Request for Comments: 3612 Old Dog Consulting
Category: Informational September 2003
Applicability Statement for Restart Mechanisms
for the Label Distribution Protocol (LDP)
Status of this Memo
This memo provides information for the Internet community. It does
not specify an Internet standard of any kind. Distribution of this
memo is unlimited.
Copyright Notice
Copyright (C) The Internet Society (2003). All Rights Reserved.
Abstract
This document provides guidance on when it is advisable to implement
some form of Label Distribution Protocol (LDP) restart mechanism and
which approach might be more suitable. The issues and extensions
described in this document are equally applicable to RFC 3212,
"Constraint-Based LSP Setup Using LDP".
1. Introduction
Multiprotocol Label Switching (MPLS) systems are used in core
networks where system downtime must be kept to a minimum. Similarly,
where MPLS is at the network edges (e.g., in Provider Edge (PE)
routers) [RFC2547], system downtime must also be kept to a minimum.
Many MPLS Label Switching Routers (LSRs) may, therefore, exploit
Fault Tolerant (FT) hardware or software to provide high availability
of the core networks.
The details of how FT is achieved for the various components of an FT
LSR, including the switching hardware and the TCP stack, are
implementation specific. How the software module itself chooses to
implement FT for the state created by the LDP is also implementation
specific. However, there are several issues in the LDP specification
[RFC3036] that make it difficult to implement an FT LSR using the LDP
protocols without some extensions to those protocols.
Proposals have been made in [RFC3478] and [RFC3479] to address these
issues.
2. Requirements of an LDP FT System
Many MPLS LSRs may exploit FT hardware or software to provide high
availability (HA) of core networks. In order to provide HA, an MPLS
system needs to be able to survive a variety of faults with minimal
disruption to the Data Plane, including the following fault types:
- failure/hot-swap of the switching fabric in an LSR,
- failure/hot-swap of a physical connection between LSRs,
- failure of the TCP or LDP stack in an LSR,
- software upgrade to the TCP or LDP stacks in an LSR.
The first two examples of faults listed above may be confined to the
Data Plane. Such faults can be handled by providing redundancy in
the Data Plane which is transparent to LDP operating in the Control
Plane. However, the failure of the switching fabric or a physical
link may have repercussions in the Control Plane since signaling may
be disrupted.
The third example may be caused by a variety of events including
processor or other hardware failure, and software failure.
Any of the last three examples may impact the Control Plane and will
require action in the Control Plane to recover. Such action should
be designed to avoid disrupting traffic in the Data Plane. Since
many recent router architectures can separate the Control and Data
Planes, it is possible that forwarding can continue unaffected by
recovery action in the Control Plane.
In other scenarios, the Data and Control Planes may be impacted by a
fault, but the needs of HA require the coordinated recovery of the
Data and Control Planes to a state that existed before the fault.
The provision of protection paths for MPLS LSP and the protection of
links, IP routes or tunnels through the use of protection LSPs is
outside the scope of this document. See [RFC3469] for further
information.
3. General Considerations
In order for the Data and Control Plane states to be successfully
recovered after a fault, procedures are required to ensure that the
state held on a pair of LDP peers (at least one of which was affected
directly by the fault) are synchronized. Such procedures must be
implemented in the Control Plane software modules on the peers using
Control Plane protocols.
The required actions may operate fully after the failure (reactive
recovery) or may contain elements that operate before the fault in
order to minimize the actions taken after the fault (proactive
recovery). It is rare to implement actions that operate solely in
advance of the failure and do not require any further processing
after the failure (preventive recovery) - this is because of the
dynamic nature of signaling protocols and the unpredictability of
fault timing.
Reactive recovery actions may include full re-signaling of state and
re-synchronization of state between peers and synchronization based
on checkpointing.
Proactive recovery actions may include hand-shaking state transitions
and checkpointing.
4. Specific Issues with the LDP Protocol
LDP uses TCP to provide reliable connections between LSRs to exchange
protocol messages to distribute labels and to set up LSPs. A pair of
LSRs that have such a connection are referred to as LDP peers.
TCP enables LDP to assume reliable transfer of protocol messages.
This means that some of the messages do not need to be acknowledged
(e.g., Label Release).
LDP is defined such that if the TCP connection fails, the LSR should
immediately tear down the LSPs associated with the session between
the LDP peers, and release any labels and resources assigned to those
LSPs.
It is notoriously difficult to provide a Fault Tolerant
implementation of TCP. To do so might involve making copies of all
data sent and received. This is an issue familiar to implementers of
other TCP applications, such as BGP.
During failover affecting the TCP or LDP stacks, therefore, the TCP
connection may be lost. Recovery from this position is made worse by
the fact that LDP control messages may have been lost during the
connection failure. Since these messages are unconfirmed, it is
possible that LSP or label state information will be lost.
At the very least, the solution to this problem must include a change
to the basic requirements of LDP so that the failure of an LDP
session does not require that associated LDP or forwarding state be
torn down.
Any changes made to LDP in support of recovery processing must meet
the following requirements:
- offer backward-compatibility with LSRs that do not implement the
extensions to LDP,
- preserve existing protocol rules described in [RFC3036] for
handling unexpected duplicate messages and for processing
unexpected messages referring to unknown LSPs/labels.
Ideally, any solution applicable to LDP should be equally applicable
to CR-LDP.
5. Summary of the Features of LDP FT
LDP Fault Tolerance extensions are described in [RFC3479]. This
approach involves:
- negotiation between LDP peers of the intent to support extensions
to LDP that facilitate recovery from failover without loss of
LSPs,
- selection of FT survival on a per LSP/label basis or for all
labels on a session,
- sequence numbering of LDP messages to facilitate acknowledgement
and checkpointing,
- acknowledgement of LDP messages to ensure that a full handshake is
performed on those messages either frequently (such as per
message) or less frequently as in checkpointing,
- solicitation of up-to-date acknowledgement (checkpointing) of
previous LDP messages to ensure the current state is secured, with
an additional option that allows an LDP partner to request that
state is flushed in both directions if graceful shutdown is
required,
- a timer to control how long LDP and forwarding state should be
retained after the LDP session failure, but before being discarded
if LDP communications are not re-established,
- exchange of checkpointing information on LDP session recovery to
establish what state has been retained by recovering LDP peers,
- re-issuing lost messages after failover to ensure that LSP/label
state is correctly recovered after reconnection of the LDP
session.
The FT procedures in [RFC3479] concentrate on the preservation of
label state for labels exchanged between a pair of adjacent LSRs when
the TCP connection between those LSRs is lost. There is no intention
within these procedures to support end-to-end protection for LSPs.
6. Summary of the Features of LDP Graceful Restart
LDP graceful restart extensions are defined in [RFC3478]. This
approach involves:
- negotiation between LDP peers of the intent to support extensions
to LDP that facilitate recovery from failover without loss of
LSPs,
- a mechanism whereby an LSR that restarts can relearn LDP state by
resynchronization with its peers,
- use of the same mechanism to allow LSRs recovering from an LDP
session failure to resynchronize LDP state with their peers
provided that at least one of the LSRs has retained state across
the failure or has itself resynchronized state with its peers,
- a timer to control how long LDP and forwarding state should be
retained after the LDP session failure, but before being discarded
if LDP communications are not re-established,
- a timer to control the length of the resynchronization period
between adjacent peers should be completed.
The procedures in [RFC3478] are applicable to all LSRs, both those
with the ability to preserve forwarding state during LDP restart and
those without. LSRs that can not preserve their MPLS forwarding
state across the LDP restart would impact MPLS traffic during
restart. However, by implementing a subset of the mechanisms in
[RFC3478] they can minimize the impact if their neighbor(s) are
capable of preserving their forwarding state across the restart of
their LDP sessions or control planes by implementing the mechanism in
[RFC3478].
7. Applicability Considerations
This section considers the applicability of fault tolerance schemes
within LDP networks and considers issues that might lead to the
choice of one method or another. Many of the points raised below
should be viewed as implementation issues rather than specific
drawbacks of either solution.
7.1. General Applicability
The procedures described in [RFC3478] and [RFC3479] are intended to
cover two distinct scenarios. In Session Failure, the LDP peers at
the ends of a session remain active, but the session fails and is
restarted. Note that session failure does not imply failure of the
data channel even when using an in-band control channel. In Node
Failure, the session fails because one of the peers has been
restarted (or at least, the LDP component of the node has been
restarted). These two scenarios have different implications for the
ease of retention of LDP state within an individual LSR, and are
described in sections below.
These techniques are only applicable in LDP networks where at least
one LSR has the capability to retain LDP signaling state and the
associated forwarding state across LDP session failure and recovery.
In [RFC3478], the LSRs retaining state do not need to be adjacent to
the failed LSR or session.
If traffic is not to be impacted, both LSRs at the ends of an LDP
session must at least preserve forwarding state. Preserving LDP
state is not a requirement to preserve traffic.
[RFC3479] requires that the LSRs at both ends of the session
implement the procedures that it describes. Thus, either traffic is
preserved and recovery resynchronizes state, or no traffic is
preserved and the LSP fails.
Further, to use the procedures of [RFC3479] to recover state on a
session, both LSRs must have a mechanism for maintaining some session
state and a way of auditing the forwarding state and the
resynhcronized control state.
[RFC3478] is scoped to support preservation of traffic if both LSRs
implement the procedures that it describes. Additionally, it
functions if only one LSR on the failed session supports retention of
forwarding state, and implements the mechanisms in the document. In
this case, traffic will be impacted by the session failure, but the
forwarding state will be recovered on session recovery. Further, in
the event of simultaneous failures, [RFC3478] is capable of
relearning and redistributing state across multiple LSRs by combining
its mechanisms with the usual LDP message exchanges of [RFC3036].
7.2. Session Failure
In Session Failure, an LDP session between two peers fails and is
restarted. There is no restart of the LSRs at either end of the
session and LDP continues to function on those nodes.
In these cases, it is simple for LDP implementations to retain the
LDP state associated with the failed session and to associate the
state with the new session when it is established. Housekeeping may
be applied to determine that the failed session is not returning and
to release the old LDP state. Both [RFC3478] and [RFC3479] handle
this case.
Applicability of [RFC3478] and [RFC3479] to the Session Failure
scenario should be considered with respect to the availability of the
data plane.
In some cases the failure of the LDP session may be independent of
any failure of the physical (or virtual) link(s) between adjacent
peers; for example, it might represent a failure of the TCP/IP stack.
In these cases, the data plane is not impacted and both [RFC3478] and
[RFC3479] are applicable to preserve or restore LDP state.
LDP signaling may also operate out of band; that is, it may use
different links from the data plane. In this case, a failure of the
LDP session may be a result of a failure of the control channel, but
there is no implied failure of the data plane. For this scenario
[RFC3478] and [RFC3479] are both applicable to preserve or restore
LDP state.
In the case where the failure of the LDP session also implies the
failure of the data plane, it may be an implementation decision
whether LDP peers retain forwarding state, and for how long. In such
situations, if forwarding state is retained, and if the LDP session
is re-established, both [RFC3478] and [RFC3479] are applicable to
preserve or restore LDP state.
When the data plane has been disrupted an objective of a recovery
implementation might be to restore data traffic as quickly as
possible.
7.3. Controlled Session Failure
In some circumstances, the LSRs may know in advance that an LDP
session is going fail (e.g., perhaps a link is going to be taken out
of service).
[RFC3036] includes provision for controlled shutdown of a session.
[RFC3478] and [RFC3479] allow resynchronization of LDP state upon
re-establishment of the session.
[RFC3479] offers the facility to both checkpoint all LDP states
before the shut-down, and to quiesce the session so that no new state
changes are attempted between the checkpoint and the shut-down. This
means that on recovery, resynchronization is simple and fast.
[RFC3478] resynchronizes all state on recovery regardless of the
nature of the shut-down.
7.4. Node Failure
Node Failure describes events where a whole node is restarted or
where the component responsible for LDP signaling is restarted. Such
an event will be perceived by the LSR’s peers as session failure, but
the restarting node sees the restart as full re-initialization.
The basic requirement is that the forwarding state is retained,
otherwise the data plane will necessarily be interrupted. If
forwarding state is not retained, it may be relearned from the saved
control state in [RFC3479]. [RFC3478] does not utilize or expect a
saved control state. If a node restarts without preserved forwarding
state it informs its neighbors, which immediately delete all label-
FEC bindings previously received from the restarted node.
The ways to retain a forwarding and control state are numerous and
implementation specific. It is not the purpose of this document to
espouse one mechanism or another, nor even to suggest how this might
be done. If state has been preserved across the restart,
synchronization with peers can be carried out as though recovering
from Session Failure as in the previous section. Both [RFC3478] and
[RFC3479] support this case.
How much control state is retained is largely an implementation
choice, but [RFC3479] requires that at least small amount of per-
session control state be retained. [RFC3478] does not require or
expect control state to be retained.
It is also possible that the restarting LSR has not preserved any
state. In this case, [RFC3479] is of no help. [RFC3478] however,
allows the restarting LSR to relearn state from each adjacent peer
through the processes for resynchronizing after Session Failure.
Further, in the event of simultaneous failure of multiple adjacent
nodes, the nodes at the edge of the failure zone can recover state
from their active neighbors and distribute it to the other recovering
LSRs without any failed LSR having to have saved state.
7.5. Controlled Node Failure
In some cases (hardware repair, software upgrade, etc.), node failure
may be predictable. In these cases all sessions with peers may be
shutdown and existing state retention may be enhanced by special
actions.
[RFC3479] checkpointing and quiesce may be applied to all sessions so
that state is up-to-date.
As above, [RFC3478] does not require that state is retained by the
restarting node, but can utilize it if it is.
7.6. Speed of Recovery
Speed of recovery is impacted by the amount of signaling required.
If forwarding state is preserved on both LSRs on the failed session,
then the recovery time is constrained by the time to resynchronize
the state between the two LSRs.
[RFC3479] may resynchronize very quickly. In a stable network, this
resolves to a handshake of a checkpoint. At the most,
resynchronization involves this handshake plus an exchange of
messages to handle state changes since the checkpoint was taken.
Implementations that support only the periodic checkpointing subset
of [RFC3479] are more likely to have additional state to
resynchronize.
[RFC3478] must resynchronize state for all label mappings that have
been retained. At the same time, resources that have been retained
by a restarting upstream LSR but are not actually required, because
they have been released by the downstream LSR (perhaps because it was
in the process of releasing the state), they must be held for the
full resynchronization time to ensure that they are not needed.
The impact of recovery time will vary according to the use of the
network. Both [RFC3478] and [RFC3479] allow advertisement of new
labels while resynchronization is in progress. Issues to consider
are re-availability of falsely retained resources and conflict
between retained label mappings and newly advertised ones. This may
cause incorrect forwarding of data (since labels are advertised from
downstream), an LSR upstream of a failure may continue to forward
data for one FEC on an old label while the recovering downstream LSR
might re-assign that label to another FEC and advertise it. For this
reason, restarting LSRs may choose to not advertise new labels until
resynchronization with their peers has completed, or may decide to
use special techniques to cover the short period of overlap between
resynchronization and new LSP setup.
7.7. Scalability
Scalability is largely the same issue as speed of recovery and is
governed by the number of LSPs managed through the failed session(s).
Note that there are limits to how small the resynchronization time in
[RFC3478] may be made given the capabilities of the LSRs, the
throughput on the link between them, and the number of labels that
must be resynchronized.
Impact on normal operation should also be considered.
[RFC3479] requires acknowledgement of all messages. These
acknowledgements may be deferred as for checkpointing described in
section 4, or may be frequent. Although acknowledgements can be
piggy-backed on other state messages, an option for frequent
acknowledgement is to send a message solely for the purpose of
acknowledging a state change message. Such an implementation would
clearly be unwise in a busy network.
[RFC3478] has no impact on normal operations.
7.8. Rate of Change of LDP State
Some networks do not show a high degree of change over time, such as
those using targeted LDP sessions; others change the LDP forwarding
state frequently, perhaps reacting to changes in routing information
on LDP discovery sessions.
Rate of change of LDP state exchanged over an LDP session depends on
the application for which the LDP session is being used. LDP
sessions used for exchanging <FEC, label> bindings for establishing
hop by hop LSPs will typically exchange state reacting to IGP
changes. Such exchanges could be frequent. On the other hand, LDP
sessions established for exchanging MPLS Layer 2 VPN FECs will
typically exhibit a smaller rate of state exchange.
In [RFC3479], two options exist. The first uses a frequent (up to
per-message) acknowledgement system which is most likely to be
applicable in a more dynamic system where it is desirable to preserve
the maximum amount of state over a failure to reduce the level of
resynchronization required and to speed the recovery time.
The second option in [RFC3479] uses a less-frequent acknowledgement
scheme known as checkpointing. This is particularly suitable to
networks where changes are infrequent or bursty.
[RFC3478] resynchronizes all state on recovery regardless of the rate
of change of the network before the failure. This consideration is
thus not relevant to the choice of [RFC3478].
7.9. Label Distribution Modes
Both [RFC3478] and [RFC3479] are suitable for use with Downstream
Unsolicited label distribution.
[RFC3478] describes Downstream-On-Demand as an area for future study
and is therefore not applicable for a network in which this label
distribution mode is used. It is possible that future examination of
this issue will reveal that once a label has been distributed in
either distribution mode, it can be redistributed by [RFC3478] upon
session recovery.
[RFC3479] is suitable for use in a network that uses Downstream-On-
Demand label distribution.
In theory, and according to [RFC3036], even in networks configured to
utilize Downstream Unsolicited label distribution, there may be
occasions when the use of Downstream-On-Deman distribution is
desirable. The use of the Label Request message is not prohibited in
a Downstream Unsolicited label distribution LDP network.
Opinion varies as to whether there is a practical requirement for the
use of the Label Request message in a Downstream Unsolicited label
distribution LDP network. Current deployment experience suggests
that there is no requirement.
7.10. Implementation Complexity
Implementation complexity has consequences for the implementer and
also for the deployer since complex software is more error prone and
harder to manage.
[RFC3479] is a more complex solution than [RFC3478]. In particular,
[RFC3478] does not require any modification to the normal signaling
and processing of LDP state changing messages.
[RFC3479] implementations may be simplified by implementing only the
checkpointing subset of the functionality.
7.11. Implementation Robustness
In addition to the implication for robustness associated with
complexity of the solutions, consideration should be given to the
effects of state preservation on robustness.
If state has become incorrect for whatever reason, then state
preservation may retain incorrect state. In extreme cases, it may be
that the incorrect state is the cause of the failure in which case
preserving that state would be inappropriate.
When state is preserved, the precise amount that is retained is an
implementation issue. The basic requirement is that forwarding state
is retained (to preserve the data path) and that that state can be
accessed by the LDP software component.
In both solutions, if the forwarding state is incorrect and is