Rfc | 7031 |
Title | DHCPv6 Failover Requirements |
Author | T. Mrugalski, K. Kinnear |
Date | September
2013 |
Format: | TXT, HTML |
Status: | INFORMATIONAL |
|
Internet Engineering Task Force (IETF) T. Mrugalski
Request for Comments: 7031 ISC
Category: Informational K. Kinnear
ISSN: 2070-1721 Cisco
September 2013
DHCPv6 Failover Requirements
Abstract
The DHCPv6 protocol, defined in RFC 3315, allows for multiple servers
to operate on a single network; however, it does not define any way
the servers could share information about currently active clients
and their leases. Some sites are interested in running multiple
servers in such a way as to provide increased availability in case of
server failure. In order for this to work reliably, the cooperating
primary and secondary servers must maintain a consistent database of
the lease information. RFC 3315 allows for, but does not define, any
redundancy or failover mechanisms. This document outlines
requirements for DHCPv6 failover, enumerates related problems, and
discusses the proposed scope of work to be conducted. This document
does not define a DHCPv6 failover protocol.
Status of This Memo
This document is not an Internet Standards Track specification; it is
published for informational purposes.
This document is a product of the Internet Engineering Task Force
(IETF). It represents the consensus of the IETF community. It has
received public review and has been approved for publication by the
Internet Engineering Steering Group (IESG). Not all documents
approved by the IESG are a candidate for any level of Internet
Standard; see Section 2 of RFC 5741.
Information about the current status of this document, any errata,
and how to provide feedback on it may be obtained at
http://www.rfc-editor.org/info/rfc7031.
Copyright Notice
Copyright (c) 2013 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 ....................................................3
2. Definitions .....................................................3
3. Scope of Work ...................................................5
3.1. Alternatives to Failover ...................................5
3.1.1. Short-Lived Addresses ...............................5
3.1.2. Redundant Servers ...................................6
3.1.3. Distributed Databases ...............................6
3.1.4. Load Balancing ......................................7
4. Failover Scenarios ..............................................7
4.1. Hot Standby Model ..........................................7
4.2. Geographically Distributed Failover ........................7
4.3. Load Balancing .............................................8
4.4. 1-to-1, m-to-1, and m-to-n Models ..........................8
4.5. Split Prefixes .............................................8
4.6. Long-Lived Connections .....................................8
4.7. Partial Server Communication Loss ..........................9
5. Principles of DHCPv6 Failover ...................................9
5.1. Failure Modes ..............................................9
5.1.1. Server Failure .....................................10
5.1.2. Network Partition ..................................10
5.2. Synchronization Mechanisms ................................11
5.2.1. Lockstep ...........................................11
5.2.2. Lazy Updates .......................................12
6. DHCPv4 and DHCPv6 Failover Comparison ..........................12
7. DHCPv6 Failover Requirements ...................................13
7.1. Features out of Scope .....................................14
8. Security Considerations ........................................15
9. Acknowledgements ...............................................15
10. References ....................................................16
10.1. Normative References .....................................16
10.2. Informative References ...................................16
1. Introduction
The DHCPv6 protocol, defined in [RFC3315], allows for multiple
servers to be operating on a single network; however, it does not
define how the servers can share the same address and prefix
delegation pools and allow a client to seamlessly extend its existing
leases when the original server is down. [RFC3315] provides for
these capabilities but does not document how the servers cooperate
and communicate to provide this capability. Some sites are
interested in running multiple servers in such a way as to provide
redundancy in case of server failure. In order for this to work
reliably, the cooperating primary and secondary servers must maintain
a consistent database of the lease information.
This document discusses failover implementations scenarios, failure
modes, and synchronization approaches to provide background to the
list of requirements for a DHCPv6 failover protocol. It then defines
a minimum set of requirements that failover must provide to be
useful, while acknowledging that additional features may be specified
as extensions. This document does not define a DHCPv6 failover
protocol.
The failover model, to which these requirements apply, will initially
be a pairwise "hot standby" model (see Section 4.1) with a primary
server used in normal operation switching over to a backup secondary
server in the event of failure. Optionally, a secondary server may
provide failover service for multiple primary servers. However, the
requirements will not preclude a future load-balancing extension
where there is a symmetric failover relationship.
The DHCPv6 failover concept borrows heavily from its DHCPv4
counterpart [DHCPV4-FAILOVER] that never completed the
standardization process but has several successful, operationally
proven vendor-specific implementations. For a discussion about
commonalities and differences, see Section 6.
2. Definitions
This section defines terms that are relevant to DHCPv6 failover.
Definitions from [RFC3315] are included by reference. In particular,
"client" means any device, e.g., end-user host, CPE (Customer
Premises Equipment), or other router that implements client
functionality of the DHCPv6 protocol. A "server" is a DHCPv6 server,
unless explicitly noted otherwise. A "relay" is a DHCPv6 relay.
Binding (or client binding): a group of server data records
containing the information the server has about the addresses in
an IA (Identity Association, see Section 10 of [RFC3315]) or
configuration information explicitly assigned to the client.
Configuration information that has been returned to a client
through a policy -- for example, the information returned to all
clients on the same link -- does not require a binding.
DNS Update: the capability to update a DNS server's name database
using the on-the-wire protocol defined in [RFC2136]. Clients and
servers can negotiate the scope of such updates as defined in
[RFC4704].
Failover: the ability of one partner to continue offering services
provided by another partner, with minimal or no impact on clients.
FQDN: a fully qualified domain name. A fully qualified domain name
generally is a host name with at least one domain label under the
top-level domain. For example, "dhcp.example.org" is a fully
qualified domain name.
High Availability: a desired property of DHCPv6 servers to continue
providing services despite experiencing unwanted events such as
server crashes, link failures, or network partitions.
Load Balancing: the ability for two or more servers to each process
some portion of the client request traffic in a conflict-free
fashion.
Lease: an IPv6 address, an IPv6 prefix, or other resource that was
assigned ("leased") by a server to a specific client. A lease may
include additional information, like associated fully qualified
domain name (FQDN) and/or information about associated DNS
updates. A client obtains a lease for a specified period of time
(valid lifetime).
Partner: A "partner", for the purpose of this document, refers to a
failover server, typically the other failover server in a failover
relationship.
Stable Storage: each DHCP server is required to keep its lease
database in some form of storage (known as "stable storage") that
will be consistent throughout reboots, crashes, and power
failures.
Partner Failure: A power outage, unexpected shutdown, crash, or
other type of failure that renders a partner unable to continue
its operation.
3. Scope of Work
In order to fit within the IETF process effectively and efficiently,
the standardization effort for DHCPv6 failover is expected to proceed
with the creation of documents of increasing specificity.
Requirements document:
It begins with this document specifying the requirements for
DHCPv6 failover.
Design document:
A later document is expected to address the design of the DHCPv6
failover protocol.
Protocol document:
If sufficient interest exists, a later document is expected to
address the protocol details required to implement the DHCPv6
failover protocol itself.
The goal of this partitioning is, in part, to ease the validation,
review, and approval of the DHCPv6 failover protocol by presenting it
in comprehensible parts to the larger community.
Additional documents describing extensions may also be defined.
DHCPv6 failover requirements are presented in Section 7.
3.1. Alternatives to Failover
There are many scenarios in which a failover capability would be
useful. However, there are often much simpler approaches that will
meet the required goals. This section documents examples where
failover is not really needed.
3.1.1. Short-Lived Addresses
There are cases when IPv6 addresses are used only for a short time,
but there is a need to have high degree of confidence that those
addresses will be served. A notable example is PXE (Preboot
eXecution Environment) [RFC5970]. This is a mechanism for obtaining
configuration early in the process of bootstrapping over the network.
The PXE BIOS acquires an address in order to load the operating
system image and continue booting. Address and possibly other
configuration parameters are used during the boot process and are
discarded thereafter. Any lack of available DHCPv6 service at this
time will prevent such devices from booting.
Instead of deploying failover, it is better to use the much simpler
preference mechanism, defined in [RFC3315]. For example, consider
two or more servers with each having a distinct preference set (e.g.,
10 and 20). Both will answer a client's request. The client should
choose the one with the larger preference value. In case of failure
of the most preferred server, the next server will keep responding to
clients' queries. This approach is simple to deploy but does not
offer lease stability, i.e., in case of server failure, clients'
addresses and prefixes will change.
3.1.2. Redundant Servers
In some cases, the desire to deploy failover is motivated by high
availability, i.e., to continue providing services despite server
failure. If there are no additional requirements, that goal may be
fulfilled with simply deploying two or more independent servers on
the same link.
There are several well-documented approaches showing how such a
deployment could work. They are discussed in detail in [RFC6853].
Each of those approaches is simpler to deploy and maintain than full
failover.
3.1.3. Distributed Databases
Some servers may allow their lease database to be stored in external
databases. Another possible alternative to failover is to configure
two servers to connect to the same distributed database.
Care should be taken to understand how inconsistencies are solved in
such database backends and how such conflict resolutions affect
DHCPv6 server operation.
It is also essential to use only a database that provides equivalent
reliability and failover capability. Otherwise, the single point of
failure is only moved to a different location (database rather than
DHCPv6 server). Such a configuration does not improve redundancy but
significantly complicates deployment.
A common misconception regarding database-based redundancy is the
assumption that a conflict resolution after recovering from a network
partition is not necessary. To explain that fallacy, let's consider
an example where there is a very small pool with only one address.
There are two servers, each connected to a co-located database node
(i.e., running on the same hardware). Network partition occurs.
Each server is operating but has lost connection to its partner. Two
clients request an address, one from each server. Each server
consults its database and discovers that only one address is
available, so it is assigned to the client. Unfortunately, each
server assigned the same address to a different client. Making the
scenario more realistic (millions of addresses rather than one) just
decreased failure probability, but it did not eliminate the
underlying issue.
Any solution that involves a distributed database implementation of
DHCPv6 failover must take into account the requirements for security.
See Section 8 for additional information.
3.1.4. Load Balancing
Sometimes the desire to deploy more than one server is based on the
assumption that they will share the client traffic. Administrators
that are interested in such a capability are advised to deploy a
load-balancing mechanism, defined in [LOAD-BALANCING].
4. Failover Scenarios
The following sections provide several examples of deployment
scenarios and use cases that may be associated with capabilities
commonly referred to as "failover". These scenarios may be in or out
of scope for the DHCPv6 failover protocol to which this document's
requirements apply; they are enumerated here to provide a common
basis for discussion.
4.1. Hot Standby Model
In the simplest case, there are two partners that are connected to
the same network. Only one of the partners ("primary") provides
services to clients. In case of its failure, the second partner
("secondary") continues handling services previously handled by the
first partner. As both servers are connected to the same network, a
partner that fails to communicate with its partner while also
receiving requests from clients may assume with high probability that
its partner is down and the network is functional. This assumption
may affect its operation.
4.2. Geographically Distributed Failover
Servers may be physically located in separate locations. A common
example of such a topology is where a service provider has at least a
regional high performance network between geographically distributed
data centers. In such a scenario, one server is located in one data
center, and its failover partner is located in another remote data
center. In this scenario, when one partner finds that it cannot
communicate with the other partner, it does not necessarily mean that
the other partner is down.
4.3. Load Balancing
A desire to have more than one server in a network may also be
created by the desire to have incoming traffic be handled by several
servers. This decreases the load each server must endure when all
servers are operational. Although such a capability does not,
strictly, require failover -- it is clear that failover makes such an
architecture more straightforward.
Note that in a load-balancing situation that includes failover, each
individual server must be able to handle the full load normally
handled by both servers working together, or there is not a true
increase in availability.
4.4. 1-to-1, m-to-1, and m-to-n Models
A failover relationship for a specific network is provided by two
failover partners. Those partners communicate with each other and
back up all pools. This scenario is sometimes referred to as the
1-to-1 model and is considered relatively simple. In larger
networks, one server may be participating in several failover
relationships, i.e., it provides failover for several address or
prefix pools, each served by separate partners. Such a scenario can
be referred to as m-to-1. The most complex scenario, m-to-n, assumes
that each partner participates in multiple failover relationships.
4.5. Split Prefixes
Due to the extensive IPv6 address space, it is possible to provide
semi-redundant service by splitting the available pool of addressees
into two or more non-overlapping pools, with each server handling its
own smaller pool. Several versions of such a scenario are discussed
in [RFC6853].
4.6. Long-Lived Connections
Certain nodes may maintain long-lived connections. Since the IPv6
address space is large, techniques exist (e.g., [RFC6853]) that use
the easy availability of IPv6 addresses in order to provide increased
DHCPv6 availability. However, these approaches do not generally
provide for stable IPv6 addresses for DHCPv6 clients should the
server with which the client is interacting become unavailable.
The obvious benefit of stable addresses is the ability to update DNS
infrequently. While DNS can be updated every time an IPv6 address
changes, it introduces delays, and (depending on DNS configuration)
old entries may be cached for prolonged periods of time.
The other benefit of having a stable address is that many monitoring
solutions provide statistics on a per-IP basis, so IP changes make
measuring characteristics of a given box more difficult.
4.7. Partial Server Communication Loss
There is a scenario where the DHCPv6 server may be configured to
serve clients on one network adapter and communicate with a partner
server (server-to-server traffic) on a different network adapter. In
this scenario, if the server loses connectivity on the network
adapter used to communicate with the clients because of network
adapter (hardware) failure, there is no intimation of the loss of
service to the partner in the DHCPv6 failover protocol. Since the
servers are able to communicate with each other, the partner remains
ignorant of the loss of service to clients.
5. Principles of DHCPv6 Failover
This section describes important issues that will affect any DHCPv6
failover protocol. This section is not intended to define
implementation details but rather describes high-level concepts and
issues that are important to DHCPv6 failover. These issues form a
basis for later documents that will deal with the solutions to these
issues.
The general failover concept assumes that there are backup servers
that can provide service in case of a primary server failure. In
theory, there could be more than one backup server that could take up
the role if such a need arises. However, having more than two
servers introduces a very difficult issue of synchronizing between
partners. In the case of just a pair of cooperating servers, the
notification and processes can result in only one of two states:
fully successful (got response from a partner) and total failure (no
response, failure event occurred). Were there more than two partners
participating in a relationship, there would be intermediate,
inconsistent states where some partners had updated their state and
some had not. This would greatly complicate the protocol design and
would give little advantage in return. Therefore, an approach that
assumes a pair of cooperating servers was chosen.
5.1. Failure Modes
This section documents failure modes. This requirements document
does not make any claims whether those two failures are
distinguishable by a server.
5.1.1. Server Failure
Servers may become unresponsive due to a software crash, hardware
failure, power outage, or any number of other reasons. The failover
partner will detect such an event due to lack of responses from the
down partner. In this failure mode, the assumption is that the
server is the only equipment that is off-line and all other network
equipment is operating normally. In particular, communication
between other nodes is not interrupted.
When working under the assumption that this is the only type of
failure that can happen, the server may safely assume that its
partner unreachability means that it is down, so other nodes
(clients, in particular) are not able to reach it either, and no
services are provided.
It should be noted that recovery after the failed server is brought
back on-line is straightforward, due to the fact that it just needs
to download current information from the lease database of the
healthy partner and there is no conflict resolution required.
This is by far the most common failure mode between two failover
partners.
When the two servers are located physically close to each other,
possibly in the same room, the probability that a failure to
communicate between failover partners is due to server failure is
increased.
5.1.2. Network Partition
Another possible cause of partner unreachability is a failure in the
network that connects the two servers. This may be caused by failure
of any kind of network equipment: router, switch, physical cables, or
optic fibers. As a result of such a failure, the network is split
into two or more disjoint sections (partitions) that are not able to
communicate with each other. Such an event is called "network
partition". If failover partners are located in different
partitions, they won't be able to communicate with each other.
Nevertheless, each partner may still be able to serve clients that
happen to be part of the same partition.
If this failure mode is taken into consideration, a server can't
assume that partner unreachability automatically means that its
partner is down. They must consider the fact that the partner may
continue operating and interacting with a subset of the clients. The
only valid assumption is that the partner also detected the network
partition event and follows procedures specified for such a
situation.
It should be noted that recovery after a partitioned network is
rejoined is significantly more complicated than recovery from a
server failure event. As both servers may have kept serving clients,
they have two separate lease databases, and they need to agree on the
state of each lease (or follow any other algorithm to bring their
lease databases into agreement).
This failure mode is more likely (though still rare) in the situation
where two servers are in physically distant locations with multiple
network elements between them. This is the case in geographically
distributed failover (see Section 4.2).
5.2. Synchronization Mechanisms
Partners must exchange information about changes made to the lease
database. There are at least two types of synchronization methods
that may be used (see Sections 5.2.1 and 5.2.2). These concepts are
related to distributed databases, so some familiarity with
distributed database technology is useful to better understand this
topic.
5.2.1. Lockstep
When a server receives a REQUEST message from a client, it consults
its lease database and assigns requested addresses and/or prefixes.
To make sure that its partner maintains a consistent database, it
then sends information about a new or just updated lease to the
partner and waits for the partner's response. After the response
from its partner is received, the REPLY message is transmitted to the
client.
This approach has the benefit of having a completely consistent lease
database between partners at all times. Unfortunately, there is
typically a significant performance penalty for this approach as each
response sent to a client is delayed by the total sum of the delays
caused by two transmissions between partners and the processing by
the second partner. The second partner is expected to update its own
copy of the lease database in permanent storage, so this delay is not
negligible, even in fast networks.
Due to the advent of fast SSD (solid state disk) and battery-backed
RAM (random access memory) disk technology, this write performance
penalty can be limited to some degree.
5.2.2. Lazy Updates
Another approach to synchronizing the lease databases is to transmit
the REPLY message to the client before completing the update to the
partner. The server sends the REPLY to the client immediately after
assigning appropriate addresses and/or prefixes and initiates the
partner update at a later time, depending on the algorithm chosen.
Another variation of this approach is to initiate transmission to the
partner but not wait for its response before sending the REPLY to the
client.
This approach has the benefit of a minimal impact on server response
times; it is thus much better from a performance perspective.
However, it makes the lease databases loosely synchronized between
partners. This makes the synchronization more complex (particularly
the re-integration after a network partition event), as there may be
cases where one client has been given a lease on an address or prefix
of which the partner is not aware (e.g., if the server crashes after
sending the REPLY to the client but before sending update information
to its partner).
6. DHCPv4 and DHCPv6 Failover Comparison
There are significant similarities between existing DHCPv4 and
envisaged DHCPv6 failovers. In particular, both serve IP addresses
to clients, require maintaining consistent databases among partners,
need to perform consistent DNS updates, must be able take over
bindings offered by a failed partner, and must be able to resume
operation after the partner is recovered. DNS conflict resolution
works on the same principles in both DHCPv4 and DHCPv6.
Nevertheless, there are significant differences. IPv6 introduced
prefix delegation [RFC3633], which is a crucial element of the DHCPv6
protocol. IPv6 also introduced the concept of deprecated addresses
with separate preferred and valid lifetimes, both configured via
DHCPv6. Negative response (NACK) in DHCPv4 has been replaced with
the ability in DHCPv6 to provide a corrected response in a REPLY
message, which differs from a REQUEST.
Also, the typical large address space (close to 2^64 addresses on /64
prefixes expected to be available on most networks) may make managing
address assignment significantly different from DHCPv4 failover. In
DHCPv4, it was not possible to use a hash or calculated technique to
divide the significantly more limited address space, and therefore,
much of the protocol that deals with pool balancing and rebalancing
might not be necessary and can be done mathematically. Also, because
of the much lower degree of contention for IP addresses, the DHCPv6
failover protocol does not need to be tuned to support rapid
reclamation of IPv6 addresses following the loss of a failover peer's
database.
However, DHCPv6 prefix delegation is similar to IPv4 addressing in
terms of the number of available leases, and therefore, techniques
for pool balancing and rebalancing and more rapid reclamation of
prefixes allocated by a failed peer will be needed.
7. DHCPv6 Failover Requirements
This section summarizes the requirements for DHCPv6 failover.
Certain capabilities may be useful in some, but not all, scenarios.
Such additional features will be considered optional parts of
failover and will be defined in separate documents. As such, this
document can be considered an attempt to define requirements for the
DHCPv6 failover "core" protocol.
The core of the DHCPv6 failover protocol is expected to provide the
following properties:
1. The number of supported partners must be exactly two, i.e., there
are at most two servers that are aware of a specific lease.
2. For each prefix or address pool, a server must not participate in
more than one failover relationship.
3. The defined protocol must support the m-to-1 model (i.e., one
server may form more than one relationship), but an
implementation may choose to implement only the 1-to-1 model
(i.e., everything from one server is backed on another).
4. One partner must be able to continue serving leases offered by
the other partner. This property is also sometimes called "lease
stability". The failure of either failover partner should have
minimal or no impact on client connectivity. In particular, it
must not force the client to change addresses and/or change
prefixes delegated to it. Lease stability has the aim of
avoiding disturbance to long-lived connections.
5. Prefix delegation must be supported.
6. Use of the failover protocol must not introduce significant
performance impact on server response times. Therefore,
synchronization between partners must be done using some form of
lazy updates (see Section 5.2.2).
7. The pair of failover servers must be able to recover from a
server down failure (see Section 5.1.1).
8. The pair of failover servers must be able to recover from a
network partition event (see Section 5.1.2).
9. The design must allow secure communication between the failover
partners.
10. The definition of extensions to this core protocol should be
allowed, when possible.
Depending on the specific nature of the failure, the recovery
procedures mentioned in points 7 and 8 may require manual
intervention.
High availability is a property of the protocol that allows clients
to receive DHCPv6 services despite the failure of individual DHCPv6
servers. In particular, it means the server that takes over
providing service to clients from its failed partner will continue
serving the same addresses and/or prefixes. This property is also
called "lease stability".
Although progress on a standardized interoperable DHCPv4 failover
protocol has stalled, vendor-specific DHCPv4 failover protocols have
been deployed that meet these requirements to a large extent.
Accordingly, it would be appropriate to take into account the likely
coexistence of DHCPv4 and DHCPv6 failover solutions. In particular,
certain features that are common to both IPv4 and IPv6
implementations, such as the DNS Update mechanism, should be taken
into consideration to ensure compatible operation.
7.1. Features out of Scope
The following features are explicitly out of scope.
1. Load Balancing - This capability is considered an extension and
may be defined in a separate document. It must not be part of
the core protocol but rather defined as an extension. The
primary reason for this the desire to limit the complexity of the
core protocol. See [LOAD-BALANCING].
2. Configuration synchronization - Two failover partners are
expected to maintain the same configuration. Mismatched
configuration between partners is a frequent problem in failover
solutions. Unfortunately, that is an open-ended problem, since
different servers have very different configuration data models.
3. m-to-n model (see Section 4.4).
4. Servers participating in multiple failover relationships for any
given prefix or address pool.
8. Security Considerations
The design must provide a mechanism whereby each peer in a failover
relationship can identify the other peer, authenticate that
identification, and validate that the identified peer is the one with
which communication is intended. This mechanism should also
optionally provide support for confidentiality.
The protocol specification, when it is written, should provide
operational guidelines in the case of authentication mechanisms that
require access to network servers that have the potential to be
unreachable (e.g., what to do if a partner is reachable but the
remote Certificate Authority is unreachable due to a network
partition event).
The security considerations for the design itself will be discussed
in the design document.
9. Acknowledgements
This document extensively uses concepts, definitions and other parts
of [DHCPV4-FAILOVER]. Thanks to Bernie Volz and Shawn Routhier for
their frequent reviews and substantial contributions. The authors
would also like to thank Qin Wu, Jean-Francois Tremblay, Frank
Sweetser, Jiang Sheng, Yu Fu, Greg Rabil, Vithalprasad Gaitonde,
Krzysztof Nowicki, Steinar Haug, Elwyn Davies, Ted Lemon, Benoit
Claise, Stephen Farrell, Michal Hoeft, and Krzysztof Gierlowski for
their comments and feedback.
This work has been partially supported by the Department of Computer
Communications (a division of Gdansk University of Technology) and
the National Centre for Research and Development (Poland) under the
European Regional Development Fund, Grant No. POIG.01.01.02-00-045 /
09-00 (Future Internet Engineering Project).
10. References
10.1. Normative References
[RFC3315] Droms, R., Bound, J., Volz, B., Lemon, T., Perkins, C.,
and M. Carney, "Dynamic Host Configuration Protocol for
IPv6 (DHCPv6)", RFC 3315, July 2003.
[RFC3633] Troan, O. and R. Droms, "IPv6 Prefix Options for Dynamic
Host Configuration Protocol (DHCP) version 6", RFC 3633,
December 2003.
[RFC4704] Volz, B., "The Dynamic Host Configuration Protocol for
IPv6 (DHCPv6) Client Fully Qualified Domain Name (FQDN)
Option", RFC 4704, October 2006.
10.2. Informative References
[DHCPV4-FAILOVER]
Droms, R., Kinnear, K., Stapp, M., Volz, B., Gonczi, S.,
Rabil, G., Dooley, M., and A. Kapur, "DHCP Failover
Protocol", Work in Progress, March 2003.
[LOAD-BALANCING]
Kostur, A., "DHC Load Balancing Algorithm for DHCPv6",
Work in Progress, December 2012.
[RFC2136] Vixie, P., Thomson, S., Rekhter, Y., and J. Bound,
"Dynamic Updates in the Domain Name System (DNS UPDATE)",
RFC 2136, April 1997.
[RFC5970] Huth, T., Freimann, J., Zimmer, V., and D. Thaler, "DHCPv6
Options for Network Boot", RFC 5970, September 2010.
[RFC6853] Brzozowski, J., Tremblay, J., Chen, J., and T. Mrugalski,
"DHCPv6 Redundancy Deployment Considerations", BCP 180,
RFC 6853, February 2013.
Authors' Addresses
Tomek Mrugalski
Internet Systems Consortium, Inc.
950 Charter Street
Redwood City, CA 94063
USA
Phone: +1 650 423 1345
EMail: tomasz.mrugalski@gmail.com
Kim Kinnear
Cisco Systems, Inc.
1414 Massachusetts Ave.
Boxborough, Massachusetts 01719
USA
Phone: +1 (978) 936-0000
EMail: kkinnear@cisco.com