling
tunnels. In addition to optimally handling both ECMP and Non-Equal
Cost Multi Path load balancing, HSDN enables any-to-any, end-to-end,
server-to-server TE at scale. With HSDN, the "cost" of establishing a
tunnel is essentially eliminated, since the "tunnels" are pre-
established in the network, and the TE task becomes one of path
assignment and bandwidth allocation to the flows. As a larger portion
of the traffic can be engineered effectively, the network can be run
at a higher utilization using comparatively smaller buffers at the
nodes.
The HSDN forwarding architecture in the underlay network is based on
four main concepts: 1. Dividing the DC and DCI in a hierarchically-
partitioned structure; 2. Assigning groups of Underlay Border Nodes
in charge of forwarding within each partition; 3. Constructing HSDN
MPLS label stacks to identify the end points according to the HSDN
structure; and 4. Forwarding using the HSDN MPLS labels.
HSDN is designed to allow the physical decoupling of control and
forwarding, and have the LFIBs configured by a controller according
to a full SDN approach. The controller-centric approach is described
in this document. In this context, "MPLS forwarding" in HSDN simply
means using MPLS labels to forward the packets, since there is no
need for label distribution protocols.
However, the HSDN control plane can also be built using a hybrid
approach, in which a routing or label distribution protocol is used
to distribute the labels, in conjunction with a SDN controller. This
hybrid approach may be particularly useful during technology
migration. The use of BGP Labeled Unicast (BGP-LU) for label
distribution and LFIB configuration in a HSDN architecture is
described in [I-D.fang-idr-bgplu-for-hsdn].
1.1. Terminology
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].
Term Definition
----------- --------------------------------------------------
BGP Border Gateway Protocol
BGP-LU Border Gateway Protocol Labeled Unicast
Fang et al. Expires <January 7, 2016> [Page 6]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
DC Data Center
DCGW DC Gateway (Border Leaf)
DCI Data Center Interconnect
DID Destination Identifier
ECMP Equal Cost MultiPathing
FIB Forwarding Information Base
HSDN Hierarchical SDN
LDP Label Distribution Protocol
LFIB Label Forwarding Information Base
LN Leaf Node
MPLS Multi-Protocol Label Switching
NIC Network Interface Card
PID Path Identifier
SDN Software Defined Network
SN Spine Node
SVR Server
UP Underlay Partition
UPBG Underlay Partition Border Group
UPBN Underlay Partition Border Node
TE Traffic Engineering
ToR Top-of-Rack switch
TR Top-of-Rack switch (used in figures)
VN Virtual Network
VM Virtual Machine
VNF Virtualized Network Function
WAN Wide Area Network
In this document, we also use the following terms.
o End device: A physical device attached to the DC/DCI network.
Examples of end devices include bare metal servers, virtualized
servers, network appliances, etc.
o Level: A layer in the hierarchy of underlay partitions in the HSDN
architecture.
o Overlay Network (ON): A virtualized network that provides Layer 2
or Layer 3 virtual network services to multiple tenants. It is
implemented over the underlay network.
o Path Label (PL): A label used for MPLS-based HSDN forwarding in
the underlay network.
o Row: A row of racks where end devices reside in a DC.
o Tier: One of the layers of network nodes in a Clos-based topology.
o Underlay Network (UN): The physical network that provides the
Fang et al. Expires <January 7, 2016> [Page 7]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
connectivity among physical end devices. It provides transport for
the overlay network traffic.
o Underlay Partition (UP): A logical portion of the underlay network
designed according to the HSDN architecture. Underlay partitions
are arranged in a hierarchy consisting of multiple levels.
o VN Label (VL): A label carrying overlay network traffic. It is
encapsulated in the underlay network in a stack of path labels
constructed according to the HSDN forwarding scheme.
1.2. DC and DCI Reference Model
Here we show the typical structure of the DC and DCI, which we use in
the rest of this document to describe the HSDN architecture. We also
introduce a few commonly used terms to assist in the explanation.
Figure 1 illustrates multiple DCs interconnected by the DCI/WAN.
+-------------+
| |
| DC |
| |
+-------------+
\-----.
( ')
.--(. '.---.
( ' ' )
( DCI/WAN )
(. .)
( ( .) \
/'--' '-''---' +-------------+
+-------------+ | |
| | | DC |
| DC | | |
| | +-------------+
+-------------+
Figure 1. DCIWAN interconnecting multiple DCs.
Figure 2 below illustrates the typical structure of a Clos-based DC
fabric.
Fang et al. Expires <January 7, 2016> [Page 8]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
+----+ +----+
DCGW +---------+ GW + + GW +---------+
| ++--++\ /++--++ |
| +--------|--|--\--\ | |
| | | \-/-\-|--|--------+ |
+-+-++ +-+--+/ \+--+-+ ++-+-+
Spine +---------+ SN + | SN | | SN + + SN +---------+
| +----+\ ++--++ ++--++\ /+----+ |
| +---------------\---+ | | +---/---------------+ |
| | ... \ | | / ... | |
++-+-+ +----+ +----+ ++--++ ++--++ +----+ +----+ +-+-++
Leaf | LN | | LN | | LN | | LN | ||LN | | LN | | LN | | LN |
++--++ ++--++ +---++ ++---+ +---++ ++--++ +---++ ++--++
| \ / | | \ / | | \ / | | \ / |
| / \ | | / \ | | / \ | | / \ |
| / ... \ | | / ... \ | | / ... \ | | / ... \ |
++++ +--+ ++++ ++++ +--+ ++++ ++++ +--+ ++++ ++++ +--+ ++++
ToR |TR| |TR| |TR| |TR| |TR| |TR| |TR| |TR| |TR| |TR| |TR| |TR|
+--+ +--+ +--+ +--+ +--+ +--+ +--+ +--+ +--+ +--+ +--+ +--+
| | | | | | | | | | | |
+--+ +--+ +--+ +--+ +--+ +--+ +--+ +--+ +--+ +--+ +--+ +--+
Server +--+ +--+ +--+ +--+ +--+ +--+ +--+ +--+ +--+ +--+ +--+ +--+
rack +--+ +--+ +--+ +--+ +--+ +--+ +--+ +--+ +--+ +--+ +--+ +--+
+--+ +--+ +--+ +--+ +--+ +--+ +--+ +--+ +--+ +--+ +--+ +--+
Figure 2. Typical Clos-based DC fabric topology.
Note: Not all nodes and links are shown in Figure 2.
The DC fabric shown in Figure 2 uses what is known as a spine and
leaf architecture with a multi-stage Clos-based topology
interconnecting multiple tiers of network nodes. The DC Gateways
(DCGWs) connect the DC to the DCI/WAN. The DCGW connect to the Spine
Nodes (SNs), which in turn connect to the Leaf Nodes (LFs). The Leaf
Nodes connect to the ToRs. Each ToR typically resides in a rack
(hence the name) accommodating a number of servers connected to their
respective ToR. The servers may be bare metal or virtualized.
Each tier of switches and the connectivity between switches is
designed to offer a desired capacity and provide sufficient bandwidth
to the servers and end devices.
Figure 2 is not meant to represent the precise topology of the DC. In
fact, the precise topology and connectivity between the tiers of
switches depends on the specific design of the DC. More or less tiers
of switches (spines or leafs) or asymmetric topologies, not shown in
the figure, may be used. A precise description of the possible
topologies and related design criteria is out of the scope of this
Fang et al. Expires <January 7, 2016> [Page 9]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
document.
What is relevant for this document is the fact that a typical large-
scale DC topology does not have all the tiers fully connected to the
adjacent tier (i.e., not all network nodes in a tier are necessarily
connected to all network nodes in the adjacent tiers). This is
especially true for the tiers closer to the endpoints, and is due to
the sheer number of connections and devices (in other words, in a
large, fat Clos there are too many network nodes in some tiers for
all network nodes to connect to one another), and to the physical
constraints of the DC (i.e., the network nodes may be located
physically apart in separate rooms or buildings, and full
connectivity may become too costly).
In a typical DC, the racks of servers are physically organized in
"clusters" of racks, and dedicated banks of leaf switches may serve
the ToRs in each cluster. For example, the racks may be physically
placed in rows of racks, and a cluster of racks may correspond to a
portion of a row, an entire row, or multiple rows of racks. Indeed,
the leaf nodes are sometimes called "middle (or end) of the row
switches" because they are physically located in a rack in the middle
(or end) of a row of racks of servers. In turn, leaf nodes may also
be organized in "zones" (we use "clusters" and "zones" as generic
terms, but other terms may be used in the industry to refer to
similar concepts), and banks of spines may be assigned to serve each
zone. For example, a zone may include all the banks of leaf nodes
that are in a room or in a building in the DC.
The actual connectivity is typically organized following an
aggregation/multiplexing connectivity architecture that consolidates
traffic from the edges into the leafs and spines, while allowing for
over-subscription in order to strike a reasonable trade-off between
cost and available capacity. The connectivity between each tier may
use some form of shuffle-exchange topology that attempts to "mix" the
available paths while taking in account the physical constraints.
The key observation is that it is impractical, uneconomical, and
ultimately unnecessary to use a fully connected Clos-based topology
in a large scale DC. Because of the physical constraints, the
topology of a large DC is not a flat, fully-connected Clos, but
rather has a dertain hierarchy. The HSDN architecture recognizes this
fact, and uses it to dramatically simplify forwarding and control
planes using an approach that is also hierarchical.
2. Requirements
2.1. MPLS-Based HSDN Design Requirements
Fang et al. Expires <January 7, 2016> [Page 10]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
The following are the key design requirements for MPLS-based HSDN
solutions.
1) MUST support millions to tens of millions of underlay network
endpoints in the DC/DCI.
2) MUST use very small LFIB sizes (e.g., 16K or 32K LFIB entries) in
all network nodes.
3) MUST support both ECMP load-balanced traffic and any-to-any, end-
to-end, server-to-server TE traffic.
4) MUST support ECMP traffic load balancing using a single forwarding
entry in the LFIBs per ECMP group.
5) MUST require IP lookup only at the network edges.
6) MUST support encapsulation of overlay network traffic, and support
any network virtualization overlay technology.
7) MUST support control plane using both full SDN controller
approach, and traditional distributed control plane approach using
any label distribution protocols.
2.2. Hardware Requirements
The following are the hardware requirements to support HSDN.
1) The server NICs MUST be able to push a HSDN label stack consisting
of as many path labels as levels in the HSDN hierarchical
partition (e.g., 3 path labels).
2) The network nodes MUST support MPLS forwarding.
3) The network nodes MUST be able perform ECMP load balancing on
packets carrying a label stack consisting of as many path labels
as levels in the HSDN hierarchical partition, plus one or more VN
label/header for the overlay network (e.g., 3 path labels + 1 VN
label/header). For example, if the hash function used for ECMP
forwarding is based on the IP 5-tuple, as is often the case, this
requirement implies that the network nodes MUST be able to lookup
the 5-tuple inside up to four labels.
3. HSDN Architecture - Forwarding Plane
As mentioned above, a primary design requirement for HSDN is to
enable scalability of the forwarding plane to tens of millions of
network endpoints using very small LFIB sizes in all network nodes in
Fang et al. Expires <January 7, 2016> [Page 11]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
the DC/DCI, while supporting both ECMP and any-to-any server-to-
server TE traffic.
The driving principle of the HSDN forwarding plane is "divide and
conquer" by partitioning the forwarding task into local and
independent forwarding. When designed properly, such an approach
enables extreme horizontal scaling of the DC/DCI.
HSDN is based on four concepts:
1) Dividing the underlay network in a hierarchy of partitions;
2) Assigning groups of Underlay Partition Border Nodes (UPBN) to each
partition, in charge of forwarding within the corresponding
partition;
3) Constructing HSDN label stacks for the endpoint Forward
Equivalency Classes (FECs) in accordance with the underlay network
partition hierarchy;
4) Configuring the LFIBs in all network nodes and forwarding using
the label stacks.
As explained in Section 3.3.1, the HSDN label stacks can be used to
identify entire paths to each endpoint, rather than simply the
destination endpoint itself. As a matter of fact, the HSDN solution
is meant to be configured with all possible paths in the network pre-
established in the LFIBs in the network nodes. In this case, a FEC
per path to each endpoint is defined. However, because of the way the
HSDN architecture is designed, the required local number of entries
in the LFIB of each network node remains surprisingly small.
In this section, we explain in detail each of these concepts.
Scalability analysis for both ECMP load-balanced and TE traffic is
presented in Section 4. In Section 5, we describe a possible label
stack assignment scheme for HSDN.
3.1. Hierarchical Underlay Partitioning
HSDN is based on dividing the DC/DCI underlay network into logical
partitions arranged in a multi-level hierarchy.
The HSDN hierarchical partitioning is illustrated in Figure 3.
Fang et al. Expires <January 7, 2016> [Page 12]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
+----------------------------------------------------------+
UP | UP0 |
Level 0 | |
|+---------------------------------+ +------------------+|
|| |...| ||
+|---------------------------------|---|------------------|+
UP | UP1-0 | | UP1-N |
Level 1 | | | |
|+-------------+ +-------------+| | +-------------+|
|| |...| || |...| ||
+|-------------|---|-------------|+ +---|-------------|+
UP | UP2-0-0 | | UP2-0-N | | UP2-N-N |
Level 2 | | | | | |
+-----+ +-----+ +-----+ +-----+ +-----+ +-----+
| SVR | | SVR | | SVR | | SVR | | SVR | | SVR |
|-----|-|-----| |-----|-|-----| |-----|-|-----|
Overlay |VM|VM| |VM|VM| |VM|VM| | | |VM|VM| | |
Level |-----| |-----| |-----| | | |-----| | |
|VM|VM| |VM|VM| |VM|VM| | | |VM|VM| | |
+-----+ +-----+ +-----+ +-----+ +-----+ +-----+
Figure 3. HSDN underlay network hierarchical partitioning of DC/DCI.
The hierarchy consists of multiple levels of Underlay Partitions
(UPs). For simplicity, we describe HSDN using three levels of
partitioning, but more or less levels can be used, depending on the
size and architecture of the overall network, using similar design
principles (as shown in Section 4, three levels of partitions are
sufficient to achieve scalability to tens of millions servers using
very small LFIBs).
The levels of partitions are nested into a hierarchical structure. At
each level, the combination of all partitions covers the entire
DC/DCI topology. In general, within each level, the UPs do not
overlap, although there may be design scenarios in which overlapping
UPs within a level may be used. The top level (Level 0) consists of a
single underlay partition UP0 (the HSDN concept can be extended to
multi-partitioned Level 0).
We use the following naming convention for the UPs:
- Partitions at Level i are referred to as UPi (e.g., UP0 for Level
0, UP1 for Level 1, UP2 for Level2, and so on).
- Within each level, partitions are identified by a rightmost
sequential number (starting from 1) referring to the corresponding
level and a set of sequential number(s) for each partition in a
Fang et al. Expires <January 7, 2016> [Page 13]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
higher level that the specific partition is nested into.
For example, at Level 1, there are N partitions, referred to as
UP1-1 to UP1-N.
Similarly, at Level 2, there are M partitions for each Level 1
partitions, for a total of NxM partitions. For example, the Level
2 partitions nested into Level 1 partition UP1-1 are UP2-1-1 to
UP2-1-M, while the ones nested into UP1-N are UP2-N-1 to UP2-N-M.
- Note that for simplicity in illustrating the partitioning, we
assume a symmetrical arrangement of the partitions, where the
number of partitions nested into each partition at a higher level
is the same (e.g., all UP1 partitions have M UP2 partitions). In
practice, this is rarely the case, and the naming convention can
be adapted accordingly for different numbers of partitions nesting
into each higher level partition (e.g., partition UP1-1 has M1 UP2
partitions, partition UP1-2 has M2 UP2 partitions, and so on).
The following considerations complete the description of Figure 3.
o The servers (bare metal or virtualized) are attached to the bottom
UP level (in our case, Level 2). A similar naming convention as
the one used for the partitions may be used.
o In Figure 3, we also show an additional Overlay Level. This
corresponds to the virtualized overlay network (if any) providing
Virtual Networks (VN) connecting Virtual Machines (VMs) and other
overlay network endpoints. Overlay network traffic is encapsulated
by the HSDN underlay network. As mentioned in the Introduction,
the HSDN underlay network is suited to support any Layer 2 or
Layer 3 virtualized overlay network technology, such as BGP/MPLS
IP VPNs [RFC4364], EVPN [RFC7432], VXLAN [RFC7348], NVGRE
[I-D.sridharan-virtualization-nvgre], Geneve
[I-D.draft-gross-geneve], and other technologies. A full
description of the encapsulation of these technologies into the
HSDN underlay label stack is out of scope of this document and
will be addressed in a separate document.
The UPs are designed to contain one or more tiers of switches in the
DC topology or nodes in the DCI. The key design criteria in defining
the partitions at each level is that they need to follow the
"natural" connectivity implemented in the DC/DCI topology. An example
is given in Section 3.2.3 to further clarify how the partitions are
designed.
3.2. Underlay Partition Border Nodes
Fang et al. Expires <January 7, 2016> [Page 14]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
Once the HSDN hierarchical partitioning is defined, Underlay
Partition Border Nodes (UPBNs) are assigned to each UP. This is
illustrated in Figure 4.
+++++++++++++++++++++++++++++++++++++++++++++++++
| UP0 <--------|-----+
| | |
| +++++++++++++++++++++++++++++++++++++++++++++ | |
| | +-----------------------------------+ | | |
| | | +---------+ UPBG1-i +---------+ | | | |
| | | | UPBN1-i | ... | UPBN1-i | | | | |
| | | +---------+ +---------+ | | | |
| | +-----------------------------------+ | | |
++|+++++++++++++++++++++++++++++++++++++++++++|++ |
| UP1 <------|----+ | +---------+
| +++++++++++++++++++++++++++++++++++++++++ | | +---| PL0 |
| | +-----------------------------------+ | | | +---------+
| | | +---------+ UPBG2-i-j +---------+ | | | +------| PL1 |
| | | |UPBN2-i-j| ... |UPBN2-i-j| | | | +---------+
| | | +---------+ +---------+ | | | +------| PL2 |
| | +-----------------------------------+ | | | +---------+
++|+++++++++++++++++++++++++++++++++++++++|++ | +---| VL |
| UP2 <----|------+ | +---------+
| +---------------+ | |
| | SVR-i-j-k | | |
| | +-----------+ | | |
| | | NVE |<-------------|---------+
| | +-----+-----+ | |
+++++++++++| | VM | VM | |+++++++++++++
| +-----+-----+ |
| | VM | VM | |
+-+-----+-----+-+
Figure 4. UBPNs, UBPGs, and label stack assignment.
The UPBNs serve as the connecting nodes between adjacent partitions.
As such, the UPBNs belong to two partitions in adjacent levels in the
hierarchy and they constitute the entry points for traffic from the
higher level partition destined to the corresponding lower level
partition (and vice-versa, they are the exit points for traffic from
a lower level partition to a higher level partition). As such, they
constitute the forwarding end destinations within each partition.
In order to provide sufficient capacity and support traffic load
balancing between the levels in the hierarchy, multiple UPBNs are
assigned to each partition. The UPBNs for each partition are grouped
into an Underlay Partition Border Group (UPBG). As shown in Section
Fang et al. Expires <January 7, 2016> [Page 15]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
5, using an appropriate Label Stack Assignment scheme all UPBNs in a
UPBG can be made identical for ECMP traffic forwarding (i.e., the
ECMP entries in the LFIBs in all UPBNs in a UPBG are identical).
Thus, for ECMP traffic load balancing, all UPBNs belong to the same
FEC as far as the higher level partition is concerned. For TE
traffic, a desired UPBN within a UPBG group may need to be specified,
and thus the UPBNs in a UPBG are not forwarding-wise equivalent.
In practice, the UPs are designed by finding the most advantageous
way to partition the DC Clos-based topology and the DCI topology. As
mentioned above, the connectivity of any large-scale DC is not fully
flat, but rather contains some sort of hierarchical organization.
Recognizing the hierarchy of the physical connectivity is an
important starting point in the design of the partitions.
Within the DC, the UPBNs in each level are subsets of the network
nodes in one of the tiers that form the multi-stage Clos
architecture.
In general, in addition to the UPBNs, the UPs may internally contain
tiers of network nodes that are not UPBNs. A specific design example
to further illustrate the HSDN partitioning is provided in Section
3.2.3.
As explained in more detail in Section 3.3, for forwarding purposes,
by partitioning the DC/DCI in this manner and using HSDN forwarding,
the UPBNs need to have entries in their LFIBs only to reach
destinations in the two partitions to which they belong to (i.e.,
their own corresponding lower-level partition and the higher-level
partition to which they nested to). The network nodes inside the UPs
only need to have entries in their LFIB to reach the destinations in
their partition.
Similarly, in order to establish all possible paths in the entire
network, the UPBNs need to have entries in their LFIBs only for all
possible paths to the destinations in the two partitions to which
they belong to.
From these considerations, a first design heuristic for choosing the
partitioning structure is to keep the number of partitions nested at
each level into the higher level relatively small for all levels. For
the lowest level, the number of endpoints (servers) in each partition
should also be kept to manageable levels.
Clearly, the design tradeoff is between the size and the number of
partitions at each level. Although finding the optimal design choice
may require a little trial-and-error computation of different
options,fortunately, for most practical deployments, it is relatively
Fang et al. Expires <January 7, 2016> [Page 16]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
simple to find a good tradeoff that achieves the desired scalability
to millions or tens of millions of endpoints.
3.2.1. UPBN and UPBG Naming Convention
We use a similar naming convention for the UPBNs and UPBGs as the one
used for the UPs:
- UPBNi is a UPBN between partitions at Level(i) and Level(i-1).
Similarly for UPBG.
- Within each level, the UPBNs are identified by a set of sequential
number(s) equal to the corresponding sequential number(s) of the
corresponding partition within that level.
For example, at Level 1, UPBN1-1 corresponds to partition UP1-1,
and connects UP0 with UP1-1. UPBN1-N corresponds to partition UP1-
N and connects UP0 with UP1-N, and so on. Similarly for UPBG.
At Level 2, UPBN2-1-1 corresponds to partition UP2-1-1 and
connects UP1-1 with UP2-1-1, and so on. Similarly for UPBG.
Note that the UPBNs within an UPBGs can be further distinguished
using an appropriate naming convention (for example, using an
additional sequential number within the UPBG), which for
simplicity is not shown here. This more granular naming convention
is needed to configure the paths and the TE tunnels.
3.2.2. HSDN Label Stack
In MPLS-Based HSDN, an MPLS label stack is defined and used for
forwarding. The key notion in HSDN is that the label stack is defined
and the labels are assigned in accordance with the hierarchical
partitioned structure defined above.
The label stack, shown in Figure 4 above, is constructed as follows.
- The label stack contains as many Path Labels (PLs) as levels in
the partitioning hierarchy.
- Each PL in the label stack is associated to a corresponding level
in the partition hierarchy and is used for forwarding at that
level.
In the scenario of Figure 4, PL0 is associated to Level 0 and is
used to forward to destinations in UP0, PL1 is associated to Level
1 and is used to forward to destinations in any UP1 partitions,
and PL2 is associated to Level 2 and is used to forward to
Fang et al. Expires <January 7, 2016> [Page 17]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
destinations in any UP2 partitions.
- A VN Label (VL) is also shown in the label stack in Figure 4. This
label is associated to the Overlay Level and is used to forward in
the overlay network. The VL is simply encapsulated in the label
stack and transported in the HSDN underlay network. As mentioned
above, the HSDN underlay network is suited to support any Layer 2
or Layer 3 virtualized overlay network technology, and thus the VL
may be a label, a tag, or some other identifier, depending on the
overlay technology used. The details of the VL encapsulation and
processing for different overlay technologies are out of scope of
this document.
Each endpoint in the DC/DCI is identified by a corresponding label
stack. For a given endpoint, the label stack is constructed in such a
way that the PLO specifies the UP1 to which the endpoint is attached
to, the PL1 specifies the UP2 to which the endpoint is attached to,
and the PL2 specifies the FEC in the UP2 corresponding to the
endpoint.
The labels in the HSDN label stack can identify entire paths, rather
than simply the end destination within the corresponding partition.
This can be used to bring dramatic simplifications in handling
tunnels and TE traffic in particular, as further explained in Section
3.3.2.
As mentioned above, in this draft we describe end-to-end MPLS-based
HSDN forwarding, where the entire HSDN label stacks from the sources
to the destinations are inserted at the server's NICs. In this
scenario, the label stack imposed at a server points all the way to
the end destination of the packet, which may be in a different DC.
With any-to-any, end-to-end TE, the HSDN label stack identifies the
entire path to the destination. For inter-DC traffic, there may be
cases where the path through the remote DC would be preferably
determined when the packet arrives at that DC, or when the packet
leaves the source DC, so it may be desirable that part of the label
stack be imposed inside the network rather than at the server.
Nothing precludes this design choice, and a lookup may be added where
desired in the HSDN network.
A scheme to assign the PL labels in the HSDN label stack is described
in Section 5.
3.2.3. HSDN Design Example
We use an example to further explain the HSDN design criteria to
define the hierarchically-partitioned structure of the DC/DCI. We use
the same design example in the Scalability Analysis section (Section
Fang et al. Expires <January 7, 2016> [Page 18]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
4) to show the LFIB sizing with ECMP and TE traffic.
To summarize some of the design heuristics for the HSDN underlay
partitions:
- The UPs should be designed to follow the "natural" connectivity
topology in the DC/DCI.
- The number of partitions at each level nested into the higher
level should be relatively small (since they are FEC entries in
the LFIBs in the network nodes in the corresponding levels).
- The number of endpoints (servers) in each partition in the lowest
level should be relatively small (since they are FEC entries in
the LFIBs in the network nodes in the lowest level).
- The number of levels should be kept small (since it corresponds to
the number of path labels in the stack).
- The number of tiers in each partition in each level should be kept
small. This is due to the multiplicative fanout effect for TE
traffic (explained in Section 4.2), which has a major impact on
the LFIB size needed to support any-to-any server-to-server TE.
The HSDN forwarding plane design consists in finding the best
tradeoff among these conflicting objectives. Although the optimal
design choices ultimately depend on the specific deployment,
fortunately, it is generally rather straightforward to identify
design choices that can support scalability to millions or tens of
millions of servers.
Here we describe a design example to illustrate that a three-level
HSDN hierarchy is sufficient to scale the DC/DCI to tens of millions
of servers.
With three levels, a possible design choice for the UP1s is to have
each UP1 correspond to a DC. With this choice, the UP0 corresponds to
the DCI and the UPBN1s are the DCGWs in each DC (the UPBG1s group the
DCGWs in each DC).
Once the UP1s are chosen this way, a possible design choice for the
UP2s is to have each UP2 correspond to a group of racks, where each
group of racks may correspond to a portion of a row of racks, an
entire row of racks, or multiple rows of racks. The specific best
choice of how many racks should be in a group of racks corresponding
to each UP2 ultimately depends on the specific connectivity in the DC
and the number of servers per racks.
Fang et al. Expires <January 7, 2016> [Page 19]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
While precise numbers depend on the specific technologies used in
each deployment, here and in the Scalability Analysis section
(Section 4) we want to give some ideas of the scaling capabilities of
HSDN. For this purpose, we use some hypothetical yet reasonable
numbers to characterize the partitioning design example.
Assume the following: a) 20 DCs connected via the DCI/WAN; b) 50
servers per rack; c) 20 racks per group of racks; d) 50 groups of
racks per DC.
With these numbers, there are 500K servers per DC, for a total of 10M
underlay network endpoints in the DC/DCI.
In the HSDN structure in this example, there are 20 UP1s, 500 UP2s
per UP1, and 1000 servers per UP2.
3.3. MPLS-Based HSDN Forwarding
The hierarchically partitioned structure and the corresponding label
stack are used in HSDN to scale the forwarding plane horizontally
while using LFIBs of surprising small sizes in the network nodes.
As explained above, each label in the HSDN label stack is associated
with one of the levels in the hierarchy and is used to forward to
destinations in the underlay partitions at that level.
With HSDN, by superimposing a hierarchically-partitioned structure
and using a label stack constructed according to such a structure, we
are able to impose a forwarding scheme that is aggregated by
construction. This translates in dramatic reductions in the size of
the LFIBs in the network nodes, since each node only needs to know a
limited portion of the forwarding space.
HSDN supports any label assignment scheme to generate the labels in
the label stack. However, if a label assignment scheme that is
consistent with the HSDN structure is used, additional
simplifications of the LFIBs and the control plane can be achieved.
In Section 5 below, we present one example of such a scheme, where
the labels in the label stack represent the "physical" location of
the endpoint, expressed according to the HSDN structure. For TE
traffic, the labels represent a specific path towards the desired
destination through the HSDN structure.
In the Scalability Analysis section (Section 4) and in the Control
Plane section (Section 6) we assume that such a Label Assignment
scheme is used.
Fang et al. Expires <January 7, 2016> [Page 20]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
In the rest of this section, we describe the life of a packet in the
HSDN DC/DCI. We use the specific design example described in Section
3.2.3 above to help in the explanation, but of course the forwarding
would be similar for other design choices.
3.3.1 Non-TE Traffic
We first describe the behavior for ECMP load-balanced, non-TE
traffic. In the HSDN DC/DCI, for a packet that needs to be forwarded
to a specific endpoint in the underlay network, the outer label PL0
specifies which UP1 contains the endpoint. Let's refer to this UP1 as
UP1-a. For ECMP traffic, the PL0 binding is with a FEC corresponding
to the UPBG1-a associated with UP1-a. Note that all the endpoints
reachable via UP1-a are forwarded using the same FEC entry for Level
0 in the hierarchical partitioning.
Once the packet reaches one of the network nodes UPBN1-a in the
UPBG1-a group (the upstream network nodes perform ECMP load
balancing, thus the packet may enter UP1-a via any of the UPBN1-a
nodes), the PL0 is popped and the PL1 is used for forwarding in the
UP1-a (to be precise, because of penultimate hop popping, it is the
network node immediately upstream of the chosen UPBN1-a that pops the
label P0).
The PL1 is used within UP1-a to reach the UP2 which contains the
endpoint. Let's refer to this UP2 as UP2-a. In the UP2 network nodes
the PL1 binding is with a FEC corresponding to the UPBG2-a associated
with UP2-a. Similarly as above, note that all the endpoints reachable
via UP2-a are forwarded using the same FEC entry for Level 1 in the
hierarchical partitioning.
Once the packet reaches one of the network nodes UPBN2-a in the
UPBG2-a group (once again, the upstream network nodes perform ECMP
load balancing, so the packet may transit to any of the UPBN2-a
nodes), the PL1 is popped and the PL2 is used for the rest of the
forwarding (again, to be precise, the penultimate network node
upstream of UPBN2-a is the one popping the PL1 label).
The PL2 is used within UP2-a to reach the desired endpoint. Note that
the UPBN2 nodes and the network nodes in the UP2s have entries in
their LFIBs only to reach endpoints within their UP2. They can reach
endpoints in other UP2s by using a FEC entry corresponding to the UP2
containing the destination endpoint, identified by PL1.
The following two observations help in further clarifying the
forwarding operation above.
- The PL0 is used for forwarding from the source to the UPBN1-a. For
Fang et al. Expires <January 7, 2016> [Page 21]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
a packet originating from an endpoint attached to a certain UP2,
say UP2-b, nested to a different UP1, say UP1-b, PL0 is used for
forwarding in all network nodes that the packet transits until it
reaches the UPBN1-a. This includes network nodes in UP2-b and UP1-
b (i.e., "on the way up" from UP2). It also includes one of the
UPBN1-b nodes.
It is important to note, however, that the PL0 is not popped at
the UPBN1-b, since it is used for forwarding to the destination
UPBN1-a.
- It should be pointed out that an important requirement for HSDN is
to achieve route optimization for ECMP traffic, meaning that the
hierarchy should forward a packet from any source to any
destination using the same number of hops and without introducing
any additional latency compared to a flat architecture. For
example, a packet originating from an endpoint in UP2,N,M, and
destined to an endpoint in the same UP2,N,M should not be
forwarded all the way to the highest level in the hierarchy and
back, but should be forwarded to the desired endpoint by "turning
it around" towards the destination at the first node in the
UP2,N,M that contains an entry to that desired endpoint. Indeed,
if the packet turns around at the proper node, it will go through
the same number of hops as it would have gone through in a flat
architecture. This should hold true even in the case where the
UP1s and/or UP2s contain intermediate tiers of switches and the
packet needs to be turned around in the intermediate nodes.
Route optimization is easily achieved in HSDN by simply having the
packets only carry the portion of the label stack that is needed
to reach the destination using the appropriate turn around node.
Continuing with our example, the packet above only needs PL2 to be
optimally forwarded, since it should never "go out" of UP2,N,M.
Thus, PL0 and PL1 should not be included in its label stack, to
avoid an unnecessary round trip up and down the DC through all the
levels in the hierarchy. Similarly, a packet originating and
terminating in the same UP1, but in different UP2s, only needs PL1
and PL2 to be forwarded.
In this case, a network node would have to process different
labels for traffic going up and out the partition versus traffic
staying in the partition ("going up" and "coming down" refer to
the direction of traffic in Figure 3). Since the label spaces for
the two path labels may overlap, ambiguity would result. Depending
on the LFIB configuration, the two Most Significant Bits (MSBs) in
each label in Figure 4 may be reserved for identifying the layer
(i.e., whether the label is PL0, PL1, or PL2) and resolve
ambiguity.
Fang et al. Expires <January 7, 2016> [Page 22]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
A better solution to achieve the same without using precious bits
in the labels is to use a "turn around entry" in the LFIBs, which
flags that the packet needs to turn around at that node and the
relevant label is not the outer label (as it would be for traffic
going up or coming down, for which the outer label just needs to
pass through), but is the one underneath (thus, the outer label
needs to be popped to expose the relevant label). In our example,
the packet destined to an endpoint in the same UP2,N,M of the
originating server may carry a PL1 corresponding to the "turn
around" label value and a PL2 corresponding to the desired
endpoint within UP2, and does not need a PL0.
In the case of ECMP load-balanced non-TE traffic, the labels in the
HSDN label stack identify ECMP groups for each destination in the
corresponding partition. In this way, at each node in the partition,
the outgoing label is the same for all paths belonging to the same
ECMP group. A label allocation scheme for this is described in
Section 5.
3.3.2 TE Traffic
Handling TE traffic in the hyper-scale DC/DCI presents major
scalability challenges, since each TE tunnel contributes one entry in
the forwarding tables, and the TE path and bandwidth allocation
computation is a NP-complete problem.
HSDN introduces radical simplifications in establishing and handling
tunnels, and in supporting TE in particular.
In HSDN, all paths in the network can be pre-established in the
LFIBs. Because of the way the HSDN architecture is constructed, the
number of entries that have to be stored in the local LFIB in each
network node remains surprisingly small.
In this case, the labels in the HSDN stack identify entire paths, or
groups of paths, to each destination in each partition, rather than
just the destination itself.
With HSDN, since the "cost" of establishing a tunnel is essentially
eliminated (all "tunnels" are pre-established in the network), and
the TE task becomes one of path assignment and bandwidth allocation
to the flows. Furthermore, the hierarchical structure of HSDN makes
it possible to devise algorithms and heuristics for path and
bandwidth allocation computation that operate largely independently
in each partition, and are therefore computationally feasible even at
large scale. A description of such algorithms is out of scope of this
document. As a larger portion of the traffic can be engineered
effectively, the network can be run at a higher utilization using
Fang et al. Expires <January 7, 2016> [Page 23]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
comparatively smaller buffers at the nodes.
Since all paths can be accommodated in the LFIBs, HSDN makes it
possible to support "TE Max Case" with small LFIB sizes. In TE Max
Case, all sources are connected to all destinations (e.g., server to
server) with TE tunnels, the tunnels using all possible distinct
paths in the network. TE Max Case gives therefore an upper bound to
the number of TE tunnels (and consequently, LFIB entries) in the
network.
The fact that the LFIBs remain relatively small even when all
possible paths are configured is the consequence of two desirable
properties of HSDN.
First, since in HSDN the individual UPs are designed in such a way to
be relatively small, the number of paths in each partition can be
kept to a manageable number.
Second, the hierarchical structure of HSDN makes it possible to use
the partitioning astutely to break the "TE Fanout Multiplicative
Effect," which defines the number of paths to a destination, and can
easily contribute to the LFIB explosion as the number of hops and the
fanout of each hop to each destination in the network increases. As
explained in Section 4.2, with the hierarchical structure, the TE
Fanout Multiplicative Effect is only multiplicative within each level
in the hierarchy. Thus, by properly designing the partitioning, the
multiplicative effect can be kept to a manageable level.
In the case of TE traffic, the processing of the different labels in
the label stack is similar to what described above for ECMP load-
balanced non-TE traffic. However, the labels are bound to FECs
identifying a specific path within each UPs that is traversed.
4. Scalability Analysis
In this section, we compute the maximum size of the LFIBs for non-
TE/ECMP traffic and any-to-any server-to-server TE traffic.
4.1. LFIB Sizing - ECMP
For ECMP traffic, at each level, all destinations belonging to the
same partition at a lower level are forwarded using the same FEC
entry in the LFIB, which identifies the destination UPBG for that
level, or the destination endpoint at the lower level. Since the UPs
are designed in such a way to keep the number of destinations small
in all UPs, and the network nodes only need to know how to reach
destinations in their own UP and in the adjacent UP at the higher
level in the hierarchy, this translate to the fact that hyper scale
Fang et al. Expires <January 7, 2016> [Page 24]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
of the DC/DCI can be achieved with very small LFIB sizes in all the
individual network nodes.
A detailed explanation of how the LFIB size can be computed in all
the nodes of an HSDN network is given in [HSDNSOSR15]. The worst case
for the LFIB size occurs at one of the network nodes that serve as
UPBNs for one of the levels of UPs in the hierarchy. The level where
the LFIB size occurs depend on the specific choice of the
partitioning design.
4.2. LFIB Sizing - TE
As noted above, TE traffic may add a considerable number of entries
to LFIB, since it creates one new FEC per TE tunnel to each
destination.
HSDN provides a solution to this problem, and in fact, HSDN can
support any-to-any server-to-server "TE Max Case" with small LFIB
sizes.
In a Clos Topology (the analysis can be extended to generic
topologies), the number of paths in a UP with N destination can be
easily computed. The number of paths (and the maximum number of LFIB
entries) is equal to the products of the switch fanout in each tier
traversed from the source to the destination in that UP. This is the
"TE Fanout Multiplicative Effect" mentioned above, which is
illustrated in Figure 5. Accordingly:
Total # LFIB Entries for TE Max Case = N * F1 * F2 * ... * F(M-1)
Where Fi is the fanout of a switch in each tier traversed to the
destination, M is the number of tiers in the UP, and N is the number
of destinations in the UP.
Once again, by properly designing the UPs, the TE Fanout
Multiplicative Effect can be kept under control, since the path
computation is local for each of the UPs. HSDN breaks the
multiplicative effect, since the TE Fanout Multiplicative Effect is
multiplicative only within each UP, rather than in the entire
network, and the "multiplication" restarts at each level of the
hierarchy. A detailed description of the LFIB computation in all
network nodes to support TE Max Case is given in [HSDNSOSR15].
Fang et al. Expires <January 7, 2016> [Page 25]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
+-------+
Source | Src |
| Node |
+-------+ F1=3
/ | \
/ | \
+-------+ +-------+ +-------+
Tier 1 | | | | | |
| Node | | Node | | Node |
+-+-----+ +---+---+ +-----+-+ F2=3
| \ \ / | \ / / |
| \/ \---|---/ \/ |
| / \ /---|---\ / \ |
| / / \ | / \ \ |
+-+-----+ +---+---+ +-----+-+
Tier 2 | | | | | |
| Node | | Node | | Node |
+-------+ +-------+ +-------+ F3=2
\ \ / \ / /
\ / \ / \ /
\ / / \ \ /
+-------+ +-------+
| | | |
Tier 3 | Node | | Node |
+-------+ +-------+ F4=1
\ /
\ /
+-------+
| Dest |
Destination | Node |
+-------+
Figure 5. Fan out multiplicative effect with TE.
5. HSDN Label Stack Assignment Scheme
HSDN can use any scheme to assign the labels in the label stack.
However, if a label assignment scheme which assigns labels in a way
consistent with the HSDN structure, important simplifications can be
achieved in the control plane and in the LFIBs.
For non-TE FECs, the HSDN label assignment scheme assigns labels
according to the "physical" location of the endpoint in the HSDN
structure. Continuing our design example from above, for an endpoint
X in UP2-a, PL0 would identify the DC in which the endpoint is
located, PL1 would identify the group of racks in which the endpoint
is located within the DC, and PL2 would identify the endpoint within
Fang et al. Expires <January 7, 2016> [Page 26]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
the group of servers within the DC.
For TE FECs, the HSDN label assignment scheme assigns labels to
identify a specific path in each UP that is traversed. In our
example, for a specific TE tunnel to endpoint X, PL0 would identify
the specific path that should be followed in the DCI, PL1 would
identify the path that should be followed within the DC to reach the
group of racks, and PL2 would identify the path to reach the endpoint
within the group of racks (if there are multiple paths).
In order to assign labels to both non-TE traffic and TE traffic, HSDN
uses a label format in which the labels are divided into two logical
sub-fields, one identifying the destination within the UP, called
Destination Identifier (DID), and one identifying the path, called
Path Identifier (PID). The Path Identifier is only relevant for TE
traffic, and can be zero for non-TE traffic. The HSDN Label format is
illustrated in Figure 6.
0 d 19
+++++++++++++++++++++++++++++++++++++++++++++++++++++
| Destination Identifier | Path Identifier |
+++++++++++++++++++++++++++++++++++++++++++++++++++++
|<------- (d) bits ------->|<---- (20-d) bits ----->|
LSB MSB
Figure 6. HSDN Label format.
In this label assignment scheme, the path labels associated with a
partition are globally unique within that partition, meaning that
different partitions at the same level can use the same label space.
For PL0, the path labels are globally unique within the entire
network, since there is only one UP0. Neither of these is a scaling
limitation, since all partitions are relatively small.
The bits in the DID for each level must be sufficient to represent
the distinct destinations that need to be known in the UPs at that
level, and the bits in the PID for each level must be sufficient to
represent all the distinct paths that need to be defined in the UPs
at that level (closed-form expressions for both these numbers are
given in [HSDNSOSR15]). In practice, this is not a significant
scalability constraint: with three MPLS labels in the stack,
partitioning architectures and label formats according to this scheme
can be found to scale to tens of millions of servers.
Depending on the LFIB configuration, the two MSBs may be reserved for
Fang et al. Expires <January 7, 2016> [Page 27]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
identifying the level (i.e., whether the label is PL0, PL1, or PL2)
to resolve ambiguity (not shown in Figure 6). Note, however, that
this is not strictly necessary and the same function of identifying
the level can be achieved by simply allocating "turn around" entries
in the nodes, as explained in Section 3.3.1, so an individual node
always sees the same label in the stack.
By properly designing the UPs, this label assignment scheme can
support the desired scalability and the support of end-to-end TE
traffic.
Note that by using this type of label assignment scheme important
benefits can be achieved, including:
- The LFIBs become rather "static," since the FECs are tied to
"physical" locations and paths, which change infrequently. This
simplifies the use of the SDN approach to configure the LFIBs via
a controller.
- All paths in each ECMP group use the same outgoing labels. This
guarantees that a single LFIB entry can be used for each ECMP
group.
The label stack needs to be imposed at the entry points. For an
endpoint, this implies that the server NIC must be able to push a
three-label stack of path labels (in addition to possibly push one
additional VL label for the overlay network).
6. HSDN Architecture - Control Plane
HSDN has been designed to support the controller-centric SDN approach
in a scalable fashion. HSDN also supports the traditional distributed
control plane approach.
HSDN introduces important simplifications in the control plane and in
the network state as well.
6.1. The SDN Approach
In the controller-centric SDN approach, the SDN controller configures
the LFIBs in all the network nodes. With HSDN, the hierarchical
partitioned structure offers a natural framework for a distributed
implementation of the SDN controller, since the control plane in each
UP is largely independent from other UPs. The individual UP control
planes operate in parallel, with loose synchronization among one
another.
Therefore, the HSDN control plane is logically partitioned in a way
Fang et al. Expires <January 7, 2016> [Page 28]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
that is consistent with the forwarding plane partitioning. Each UP is
assigned a corresponding UP controller, which configures the LFIBs in
the network nodes in the corresponding UP. The individual UP
controllers communicate with one another to exchange the labels and
construct the label stacks. In HSDN, configuring the LFIBs in the
network nodes is not a difficult task, since the labels are static
and configuration updates are needed only when the physical topology
changes or endpoints are added or permanently removed, and thus they
are not too frequent.
Each UP controller at the lowest level of the hierarchy is also in
charge of providing the label stacks to the server's NICs in the
corresponding partition. For this purpose, a number of label servers,
which may also be arranged in a hierarchy, are used to provide the
mappings between IP addresses and label stacks.
Redundancy is superimposed to the structure of UP controllers, with
each UP controller shadowing UP controllers in other UPs.
The HSDN UP controllers may also be in charge of TE computation. HSDN
TE path computation algorithms that perform for the most part
partition-local computation (so the computation is also horizontally
scalable) but still approach global optimality using inter-UP-
controller synchronization at a different time scale, can be devised.
6.2. HSDN Distributed Control Plane
The HSDN control plane can also be built using a hybrid approach, in
which a routing or label distribution protocol is used to distribute
the labels, in conjunction with a controller. An example using BGP-
LU [RFC3107] is presented in [I-D.fang-idr-bgplu-for-hsdn].
7. Security Considerations
When the SDN approach is used, the protocols used to configure the
LFIBs in the network nodes MUST be mutually authenticated.
For general MPLS/GMPLS security considerations, refer to [RFC5920].
Given the potentially very large scale and the dynamic nature in the
cloud/DC environment, the choice of key management mechanisms need to
be further studied.
To be completed.
8. IANA Considerations
TBD.
Fang et al. Expires <January 7, 2016> [Page 29]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
9. Acknowledgments
We would like to acknowledge Yakov Rekhter for many discussions
related to HSDN.
10. Contributors
Vijay Gill
Salesforce
Email: vgill@salesforce.com
Linda Dunbar
Huawei Technologies
5430 Legacy Drive, Suite #175
Plano, TX 75024
Email: linda.dunbar@huawei.com
Andrew Qu
MediaTek
2860 Junction Ave.
San Jose, CA 95134
Email: andrew.qu@mediatek.com
Jeff Tantsura
Ericsson
200 Holger Way
San Jose, CA 95134
Email: jeff.tantsura@ericsson.com
Wen Wang
Century Link
2355 Dulles Corner Blvd.
Herndon, VA 20171
Email: wen.wang@centurylink.com
Himanshu Shah
Ciena
3939 North 1st Street
San Jose, CA 95112
Email: hshah@ciena.com
Ramki Krishnan
Dell
Email: ramki_krishnan@dell.com
Iftekhar Hussain
Infinera Corporation
140 Caspian Ct,
Fang et al. Expires <January 7, 2016> [Page 30]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
Sunnyvale, CA 94089
Email: ihussain@infinera.com
11. References
11.1 Normative References
[RFC2119] Bradner, S., "Key words for use in RFCs to Indicate
Requirement Levels", BCP 14, RFC 2119, March 1997.
[RFC3031] Rosen, E., Viswanathan, A., and R. Callon, "Multiprotocol
Label Switching Architecture", RFC 3031, January 2001.
[RFC3032] Rosen, E., Tappan, D., Fedorkow, G., Rekhter, Y.,
Farinacci, D., Li, T., and A. Conta, "MPLS Label Stack
Encoding", RFC 3032, January 2001.
[RFC3107] Rekhter, Y. and E. Rosen, "Carrying Label Information in
BGP-4", RFC 3107, May 2001.
[RFC4364] Rosen, E. and Y. Rekhter, "BGP/MPLS IP Virtual Private
Networks (VPNs)", RFC 4364, February 2006.
[RFC7432] Sajassi et al., "BGP MPLS Based Ethernet VPN", RFC 7432,
February 2015.
11.2 Informative References
[RFC5920] Fang, L., Ed., "Security Framework for MPLS and GMPLS
Networks", RFC 5920, July 2010.
[RFC7348] M. Mahalingam et al., "Virtual eXtensible Local Area
Network (VXLAN): A Framework for Overlaying Virtualized
Layer 2 Networks over Layer 3 Networks", RFC 7348, August
2014.
[I-D.fang-idr-bgplu-for-hsdn] L. Fang et al., "BGP-LU for HSDN Label
Distribution", draft-fang-idr-bgplu-for-hsdn-01 (work in
progress), July 2015.
[I-D.sridharan-virtualization-nvgre] M. Sridharan et al., "NVGRE:
Network Virtualization using Generic Routing
Encapsulation", draft-sridharan-virtualization-nvgre-08
(work in progress), April 2015.
[I-D.draft-gross-geneve] J. Gross et al., "Geneve: Generic Network
Virtualization Encapsulation", draft-gross-geneve-02 (work
in progress), October 2014.
Fang et al. Expires <January 7, 2016> [Page 31]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
[HSDNSOSR15] L. Fang et al., "Hierarchical SDN for the Hyper-Scale,
Hyper-Elastic Data Center and Cloud", ACM SIGCOMM
Symposium on SDN Research 2015, Santa Clara, CA, June
2015.
Authors' Addresses
Luyuan Fang
Microsoft
5600 148th Ave NE
Redmond, WA 98052
Email: lufang@microsoft.com
Deepak Bansal
Microsoft
15590 NE 31st St.
Redmond, WA 98052
Email: dbansal@microsoft.com
Fabio Chiussi
Cisco Systems
500 108th Avenue N.E., Suite 500
Bellevue, WA 98004
Email: fchiussi@cisco.com
Chandra Ramachandran
Juniper Networks
Electra, Exora Business Park Marathahalli - Sarjapur Outer Ring Road
Bangalore, KA 560103, India
Email: csekar@juniper.net
Ebben Aries
Facebook
1601 Willow Road
Menlo Park, CA 94025
Email: exa@fb.com
Shahram Davari
Broadcom
3151 Zanker Road
San Jose, CA 95134
Email: davari@broadcom.com
Barak Gafni
Mellanox
6 Habarzel St.
Tel Aviv, Israel
Fang et al. Expires <January 7, 2016> [Page 32]
INTERNET DRAFT MPLS-Based HSDN July 6, 2015
Email: gbarak@mellanox.com
Daniel Voyer
Bell Canada
Email: daniel.voyer@bell.ca
Fang et al. Expires <January 7, 2016> [Page 33]