644 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 22, NO. 4, MAY 2004
Most production-mode wireless LANs today are config-
ured to operate in infrastructure mode, which poses unique
problems in reducing the handoff latency for mobile IP. This
paper presents the design, implementation, and evaluation of
a low-latency handoff mechanism that reduces the mobile IP
handoff latency of wireless LANs operating in infrastructure
mode by a factor of five, compared with standard mobile IP
Fig. 1. Generic wired and wireless network topology. Wireless LANs lie at the The remainder of this paper is organized as follows. Sec-
edge of the wired infrastructure. Wireless access points provide internet access tion II elaborates on the handoff problem in Wi-Fi networks run-
to mobile nodes. Mobile nodes freely roam across various subnets using mobile ning in infrastructure mode. Section III reviews related research
on handoff latency optimizations in different network setups.
Section IV describes the proposed scheme and its implemen-
corresponding foreign agent. The mobile IP software running tation. Section V describes how this optimization can be inte-
on the mobile node intercepts these advertisements and sends a grated with quality-of-service (QoS) mechanisms. Section VI
registration request to the newly discovered foreign agent. After presents the evaluation of the performance optimization of the
due authentication an IP-over-IP tunnel is established between implemented scheme. Section VII concludes this paper with a
the home agent and the foreign agent. From this point onwards, summary of the main results and an outline of future work.
the home agent acts as a proxy for the mobile node, intercepts all
packets intended for the mobile node and transmits them over
the tunnel. The foreign agent takes care of de-encapsulating the II. PROBLEM DESCRIPTION
packets coming from the tunnel and forward them to the mo-
bile node. Similarly, all packets that a mobile node transmits When an IEEE 802.11b-based wireless network is configured
are first received by the foreign agent and are tunneled over to in infrastructure mode, each mobile node is associated with the
its home agent, which further routes them to the true destina- access point of the wireless LAN segment in which it currently
tion. This process is known as bidirectional tunneling. This is resides. Whenever a mobile node moves from one segment to
the most preferred mode of routing to avoid various issues like another, its network interface card compares the difference be-
ingress and egress filtering at the routers and firewalls. When- tween the signal strengths of the two access points and initi-
ever a mobile node migrates to a new foreign subnet, it needs to ates a link-layer handoff. In all known IEEE 802.11b cards, this
bind with the foreign agent of the new foreign subnet, and needs link-layer handoff logic is built into the firmware of the network
to tear down the association with the foreign agent in the old interface card, and does not generate any interrupts to notify the
subnet. When the mobile node returns to the home subnet, stan- higher-layer software. That is the network interface card (NIC)
dard routing is resumed. The entire process of switching from enacts this change in access point unilaterally without notifying
one MA to another as a mobile node moves across adjacent wire- the higher layer system software. If the new network segment
belongs to the same IP subnet as the old segment, then to the
less IP subnets is called mobile IP handoff.
IP layer and above on the mobile node there is no change in
After a mobile node moves to a new wireless network
connectivity and the network applications continue without any
segment but before the associated mobile IP handoff completes,
disruptions. However, if the new network segment belongs to a
the mobile node is essentially cut off from the wired network.
different IP subnet, then the mobile node can no longer commu-
Therefore, it is critical to reduce this handoff latency. Handoff nicate with wired network nodes until a network layer handoff is
latency has different effects on different kinds of applications. completed. In this case, according to the mobile IP standard, the
For time-insensitive applications such as file transfers, web mobile node would eventually receive an advertisement from the
browsing, etc., short glitches in connectivity do not have any foreign agent associated with the new IP subnet, and then would
serious impact on their functionality. However, for time-critical reestablish its connectivity to the wired network by requesting
applications such as media conferencing, media streaming, a tunnel to be established between its home agent and the new
etc., even momentary discontinuity in connectivity could have foreign agent.
pronounced effects on the perceived quality. The link-layer handoff implementation in infrastruc-
The most dominant wireless LAN technology is based on the ture-mode wireless LANs poses two problems in reducing
Wi-Fi  alliance’s specification, which follows IEEE 802.11b mobile IP handoff latency. First, because link-layer handoff
standard . These networks can be configured in two modes, is transparent to software, it is not possible for mobile IP
namely, infrastructure (or access point) mode, and ad hoc (or code to detect its occurrence immediately. Second, because a
peer to peer) mode. The ad hoc mode configuration is used mobile node can only receive packets from one access point at
when the communication is mainly amongst mobile nodes and a time in infrastructure mode, mobile IP code can only receive
there is no need for a dedicated access point to relay packets advertisements from the current foreign agent, but not those
among them. In contrast, the infrastructure mode configuration from neighboring foreign agents. As a result, even if a mobile
is commonly used when the mobile nodes need connectivity to node can detect the link-layer handoff immediately, without
the wired networks. This connectivity is established through ac- advertisements from the new foreign agent it still does not know
cess points that serve as bridges between wired and wireless net- which foreign agent to contact to facilitate the network-layer
works. handoff process.
SHARMA et al.: LOW-LATENCY MOBILE IP HANDOFF FOR INFRASTRUCTURE-MODE WIRELESS LANS 645
For mobile IP software, link-layer handoff in infrastruc- Tan et al.  proposed an improvement over the proposal
ture-mode wireless LANs is both hard and forward. A handoff by Seshan et al.  with a domain foreign agent that multicasts
is hard if a mobile node can communicate with exactly one data across multiple cells. Although this proposal assumes that
access point before and after a handoff, and it is forward if the the handoff is a forward handoff, it implicitly assumes that the
mobile node cannot communicate with the old agent during the handoff is soft as well. Thus, although the scheme is immedi-
handoff and has to carry out the handoff by reestablishing a ately applicable to the networks using mobile IP, the suggested
connection with the new foreign agent . Both properties are handoff performance may no longer be possible when link-layer
detrimental to fast mobile IP handoff. Because the link-layer handoff is hard.
handoff is hard, a mobile node cannot obtain the new foreign Snoeren and Balakrishnan  propose an end-to-end ap-
agent’s address before the handoff. Because the link-layer proach to host mobility which uses the domain name system
handoff is forward, a mobile node cannot contact the old agent (DNS) to track the location of the mobile nodes and, hence,
during the handoff and request for additional buffering to manage the mobility. The proposal is to provide migration se-
minimize the data loss during the handoff period. mantics for TCP connections in the protocol stack on the end
hosts. This approach tries to avoid the complexity and cost of the
transparent solution provided by mobile IP. The argument given
III. RELATED WORK is that not all the applications require the level of transparency
and generality provided by mobile IP. This approach exploits
The handoff research finds its roots in the cellular Global the ability of DNS to support secure dynamic updates to the host
System for Mobile Communication (GSM) networking. name to address mapping. For applications like web browsers,
Research in handoff optimization is mainly driven by the etc. the network transparency is not highly important. These ap-
needs of time-critical QoS enabled applications such as media plications are resilient to connection failures. For applications
conferencing and voice-over-IP (VoIP) to have a lossless and that require transparent TCP connectivity, a TCP connection is
low-latency handoff mechanism. Pollini  gives an overview migrated to a new address that is obtained after handoff using
of research on handoff performance and control. He also the proposed connection migration semantics. The advantage of
discusses the trends in handoff research specific to wireless this approach is in cutting down on the extra resources that are
telecommunication networks before the advent of wireless IP required to support the host mobility. The latency of connection
networks. migration is on the order of the round-trip time between the end
The handoff problem in IP networks can be considered as a hosts. But the overall latency to complete the handoff includes
special instance of the broader mobility management problem, the latency of the link-layer handoff and the delay required to
which arises because of the change in a mobile node’s point of obtain the new address from the new point of attachment. As a
attachment to the network as it moves around. There are several result, the entire handoff may require up to multiple seconds.
mobility management solutions emphasizing on enabling Shim and Gitlin  propose a fast handoff mechanism termed
network routing to and from alternate points of attachment for NeighborCasting. This mechanism is based on utilizing and
a mobile node. Design issues involved in mobility management sometimes even wasting the wired bandwidth in order to min-
are – changes or enhancements to the existing protocols, the imize the number of message exchanges between the wireless
resource requirement to support mobility, signaling involved interfaces. This proposal requires modifications to mobile IP to
during handoff, data loss incurred during handoff, and the support the discovery of neighboring MAs. During the handoff,
mechanisms to overcome this data loss, etc. Also, every each mobile node informs the new MA about the previous MA,
mobility management solution makes specific assumptions helping it to build a map of its neighbors. This proposal as-
regarding whether the handoff is forward or backward and soft sumes the availability of a notification mechanism from link
or hard, and regarding the capabilities from the underlying layer to the mobile IP layer about the impending handoff. In
hardware. Different proposed mobility management solutions stable state, each MA utilizes the neighbor map to forward the
differ in terms of approaches to address these issues and of data to all the neighboring MAs whenever a notification from
their assumptions. the link layer is received. In effect, this proposal changes the for-
One of the prominent works is the handoff scheme proposed ward link-layer handoff in infrastructure Wi-Fi networks into a
by Seshan et al.  and implemented as a part of the Daedalus backward handoff by making use of a notification mechanism,
project. This scheme uses IP multicast and buffering to elim- which certainly is a desirable feature for the mobile IP software.
inate data loss and reduce the handoff delay. However, this Baker et al.  suggest a mobility solution as part of the
scheme is based on anticipating a handoff using the information MosquitoNet project. The inherent assumption in this setup is
about the signal strengths of the communication between the that the mobility management solution that is deployed should
base stations and the mobile nodes. Moreover, the handoffs are not expect any additional infrastructure support from any for-
assumed to be initiated by software on the mobile nodes, which eign network. The minimal support that is expected is alloca-
is not the case for Wi-Fi networks, where link-layer handoff is tion of an IP address. The setup is based on conventional mo-
carried out by the network interface cards without notifying the bile IP setup with a difference that the mobile host acts as a
software. Also, this scheme does not work directly with mobile foreign agent when the host is in some foreign subnet. The for-
IP, which is currently the only standard for handling host eign agent module on the mobile node acquires the newly allo-
mobility. However, this scheme gives very low handoff latency cated address and the mobile node is connected to the network
(8–15 ms) and, thus, supports handoff without any data loss. through the foreign agent module with which it communicates
646 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 22, NO. 4, MAY 2004
using a loop-back virtual interface. This solution is an elegant to utilize L2 triggers to setup a bi-directional tunnel between
solution which eliminates the need of mobile IP support at the the old agent and the new agent that allows the mobile nodes
foreign network. Although this solution cuts down on the la- to keep using the old agent even when the mobile nodes move
tency that is incurred due to the signaling between the mobile to a new subnet. This requires some modifications to the mo-
node and the foreign agent there is a new delay factor intro- bile IP specifications. A similar approach of L2 triggers is used
duced because of the additional requirement of acquiring an ad- by proactive handoff scheme proposed by Calhoun et al. .
dress from the new subnet. Usually, this IP address allocation This scheme relies on foreign agent assistance to perform the
is done by some servers like dynamic host configuration pro- handoff, once a link layer handoff is detected, unlike the mobile
tocol (DHCP) servers, etc. The handoff latency of this scheme node driven handoff in mobile IP.
is a combination of the link-layer handoff, the address acquisi- Koodli  discusses handoff initiation in IPV6 networks in
tion delay, and the mobility management latency, and is, thus, the context of nonanticipated link handoffs. The proposal is to
expected to be no smaller than the generic mobile IP handoff send a fast binding update to the previous access router [(PAR)
latency,which does not involve new address allocation. home agent or old foreign agent] as soon a connectivity with
Cáceres and Padmanabhan  describe a fast and scalable the new access router [(NAR) new foreign agent] is established.
handoff mechanism for wireless internetworks (FSHWI). They This binding update is used to setup a tunnel between the access
also propose a hierarchical mobility management extension to routers. This does not impact the handoff latency in a significant
mobile IP, which is based on the observation that when a user way since the main component of handoff latency comes from
is visiting a foreign administrative domain, there is little need discovery of NAR.
to expose motion within that domain to the home agent or to The handoff schemes proposed in previous research almost
correspondent hosts in other domains. This scheme requires the invariably assume that mobile nodes can anticipate link-layer
base stations to be specialized routers which reside at the lowest handoffs and maintain connectivity with the old as well as the
level of the hierarchy. The handoffs in this scenario are soft new MAs. That is, these optimizations are applicable mainly to
handoffs. A similar hierarchical mobility management is pro- soft and backward handoffs. While such assumptions are valid
posed by Gustafsson et al. , where a foreign administrative for wireless LANs operating in the ad hoc or peer-to-peer mode,
domain is covered by several foreign agents organized hierar- they do not hold for wireless LANs running in the infrastruc-
chically. The local micromobility of mobile nodes is masked ture mode, where link-layer handoffs are hard and forward. In
from home network exposing only a top level gateway foreign infrastructure mode, a mobile node is associated with only one
agent to the home network. Many mobile IP implementations access point at a time. Although the IEEE 802.11b NIC on a
such as dynamics  support this kind of hierarchical mo- mobile node may be able to access the signal strength informa-
bility management. This approach improves the handoff per- tion for all neighboring access points, such information is not
formance by avoiding tunnel setup over the internet whenever available to mobility management software. As a result, pre-
local handoffs occur. This scheme is complementary to the pro- vious proposals on fast handoff that rely on the ability to an-
posed scheme since it optimizes the tunnel setup time as op- ticipate an imminent link-layer handoff through signal strength
posed to link handoff detection and network handoff expedi- comparison cannot be applied to IEEE 802.11 networks oper-
ation. Cáceres and Iftode  analyze the effects of handoffs on ating in infrastructure mode. Moreover, since the mobile nodes
TCP traffic in wireless environment and propose an end-to-end cannot receive packets transmitted from other access points, it
fast retransmission scheme to avoid the problems that TCP en- is not possible to receive multiple foreign agent advertisements
counters in wireless environments because of misinterpretation and maintain a list of neighboring foreign agents a priori for
of data loss as congestion. future use. That is, a mobile node needs to identify the foreign
Forsberg  discusses issues in communication availability agent of a new cell before it can switch to the new cell.
under signal quality-based handoff management, while using An overall comparison of all the above work is summarized
mobile IP software. The discussion assumes the ad hoc mode of in Table I. The unique contribution of the low-latency handoff
operation for the wireless LANs. The advantage of ad hoc mode scheme proposed in this paper is that it is designed specifically
is that the link-layer handoffs are soft and, thus, it is possible for wireless LANs running in the infrastructure mode. To the
to monitor the signal quality of multiple-access points simul- best of the authors’ knowledge, the scheme proposed in this
taneously. The work also supports policy-based handoff man- work achieves the lowest mobile IP handoff latency on such
agement in HUT dynamics mobile IP software implementation networks. Since the optimization is targeted to a specific but
, but most production mode wireless LANs are set up to highly relevant network setup the above comparison with other
operate in the infrastructure mode for performance and man- schemes is necessarily limited to only implemented systems or
ageability reasons. Thus, this solution cannot be immediately very closely related efforts.
applied to general wireless LANs.
Malki et al.  propose to reduce the delay and the data
IV. HANDOFF LATENCY OPTIMIZATION
loss during handoff by tackling it in two ways. They propose a
PREREGISTRATION handoff method by enabling the mobile Mobile IP provides macromobility and not micromobility. But
nodes to communicate with the (possible) new foreign agent, it is not difficult to enhance the mobility software to support
while still connected to the old agent when a handoff is antic- micromobility. The policy and signal quality-based handoffs in
ipated. This anticipation can come in the form of L2 triggers. HUT dynamics mobile IP ,  have shown this to be the
They also propose a POST-REGISTRATION handoff method case in ad hoc mode. In order to have a smooth and a low-latency
SHARMA et al.: LOW-LATENCY MOBILE IP HANDOFF FOR INFRASTRUCTURE-MODE WIRELESS LANS 647
COMPARISON OF DIFFERENT MOBILITY MANAGEMENT SCHEMES. THE PROPOSED HANDOFF SCHEME COMPLETELY COMPLIES WITH THE
MOBILE IP STANDARD, IS NOT DEPENDENT ON THE ABILITY OF ANTICIPATING LINK-LAYER HANDOFF AND IS DESIGNED FOR (HARD)
INFRASTRUCTURE-MODE WIRELESS LANS WHERE SIMULTANEOUS COMMUNICATION WITH MULTIPLE BASE STATIONS IS NOT POSSIBLE
handoff, it is essential for the network interface cards to notify handoff at a higher frequency than once per several millisec-
the higher protocol layers about the the impending handoffs, onds. We, therefore, decided to piggyback this software probe
and provide a facility to maintain the channel connectivity with with the system timer service routine, which is invoked once
both old and new access points. Unfortunately, existing IEEE every 10 ms. The cost of each probe for the access point identi-
802.11b network interface cards support neither. fication takes only a few CPU clock cycles and, therefore, does
To reduce mobile IP handoff latency, let us first identify the not add any noticeable performance overhead to the timer pro-
individual delay components involved. A mobile IP handoff pe- cessing routine.
riod for an infrastructure-mode wireless LAN can be divided While the Wi-Fi cards support access point ID probing, the
into four distinct subperiods: value that is returned in response to a probe is not always re-
1) time between when a link-layer handoff takes place and liable. Sometimes the returned value is obviously ill-formed
when it is detected by the mobile IP software; and sometimes it appears perfectly proper but turns out to be
2) time from link-layer handoff detection to the reception of incorrect. Those proper-looking returned values lead to false
first mobile IP advertisement from the new foreign agent; positives and, thus, trigger two successive redundant mobile IP
3) time required for a mobile node to register with the new handoffs when in fact none should be triggered. Our experiences
foreign agent after receiving the first advertisement; show that the number of false positives encountered increases
4) time between request sent to and response returned from with the probe frequency.
the new foreign agent. After analyzing the tradeoff between the probe frequency and
To minimize the overall handoff latency, each of these delay the reliability of return value, it was observed that one can main-
components needs to be reduced as much as possible. tain a high probe frequency without compromising the relia-
bility of probed value by keeping a history of returned values
A. Link-Layer Handoff Detection from consecutive probes. More specifically, when two or three
Although the Wi-Fi network interface cards do not provide consecutive probes return the same value, it is very unlikely that
any mechanism to notify the software when link-layer handoff the returned value is a false positive. Given the above analysis,
takes place, there is a hardware control functionality that allows the algorithm for detecting a link-layer handoff is given in Fig. 2.
software to probe for the identity of the access point with which Since all the operations involved in this algorithm are simple
a Wi-Fi card is currently associated. The wireless LAN medium read, compare and increment operations, the additional over-
accfess control (MAC) address of the access point is the iden- head added to a standard timer interrupt service routine is very
tity reported by the Wi-Fi card. With this mechanism, software small, around 60 000 clock cycles.
can detect change in the access point (and, thus, the occurrence
of a link-layer handoff) by comparing results from consecutive B. Foreign Agent Discovery
probes. A mismatch from such comparison can then be used to Once a mobile node successfully determines that it has moved
trigger a mobile IP handoff. into a new wireless network segment, the next critical step is to
These software probes themselves need to be triggered by cer- discover the MA serving the new segment, which could be either
tain events. The probing frequency determines how fast the oc- a home agent or a foreign agent. Without loss of generality, we
currence of a link-layer handoff can be detected. Because our can safely assume that each new wireless segment belongs to a
empirical observations show that each link-layer handoff takes a different IP segment and, hence, each new wireless segment has
few milliseconds, it does not make sense to probe for link-layer a different designated MA. The mobile IP specification specifies
648 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 22, NO. 4, MAY 2004
the average time to send out a solicitation would be 1.5 s after
the link-layer handoff. Therefore, with agent solicitation, the
network-layer handoff time is still on the order of multiple
Because the mobile IP specification permits the use of any
other appropriate link-layer mechanism to discover new MA,
we propose a caching and replaying approach to MA solici-
tation using only unicast. In this scheme, each subnet is aug-
mented with a host on the wired segment acting as a caching
agent which keeps the most recent MA advertisement cached in
its memory. In addition, this caching agent periodically sends
dummy Ethernet frames on the network with the source ad-
dress faked as that of a well-known MAC address, say a:b:c:d:e:f.
This address is made known to all the mobile nodes and all the
Fig. 2. Algorithm for detecting link-layer handoff. Threshold is usually set to caching agents. These dummy Ethernet frames are required to
2 or 3 in practice.
trick the switches and bridges, including the access point, to
relay all packets with a:b:c:d:e:f as the destination address to
two mechanisms for MA discovery, namely, MA advertisement the caching agent. This is possible because switches and bridges
and MA solicitation. use source learning to construct the link-layer forwarding ta-
MAs are required to periodically broadcast Advertisement bles. Whenever a mobile node detects a link-layer handoff, it
messages on the associated subnet to announce their presence. can send a unicast solicitation packet with destination address
When a mobile node enters a new wireless subnet, it can inter- set to a:b:c:d:e:f to request for MA advertisement. This solic-
cept these messages and identify the MA serving that subnet. itation would eventually be picked up by the caching agent.
The latency of this MA discovery mechanism is dependent on The caching agent, upon receiving a solicitation, can replay the
the frequency of the advertisements. Since these messages are cached advertisement to the mobile node, which then uses this
broadcast messages, it is advisable not to send advertisements advertisement as an indication of presence of a new agent.
at a high frequency to avoid performance problems. The mobile The above unicast-based MA solicitation scheme can help re-
IP specification enforces a maximum limit of one advertisement duce the MA advertisement frequency without increasing the
every second. Therefore, the latency of handoffs based on MA handoff latency or incurring extra broadcast traffic on the wire-
advertisements is on the order of seconds, which is clearly not less LANs. Further, this scheme is independent of the mobile IP
acceptable for time-critical applications. Alternatively, mobile specification since the solicitation is for replay only and not for
nodes, can send agent solicitation requests. These solicitation the advertisements. Since this scheme is transparent to the mo-
are broadcast messages. On receiving these solicitation requests, bile IP software on the mobile nodes, even the implementations
the MAs reply by sending unicast advertisement messages to the that are completely independent and unaware of the link-layer
mobile nodes. details can take advantage of this scheme.
There are three disadvantages with this approach. First, Advertisement soliciting, caching and replaying can all be
broadcast messages on infrastructure Wi-Fi networks occupy done in a way completely transparent to mobile IP. The net re-
two packet transmission slots, once upstream to the access sult is a transparent and independent MA discovery process. The
point, and next downstream to all the wireless nodes from the caching agent need not be an independent host on the wired
access point. The downstream transmission of the solicitation network. The caching and replay module may be run on the
request is redundant since the MAs typically reside on wired MA machine itself. The only requirement is that the host par-
nodes. Second, according to IEEE 802.11b  specifica- ticipating in caching and replaying should set its network inter-
tions, broadcast packets are categorized as asynchronous data face in the promiscuous mode so that it can receive the Ethernet
services, which may experience lower QoS compared with frames addressed to the well-known MAC address.
other packets. Further, the specifications allow reordering of
the broadcast packet transmissions to improve the reliability C. Mobile Node Registration
of other packets. This limits the performance of any broad-
cast-based transmissions over wireless segments. Third, the To avoid the ping pong effect of oscillating between adja-
mobile IP standard states that the mobile nodes should carry cent wireless IP subnets, mobile nodes usually do not imme-
out solicitation only in the absence of agent advertisements diately register with the newly discovered MA if the old agent’s
and when the agent address could not be determined through advertisement has not expired. While this policy is beneficial
any appropriate link-layer mechanism. The absence of an agent when the network is operating in ad hoc mode, on infrastructure
advertisement can be determined only when the lifetime of mode networks it increases the handoff latency by the lifetime
all the previously received advertisements has expired. The of the advertisements. Because of advertisement caching and re-
standard suggests that the lifetime of an advertisement should playing, the MA advertisement frequency can be reduced, which
be three times that of the advertisement interval. Thus, the further increase the advertisement’s life time. Moreover, in the
lifetime of any advertisement would be at least 3 s. If a mobile infrastructure mode of operation the ping-pong effect is avoided
node employs the solicitation scheme to discover the MAs then at the link-layer handoff itself. The intelligence to avoid this is
SHARMA et al.: LOW-LATENCY MOBILE IP HANDOFF FOR INFRASTRUCTURE-MODE WIRELESS LANS 649
Fig. 4. Integration of fast handoff scheme with wireless Rether. The WRS acts
as the caching agent. The arrows indicate the messages exchanged between
different nodes. The initial solicitation requests are piggybacked over Rether
messages. Subsequent messages are treated as urgent messages by Rether.
Fig. 3. Message exchanges required for mobile IP optimization. After Wireless Rether  is one such bandwidth guarantee
detecting the link-layer handoff the mobile node contacts the caching agent mechanism which provides QoS support on wireless networks
using well-known MAC address. Other exchanges follow the standard mobile
IP protocol. to legacy applications by making use of port-based signaling
policies. Wireless Rether employs a centralized software token
embedded in the network interface card firmware rendering any passing protocol to avoid collisions on the wireless channel.
higher level solution ineffective. The token is distributed across wireless nodes in an exclusive
Many mobile IP software implementations like HUT dy- fashion by a central node called wireless Rether server (WRS).
namics system  provide policy-based handoff schemes , All the wireless nodes are Wireless Rether clients (WRCs)
where one can specify a high-level system policy for handoff which register with the WRS and request bandwidth guarantees.
management on the mobile node. Some examples of the policies We integrated the proposed scheme with wireless Rether in
used in the above mentioned software are Early-Expire, which order to provide QoS with roaming support. In this integrated
specifies a different expiration-time than the value specified setup, the WRS acts as the caching agent. Whenever a mobile
in the advertisement, Newest-FA, which forces a mobile node node (WRC) migrates to a new wireless subnet, it registers with
to choose the most recently discovered MA for registration, the WRS responsible for the new subnet by sending a rether
etc. A proper combination of these policies, (Newest-FA ON registration request to the well-known MAC address. The solic-
and Early-Expire ON) forces the mobile nodes to send out a itation for the cached mobile IP advertisement is piggybacked
registration request immediately after the discovery of the new with the registration request. The WRS then piggybacks the
MA. cached advertisement with the rether registration reply. Also,
the QoS mechanism treats the mobile IP registration and reply
D. Tunnel Setup Time messages as urgent messages in order to expedite the mobile
IP processing. After registering with the new WRS, the WRC
After receiving the registration request from a mobile node,
reestablishes bandwidth reservations for each of its active ap-
the MA processes the registration message and establishes a
tunnel with the mobile node’s home agent. Because these activ- plication. The system architecture of this integrated system is
ities are part of mobile IP code and we aim at performance en- shown in Fig. 4.
hancements that are independent of mobile IP implementations,
we chose not to perform any optimization in this step. Moreover, VI. PERFORMANCE MEASUREMENTS
performance measurements show that the relative cost of this A. Testbed Setup
step compared with the previous three steps is comparatively
The proposed low-latency handoff scheme has been imple-
mented on the Linux kernel version 2.2.16 and the user environ-
Fig. 3 shows the complete flow of the proposed mobile IP
ment Redhat 7.0 and Redhat 6.2. The caching agent/WRS is a
handoff process and the message exchanges that take place
Pentium II–400 MHz machine with 128 MB of RAM. The mo-
among the different entities on a wireless access network.
bile nodes are portable computers with Pentium III–650 MHz
and 64 MB of RAM. The network interface cards on the porta-
V. INTEGRATING LOW-LATENCY HANDOFF bles are PCMCIA Agere Systems’ Orinoco wireless cards. The
WITH QOS GUARANTEES driver used for the Orinoco cards is a modified version of Agere
Time-critical applications running on mobile nodes may em- Systems’ driver for Orinoco. The modification is to provide
ploy some QoS or bandwidth guarantee mechanism to ensure the probe functionality to retrieve the ID of the current access
that the network traffic is prioritized and remains unaffected point. The access points used in the testbed are Agere Systems’
because of unexpected network load. In order to provide QoS AP-1000 with wired interfaces configured at 100 Mb/s and wire-
and roaming support together, the QoS mechanism and the mo- less interfaces configured at 11 Mb/s. The MAs are Pentium
bility software need to interoperate with each other so that the III–650 MHz desktops. The mobile IP software used for the
QoS policies do not delay the messages exchanged between MAs and the mobile nodes is HUT mobile IP Dynamics ver-
the caching agent, mobile node, and the MA. Also, the policies sion 0.7.5.
across different subnets should be consistent with each other so During measurements the policy configurations on mobile
that the same level of QoS is available to mobile applications nodes, namely, Newest-FA, Newest-ADV, Early-expire, and
even while they are roaming. Eager-switching, were set to ON. The MA advertisement period
650 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 22, NO. 4, MAY 2004
was set to 15 s when the proposed optimizations were activated,
and was set to 1 s when the optimizations were turned off.
B. Probing Link-Layer Handoff
The probing interval for the current access point ID was set
to the lowest possible value of 10 ms. Each probing uses only
simple operations like read and comparison. The probe rou-
tine takes less than 100 s per invocation. This translates to
less than 1% overhead of the end-to-end handoff latency. Al-
though this overhead is insignificant, it can be further reduced
by turning off the probe and link-layer handoff detection im-
mediately after a successful handoff for some period during
which another handoff is not anticipated. Also, if there exists
any mechanism to monitor the link quality, then probing can be
selectively turned on when some significant variation in the link
quality is observed.
Probe reliability is a factor which forces repeated probing for
handoff confirmation. The ratio of incorrect return value to the
total number of probes is around 1:250. In other words, the prob-
ability of a probe returning a wrong value is 0.4%. It was never
observed that successive probes with the same return value ac- Fig. 5. Histogram of link-layer handoff time. The beginning and the end of
handoff period are marked by reception of the last packet on the old subnet and
tually lead to an incorrect access ID. Thus, a positive link-layer the first packet on the new subnet.
handoff can be determined using only two successive probes at
a probability of 99.9984%, with a possible complete elimination
of false positives.
The wireless LAN card performs link-layer handoff without
any software intervention. However, the amount of time that
a link-layer handoff takes is a component of the end-to-end
handoff latency, and it depends upon various factors, including
density of access points, the instantaneous network traffic,
channel radio characteristics and interference, load on in-
dividual access points, various threshold values set for the
network interface cards such as signal to noise ratio threshold,
cell search threshold, and roaming threshold. Since there are
numerous factors, the link-layer handoff time varies over a
wide range. For any given pair of access points this time can be
anywhere between a few microseconds to tens of milliseconds.
In our testbed setup, we observed different values ranging
from 3 to 90 ms with no apparent distribution. However, most
of the link-layer handoffs were observed to be completed in
the range of 20–25 ms. The histogram of link-layer handoff
time measurements is shown in Fig. 5. The link-layer handoff
time, which sets a lower bound for handoff detection time, was
measured by taking the time-stamp difference on a mobile node
between the last packet received on the old subnet and the first
packet received on the new subnet, when both subnets are fully Fig. 6. Histogram of link-layer handoff detection period. This period includes
saturated with downstream traffic. the handoff time and the handoff detection and confirmation by probing the
network interface card internal registers/buffers.
The time to detect the link-layer handoff is determined by the
probe interval and the refresh interval of the Wi-Fi card’s in-
ternal registers/buffers. Since our detection algorithm requires cess point’s MAC address after a fixed interval since the begin-
at least two probes to confirm any handoff, each detection takes ning of the most recent handoff. Thus, the total sum of the probe
at least two invocations of the probe routine. Compared with the invocations and the (supposed) refresh period is uniform. The
irregular amount of time a link-layer handoff itself takes, the handoff detection time, measured from the beginning of the last
detection time of our algorithm is surprisingly uniform. We be- handoff, is within a relatively short range of 50–70 ms. The his-
lieve that this is probably because the internal registers/buffers togram of observed link-layer handoff detection time is shown
holding the current access point’s MAC address which are actu- in Fig. 6. The beginning of a link-layer handoff is marked by the
ally read during the probe are refreshed to reflect the correct ac- time-stamp of the last packet received on the previous subnet.
SHARMA et al.: LOW-LATENCY MOBILE IP HANDOFF FOR INFRASTRUCTURE-MODE WIRELESS LANS 651
DATA-LOSS EXPERIENCED BY VARIOUS STREAMS
C. Mobile IP Handoff and Data Loss
Immediately after a link-layer handoff detection, a mobile
node can send out a unicast solicitation for the new MA. The
response to this solicitation is received after a processing and
message round-trip delay of 4 ms. This solicitation response
triggers the mobile IP registration, which is received by the MA
within the next 2 ms. Further authentication and tunnel setup
was observed to take around 36 ms in LAN setup. Therefore,
the amount of time required for a mobile IP handoff is about
100 ms in the average case ( ).
In contrast to this optimized scheme, the standard mobile IP
handoff implementation initiates a network-layer handoff only
upon reception of a MA advertisement. If the advertisement pe-
riod is set to the allowed minimum value of 1 s, the average delay
between the completion of a link-layer handoff on a mobile node
and reception of the next MA advertisement is around 500 ms.
Subsequent mobile IP processing may take around 35–50 ms.
Thus, the average mobile IP handoff latency measured from the
reception of the last packet on previous subnet is around 550
ms. The mobile IP tunnel setup time may vary depending on the Fig. 7. Gant chart for the mobile IP handoff with and without detection
network delays. However, this does not impact the optimization optimization. In the optimized case, the handoff time is around 100 ms,
whereas, in unoptimized case the average handoff time is around 550 ms and
as it is not dependent on tunnel setup. The proposed optimiza- can extend up to several seconds.
tion reduces the average handoff latency by more than 450 ms,
which is significant for time-critical applications. The deviation
While responding to this solicitation request, the caching agent
in the handoff latency for the unoptimized case is around 500
can inform the buffering module on the home agent to replay/re-
ms whereas, for the optimized case it is only 10 ms. Thus, the
send the data that might have been lost during the handoff. The
upper bound on the mobile IP handoff latency is of the order of
buffering module can then replay the packets over the newly es-
seconds in the unoptimized case and is still below 110 ms in the
tablished tunnel. Thus, by intelligent data replaying, the mobile
optimized case. Fig. 7 shows the timing diagram that contrasts
node would experience only delay in the data stream but no data
the optimized and unoptimized implementations for mobile IP
loss. Since minor delays and packet reordering are acceptable
for applications, completely avoiding the data loss is of great
The data loss experienced by a stream is directly dependent
advantage. This way the downstream losses can be eliminated
on the mobile IP handoff latency and the data rate of the stream.
completely. For the packets that are transmitted by the mobile
For an MPEG1 media stream at the frame rate of 30 frames/s,
nodes, adding a small buffering, replay, and delay in transmis-
the data loss in the optimized scenario is around 3–4 frames,
sion till the mobile IP registration reply is received can eliminate
and in unoptimized scenario it is around 15–30 frames. Table II
all possible data loss for upstream traffic.
shows the observed data loss for various streams.
It is possible to completely eliminate this data loss by using
buffering and data replaying. The home agent can also act as a
buffering and data replaying agent as it can snoop on the packets In this paper, we presented the design and implementation
that are being transmitted through the access point. When a mo- details of a low-latency mobile IP handoff scheme for IEEE
bile node is at home, the data packets meant for the mobile 802.11b compliant wireless LANs that are operating in infra-
node will be handed over to the access point, and when the mo- structure mode, which is the predominant configuration used
bile node is in a foreign subnet, these will be handed over to in most production-grade wireless LAN-based networks. The
the home agent. In both cases, the buffering module would be central problem associated with such networks is that their link-
able to keep track of the data traffic meant for any given mo- layer handoff is both hard and forward, which makes it difficult,
bile node from its subnet. The data corresponding to the worst if not impossible, to apply the handoff latency reduction tech-
case handoff delay period can be buffered in individual queues niques proposed in the existing literature, such as anticipation
differentiated by the destination host IP addresses. Whenever a and multicasting. The proposed scheme solves this problem by
mobile node migrates from one subnet to another, the caching a timer-driven software probing technique to facilitate the de-
agent of the new subnet would receive a solicitation request. tection of the occurrence of link-layer handoff and a mobile IP
652 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 22, NO. 4, MAY 2004
advertisement caching and replay proxy to quickly discover new  E. Gustafsson, A. Jonsson, and C. Perkins. (2001) ”Mobile IP regional
MAs using only unicast communication. registration,“ Internet Draft. [Online]. Available: draft-ietf-mobileipreg-
The performance measurements show that the proposed  P. Calhoun, T. Hiller, J. Kempf, P. McCann, C. Pairla, A. Singh, and S.
scheme can achieve a factor of 5 reduction in mobile IP handoff Thalanany. (2000) “Foreign Agent Assisted Hand-Off,” Internet Draft.
latency compared with the best mobile IP schemes known to [Online]. Available: draft-ietf-mobileip-proactive-fa-03.txt
 M. Baker, X. Zhao, S. Cheshire, and J. Stone, “Supporting mobility in
work on such networks. The additional run-time performance MosquitoNet,” presented at the USENIX Tech. Conf., San Diego, CA,
overhead is negligible. Moreover, the proposed scheme can Jan. 1996.
interoperate with existing mobile IP implementations without  R. Cáceres and V. N. Padmanabhan, “Fast and scalable handoffs for
wireless internetworks,” presented at the MOBICOM ’96, Rye, NY,
any modification and with only minimal network configu- Nov. 1996.
ration effort. Further, this scheme can interoperate with a  R. Cáceres and L. Iftode, “Improving the performance of reliable
link-layer QoS mechanism for wireless networks to provide transport protocols in mobile computing environments,” IEEE J. Select.
Areas Commun., vol. 13, pp. 850–857, June 1995.
bandwidth guarantees even while roaming. With low-latency  A. C. Snoeren and H. Balakrishnan, “An end-to-end approach to host
handoff and bandwidth guarantee mechanisms in place, mobility,” presented at the 6th ACM/IEEE Int. Conf. Mobile Computing
IEEE 802.11b-based wireless LANs can truly compete with Networking, Boston, MA, 2000.
 J. Inouye, J. Binkley, and J. Walpole, “Dynamic network reconfigura-
third-generation (3G) cellular networks. tion support for mobile computers,” presented at the 3rd ACM/IEEE
Although currently this scheme is verified only with IEEE Int. Conf. Mobile Computing and Networking, Budapest, Hungary, Sept.
802.11b-based (Wi-Fi) hardware, it does not depend on any 1997.
 S. Sharma, K. Gopalan, N. Zhu, G. Peng, P. De, and T. Chiueh, “Im-
specific protocol details. It only relies on the capability of plementation experiences of bandwidth guarantee on a wireless LAN,”
wireless network interface cards to provide a mechanism to presented at the SPIE Multimedia Computing and Networking, San Jose,
probe the access point’s SSID. Thus, the proposed scheme CA, Jan. 2002.
is immediately applicable to any wireless equipment which
operates in infrastructure mode such as IEEE 802.11a-based
hardware or forthcoming IEEE 802.11g-based hardware. In the
future, we plan to extend the proposed handoff scheme with Srikant Sharma (S’03) received the B.Tech. degree
buffering, snooping, and data replaying mechanisms to achieve in computer science from Regional Engineering Col-
lege, Calicut, India, in 1996. He is currently working
smooth handoff without any data loss, which will enable mobile toward the Ph.D. degree in computer science at Stony
multi-cell wireless LANs to support real-time multimedia and Brook University, Stony Brook, NY.
voice over IP applications. He was with Wipro Corporation, Karnataka, India,
as a Senior Research Engineer. His research interests
are in gigabit metro networks and wireless networks.
 Gartner Group, [Online]. Available: http://www.gartner.com.
 C. Perkins, Ed., “IP Mobility Support,”, RFC 2002, 1996.
 J. Sevanto, M. Liljeberg, and K. Raatikainen, “Introducing quality of
service and traffic classes into wireless mobile networks,” presented at
the 1st ACM Int. Workshop on Wireless Mobile Multimedia, Dallas, TX,
 IEEE Standard for Wireless LAN Medium Access Control (MAC) and Ningning Zhu (S’03) received the B.S. and M.S.
Physical Layer (PHY) Specifications, Nov. 1999. degrees in computer science from Peking University,
 [Online]. Available: http://www.wi-fi.org Beijing, China, in 1992 and 1995, respectively. She
 G. Pollini, “Trends in handover design,” IEEE Commun. Mag., vol. 34, is currently working toward the Ph.D. degree in
pp. 82–90, Mar. 1996. computer science at Stony Brook University, Stony
 S. Seshan, H. Balakrishnan, and R. Katz, “Handoffs in cellular wireless Brook, NY.
networks: The Daedalus implementation and experience,” Kluwer Int. From 1995 to 1999, she was with the National
J. Wireless Commun. Syst., vol. 4, no. 2, pp. 141–162, Jan. 1997. Center for Intelligent Computing, Beijing, China,
 H. Balakrishnan, S. Seshan, and R. Katz, “Improving reliable transport as a Research Engineer. Her research interest is on
and handoff performance in cellular wireless networks,” ACM Wireless storage system and wireless networking.
Networks, vol. 1, no. 4, pp. 469–481, 1995.
 E. Shim and R. Gitlin, “NeighborCasting: A fast handoff mechanism in
wireless IP using neighboring foreign agent information,” presented at
the New York Metro Area Networking Workshop, New York, 2001.
 C. Tan, S. Pink, and K. Lye, “A fast handoff scheme for wireless net-
works,” presented at the 2nd ACM Int. Workshop on Wireless Mobile Tzi-cker Chiueh (S’89–M’92) received the B.S. de-
Multimedia, Seattle, WA, 1999. gree in electrical engineering from National Taiwan
 D. Forsberg, “Communication availability with mobile IP in wireless University, Taipei, Taiwan, R.O.C., the M.S. degree
LANs,” Masters thesis, Helsinki Univ. Technol., Helsinki, Finland, in computer science from Stanford University, Stan-
2000. ford, CA, and the Ph.D. degree in computer science
 K. Malki, P. Calhoun, T. Hiller, J. Kempf, P. McCann, A. Singh, H. from University of California at Berkeley, in 1984,
Soliman, and S. Thalanany, ““Low Latency Handoffs in Mobile IPv4,” 1988, and 1992, respectively.
Internet Draft,” Mobile IP Working Group, 2002. He is currently an Associate Professor in the Com-
 R. Koodli, ““Fast handovers for mobile IPv6,” Internet Draft,” Mobile puter Science Department, Stony Brook University,
IP Working Group, 2002. Stony Brook, NY, and the Chief Scientist of Rether
 D. Forsberg, J. Malinen, J. Malinen, T. Weckstrom, and M. Tiusanen, Networks, Inc., Centereach, NY. His research interest
“Distributing mobility agents hierarchically under frequent location up- is on computer security, network/storage QoS, and wireless networking.
dates,” in Proc. 6th IEEE Int. Workshop on Mobile Computing Systems Dr. Chiueh received the National Science Foundation (NSF) CAREER Award
Applications, San Diego, CA, Feb. 1999, pp. 159–168. in 1995.