You are on page 1of 6

An Analysis of Convergence Delay Caused by

Link Failures in Autonomous Systems


John Dwyer, Hetal Jasani
Department of Computer Science, Northern Kentucky University
Highland Heights, KY 41099
{dwyerj1, jasanih1}@nku.edu

Abstract–With internal networks growing larger,


between any of the routers, and it must reflect the real
dynamic routing protocols are becoming more prevalent state of the network [1].
in autonomous systems. Due to the importance of Dynamic routing protocols can be split into three
computers and network resources, many organizations distinct types: distance vector algorithms, link-state
implement redundant links to allow failover in the event
algorithms, and path vector protocol. Distance vector
of a link failure. When this failover occurs, the network
must converge before traffic will be able to pass to and
algorithms assign a cost to each link between nodes in a
from the network segment that incurred a link failure. network. They will then pick routes based on the
The time that is required for the network to converge is lowest cost path by computing the sum of the cost of
referred to as convergence delay. This paper will each link the network traffic must traverse to reach its
compare the convergence delay caused by link failures in destination. Link-state algorithms specify that each
both EIGRP and OSPF. node floods the entire network with information about
Index Terms– Convergence, Enhanced Integrated what other nodes it can connect to. This allows each
Gateway Routing Protocol, Open Shortest Path First, node to independently assemble this information into a
Dynamic Routing Protocol. network map that can then be used to determine least
cost paths by using a shortest path algorithm. Path
vector protocol is used for routing between autonomous
I. INTRODUCTION systems and is very similar to distance vector routing.
Path vector protocol uses a node in each autonomous
Routing is the process of selecting paths in a network system, called a speaker node, to advertise its routing
along which to send network traffic. Most routing is table to other neighboring speaker nodes. In this form
performed by routers which operate at layer 3 of the of dynamic routing, networks are advertised as
OSI model. These routers use either dynamic routing destination addresses and path descriptions to reach
protocols or static routes to make decisions on where to those destinations [1]. Some examples of dynamic
send network traffic to reach its destination [1]. Static routing protocols are:
routes are routes configured by an administrator which
tell a router how to communicate with a network that is • IGRP (Interior Gateway Routing Protocol)
not directly connected. Dynamic routing protocols • EIGRP (Enhanced Interior Gateway Routing
allow routers to advertise the networks with which they Protocol)
know how to communicate. This information is
received by other routers on the network and these other • OSPF (Open Shortest Path First)
routers utilize this data to populate their routing table • RIP (Routing Information Protocol)
with information to connect to each network within the
• IS-IS (Intermediate System to Intermediate
same autonomous system. This process of propagating
System) [1]
topology information throughout a network is called
convergence. When a set of routers has converged, the EIGRP is a Cisco proprietary distance vector routing
network topology information must not contradict protocol. It was designed to replace IGRP and has the
ability to communicate with older nodes that are still

978-1-4673-1375-9/12/$31.00 ©2012 IEEE


using IGRP even though it has been extensively the routing table using the successor as the next-hop
changed. EIGRP was optimized to minimize routing router and EIGRP goes into passive state. A feasible
instability incurred after topology changes and bases successor is a next-hop router for a destination that is
this and other routing optimizations on the Diffusing guaranteed to not be part of a routing loop. When a
Update Algorithm (DUAL), which guarantees loop-free successor route no longer meets the feasibility
operation and provides for fast convergence [2]. condition, the feasible successor with the lowest cost
EIGRP stores data in three tables. The neighbor table is will take over as the successor for the destination. If no
used to store information about directly connected feasible successor is available at this point, EIGRP will
routers. The topology table contains only an go into active state until a successor is found [2].
aggregation of the routing tables gathered from all
OSPF, a link-state routing protocol, is perhaps the
directly connected neighbors. It does not store an
most commonly-used interior gateway protocol in large
overview of the entire network topology as its name
network environments. OSPF has built-in optimizations
leads many to believe. The routing table is populated
to allow it to quickly recognize topology changes and
by information from the topology table and stores the
converge in an efficient manner. OSPF may be
actual routes to all destinations [2].
subdivided into areas to allow for easier administration
As a distance vector routing protocol, EIGRP assigns and optimized traffic flow [3]. As is typical with many
costs to each link between nodes in the network. routing protocols, OSPF implements route flooding to
EIGRP associates six different vector metrics with each multicast addresses over a broadcast network link.
route but only considers four of the metrics in OSPF gathers link state information about available
computing the composite metric. The six vector metrics routers by its implementation of route flooding. It uses
are bandwidth, load, delay, reliability, MTU, and hop this link state information to build a topology map of
count. In the composite equation, there are five K the network. This map allows OSPF to use methods
values – K1 through K5 [2]. The K values only act as based on a shortest path first algorithm to construct a
multipliers or modifiers in the composite equation and shortest path tree of the network [3]. Within an OSPF
do not reflect the values of the vector metrics. The K autonomous system, each individual router utilizes a
values can be changed to produce varying behavior in link-state database (LSDB) which contains a tree-image
the calculation. The following is the calculation used of the complete network topology. Periodically, the
by EIGRP to determine the composite metric. routers will flood identical copies of their LSDB to all
other routers within the OSPF autonomous system to
maintain consistency in the LSDB between all routers.
During the convergence process, OSPF routers will
create a two-way state neighbor relationship with
routers that are connected in the same broadcast domain
The default values for K1 and K3 are 1 and all other K
[4]. These neighbor relationships are stored in the
values are to be set to zero by default. This effectively
adjacency database on each router. Two routers can
reduces the above formula to (Bandwidth + Delay) *
form a neighbor relationship only if their directly
256. Failure to set the same K values on all routers in
connected interfaces are within the same OSPF area.
an EIGRP autonomous system may result in permanent
routing loops [2]. While hop count is not used in the For an OSPF autonomous system to converge and
composite metric equation, you may set a maximum function properly, a Designated Router (DR) must be
hop count in EIGRP. If a route passes the maximum elected. During each election, the router with the
hop count, it will not be considered a valid route. second highest priority will be appointed as the backup
designated router (BDR) [3]. If the DR fails, the BDR
During convergence, EIGRP receives routing tables
will take over as the DR and an election will be held to
from all neighboring routers and places this routing
determine the new BDR. An OSPF router can never
information into the topology table. EIGRP uses this
become the DR if its priority setting is set to 0 [4]. A
information to find successors for destination networks.
DR exists for the purpose of reducing network traffic by
For a next-hop router to be a successor, it must provide
having a single source provide routing updates to all
the least cost route to the destination and it must not be
other routers in the OSPF area. Other OSPF routers in
a part of a routing loop [2]. When a successor is
the same OSPF area will only create adjacencies with
determined, EIGRP places a route to the destination in
the DR and BDR. The DR will then send out an
updated network map to all nodes in the OSPF area and it is noted that CPU utilization can be a problem
using multicast addressing; thus eliminating the need with OSPF. Both of these papers differ from this
for nodes to update each other which reduces network analysis in the fact that the testing environment contains
traffic drastically [3]. physical hardware and this paper is testing a different
form of convergence.
OSPF is highly configurable when it comes to the
task of picking a shortest path. The basic routing metric IV. IMPLEMENTATION
in OSPF is path cost. As defined by the standard, path
This paper, requiring a hands-on lab environment,
cost should not equate to any specific value such as
has taken a considerable amount of time, planning and
speed, this allows the network administrator to choose a
work to set up so that the goals of this research project
metric important to the design of the network [4]. In
could be achieved and a conclusion could be reached.
practice, the path cost is almost always determined by
The testing environment for this work has been
the speed (bandwidth) of the interface addressing the
configured with the following hardware:
given route. As an example, Cisco uses a metric like
108/bandwidth for path cost where 108 is a value • 3 Cisco 2811 routers
configurable by the network designer [3].
o 2 physical interfaces
II. PROBLEM STATEMENT o 1 Loopback interface
With the importance that large enterprise companies o IOS 15.1.3T2
are placing on data and network connectivity, many of
• Cisco 2960 switch
these companies are implementing redundant
technology to provide failure tolerance on the network o 24 physical interfaces
level. When a link fails, it is important that the dynamic
• HP Proliant ML 330 G3 server
routing protocol recognizes that failure, and converges
upon a new topology to allow for the network segment o VMWare ESXi 3.5
to still be online. This paper tests the scenario of a link
o Windows Server 2003 Standard Edition
failure and quantifies the convergence delay caused by
(Guest OS)
the link failure event. As security is also a major
concern for large enterprises, this paper will also o Kiwi Syslog Daemon
implement key pair authentication with each protocol
• Lenovo x61 laptop
and quantify the convergence delay added by the
addition of authentication. o Windows XP Pro

III. RELATED WORK


To date, many studies exist on both OSPF and
EIGRP. In [5], EIGRP and OSPF are compared with
RIP by configuring each protocol in a network
simulation. The metrics compared in these tests were
network convergence activity, network convergence
duration, routing protocol traffic, CPU utilization,
network bandwidth utilization, throughput and queuing
delay. This paper closely relates to the tests for network
convergence duration, but it testing a different form of
convergence and is using physical devices rather than a
simulation.
EIGRP convergence was tested thoroughly in [6]. A
simulation was used to allow testing of multiple metrics
on the test network. The results collected were
convergence time, TCP performance and EIGRP Figure 1: Initial Network Design
overhead. In [7], OSPF scalability is studied in depth
The three Cisco routers were used as the testing The testing procedure consisted of a precise series of
devices. They were connected to each other through the steps. First the routers were configured with one of the
Cisco switch due to a limitation in the number on routing protocols being tested. At this point, the FA 0/1
physical interfaces. The network design, shown in interface of Router 2 was disconnected from the switch.
Figure 1, utilizes three networks on each device and The time required to recognize the link failure and
requires that multiple IP addresses be created on one of converge upon a consistent topology was monitored and
the interfaces. A fourth network, 192.168.66.0/24 was recorded using Kiwi Syslog Daemon. This test was
used for configuration as can be seen in Figure 2. The performed 10 times on each protocol with and without
HP server was used for data collection using Kiwi the added key pair authentication to allow for
Syslog Daemon and each router was configured to send verification of the accuracy of the results. Therefore,
debugging level syslog information to the server. there were four testing phases:
• EIGRP
• EIGRP with key pair authentication
• OSPF
• OSPF with key pair authentication
The results from these testing phases were then
compared with graphs to allow conclusions to be made
about the performance of each protocol in relation to the
Figure 2: Interface Configuration
results of the other tests. These results are shows in the
During initial testing, it was found that the original results and conclusion sections of this paper.
network design would not converge properly. With the
The first routing protocol to be tested was EIGRP.
way that OSPF and EIGRP operate, they place a focus
Each router was configured to be a part of EIGRP
on preventing routing loops. Due to the fact that there
autonomous system 101. The networks for each router
were two networks on the same physical interface, if a
were added to the autonomous system and traffic flow
packet came in the interface on one network and left the
was verified through the use of the ping command and
same interface on another network, it appeared to both
the traceroute command. At this point, both EIGRP
routing protocols that there was a routing loop. For this
autonomous system and notification debugging were
reason, the Loopback0 interface on each router had to
enabled to allow the syslog server to see topology
be configured with an IP address as can be seen in
changes and EIGRP traffic. Once the series of tests
Figure 2. This led to the new network design shown in
were finished, a key pair was created and applied to
Figure 3.
both interfaces. At this point, convergence delay was
tested again to record any performance difference
caused by adding a key pair. The EIGRP configuration
settings can be seen in Figure 4.

Figure 4: EIGRP Configuration

Figure 3: Modified Network Design


Figure 5: OSPF Configuration

Once EIGRP had been tested, it was time to


reconfigure the routers with OSPF. The EIGRP
configuration was deleted from the devices and OSPF
was configured with process 101. The networks on Figure 6: EIGRP Kiwi Syslog example
each router were added to OSPF area 101 and
convergence was verified using the ping command and
the traceroute command. OSPF adjacency and event VI. CONCLUSION
debugging were enabled to allow the OSPF events to be When comparing the results of the tests of EIGRP
captured on the syslog server. Once the testing was and OSPF without key authentication, the first
completed, an OSPF key pair was applied to each conclusion is that EIGRP vastly outperforms OSPF.
interface to allow for testing of convergence delay with This should be self evident as the average convergence
the addition of authentication. Figure 5 shows the delay for OSPF is 542ms and the average convergence
configuration applied for OSPF. delay for EIGRP is 63ms. While this is true, it is
important to remember that OSPF relies on the DR and
V. RESULTS BDR heavily during convergence. During this test, the
link between the DR and BDR is disconnected. This
For each testing phase, Kiwi Syslog Daemon was causes OSPF to perform a DR election even though
used to capture log traffic from the routers to determine neither router truly went offline. The OSPF standard
when the convergence process had completed. In Kiwi requires that after a DR election, there is a wait period
Syslog Daemon, time was measured from the alert that of no less than 100ms before LSAs are sent out from the
the Fast Ethernet interface was down to the time that the routers. The default wait period set by Cisco on all
last advertisements were sent. Since the routers were routers they distribute is 500ms. This accounts for the
set to send all debugging traffic to the syslog server, large delay caused by OSPF seen in Figure 7. The fact
interface up/down events are automatically logged in that EIGRP is Cisco proprietary is likely the reason that
the syslog. The EIGRP and OSPF debugging it is not favored over OSPF in large enterprise
commands allow the routers to send debugging environments.
information about the routing protocol they are
currently using. With EIGRP, the full process of The comparison of EIGRP with and without key pair
update, reply, and query operations can be seen in the authentication brought about rather surprising results.
syslog while the network is in the process of The results showed no or very little quantifiable
convergence. With OSPF, a DR election must occur in difference between the two over a period of 10 tests.
the network as the link between the DR and BDR was This same trend can also be seen in the comparison of
taken offline, thus the DR election and Link-State OSPF with ant without the addition of key pair
Advertisements (LSAs) can be seen in the syslog authentication. The results also showed no or very little
information. The fact that the complete convergence difference between the two series of tests. It should be
process could be recorded allowed for the total time to noted that with key pair authentication for both OSPF
be computed by subtracting the time of the first event and EIGRP, it took longer for the network to initially
from that of the last. An example of the results EIGRP converge at router startup, but convergence after a link
can be seen in Figure 6. failure did not appear to be affected. This is likely due
to the fact that no new interfaces were being added and
thus no new interfaced needed to be authenticated into VII. FUTURE WORK
the autonomous system. The results from these tests
This paper only tests two of the many interior
can be seen in Figures 8 and 9.
gateway protocols that can be used for dynamic routing.
Future tests could use RIP or IS-IS in comparison to the
600 tests shown here. There are many other results that
could be gathered, such as throughput during network
500 convergence, which would be important to enterprise
environments. Another interesting test would be to use
400
some form of packet capture software and capture the
300 Time (ms) EIGRP EIGRP and OSPF packets sent during convergence to
Time (ms) OSPF analyze the information that is sent between routers
200 during convergence. This could help to lower the
convergence time by analyzing the information and
100 finding steps that could be eliminated. These tests
could be performed on other models of Cisco routers or
0
1 2 3 4 5 6 7 8 9 10 on routers by other manufacturers as long as EIGRP is
Figure 7: EIGRP vs. OSPF results
not used as it is Cisco proprietary. EIGRP could be
tested with old IGRP nodes in the network to see if they
make a difference in convergence time. These same
80 tests could be run with a much larger network to see
70 how the results trend as the network scales to a larger
size. These tests could also be performed with older
60
version of IOS such as IOS 12.x or with any newer
50 Time (ms) EIGRP versions to check for differences in performance.
40
Time (ms) EIGRP w/
Security
REFERENCES
30
[1] Introduction; “IP Routing”;
20 http://www.cisco.com/en/US/tech/tk365/tsd_technology_suppo
10
rt_protocol_home.html
[2] Introduction to EIGRP; “IP Routing”;
0 http://www.cisco.com/en/US/tech/tk365/technologies_tech_not
1 2 3 4 5 6 7 8 9 10 11 e09186a0080093f07.shtml
[3] OSPF Design Guide; “IP Routing”;
Figure 8: EIGRP vs. EIGRP with key pair authentication results http://www.cisco.com/en/US/tech/tk365/technologies_white_pa
per09186a0080094e9e.shtml
[4] Moy, J; RFC 2328; “OSPF Version 2”;
550 http://tools.ietf.org/html/rfc2328
548 [5] Thorenoor, S.G.; Dynamic Routing Protocol Implementation
Decision between EIGRP, OSPF and RIP Based on Technical
546
Background Using OPNET Modeler; International Conference
544 on Computer and Network Technology. ICCNT 2010; pp.191 –
542 Time (ms) OSPF 195.
540 [6] Jaafar, T.M., Riley, G.F., Reddy, D., Blair, D.; Simulation-
538 Time (ms) OSPF w/ Based Routing Protocol Performance Analysis - A Case Study;
Security Winter Simulation Conference; 2006; pp. 2154-2161.
536 [7] Bi, J., Xiaoxiang Leng, Jianping Wu; OSPF Performance
534 Measurements and Scalability Study; International Conference
532 on Wireless and Optical Communications Networks; 2006.
530
1 2 3 4 5 6 7 8 9 10 11

Figure 9: OSPF vs. OSPF with key pair authentication results

You might also like