Improving End-to-End Performance of TCP over Mobile Internetworks
`
`Raj Yavatkar
`
`Namrata Bhagawat
`
`Department of Computer Science
`University of Kentucky
`Lexington, KY 40506
`raj@ dcs.uky.edu
`
`Department of Computer Science
`University of Kentucky
`Lexington, KY 40506
`namrata@dcs,uky.edu
`
`Abstract
`Reliable transport protocols such as TCP use end-to-
`endjow, congestion, and error control mechanisms to pro-
`vide reliable delivery over an internetwork. However, the
`end-to-end performance of a TCP connection can suffer
`significant degradation in the presence of a wireless link.
`We are exploring alternatives for optimizing end-to-end
`performance of TCP connections across an internetwork
`consisting of both fixed and mobile networks. The central
`idea in our approach is to transparently split an end-to-end
`connection into two separate connections; one over the
`wireless link and other over the wired path. The connec-
`tion over the wireless link m y either use regular TCP or a
`specialized transport protocol optimized for better perfor-
`mance over a wireless link. Our approach does not require
`any changes to the existing protocol software on station-
`ary hosts. Results of a systematic performance evaluation
`using both our approach and regular TCP show that our
`approach yields significant performance improvements.
`1 Introduction
`Reliable transport protocols such as TCP use end-to-
`end flow, congestion, and error control mechanisms to
`provide reliable delivery over an internetwork. However,
`co-existence of wireless links and mobile hosts with fixed
`networks poses unique problems for transport protocols. In
`particular, the following communication characteristics of
`wireless links have significant implications.
`First, Maximum Transmission Unit (MTU) on a wireless
`link is typically much smaller than that over links in the
`wired network [l, 21. Small MTU over the first link forces
`transmission of smaller packets over the entire end-to-end
`path even though wired path can accommodate much larger
`packets.
`Second, the error rates on a wireless link are much higher
`than those experienced over the links in the wired net-
`work [3,4,5]. Higher error rates (and resulting intermittent
`connectivity) over a wireless link are due to a combination
`of factors such as multipath fading, terrain and environ-
`
`mental factors, and interference from other transmissions.
`In addition, these errors often cause a burst of packets to be
`lost.
`Third, communication pauses during handoffs are also
`perceived as periods of heavy losses by transport and higher
`level protocols [61.
`These wireless transmission characteristics together
`contribute to severe degradation in performance of proto-
`cols such as TCP. Use of small packets leads to under-
`utilization of available bandwidth in the wired network
`and reduces overall end-to-end throughput of a connection.
`Higher error rates and communication pauses during hand-
`off can falsely trigger congestion control mechanism of
`TCP [7]. For example, communication pauses and packet
`losses over the wireless link cause retransmission timeouts.
`In both cases, TCP’s slow-start mechanism [8] reacts by
`drastically reducing the current transmission rate and TCP
`takes a long time to recover from such a reduction resulting
`in severe degradation in throughput.
`We are exploring alternate approaches for optimizing
`end-to-end performance of TCP connections across an in-
`temetwork consisting of both fixed and mobile networks.
`Our approach is motivated by the following principles:
`
`0 We want to achieve performance optimization wifhout
`modifying TCP and its existing flow and congestion
`control mechanisms.
`0 Given the widespread use of TCP/IP in fixed hosts,
`we would like to avoid any changes to the existing
`protocol software in machines on the wired Internet.
`0 Existing clienvserver applications should see no
`changes to the socket interface and should require no
`changes to execute across mobile internetworks.
`The central idea in our approach is to introduce a new
`session layer protocol on top of TCP at both base stations
`(also called Mobile Support Routers or MSRs) and mo-
`bile hosts. We require no changes to the protocol software
`
`0-8186-6345-6/95 $04.00 0 1995 IEEE
`
`146
`
`Page 1 of 7
`
`SAMSUNG EXHIBIT 1026
`
`

`

`_:*
`
`:*
`
`Cell
`
`(sitar)
`..‘
`Mobile Host
`Boundary ”‘* ... -,..
`
`*__.......I....
`
`Wired Internet
`
`Fixed Host
`(ICs0
`
`Base Station
`(ms.uky)
`
`Fixed Host
`(Purdue)
`
`I
`
`Socket
`
`MHP
`(Session Layer)
`
`TCP/U DP
`
`- - - - - - - - - - - - - - r
`
`IP
`Loss and Handoff I
`
`Figure 1: An example mobile internetwork.
`
`Figure 2: The protocol hierarchy assumed at base stations
`and mobile hosts.
`
`on ordinary stationary hosts. The session layer proto-
`col is designed to exploit the available knowledge about
`both wireless link characteristics and host migration and
`to compensate for highly unpredictable and unreliable link
`between a mobile host and its base station.
`An advantage of this approach is that performance degra-
`dation in TCP is limited to a “short” connection over the
`wireless hop and traffic over the “long” connection over the
`wired network can be protected from the impact of erratic
`behavior over the wireless link.
`We have considered two alternatives for improving per-
`formance of TCP over the wireless hop. The two alter-
`natives can be summarized by an example using Figure 1.
`Let us assume that a TCP connection is desired between
`sitar and icsi.
`Under the first altemative (called MTCP), the proposed
`session layer protocol, called MHP (Mobile Host Proto-
`col), establishes two TCP connections, one from sitar to
`its base station, and another from its base station to icsi
`across the fixed internetwork. An intermediate agent at the
`base station acts as a relay for traffic from the first connec-
`tion to another’. In the case of a handoff, we assume that
`the mobile IP protocol can pass on an indication of hand-
`of in progress to higher layer protocols using an upcall
`through the protocol layers. When the handoff completes,
`MHP transfers the connection state information to the new
`base station and establishes a new connection between the
`mobile host and its new base station. No changes are, how-
`ever, necessary to the connection with the remote host as
`mobile IP routing [9] takes care of routing packets through
`the new base station.
`
`If the remote host is also mobile, an additional connection must also
`be set up over the wireless link to its base station.
`
`The second altemative (called SRP) is similar to the first
`alternative except that the session layer does not use TCP as
`its transport layer for the wireless hop. We are considering
`this alternative to investigate whether one can justify use
`of a specialized transport protocol tuned and optimized for
`better performance over a wireless link. Under SRP, the
`protocol used over the wireless hop uses its own flow and
`error control mechanisms designed and optimized specifi-
`cally to tackle the lossy and erratic delay characteristics of
`the wireless link. The intermediate agent at the base sta-
`tion participates in the session layer protocol and forwards
`incoming traffic over a TCP connection to the remote host.
`The session layer hides the details of the first connection
`and provides the same application layer interface as TCP
`through the Unix socket library.
`We have compared both alternatives against the use of
`normal TCP in aL mobile internet testbed consisting of a
`simulated wireless link and the Internet. Our tests have
`yielded impressive results. The rest of this paper is orga-
`nized as follows. Section 2 describes the protocol model in
`detail. Section 3 describes the experimental setup, method-
`ology, and results of our performance evaluation. Section
`4 summarizes the related work in this work and Section 5
`provides concluding remarks.
`2 Protocol Model
`Our goal is to isolate the wired portion of the path of
`a connection from the impact of erratic behavior over the
`wireless portion and also to recover quickly from errors over
`a wireless link to obtain good end-to-end performance. We
`regard the impact of small MTU and intermittent connec-
`tivity over a wirelless link as transient errors over a transport
`level connection and we believe that the protocol software
`must protect applications by transparently recovering from
`
`147
`
`Page 2 of 7
`
`

`

`such transient errors. In the IS0 reference model [IO], the
`responsibility for session management including recovery
`and re-synchronization in data transfers lies with the session
`layer in the protocol stack. Transport layer protocols only
`provide end-to-end delivery of messages or byte streams.
`In keeping with the IS0 reference model, we introduce a
`new session layer protocol called MHP (Mobile Host Pro-
`tocol) that explicitly includes mechanisms for recovering
`from errors over the wireless link.
`Figure 2 shows the proposed protocol hierarchy for net-
`working software on mobile hosts and their base stations
`(also called Mobile Support Routers or MSRs). We as-
`sume that no changes are necessary to the existing network
`protocol software on fixed hosts. The protocol software
`on mobile hosts and their base stations is now augmented
`with a new session layer protocol called MHP (Mobile Host
`Protocol) that retains the same MI (such as BSD socket
`interface [ 1 I]) as that offered by TCP.
`2.1 Connection Establishment
`Figure 3 shows an example interaction involving a mo-
`bile host and a remote, stationary host. We assume that
`protocol software on base stations and mobile hosts con-
`sists of an MHP layer that manages the transport level con-
`nections. In the following, we describe how MHP layers
`at base stations and mobile hosts cooperate to support an
`end-to-end connection.
`
`0 When a TCP application on the mobile host X issues a
`connect call to request a connection to a remote
`destination Y at < d e s t I P a d d r , d e s t P o r t > ,
`the MHF layer at X (MHPX) intercepts the call
`and instead requests a transport level connection
`(Connectionl in Figure 3) with its peer at its cur-
`rent base station. MHP peer on the base station
`sets up a surrogate or MHP agent (MHPBSl) on
`behalf of the requested connection. The surrogate,
`MHPBS1, in turn, establishes a TCP connection
`(Connection2 in Figure 3) with Y at the address
`i d e s t IPaddr , d e s t Port>on behalf of theend-
`point on X. One endpoint of Connection2 is still
`marked as cX-IPaddr, X-srcPort> and all the
`TCP traffic from Y to X is intercepted and forwarded
`to the surrogate MHP-BS1 at the base station. As
`described in Section 2.2, MHPBS 1 simply acts as a
`relay for the traffic between X and Y in both directions.
`
`0 When a TCP application on a stationary host (Y)
`requests a TCP connection to a mobile host (X) at
`address < M H a d d r , M H P o r t > , the connection re-
`quest is intercepted and forwarded to the surrogate
`MHPBS 1 at the base station. The surrogate then com-
`pletes the TCP connection establishment with Y and
`
`BASE STATION
`852
`
`J
`
`I
`
`Mobile HOsT-Y
`I MHP-x
`,v
`,' MI crossing
`
`j
`
`1
`I
`
`Connection 2
`
`Fixed
`HOST-Y
`
`TCP/IP
`Protocol
`Slack
`
`i
`i
`
`*
`
`Figure 3: An example of connection establishment and
`handoff involving a mobile host (X), stationary destination
`(Y), and two base stations. Initially, X establishes a con-
`nection to Y through the MHP agent at base station BS1.
`After a cell handoff, a new connection is established be-
`tween X and the new base station BS2 and the endpoint of
`connection 2 is transferred to MHP agent on BS2.
`
`establishes a new connection with its peer (MHPX)
`at the given address.
`0 A TCP connection between endpoints on two mobile
`hosts is handled similarly except, in this case, three
`separate connections are established.
`2.2 Data Transfer
`Data transfer to and from the mobile host X and remote
`destination host Y proceeds as follows. When a TCP appli-
`cation on X sends data, MHPX uses the first connection to
`send that data to MHPBS 1. In particular, MHPX sends
`data in small segments to match the smaller MTU over the
`wireless link. MHPBS 1 receives the data and buffers it to
`assemble these smaller packets into larger TCP segments
`before forwarding them over the connection to Y, Simi-
`larly, when MJJPBS1 receives TCP segments from Y, it
`first breaks them into smaller fragments to match the MTU
`over the wireless link, forwards smaller TCP segments to
`X.
`
`To recover from handoffs, the MHP layer must maintain
`some state information on the segments in transit to and
`from the wireless link. Therefore, the MHP layer maintains
`state information on the segments in its buffers and also
`accesses the connection state information maintained by
`its underlying transport protocol. The state information
`
`148
`
`Page 3 of 7
`
`

`

`accessed includes connection parameters such as current
`window sizes and sequence numbers for window edges.
`2.3 Error Recovery
`To recover from errors due to high bit error rates of
`the wireless link and handoff, we have investigated two
`alternatives.
`Under the first alternative called MTCP (MultipleTCP),
`MHP uses regular TCP as the transport protocol for the
`connection over the wireless link.
`Under the second alternative called SRP (Selective Re-
`peat Protocol), MHP uses a specialized transport protocol
`designed to recover quickly from higher and sometimes
`bursty packet losses experienced over the wireless link.
`SRP uses a selective repeat algorithm in which a receiver
`returns a selective ACK (SACK) when an out of sequence
`segment is received. The SACK specifies the missing seg-
`ments using a bitmap, the sequence number of the latest
`segment received, and the sequence number of the last seg-
`ment received in sequence. On receiving a SACIK, the
`sender retransmits all the missing segments specified in the
`SACK. Using this alternative, unlike TCP, SRP can recover
`more than one segment in one round trip time and can yield
`better throughput over the wireless link.
`Section 3 compares the performance of two alternatives
`when used over a mobile internet.
`2.4 Handoff Management
`When the MH moves and crosses the current cell bound-
`ary, it gets attached to another base station (BSZ) and the
`IP datagrams for TCP segments over an existing connec-
`tion start getting forwarded to the new base station. During
`the cell handoff, we must also make sure that the existing
`transport connections get transferred to a new MHP agent.
`We assume that the MHP layer at the mobile host regis-
`ters an upcall function with its IP layer. When a handoff is
`completed, IP layer on MH informs the MHP layer of hand-
`off using the upcall function and passes the address of the
`new base station (BSZ) to it. The MHP layer (MHPX) then
`contacts its peer at the new base station to initiate a handof
`management procedure that consists of the following steps:
`
`0 On receiving the upcall, MHPX first suspends the
`ongoing data transfer across its transport connections,
`contacts its peer at BS2 giving it the address of the
`previous surrogate MHP-BS1, and then waits for a
`connection resume message from BSZ.
`
`0 The MHP peer at BS2 establishes a new MHP agent
`or surrogate (MHPBS2) for the connection. The new
`surrogate then sends a handover message to the old
`surrogate (MHPBS 1) requesting the state information
`for two connections.
`
`0 MHPBS1 responds with the connection state infor-
`mation and, in addition, also forwards the TCP seg-
`ments in transit that it has buffered for traffic in each
`direction. When MHPBS2 receives the state infor-
`mation, it re-creates the state information for connec-
`tions with MHPX and Y and then sends a connection
`resume message to MHPX.
`0 Data transfer to and from MHPX then resumes and
`the remote stationary host observes no changes in the
`state of its connection except possibly for some trans-
`port layer retransmission of data lost during handoff.
`
`3 Performance Evaluation
`We have conducted a systematic performance evaluation
`of our approach using a wireless internet testbed. In the
`following, we describe the testbed, experiments performed,
`and results obtained.
`3.1 Experimental Testbed
`The testbed coinsists of two parts. The first part consists
`of a wireless network simulated over an ethemet segment
`and Sun sparcstations running a modified SunOS kernel
`acting as mobile sparcstations. The second part consists of
`our campus network attached to the rest of the Internet over
`a T1 link. Some isparcstations on the campus network act
`as base stations.
`We have implemented MHP in two different versions.
`To test our ideas, we first implemented MHP as a user
`level library and rest of this paper reports results obtained
`using the user level MHP implementation. We also have a
`kernel implementation of MHP on mobile hosts and base
`stations that resides below the socket layer (above TCP)
`and provides the same interface as TCP through the socket
`interface.
`The IP software in the SunOS kernel of the mobile sparc-
`stations has been modified to simulate a wireless link as
`follows:
`
`1. In mobile sparcstations, we have modified the IP layer
`to use a smaller MTU (128 or 256 octets). In addition,
`the IP software simulates packet losses and handoffs.
`Delay and loss characteristics simulated are taken from
`the experiences reported in the published literature [ 1,
`4,3, 121.
`2. IP software simulates bursty losses over the wireless
`link. The busty loss simulation models the interburst
`gap using an exponential distribution around a mean
`inter-burst interval (IBG) and the size of each burst
`is modeled using a geometric distribution around a
`mean burst size (BS) value. The values of IBG and
`BS were chosen for each experiments based on the
`average packet loss desired for the experiment. We
`
`149
`
`Page 4 of 7
`
`

`

`have simulated packet losses of 0, 5, and 10% for
`different testcases.
`3. Sparcstations located on campus subnets act as base
`stations where each subnet is considered a different
`cell and subnets are separated by a campus router.
`
`4. The IP layer simulates a handoff at a mobile host by
`simply pausing for the handoff duration and dropping
`all outgoing and incoming packets during the pause.
`We have simulated handoff pauses of 1, 2.8, and 5
`second durations based on figures taken from 161.
`We have also carried out tests using multiple handoffs
`in which successive cell handoffs are simulated spaced
`at different intervals ranging from 5 to 10 seconds.
`
`3.2 Methodology
`In our experiments, we use a user-level test program
`that establishes a connection between a mobile host and a
`remote stationary host and transfers data in a file of fixed
`size to its peer at the destination. Tests were canid using
`stationary hosts either located in the local area (on a cam-
`pus subnet) or located across the Intemet at ICs1 and UC
`Berkeley, Purdue University, and Washington University in
`St. Louis.
`Once the connection is in progress, mobile IP software
`simulates a handoff pause duration starting after a fixed,
`predetermined interval (typically 4 to 8 seconds, an exper-
`imental parameter) after the connection starts and contacts
`the mobile IP software at a new base station at the end of
`the handoff pause to complete the handoff.
`For each test, we repeated the experiments over a two
`week period on weekdays between 1 and 3 pm EST to ob-
`tain results under similar Intemet traffic conditions2. Using
`samples from 40 independent runs, we carried out a con-
`fidence interval analysis with 95% probability and have
`tabulated the confidence intervals along with average val-
`ues.
`3.3 Representative Results
`Tables 1 through 3 show a representative sample of re-
`sults. We have also conducted tests involving remote hosts
`located at Purdue University and Washington University
`and have obtained similar results.
`Table 1 shows the results for the base case used for com-
`parison with our approaches. The entry in upper left hand
`corner (no pause, no losses) shows the results in the absence
`of mobility (no handoff pause, no losses due to mobile link)
`and, as can be seen clearly, performance degrades as a sin-
`gle handoff pause and packet losses due to wireless link are
`inuoduced.
`
`?-We have also conductedtests late night to evaluate performanceunder
`different Intemet traffic conditions.
`
`150
`
`Results Using Regular TCP
`
`Packet loss in Percent
`
`44.6
`[40.9,48.31
`[27.6,35.9]
`[50.5,62.71
`88.7
`52.1
`32.6
`[77.6,99.71
`[45.6,58.61
`[29.2,36.01
`99.9
`69.8
`36.7
`[34.0,39.31 160.L79.61 186.6, 113.11
`
`Pause
`
`1 sec
`
`2.8 sec
`
`5 sec
`
`Table 1: Mean time to transfer a file of size look bytes
`with a single, normal TCP connection between the mo-
`bile host sitar.dcs.uky.edu and the remote destination ic-
`sib16.icsi.berkeley.edu. The confidence Interval of 95% is
`shown in square brackets.
`
`Results using MTCP
`
`Packet loss in Percent
`
`32.6
`
`Pause
`
`2.8 sec
`
`5 sec
`
`Table 2: Results for tests carried out for the same case as
`Table 1, but using MTCP.
`
`Handoff
`Pause
`0 sec
`
`0 %
`12.7
`[11.7, 13.71
`
`Packet loss in Percent
`5 %
`19.6
`[18.7,20.5]
`
`10 96
`22.4
`[21.0,23.9]
`
`112.5, 15.31
`21.1
`
`[18.4,21.71
`27.3
`
`[24.7,28.61
`29.2
`
`2.8 sec
`
`1 [19.7,22.51 I [25.5,29.11 I [26.9,31.41 I
`Table 3: Results for tests carried out for the case same as
`for Table 1, but using SRP.
`
`Page 5 of 7
`
`

`

`Table 2 shows the results for the same case using the
`MTCP approach (Two TCP connections instead of one).
`Clearly, performance improves significantly with reduc-
`tion in transfer times varying from 20% to a factor of three
`in some cases depending on handoff pause interval lengths
`and packet loss percentages. Similar results (see Table 3)
`are obtained using the SRP approach (selective repeat al-
`gorithm over UDP) except that SRP yields slightly better
`throughput than MTCP. We have also obtained similar re-
`sults involving multiple handoffs and in cases involving
`data transfer from a fixed host to a mobile host.
`Impact of Small MTU
`3.4
`It might seem surprising that delay values for the entries
`corresponding to the (no pause, no losses) case
`in the upper left hand corner show improvement over the
`normal TCP case when, in this case, there are no additional
`packet losses introduced due to mobility and there is no
`handoff. However, the h4TLJ over the wireless link is small
`(128 octets) and the normal TCP uses the same MTU for
`the entire path to the destination. On the other hand, in our
`approach, small segments get reassembled by MHlP at the
`base station to take advantage of the larger MTU available
`over the wired Internet.
`4 Related Work
`To the best of our knowledge, Caceres and Iftode were
`the first to investigate the impact of wireless networks on the
`performance of reliable data communication. They showed
`how the performance of current TCP implementation suf-
`fers significant degradation in the presence of motion and
`suggested a fast retransmission scheme to overcome the
`problem [6]. Our work differs from their work in several
`ways.
`First, our approach does not require any modifications to
`the current TCP protocol or its implementation on mobile
`or stationary hosts. In particular, the transport protocol
`software on stationary hosts remains completely unaware
`of the presence of mobile hosts on the Internet and thus
`we retain the transparency provided by the extensions to IP
`routing [9]. However, we need to modify the TCP software
`implementation on base stations so that connection state
`information can be accessed during handoff recovery.
`Second, even in the absence of motion, small MTU over
`wireless links and the lossy, intermittent connectivity over
`wireless links can cause degradation in end-to-end perfor-
`mance of TCP over the Internet. Our approach addresses
`the problem and overcomes it by restricting the impact of
`both small MTU and unusually higher packet losses only to
`the wireless portion of the end-to-end path of a connection.
`Our approach also recovers from the effect of small MTU
`by coalescing small TCP segments across wireless link into
`larger segments across wired net to exploit larger N[Tu and
`to improver overall bandwidth utilization.
`
`Third, we have also investigated the use of a specialized
`transport protocol (SRP) on the wireless portion of the path
`to better adapt to the characteristics of the wireless link.
`Our use of a spe!ialized transport protocol is completely
`transparent to the applications and retains the BSD socket
`interface used by the TCP-based applications.
`DeSimone et al [41 have investigated the possibility of
`using link layer retransmissions to counter the higher er-
`ror rates of wireless links. They show that link layer re-
`transmissions can adversely interact with the end-to-end
`mechanisms of a reliable transport protocol reducing both
`end-toad throughput and increasing link utilization over
`the wireless segment. Instead, our approach uses an end-
`to-end approach to improve the performance of a reliable
`transport protocol.
`The idea of splitting the end-to-end TCP connection into
`two halves was allso independently discovered by Badrinath
`and others [13, 141. Our work differs from their work in
`some ways. First, like the work by Caceres and Iftode,
`their work concentrates on the effect of motion and not on
`the effect of small MTU and lossy links in the absence of
`motion. Second, we have also investigated the transparent
`combination of using TCP over the wired path and using
`a specialized transport protocol over the wireless link to
`better adapt to conditions over the wireless link.
`5 Summary
`Wireless transmission characteristics (small MTU and
`high error rates) and handoff pauses can significantly de-
`grade end-to-end performance of TCP over wireless net-
`works. We propose a new session layer protocol for base
`stations and mobile hosts to compensate for effects of wire-
`less linkcharacteristics and host migration. Thecentral idea
`in our approach is to transparently split an end-to-end con-
`nection into two separate connections; one over the wireless
`link and other over the wired path. The connection over the
`wireless link may either use regular TCP or a specialized
`transport protocol optimized for better performance over a
`wireless link. Splitting of a connection is completely trans-
`parent to an application and no changes are necessary to
`protocol software on stationary hosts.
`We have compared our approach against the use of nor-
`mal TCP in a mobile intemet testbed and results show that
`our approach yields significantly better throughput than use
`of a single, end-to-end TCP connection.
`Acknowledgements
`We would like to thank Prof. Comer of Purdue Uni-
`versity, Prof. Ferrari of UC-Berkeley, and Prof. Parulkar
`of Washington TJniversity for allowing use of machines at
`their sites for conducting experiments. Bruce Mah and Hui
`Zhang provided valuable comments on an earlier draft of
`this paper. Finally, we would also like to thank the review-
`ers for their helpful comments and suggestions.
`
`151
`
`Page 6 of 7
`
`

`

`[13] A. Bakre and B. Badrinath. I-TCP: Indirect TCP for
`Mobile Hosts. Technical Report DCS-TR-3 14, Dept.
`of Computer Science, Rutgers University., October
`1994.
`[14] B. R. Badrinath, A. Bakre, T. Imielinski, and
`R. Marantz. Handling mobile clients: A case for in-
`direct interaction. In Proceedings of the IEEE Fourth
`Workshop on Workstation Operating Systems, Octo-
`ber 1993.
`
`References
`[l] D.C. Cox. Universal Portable Radio Communica-
`tions. IEEE Communications Magazine, pages 96-
`115, December 1992.
`[2] D. Raychaoudhari and N.D.Wilson.
`ATM-based
`Transport Architecture for Multiservices Wireless
`Personal Communication Networks. IEEE Journal
`on Selected Areas in Communications, 12(8), 1994.
`[3] D. Duchamp and N. F. Reynolds. Measured perfor-
`mance of a wireless Ian. In Proc. of the 17th Con$
`on Local Computer Networks, pages 494-499. IEEE,
`September 1992.
`[4] A. DeSimone, M.C. Chuah, and 0. Yue. Through-
`put Performance of Transport-Layer Protocols over
`Wireless LANs.
`In Proceedings of 1993 IEEE
`GlobeComm, 1993.
`[5] S. Nanda, R. Ejzak, and B. T. Doshi. A Retrans-
`mission Scheme for Circuit-Mode Data on Wireless
`Links. IEEE Journal on Selected Areas in Communi-
`cations, 12(8), October 1994.
`[6] R. Caceres and L. Iftode. The effects of mobility on
`reliable transport protocols. Technical Report MITL-
`TR-73-93, Matsushita Information Technology Lab-
`oratory, Princeton, New Jersey, November 1993.
`[7] R. Caceres and L. Iftode. Improving the Performance
`of Reliable Transport Protocols in Mobile Computing
`Environments. IEEE Journal on Selected Areas in
`Communications, October 1994.
`[8] V. Jacobsen. Congestion avoidance and control. In
`Proceedings of ACM SIGCOMM ’88 Symposium,
`pages 314-329, August 1988.
`[9] A. Myles and D. Skellem. Comparison of Mobile
`Host Protocols for IP. Internetworking Research and
`Experience, 4(4), December 1993.
`[lo] H. Zimmerman.
`OS1 Reference Model-The IS0
`Model of Architecture for Open Systems Intercon-
`nections. IEEE Transactions on Communications,
`COM-28( 12):425-432, April 1980.
`[ 1 11 Samuel J Leffler, Marshall Kirk McKusick, Michael J
`Karels, and John S Quatarman. The Design and Im-
`plementation of the 4.3 BSD UNIX Operating System.
`Addsion Wesly, 1989.
`[121 S . Nanda and David Goodman. Performance of
`PRMA: A Packet Voice Protocol for Cellular Sys-
`tems. IEEE transacatios on vehicular technology,
`40(3), 1991.
`
`152
`
`Page 7 of 7
`
`