Network Working Group                                     J. Moy, Editor
Request for Comments: 1245                                 Proteon, Inc.
                                                               July 1991


                         OSPF protocol analysis



Status of this Memo

This memo provides information for the Internet community. It does not
specify any Internet standard. Distribution of this memo is unlimited.
Please send comments to ospf@trantor.umd.edu.

Abstract

This is the first of two reports on the OSPF protocol. These reports are
required by the IAB/ IESG in order for an Internet routing protocol to
advance to Draft Standard Status. OSPF is a TCP/IP routing protocol,
designed to be used internal to an Autonomous System (in other words,
OSPF is an Interior Gateway Protocol).

Version 1 of the OSPF protocol was published in RFC 1131. Since then
OSPF version 2 has been developed. Version 2 has been documented in RFC
1247.  The changes between version 1 and version 2 of the OSPF protocol
are explained in Appendix F of RFC 1247. It is OSPF Version 2 that is
the subject of this report.

This report attempts to summarize the key features of OSPF V2. It also
attempts to analyze how the protocol will perform and scale in the
Internet.

1.0  Introduction

This document addresses, for OSPF V2, the requirements set forth by the
IAB/IESG for an Internet routing protocol to advance to Draft Standard
state. This requirements are briefly summarized below. The remaining
sections of this report document how OSPF V2 satisfies these
requirements:

o  What are the key features and algorithms of the protocol?

o  How much link bandwidth, router memory and router CPU cycles does the
   protocol consume under normal conditions?

o  For these metrics, how does the usage scale as the routing
   environment grows? This should include topologies at least an order

   of magnitude larger than the current environment.

o  What are the limits of the protocol for these metrics? (I.e., when
   will the routing protocol break?)

o  For what environments is the protocol well suited, and for what is it
   not suitable?

1.1  Acknowledgments

The OSPF protocol has been developed by the OSPF Working Group of the
Internet Engineering Task Force.

2.0  Key features of the OSPF protocol

This section summarizes the key features of the OSPF protocol. OSPF is
an Internal gateway protocol; it is designed to be used internal to a
single Autonomous System. OSPF uses link-state or SPF-based technology
(as compared to the distance-vector or Bellman-Ford technology found in
routing protocols such as RIP). Individual link state advertisements
(LSAs) describe pieces of the OSPF routing domain (Autonomous System).
These LSAs are flooded throughout the routing domain, forming the link
state database. Each router has an identical link state database;
synchronization of link state databases is maintained via a reliable
flooding algorithm. From this link state database, each router builds a
routing table by calculating a shortest-path tree, with the root of the
tree being the calculating router itself. This calculation is commonly
referred to as the Dijkstra procedure.

Link state advertisements are small. Each advertisement describes a
small pieces of the OSPF routing domain, namely either: the neighborhood
of a single router, the neighborhood of a single transit network, a
single inter-area route (see below) or a single external route.

The other key features of the OSPF protocol are:

o  Adjacency bringup. Certain pairs of OSPF routers become "adjacent".
   As an adjacency is formed, the two routers synchronize their link
   state databases by exchanging database summaries in the form of OSPF
   Database Exchange packets. Adjacent routers then maintain syn-
   chronization of their link state databases through the reliable
   flooding algorithm. Routers connected by serial lines always become
   adjacent. On multi-access networks (e.g., ethernets or X.25 PDNs),
   all routers attached to the network become adjacent to both the
   Designated Router and the Backup Designated router.

o  Designated router. A Designated Router is elected on all multi-access
   networks (e.g., ethernets or X.25 PDNs). The network's Designated

   Router originates the network LSA describing the network's local
   environment. It also plays a special role in the flooding algorithm,
   since all routers on the network are synchronizing their link state
   databases by sending and receiving LSAs to/from the Designated Router
   during the flooding process.

o  Backup Designated Router. A Backup Designated Router is elected on
   multi-access networks to speed/ease the transition of Designated
   Routers when the current Designated Router disappears. In that event,
   the Backup DR takes over, and does not need to go through the
   adjacency bringup process on the LAN (since it already had done this
   in its Backup capacity). Also, even before the disappearance of the
   Designated Router is noticed, the Backup DR will enable the reliable
   flooding algorithm to proceed in the DR's absence.

o  Non-broadcast multi-access network support. OSPF treats these
   networks (e.g., X.25 PDNs) pretty much as if they were LANs (i.e., a
   DR is elected, and a network LSA is generated). Additional
   configuration information is needed however for routers attached to
   these network to initially find each other.

o  OSPF areas. OSPF allows the Autonomous Systems to be broken up into
   regions call areas.  This is useful for several reasons. First, it
   provides an extra level of routing protection: routing within an area
   is protected from all information external to the area. Second, by
   splitting an Autonomous System into areas the cost of the Dijkstra
   procedure (in terms of CPU cycles) is reduced.

o  Flexible import of external routing information. In OSPF, each
   external route is imported into the Autonomous System in a separate
   LSA. This reduces the amount of flooding traffic (since external
   routes change often, and you want to only flood the changes). It also
   enables partial routing table updates when only a single external
   route changes. OSPF external LSAs also provide the following
   features.  A forwarding address can be included in the external LSA,
   eliminating extra-hops at the edge of the Autonomous System. There
   are two levels of external metrics that can be specified, type 1 and
   type 2. Also, external routes can be tagged with a 32-bit number (the
   external route tag; commonly used as an AS number of the route's
   origin), simplifying external route management in a transit
   Autonomous System.

o  Four level routing hierarchy. OSPF has a four level routing
   hierarchy, or trust model: intra-area, inter-area, external type 1
   and external type 2 routes. This enables multiple levels of routing
   protection, and simplifies routing management in an Autonomous
   System.

o  Virtual links. By allowing the configuration of virtual links, OSPF
   removes topological restrictions on area layout in an Autonomous
   System.

o  Authentication of routing protocol exchanges. Every time an OSPF
   router receives a routing protocol packet, it authenticates the
   packet before processing it further.

o  Flexible routing metric. In OSPF, metric are assigned to outbound
   router interfaces. The cost of a path is then the sum of the path's
   component interfaces. The routing metric itself can be assigned by
   the system administrator to indicate any combination of network
   characteristics (e.g., delay, bandwidth, dollar cost, etc.).

o  Equal-cost multipath. When multiple best cost routes to a destination
   exist, OSPF finds them and they can be then used to load share
   traffic to the destination.

o  TOS-based routing. Separate sets of routes can be calculated for each
   IP type of service. For example, low delay traffic could be routed on
   one path, while high bandwidth traffic is routed on another. This is
   done by (optionally) assigning, to each outgoing router interface,
   one metric for each IP TOS.

o  Variable-length subnet support. OSPF includes support for variable-
   length subnet masks by carrying a network mask with each advertised
   destination.

o  Stub area support. To support routers having insufficient memory,
   areas can be configured as stubs. External LSAs (often making up the
   bulk of the Autonomous System) are not flooded into/throughout stub
   areas. Routing to external destinations in stub areas is based solely
   on default.

3.0  Cost of the protocol

This section attempts to analyze how the OSPF protocol will perform and
scale in the Internet. In this analysis, we will concentrate on the
following four areas:

o  Link bandwidth. In OSPF, a reliable flooding mechanism is used to
   ensure that router link state databases are remained synchronized.
   Individual components of the link state databases (the LSAs) are
   refreshed infrequently (every 30 minutes), at least in the absence of
   topological changes. Still, as the size of the database increases,
   the amount of link bandwidth used by the flooding procedure also
   increases.

o  Router memory. The size of an OSPF link state database can get quite
   large, especially in the presence of many external LSAs. This imposes
   requirements on the amount of router memory available.

o  CPU usage. In OSPF, this is dominated by the length of time it takes
   to run the shortest path calculation (Dijkstra procedure). This is a
   function of the number of routers in the OSPF system.

o  Role of the Designated Router. The Designated router receives and
   sends more packets on a multi-access networks than the other routers
   connected to the network. Also, there is some time involved in
   cutting over to a new Designated Router after the old one fails
   (especially when both the Backup Designated Router and the Designated
   Router fail at the same time). For this reason, it is possible that
   you may want to limit the number of routers connected to a single
   network.

The remaining section will analyze these areas, estimating how much
resources the OSPF protocol will consume, both now and in the future. To
aid in this analysis, the next section will present some data that have
been collected in actual OSPF field deployments.

3.1   Operational data

The OSPF protocol has been deployed in a number of places in the
Internet. For a summary of this deployment, see [1]. Some statistics
have been gathered from this operational experience, via local network
management facilities. Some of these statistics are presented in the
following table:

TABLE 1. Pertinent operational statistics


  Statistic                        BARRNet    NSI        OARnet
  ___________________________________________________________________
  Data gathering (duration)        99 hrs     277 hrs    28 hrs
  Dijkstra frequency               50 min     25 min     13 min
  External incremental frequency   1.2 min    .98 min    not gathered
  Database turnover                29.7 min   30.9 min   28.2 min
  LSAs per packet                  3.38       3.16       2.99
  Flooding retransmits             1.3%       1.4%       .7%


The first line in the above table show the length of time that
statistics were gathered on the three networks. A brief description of
the other statistics follows:

o  Dijkstra frequency. In OSPF, the Dijkstra calculation involves only
   those routers and transit networks belonging to the AS. The Dijkstra
   is run only when something in the system changes (like a serial line
   between two routers goes down). Note that in these operational
   systems, the Dijkstra process runs only infrequently (the most
   frequent being every 13 minutes).

o  External incremental frequency. In OSPF, when an external route
   changes only its entry in the routing table is recalculated. These
   are called external incremental updates. Note that these happen much
   more frequently than the Dijkstra procedure. (in other words,
   incremental updates are saving quite a bit of processor time).

o  Database turnover. In OSPF, link state advertisements are refreshed
   at a minimum of every 30 minutes. New advertisement instances are
   sent out more frequently when some part of the topology changes. The
   table shows that, even taking topological changes into account, on
   average an advertisement is updated close to only every 30 minutes.
   This statistic will be used in the link bandwidth calculations below.
   Note that NSI actually shows advertisements updated every 30.7 (> 30)
   minutes. This probably means that at one time earlier in the
   measurement period, NSI had a smaller link state database that it did
   at the end.

o  LSAs per packet. In OSPF, multiple LSAs can be included in either
   Link State Update or Link State Acknowledgment packets.The table
   shows that, on average, around 3 LSAs are carried in a single packet.
   This statistic is used when calculating the header overhead in the
   link bandwidth calculation below. This statistic was derived by
   diving the number of LSAs flooded by the number of (non-hello)
   multicasts sent.

o  Flooding retransmits. This counts both retransmission of LS Update
   packets and Link State Acknowledgment packets, as a percentage of the
   original multicast flooded packets. The table shows that flooding is
   working well, and that retransmits can be ignored in the link
   bandwidth calculation below.

3.2  Link bandwidth

In this section we attempt to calculate how much link bandwidth is
consumed by the OSPF flooding process. The amount of link bandwidth
consumed increases linearly with the number of advertisements present in
the OSPF database.We assume that the majority of advertisements in the
database will be AS external LSAs (operationally this is true, see [1]).

From the statistics presented in Section 3.1, any particular
advertisement is flooded (on average) every 30 minutes. In addition,

three advertisements fit in a single packet. (This packet could be
either a Link State Update packet or a Link State Acknowledgment packet;
in this analysis we select the Link State Update packet, which is the
larger). An AS external LSA is 36 bytes long.  Adding one third of a
packet header (IP header plus OSPF Update packet) yields 52 bytes.
Transmitting this amount of data every 30 minutes gives an average rate
of 23/100 bits/second.

If you want to limit your routing traffic to 5% of the link's total
bandwidth, you get the following maximums for database size:

TABLE 2. Database size as a function of link speed (5% utilization)


                 Speed    # external advertisements
                 _____________________________________
                 9.6 Kb   2087
                 56 Kb    12,174


Higher line speeds have not been included, because other factors will
then limit database size (like router memory) before line speed becomes
a factor. Note that in the above calculation, the size of the data link
header was not taken into account. Also, note that while the OSPF
database is likely to be mostly external LSAs, other LSAs have a size
also. As a ballpark estimate, router links and network links are
generally three times as large as an AS external link, with summary link
advertisements being the same size as external link LSAs.

OSPF consumes considerably less link bandwidth than RIP. This has been
shown experimentally in the NSI network. See Jeffrey Burgan's "NASA
Sciences Internet" report in [3].


3.3  Router memory

Memory requirements in OSPF are dominated by the size of the link state
database. As in the previous section, it is probably safe to assume that
most of the advertisements in the database are external LSAs. While an
external LSA is 36 bytes long, it is generally stored by an OSPF
implementation together with some support data. So a good estimate of
router memory consumed by an external LSA is probably 64 bytes. So a
database having 10,000 external LSAs will consume 640K bytes of router
memory. OSPF definitely requires more memory than RIP.

Using the Proteon P4200 implementation as an example, the P4200 has
2Mbytes of memory. This is shared between instruction, data and packet
buffer memory. The P4200 has enough memory to store 10, 000 external

LSAs, and still have enough packet buffer memory available to run a
reasonable number of interfaces.

Also, note that while the OSPF database is likely to be mostly external
LSAs, other LSAs have a size also. As a ballpark estimate, router links
and network links consume generally three times as much memory as an AS
external link, with summary link advertisements being the same size as
external link LSAs.


3.4  Router CPU

Assume that, as the size of the OSPF routing domain grows, the number of
interfaces per router stays bounded. Then the Dijkstra calculation is of
order (n * log (n)), where n is the number of routers in the routing
domain. (This is the complexity of the Dijkstra algorithm in a sparse
network). Of course, it is implementation specific as to how expensive
the Dijkstra really is.

We have no experimental numbers for the cost of the Dijkstra calculation
in a real OSPF implementation. However, Steve Deering presented results
for the Dijkstra calculation in the "MOSPF meeting report" in [3].
Steve's calculation was done on a DEC 5000 (10 mips processor), using
the Stanford internet as a model. His graphs are based on numbers of
networks, not number of routers. However, if we extrapolate that the
ratio of routers to networks remains the same, the time to run Dijkstra
for 200 routers in Steve's implementation was around 15 milliseconds.

This seems a reasonable cost, particularly when you notice that the
Dijkstra calculation is run very infrequently in operational
deployments. In the three networks presented in Section 3.1, Dijkstra
was run on average only every 13 to 50 minutes. Since the Dijkstra is
run so infrequently, it seems likely that OSPF overall consumes less CPU
than RIP (because of RIP's frequent updates, requiring routing table
lookups).

As another example, the routing algorithm in MILNET is SPF-based.
MILNET's current size is 230 nodes, and the routing calculation still
consumes less than 5% of the MILNET switches' processor bandwidth [4].
Because the routing algorithm in the MILNET adapts to network load, it
runs the Dijkstra process quite frequently (on the order of seconds as
compared to OSPF's minutes). However, it should be noted that the
routing algorithm in MILNET incrementally updates the SPF-tree, while
OSPF rebuilds it from scratch at each Dijkstra calculation

OSPF's Area capability provides a way to reduce Dijkstra overhead, if it
becomes a burden. The routing domain can be split into areas. The extent
of the Dijkstra calculation (and its complexity) is limited to a single

area at a time.


3.5  Role of Designated Router

This section explores the number of routers that can be attached to a
single network. As the number of routers attached to a network grows, so
does the amount of OSPF routing traffic seen on the network.  Some of
this is Hello traffic, which is generally multicast by each router every
10 seconds. This burden is borne by all routers attached to the network.
However, because of its special role in the flooding process, the
Designated router ends up sending more Link State Updates than the other
routers on the network. Also, the Designated Router receives Link State
Acknowledgments from all attached routers, while the other routers just
receive them from the DR. (Although it is important to note that the
rate of Link State Acknowledgments will generally be limited to one per
second from each router, because acknowledgments are generally delayed.)

So, if the amount of protocol traffic on the LAN becomes a limiting
factor, the limit is likely to be detected in the Designated Router
first. However, such a limit is not expected to be reached in practice.
The amount of routing protocol traffic generated by OSPF has been shown
to be small (see Section 3.2). Also, if need be OSPF's hello timers can
be configured to reduce the amount of protocol traffic on the network.
Note that more than 50 routers have been simulated attached to a single
LAN (see [1]). Also, in interoperability testing 13 routers have been
attached to a single ethernet with no problems encountered.

Another factor in the number of routers attached to a single network is
the cutover time when the Designated Router fails. OSPF has a Backup
Designated Router so that the cutover does not have to wait for the new
DR to synchronize (the adjacency bring-up process mentioned earlier)
with all the other routers on the LAN; as a Backup DR it had already
synchronized. However, in those rare cases when both DR and Backup DR
crash at the same time, the new DR will have to synchronize (via the
adjacency bring-up process) with all other routers before becoming
functional. Field experience show that this synchronization process
takes place in a timely fashion (see the OARnet report in [1]). However,
this may be an issue in systems that have many routers attached to a
single network.

In the unlikely event that the number of routers attached to a LAN
becomes a problem, either due to the amount of routing protocol traffic
or the cutover time, the LAN can be split into separate pieces (similar
to splitting up the AS into separate areas).

3.6  Summary

In summary, it seems like the most likely limitation to the size of an
OSPF system is available router memory. We have given as 10,000 as the
number of external LSAs that can be supported by the memory available in
one configuration of a particular implementation (the Proteon P4200).
Other implementations may vary; nowadays routers are being built with
more and more memory.  Note that 10,000 routes is considerably larger
than the largest field implementation (BARRNet; which at 1816 external
LSAs is still very large).

Note that there may be ways to reduce database size in a routing domain.
First, the domain can make use of default routing, reducing the number
of external routes that need to be imported. Secondly, an EGP can be
used that will transport its own information through the AS instead of
relying on the IGP (OSPF in this case) to do transfer the information
for it (the EGP). Thirdly, routers having insufficient memory may be
able to be assigned to stub areas (whose databases are drastically
smaller). Lastly, if the Internet went away from a flat address space
the amount of external information imported into an OSPF domain could be
reduced drastically.

While not as likely, there could be other issues that would limit the
size of an OSPF routing domain. If there are slow lines (like 9600
baud), the size of the database will be limited (see Section 3.2).
Dijkstra may get to be expensive when there are hundreds of routers in
the OSPF domain; although at this point the domain can be split into
areas. Finally, when there are many routers attached to a single
network, there may be undue burden imposed upon the Designated Router;
although at that point a LAN can be split into separate LANs.


4.0  Suitable environments

Suitable environments for the OSPF protocol range from large to small.
OSPF is particular suited for transit Autonomous Systems for the
following reasons. OSPF can accommodate a large number of external
routes. In OSPF the import of external information is very flexible,
having provisions for a forwarding address, two levels of external
metrics, and the ability to tag external routes with their AS number for
easy management. Also OSPF's ability to do partial updates when external
information changes is very useful on these networks.

OSPF is also suited for smaller, either stand alone or stub Autonomous
Systems, because of its wide array of features: fast convergence,
equal-cost-multipath, TOS routing, areas, etc.

5.0  Unsuitable environments

OSPF has a very limited ability to express policy. Basically, its only
policy mechanisms are in the establishment of a four level routing
hierarchy: intra-area, inter-area, type 1 and type 2 external routes.  A
system wanting more sophisticated policies would have to be split up
into separate ASes, running a policy-based EGP between them.


6.0  Reference Documents

The following documents have been referenced by this report:


[1] Moy, J., "Experience with the OSPF protocol", RFC 1246, July 1991.

[2] Moy, J., "OSPF Version 2", RFC 1247, July 1991.

[3] Corporation for National Research Initiatives, "Proceedings of the
    Eighteenth Internet Engineering Task Force", University of British
    Columbia, July 30-August 3, 1990.

Security Considerations

Security issues are not discussed in this memo.


Author's Address

John Moy
Proteon Inc.
2 Technology Drive
Westborough, MA 01581

Phone: (508) 898-2800
Email:  jmoy@proteon.com