|
MPLS essentially
attempts to overlay connection-oriented concepts onto connectionless
technologies. While providing several advantages, in a number of
instances this approach reduces the overall flexibility of the IP
protocol. Some of the conclusions that led to the research into
multilayer routing, such as that routers are too slow or routing
tables becoming too large, have been weakened by the appearance of
fast and powerful gigabit routers.
The MPLS
framework and architecture define a base-level label swapping
technology. As discussed earlier, MPLS allows for traffic to be
switched under different circumstances (topology-driven, flow-driven
etc.), using different LDPs depending on the circumstances. While
this implies that MPLS is flexible, it is likely to be
applicable only within well-managed networks, where all components
are able to provide support for MPLS and the individual distribution
protocols in use.
While the label
stack concept provides benefits, the idea of having packets carry a
number of labels is likely to increase overheads, certainly in terms
of making the MPLS header larger.
With
topology-driven label assignment (where labels are allocated and
distributed without reference to the traffic), a full mesh of labels
will be established. The overhead of this approach is large relative
to the size of the network, and has the potential to use a vast
number of labels. This can be a large overhead in instances when
labels are allocated to routes where very little traffic is flowing.
The current MPLS
architecture and framework specifications have left the topic of
multicast as an area for further study.
In terms of the
provision of varying levels of QoS, MPLS poses a number of issues.
Label assignment
based on support for traffic flows will require a path to be put in
place the moment the flow is detected, therefore implying that there
will be some latency prior to a full path being in place. In this
instance, the overhead will increase in relation to the number of
flows being supported and the duration of the flows. Label
assignment in order to support short flows implies a large overhead.
When label distribution is included as part of a reservation
protocol (e.g., RSVP), the overheads and scalability of such a
protocol must also be considered.
The ordered and
independent control of labeled paths (described earlier) are said to
be compatible approaches to path setup. However, when they
interoperate the overall behavior can only be described as
independent because, to ensure QoS, ordered control must be used
entirely from ingress to egress node.
LDPs must work in
a reliable manner given that the loss of a control message in this
instance could cause a delay in the establishment of a label path.
This constitutes a serious impediment to the support of critical
applications. As mentioned earlier, the use of TCP with a number of
LDPs offers the necessary reliability. In the case of flow-based
label assignment and the use of RSVP, reliable transmission of the
LDP information is not guaranteed due to the use of UDP.
The ability of
MPLS to support a number of link-layer technologies provides a high
degree of flexibility. However, in terms of the provision of
connections with a level of associated QoS, mechanisms are required
to ensure that the QoS specified for an LSP is maintained by the
underlying link layer. This may not be possible in some instances
(e.g., with a standard Ethernet, DPT, etc) where firm guarantees
cannot be made (because of the inherent nature of the technology.
Where ATM technology is used with MPLS, in most instances the LDP
acts as the ATM signaling protocol. This implies that a low-level
control protocol is required which is able to configure connections
with defined levels of QoS. While work is progressing in this area
within the IETF GSMP Working Group, wide scale support for this type
of protocol by major switch/router vendors is not yet evident.
Note that QoS on
the LAN/MAN based on standard MAC protocols represents a major
challenge, not so much during the predictable processes, but in
sharing the connectionless transmission media with other
users/routers in a predictable and quantifiable way. As soon as the
critical traffic (e.g., voice/video) reaches the IP network, it must
compete with electronic mail traffic, database applications, and
file transfers. |
