N. A. Fikouras, K. El Malki, S. R. Cvetkovic Page 1 of 4

PERFORMANCE ANALYSIS OF MOBILE IP HANDOFFS N. A. Fikouras, K. El Malki, S. R. Cvetkovic

Department of Computer Science, University of Sheffield

Abstract The aim of this paper is to study the performance of

Mobile IP hand-offs. It is determined that the key issue is the movement detection method (i.e. Lazy Cell Switching, Prefix-Matching, Eager Cell Switching). For this paper all three movement detection methods are studied. The study involves single agent subnetworks and simple movement patterns, i.e. straight movement from one subnetwork to another with no comeback. It is assumed that the wireless technology is unable to pass link-layer information to the IP layer and that it also restricts the Mobile Node from contemporarily participating in multiple subnetworks when in their overlap area (e.g. IEEE802.11). The principle findings are two formulas that determine the average Mobile IP hand-off delay. These formulas are verified by real Mobile IP hand-off results. It is also shown that Eager Cell Switching based Mobile IP hand-offs complete faster than their Lazy Cell Switching counterparts. Finally, it is presented that TCP communications suffer more than UDP communications because of the exponential back-off and the path maximum transmission unit discovery algorithms of TCP.

I. Introduction TCP/IP was originally designed without any

considerations for mobile computing. Its greatest resistance to mobile computing resides in the routing service of the IP layer. Mobile IP (MIP) [1] is an enhancement to the IP protocol which allows transparent routing of traffic to Mobile Nodes in the Internet. Mobile IP introduces three new entities required to support the protocol: the Home Agent, the Foreign Agent and the Mobile Node but may alternatively operate without Foreign Agents by replacing them with a dynamic host configuration service (DHCP).

According to Mobile IP, the Mobile Node is required to

notify (register) the Home Agent of its current location. For every supported Mobile Node the Home Agent creates a record of its current location and acts as its proxy into the home network. Intercepted incoming traffic is redirected through packet encapsulation by the Home Agent to the Mobile Node’s registered location.

A typical scenario is presented in Figure 1. It illustrates two networks, called Subnetwork A and Subnetwork B and a Mobile Node (MN) that moves (hands-off) between subnetworks A and B. The Home Agent that lays at the home network is responsible for redirecting incoming traffic for the Mobile Node to its current foreign network. Every time a Mobile Node immigrates from a Mobile IP network to another it must register its current location with the Home Agent. This location is represented either by the IP address of a Mobility Agent (care-of address) or is an IP address temporarily allocated to the Mobile Node through a dynamic host configuration service (collocated care-of address).

In the absence of a communication between the link and

IP layers that would enable a Mobile Node to forecast an upcoming hand-off, Mobile IP has to rely on alternative means to detect movement. Specificly, every Mobility Agent is required to advertise its existence and all available care-of addresses in fixed period intervals. A Mobile Node entering a new subnetwork will detect its movement by receipt of an "unknown" advertisement. As such, a Mobile Node entering a network with multiple advertising agents may be prone to performing several hand-offs prior to re-establishing network connectivity. Several movement detection algorithms have been defined [1,5] that provide the shortest movement detection interval for various scenarios.

It is noted that in the duration of the movement detection

and registration processes the Mobile Node is experiencing network disruption. That is, prior to the registration, the Home Agent delivers all intercepted traffic to the Mobile Node’s old registered care-of address and as such it is not delivered to the Mobile Node.

The scenarios studied in this paper involve single

Foreign Agent subnetworks and a wireless technology that does not co-operate with higher layers. Moreover, it is unable to provide connectivity to multiple wireless segments when the Mobile Node resides in their overlap. An example of such a technology is the IEEE 802.11 standard for Wireless Area Networks [2]. All of the test cases presented in this study use the Transport Control Protocol (TCP) [3] as the transport-layer protocol for the transfer of data. This was selected because the majority of Internet traffic today consists of TCP traffic as it is the protocol of choice for most World Wide Web applications. Furthermore, as it will be shown, during Mobile IP handoffs TCP communications suffer more than their User Datagram Protocol (UDP) [3] counterparts. All of the conclusions drawn in this study hold for UDP communications as well. The movement patterns studied assume straight Mobile Node movement between two subnetworks with no comeback that leads to a single Mobile IP hand-off.

ISP network

Mobile Node

ForeignAgent

Router

Correspondent Node

ForeignAgent

HomeAgent

Subnetwork A/Foreign Network

Subnetwork B/Foreign Network

Home Network

Figure 1: Testbed Setup

N. A. Fikouras, K. El Malki, S. R. Cvetkovic Page 2 of 4

II. Mobile IP Hand-offs The only means by which a Mobile Node is made aware

of its location is through evaluation by the movement detection algorithm of periodically received agent advertisements. The duration of the movement detection process is critical as it determines the start of the registration process. The duration of the two consists the Mobile IP hand-off delay:

tMIP hand-off = tmovement_detection + tregistration (1)

where tmovement_detection is the movement detection delay and tregistration is the Mobile Node registration delay. The registration delay is the time taken for a Mobile Node registration request to reach the Home Agent and the reply to return from the Home Agent to the Mobile Node. As Home Agent processing of a single registration request is considered negligible it is concluded that tregistration is equivalent to the Mobile Node – Home Agent round trip delay. Although this is insignificant for a small testbed such as the one presented in Figure 1, it may prove to be a significant overhead in the Internet. This is a problem that may only be surpassed through the introduction of Mobility Agent Hierarchies [4], that reduces the tregistration delay by managing registrations locally within the Foreign Network. From this and (1) it is derived that the duration of a Mobile IP hand-off is equivalent to the movement detection delay:

tMIP hand-off = tmovement detection (2)

A. Movement Detection Algorithms

In this paper the three known movement detection

algorithms are considered, namely: Lazy Cell Switching (LCS) [5], Prefix Matching (PM) [5] and Eager Cell Switching (ECS) [5].

The LCS method has the main characteristic that after

the Mobile Node has registered a care-of address with the Home Agent it holds on to it until the agent's lifetime is expired. This occurs when the Mobile Node misses 3 successive agent advertisements indicating that it has left the subnetwork. In that case the Mobile Node must discover a new agent and register one of its care-of addresses with the Home Agent. This method can be applied to scenarios where the Mobile Node lays within the overlap area of several Mobile IP networks. If enabled by the wireless technology it will receive several different advertisements. Assuming that Mobile Node movement is rare, this algorithm chooses not to hand-off and ignore any newly discovered agents. In the absence of the capability of the wireless technology to contemporarily participate in several Mobile IP subnetworks the laziness of this algorithm results to an unnecessary delay.

The PM method has a similar functionality to the LCS

with the only difference being that it includes a prefix-length extension to agent advertisements. This extension contains the prefix length of the agent's IP address and it enables the Mobile Node to determine whether two agent advertisement have been received from the same or different subnetworks. Although, this can be useful in

multiple-agent subnetworks the study presented in this paper is restricted to single agent subnetworks. In this case, PM tends to operate like the LCS and therefore a dedicated study of the performance of Mobile IP hand-offs with PM support has been omitted.

The ECS method tends to function in a way opposite to

that of the LCS. It assumes that Mobile Nodes change their direction of movement very slowly. That is, if they are moving forward in one direction it is unlikely that they will stop and turn back. Hence, it is appropriate for Mobile Nodes to hand-off immediately upon discovering a new agent. This method tends to reduce the movement detection interval in comparison to the LCS method and therefore manages faster Mobile IP hand-offs. The study of Mobile IP hand-offs with ECS support presented in this paper assumes only single-agent subnetworks. Research on the behaviour of the ECS in multi-agent subnetworks is under way.

B. The Movement Detection Delay

x

y

lost advertisements

x

A

BTime

z

0

Figure 2: Agent Advertisements timing graph

Figure 2, illustrates the agent advertisements timing

graph of subnetworks A and B. The areas over and below the time line illustrate the broadcasting of agent advertisements for subnetworks A and B as indicated by the corresponding letters. Both advertisements occur at the fixed period interval x and the synchronisation difference between them is y. A Mobile Node may receive the advertisements of only one subnetwork and may hand-off at any point between two consecutive advertisements of the same agent. If a Mobile Node in subnetwork A hands-off prior to time point z and enters subnetwork B, it will have to wait between 0 and y until the first new agent advertisement is received. If the Mobile Node hands-off after time point z the same delay would be between y and x. In total a Mobile Node would have to wait for the new first agent advertisement between 0 and x. Considering that the point at which the Mobile Node hands-off is completely random then it can be derived that its waiting time is a continuous random variable that follows the uniform distribution. The mean for this distribution is given by:

(3),

where x is the periodic interval, or else 1/advertisement rate. According to the description of the LCS movement

detection algorithm, a Mobile Node has to miss three consecutive agent advertisements prior to registering a new care-of address. From Figure 2 and (3) it can be derived that the average LCS movement detection delay is:

2

x)f(E =

N. A. Fikouras, K. El Malki, S. R. Cvetkovic Page 3 of 4

(4) Similarly, the average ECS movement detection delay is equivalent to (3): (5) From (2) and (4) it is derived that the duration of a Mobile IP hand-off based on LCS is given by:

tMIP hand-off = tLCS (6) Similarly, from (2) and (5) it is derived that the duration of a Mobile IP hand-off based on ECS is given by:

tMIP hand-off = tECS (7) It will be shown in the following sections that equations (6) and (7) are verified by the results of Mobile IP hand-offs.

III. Experimental Setup This paper analyses the performance of Mobile IP hand-

offs. For this only Mobile IP handoffs with LCS and ECS support are studied as in single-agent subnetworks the PM algorithm performs the same as the LCS algorithm.

The testbed setup is illustrated in Figure 1. It consist of

Pentium PCs that are running RedHat 5.2 Linux. The Mobile IP software used is SunLabs Mobile IP [6] for RedHat 5.x Linux that was modified to support all three movement detection algorithms. Network connectivity is provided with Ethernet. The Mobile Node is attached through two Ethernet links to the Foreign Agents of subnetworks A and B. By manipulating these interfaces the Mobile Node is capable of performing hand-offs. The tool used for the generation and benchmarking of TCP traffic is the public-domain ttcp benchmarking software [7]. Network monitoring was managed with the public-domain tcpdump network analyser [8]. Each LCS and ECS Mobile IP hand-off was performed for 50 times. The results presented in this study represent an average case. In every trial a TCP communication was established between the Correspondent Node and the Mobile Node.

IV. Experimental Results

Figure 3 illustrates two Mobile IP handoffs with LCS

and ECS support respectively. The testbed topology for these tests is presented in Figure 1. In every experiment the transmission rate of agent advertisements is set to the value recommended by [1] of one advertisement per second. A Mobile Node will seize to use a care-of address after missing three successive corresponding agent advertisements, that is 3 seconds.

From (6) and (7) it is derived that for x=1 the LCS

movement detection delay is tLCS=2.5 seconds and the ECS delay is tECS=0.5 seconds.

From Figure 3 it can be seen that the experimental results

verify the theoretical results. The LCS and ECS delays are 2.5 and 0.5 seconds respectivelly. It can also be seen that TCP reacts to the network disruption with successive

timeouts and packet retransmissions. The functionality of the TCP retransmission scheme is to dynamically determine the value of the TCP timeout interval with every packet transmission. In the case of a timeout, TCP doubles that interval. This functionality is called TCP exponential

6.2e+06

6.25e+06

6.3e+06

6.35e+06

6.4e+06

6.45e+06

6.5e+06

6.55e+06

6.6e+06

6.65e+06

6.7e+06

6615260

6.8e+06

6.85e+06

6.9e+06

6.95e+06

7e+06

0.6

0.7

1.4 3

Time (in Seconds)

Slow Start

6.75e+06

1.6 sec. difference in TCP handoff recovery time

completed

6.22.8e+06

2.85e+06

2.9e+06

2.95e+06

3e+06

30727563.1e+06

3.15e+06

3.2e+06

3.25e+06

3.3e+06

3.35e+06

Packet loss due to Path MTU Discovery

3.05e+06

MIP handoffstarted

3rd timeout0.8 sec

1st timeout 5th timeout0.2 sec 3.2 sec

2nd timeout0.4 sec

4th timeout1.6 sec

LCS MIP handoffcompleted

TC

P P

acke

t Seq

uen

ce N

um

be

rs o

f LC

S t

est

A Packet Trace of TCP Traffic during a MIP Handoff with ECS support

ECS MIP handoff

2.5 sec. difference between ECS and LCS MIP handoffs

3rd timeout

4th timeout

0.8 sec

1.6 sec1st timeout

0.2 sec

2nd timeout0.4 sec

0.2

0

Slow start

Packet loss due to Path MTU Discovery

TC

P P

acke

t S

equ

ence

Nu

mb

ers

of

EC

S t

est

A Packet Trace of TCP Traffic during a MIP Handoff with LCS support

2.7

Figure 3: TCP packet traces for ECS and LCS Mobile IP hand-offs

2

x5tLCS =

2

xtLCS =

CorrespondentNode

TCP Seq #: 6610880 (DF)TCP Seq #: 6615260 (DF)

TCP Seq #: 6615260 (DF) 0.2

TCP Seq #: 6615260 (DF)

TCP Seq #: 6615260 (DF)

1.4

0.6

TCP Seq #: 6615260 (DF) 3.0

TCP Seq #: 6615260 (DF)TCP Seq #: 6625480TCP Seq #: 6628400

UDP Registration Request

Foreign Agent Mobile Node

1st timeout

3rd timeout

2nd timeout

4th timeout

5th timeout

FA advertisement -0.5

0

TIM

E

UDP Registration Reply

2.5

2.7

6.2

Movement Detection Delay (t movement detection)

TCP Handoff Recovery Time (t TCP Handoff Recovery)

Registration Delay (t registration)

TCP Path MTU Discovery Recovery Time (t TCP Path MTU Discovery recovery)

Figure 4: Frame Sequence of TCP during a Mobile IP handoff with LCS support

N. A. Fikouras, K. El Malki, S. R. Cvetkovic Page 4 of 4

backoff [3]. During a Mobile IP handoff TCP experiences several successive timeouts that increase the timeout interval beyond the duration of the handoff. As a result the TCP communication remains halted even after the completion of the Mobile IP handoff. This problem can not be surpassed without applying any changed to the TCP retransmission scheme. Furthermore, it can be seen that even after the handoff an extra packet is lost due to the Path Maximum Transmission Unit (MTU) Discovery algorithm [3]. This algorithm is responsible for discovering the Path MTU so as to avoid unnecessary packet fragmentation. Although, packet loss due to Path MTU discovery may not take place in every Mobile IP handoff, it needs to be considered that either packet size increase due to packet encapsulation or the change in the communication path may often lead to packet loss due to Path MTU Discovery. Finally, the TCP communication is resumed with the slow start congestion avoidance algorithm [3]. It is noted that in the case of UDP traffic the network disruption would have only been for the duration of the Mobile IP hand-off, hence for the duration of the movement detection delay. Figure 4 illustrates the TCP frame sequence during a Mobile IP handoff based on LCS. It is illustrated that the last Foreign Agent advertisment (old agent) is received half a second before the last succesfully received TCP packet. The latter is perceived as the time of the Mobile IP handoff and the start of the movement detection process. From that point on severe packet loss is witnessed. The Mobile Node is unable to receive and transmit traffic until registration is completed. In total it can be that the LCS hand-off lasts for 2.5 seconds. From Figure 5 it can be derived that the corresponding delay for the ECS hand-off is 0.5 seconds.

V. Conclusions

In this paper the performance of Mobile IP hand-offs has been evaluated. The study assumes single-agent subnetworks and inability of Mobile Nodes to contemporarily participate in multiple subnetworks when in their overlap area. It is identified that the duration of a Mobile IP hand-off is directly dependent on the movement detection algorithm. Under the assumed conditions it is presented that the Prefix-Matching algorithm has a similar performance to the Lazy Cell Switching, therefore only the Lazy and the Eager Cell Switching algorithms have been considered. It was shown that Mobile Node movement between two mobility agents follows the uniform distribution. Two generic formulas were derived for the average delays of the Lazy Cell Switching and Eager Cell Switching algorithms. The results of these formulas also matched the results of two average Mobile IP hand-offs based on these algorithms. Finally it was shown that during a Mobile IP handoff a TCP communication suffers severally due to the exponential back-off and Path MTU Discovery algorithms. It was theoretically and experimentally determined that in a testbed with small round trip delays and an agent advertisement rate of 1 per second, on average, an LCS Mobile IP hand-off would require 2.5 seconds to complete while its ECS counterpart would require 0.5 seconds.

References

[1] Perkins C., “ IP Mobility Support.” , RFC 2002, (Oct.

1996). [2] IEEE Std 802.11-1997. “Wireless LAN Medium Access

control (MAC) and Physical Layer (PHY) specifications.” , (1997)

[3] Stevens, R., “TCP/IP Illustrated Volume 1.” , Reading: Addison Wesley Professional Computing Series, (1995).

[4] El Malki K., Fikouras N., Cvetkovic S.,”Fast Handoff Method for Real-Time Traffic over Scaleable Mobile IP Networks” , draft-elmalki-mobileip-fast-handoffs-01.txt, Work in Progress, June 1999

[5] Perkins C., “Mobile IP, Design Principles and Practices.” , Addison Wesley, (1998).

[6] Sunlabs Mobile IP, [ftp://playground.sun.com/pub/mobile-ip/]

[7] “ ttcp: Test TCP.” U. S. Army Ballistics Research Lab, (Dec 1984).

[8] tcpdump: the protocol packet capture and dumper program, Lawrence Berkeley National Laboratory, [http://ee.lbl.gov/]

CorrespondentNode

TCP Seq #: 3072756 (DF)TCP Seq #: 3073292 (DF)

TCP Seq #: 3072756 (DF) 0.2

TCP Seq #: 3072756 (DF) 0.6

TCP Seq #: 3072756 (DF)1.4

TCP Seq #: 3072756 (DF)

3.0TCP Seq #: 3072756TCP Seq #: 3077044TCP Seq #: 3077580

UDP Registration Reply0.7

UDP Registration Request

Foreign Agent Mobile Node

1st timeout

2nd timeout

3rd timeout

4th timeout

TCP Seq #: 3073828 (DF)

TCP Seq #: 3077580 (DF)

TCP Seq #: 3072220 (DF)0

0.5

FA advertisement

TIM

E

New Agent advertisement

DroppedPackets

Movement Detection Delay (t movement detection)

TCP Handoff Recovery Time (t TCP Handoff Recovery)

Registration Delay (t registration)

TCP Path MTU Discovery Recovery Time (t TCP Path MTU Discovery recovery)

Figure 5: Frame Sequence of TCP during a Mobile IP handoff with ECS support