Performance modeling and analysis of mobile Internet access via cellular networks

Transcription

1 Performance modelng and analyss of moble Internet access va cellular networks Master s thess Author: Taoyng Yuan Emal: tyuan@few.vu.nl Student ID: # Unversty: Supervsors: Vrje Unverstet Amsterdam Department of Computer Scence Prof. dr. R. D. van der Me Dr. S. Bhula

2 Acknowledgements I worked on ths thess between December 2004 and June 2005 at the Vrje Unverstet Amsterdam. I would not have been able to carry out the work wthout the followng people, whom I am deeply grateful to. Frst I would lke to thank my supervsors: Professor Rob van der Me and Dr. Sandja Bhula for all ther valuable help durng the last seven months. They have been a great source of nspraton for me and were always avalable to dscuss new deas or to answer my questons. They have helped me to get back on track several tmes. I would lke to gve my deepest thanks to my boyfrend, Rk Farenhorst, for hs enormous support and understandng throughout the last sx months. Wthout hs help, ths work would not have been possble for me. I would also lke to thank my parents for ther support and encouragement durng my work and ther endless support durng all my years n school. Taoyng Yuan Amsterdam, June 2005

3 Executve Summary Over the next few years, moble Internet access servces, (e.g. the WLAN), provdng Internet access anywhere and any tme are lkely to experence dramatc growth. A key success factor n the commercal success of these servces s performance, for example, n terms of throughput. Hence, for servce provders t s crucal to be able to predct the performance of ther networks under varous scenaros, ncludng the number of subscrbers and the locatons of the Access Ponts. Ths advocates the need for the development, valdaton and analyss of quanttatve models to descrbe and predct the performance experenced by the end user for any gven set of system parameters. In ths thess our research goal s to understand the mpact of the moblty of the moble termnal on performance of nternet access va cellular networks. Ths s accomplshed by the followng four parts. 1. Derve and valdate an approxmaton closed-form expresson for the aggregated throughput. 2. Develop the model to predct the performance of multple classes moble users n a sngle cell scenaro. And then, derve and valdate closed-form expressons for the average throughput and the total amount receved amount of receved data. 3. Extend our model to a multple-class moble users n a multple-cell - scenaro and derve the closed-form expressons of the performance. 4. Illustrate how to use those expressons to optmze the desgn of the cellular networks based on protocols.

4 Table of contents Acknowledgements. Executve Summary. Lst of Fgures Lst of Tables v 1 Introducton Introducton Motvaton for the research Research goal and approach Structure of ths thess Background Introducton Wreless networks The IEEE standard The protocol archtecture The MAC layer The Physcal Layer The Transmsson Control Protocol TCP Flow Control TCP Congeston Control Network smulator 2 (NS-2) Related work Conclusons The sngle-class sngle-cell model Introducton The sngle moble staton model Smulaton prelmnares Modelng the sngle moble staton case Numercal valdaton The multple statons sngle-class model Smulaton prelmnares Hdden node and RTS/CTS ssue Farness ssue Modelng multple statons case Smulaton valdaton Conclusons... 45

5 4 The mult-class sngle-cell model Introducton Model descrpton and analyss Smulaton valdaton Optmzaton of the model Fnd optmal locaton of Access Ponts Connecton Admsson control (CAC) Conclusons The multple-cell model Introducton The model of multple-cell and analyss Applcatons of the multple-cell model Performance evaluaton Optmal placement of Access Ponts Conclusons Conclusons Introducton Summary Contrbutons Future work.. 70 A Glossary 72 Appendx 74 Bblography 77 v

6 Lst of Fgures Fgure 1.1 Current wreless network systems. 2 Fgure 1.2 WLAN total publc users and total market penetraton. 2 Fgure 2.1 The protocol model 7 Fgure 2.2 Components of an WLAN system. 8 Fgure 2.3 Basc access mechansm 9 Fgure 2.4 InterFrame spacng n Fgure 2.5 The RTS/CTS access mechansm. 10 Fgure 2.6 Packetzaton on WLAN PHY layer wth Long Preamble and Header 11 Fgure 2.7 TCP congeston control mechansms 14 Fgure 3.1 Data transmsson n Fgure 3.2 TCP over WLAN connected to a Web Server Fgure 3.3 Throughput from smulaton results.. 25 Fgure 3.4 Average throughput: analyss versus smulaton Fgure 3.5 Hdden node problem n Fgure 3.6 No hdden nodes problem n NS Fgure 3.7 The moble staton s throughput as a functon of tme 30 Fgure 3.8 The statc staton s throughput as a functon of tme.. 30 Fgure 3.9 The total throughput.. 31 Fgure 3.10 Fracton of mean throughput Fgure 3.11 The three STAs ndvdual throughput as functon of tme.. 31 Fgure 3.12 The throughput sharng fracton between the statons 32 Fgure 3.13 The average throughput sharng fracton 32 Fgure 3.14 The moble staton s throughput. 33 Fgure 3.15 The total throughput of the AP. 33 Fgure 3.16 The average throughput sharng fracton 33 Fgure 3.17 Analyss system model 34 Fgure 3.18 State- transton-rate dagram for the nfnte-server case M/G/. 37 Fgure 3.19 The aggregated throughput from one run 39 Fgure 3.20 The average aggregated throughput of multple runs.. 39 Fgure 3.21 The average aggregated throughput of dfferent veloctes Fgure 3.22 Setup tme of each ndvdual staton.. 41 Fgure 3.23 Sojourn tme of each ndvdual staton Fgure 3.24 Setup tmes of each ndvdual staton wth dfferent veloctes. 42 Fgure 3.25 The setup tme as a functon of the average velocty.. 43 v

7 Fgure 3.26 Average throughput for each ndvdual staton wth dfferent veloctes Fgure 3.27 Model and smulaton s average throughput as a functon of the velocty.. 44 Fgure 3.28 The amount of receved data as a functon of the velocty.. 45 Fgure 4.1 Sngle access pont cell model 47 Fgure 4.2 The nstant aggregated throughput of two classes 51 Fgure 4.3 Setup tmes of each staton of two classes wth the same velocty Fgure 4.4 Setup tmes of each staton of two classes wth the dfferent veloctes Fgure 4.5 Average throughput for each staton wth velocty of 2 or Fgure 4.6 Average throughput for each staton wth velocty of 2 or Fgure 4.7 State- transton-rate dagram for the Erlang loss model Fgure 5.1 Example of the system topology of a multple cell scenaro. 59 Fgure 5.2 Cell dmensons. 63 Fgure 5.3 The shape of functon Fgure 5.4 E B ) as functon of λ E(S) for dfferent veloctes. 65 ( total Fgure 5.5 The total amount of receved data as functon of the number of cells (N) for dfferent veloctes. 66 v

8 Lst of Tables Table 1.1 WLAN standards versons.. 3 Table 3.1 The most mportant parameters for IEEE b. 25 Table 3.2 Comparson analyss vs. smulaton 26 Table 3.3 RTS/CTS solves the hdden node problem Table 3.4 Setup tme selectng rules Table 3.5 The parameters for smulatons.. 39 Table 3.6 The average aggregated throughput of dfferent veloctes 40 Table 3.7 The smulaton results of average setup tme for dfferent veloctes. 42 Table 3.8 The smulaton results vs. the model predcton of the average throughput Table 4.1 The parameters for smulatons and numercal results of the multple-class case. 50 Table 4.2 The aggregated throughput of the two classes case Table 4.3 The smulaton results vs. the model predcton of the average throughput Table 4.4 The smulaton results vs. the model predcton of total amount of receved data.. 55 Table 4.5 The total receved data wth dfferent gven AP locaton.. 55 Table 5.1 Numercal results of the optmzaton. 66 v

9 Chapter 1 Introducton 1.1 Introducton Ths chapter descrbes the global contents and context of ths Master s thess. The am of ths thess s to understand the mpact of the moblty of the moble termnal on performance of moble nternet servce va WLAN networks. Models are developed that measure performance as experenced by the end users n ths doman (n terms of throughput and total amount of receved data), for a gven set of desgn choces and realstc user-behavor scenaros. The followng paragraphs frst explan the research motvaton, after whch the research goals and the relevant research questons are presented. Ths chapter concludes by dscussng the structure of the remander of ths thess. 1.2 Motvaton for the research Over recent years, there has been an ncreasng trend towards personal computers and workstatons becomng portable and moble, and wreless technology now reaches or s capable of reachng vrtually every locaton on the face of the earth. Usng wreless network nterfaces, moble devces can be connected to the publc telephone network n the same way as wred telecommuncatons, or to the Internet n the same way as desktop computers are connected, usng the Ethernet, a token rng, or pont-to-pont lnks. Mllons of people exchange nformaton every day usng cellular telephones, and other wreless communcaton devces. The market for wreless communcatons has enjoyed tremendous growth. Wth the tremendous success of wreless telephony and messagng servces, t s hardly surprsng that moble Internet access servces (e.g., va cellular networks such as GPRS, UMTS, WLAN) whch provde Internet access anywhere and any tme are lkely to experence dramatc growth. Nowadays, there are three man wreless access technologes: Bluetooth, WLAN and 3G Wreless networks (e.g., GSM/GPRS, UMTS). Fgure 1.1 shows the comparson between them [Cesana 03]. The coverage of a transmttng node n a Bluetooth network s narrow (typcal maxmum range 10m), the coverage of a base staton n 3G Wreless networks s wde, (typcal maxmum range 1000m), and the coverage of an access pont n a Wreless Local Area Network (e.g., WLAN ) s medum, (typcal maxmum range 300m n outdoor scenaros). The typcal maxmum peak speed of Bluetooth s 1Mbps, the one of 3G s 115 Kbps, the one of b s 2 to 4 Mbps, and the one of a s 24 to 45 Mbps. Bluetooth s targeted on low costs, 10US$ n 2001 evolvng towards 5 US$ n Snce 3G moble cellular systems work n lcensed bands (e.g., 900, MHz), t s usually expensve for the servce provders to get the lcense. However, WLAN has moderate costs, 35 to 40 US$, because t works n the non-lcensed spectrum (2.4 and 5 GHz). Fast development of b WLANs has made b WLAN cards currently even cheaper than Bluetooth cards. Although 3G Wreless networks can provde hgh ubquty and moblty to customers and moble cellular technologes are evolvng from GPRS to UMTS, WLAN hotspots or networks are stll spreadng rapdly. They are cheap to nstall and offer much hgher data speeds than UMTS, although n a lmted area. 1

10 Fgure 1.1 Current wreless network systems WLAN technology was ntally desgned and mplemented for use nsde companes and buldngs, but later WLAN networks have spread out to publc places such as arports, hotels and cafes, whch see a hgh passage of people. The hotspots are deployed as network slands but n many places we are seeng hotspots mergng or expandng to cover parts of ctes or larger regons. W-F servces are even already avalable on nternatonal flghts operated by Lufthansa and Japan Arlnes. That s, passengers are able to check e- mal, send nstant messages and surf the Web at 30,000 feet. [Peters 05] Fgure 1.2 shows the number of WLAN total publc users, and the total market penetraton [Alonso 02]. It s predcted that there wll be n total 20.5 mllons WLAN users and the total market penetraton wll be 2.8 bllons US$. WLAN Total Publc Users WLAN Total Market Penetraton mllons bllons (US$) Fgure 1.2 WLAN total publc users and total market penetraton The wreless LAN ndustry s today one of the fastest growng segments of the communcatons ndustry. The frst generaton of wreless LAN specfcatons, referred to as , was developed for the 2.4 GHz mcrowave band and supported data speeds of 1-2Mbp/s. The ntroducton of the b specfcatons, known as "WF", sparked the launch of cheaper wreless LAN products, offerng data speeds of up to 11Mbp/s (although, 2

11 as wth all wreless LAN standards, data speeds dmnsh as more users work concurrently n each WLAN cell) [Pareek 05]. Today, the next generatons of a, g, and are amed at varous stages of development and mplementaton, and at ncreasng the speed, securty and qualty of wreless LANs. Ths s shown n Table 1.1 Untl now, all the standards from a to m have been released. Accordng to the Industral Economcs and Knowledge Center (IEK) of the Industral Technology Research Insttute (ITRI), n products wll become avalable n 2006 and start to take off n However, frst-generaton devces wll start emergng ths year. Supplers show keen nterest n the development of For nstance, Planex Communcaton Inc. has ntroduced a pre n-compatble WLAN card and a WLAN Access Pont (AP) n cooperaton wth Argo of the Unted States. Gemtek Technology Co. Ltd plans to ncrease ts R&D efforts on n and WMAX products durng CyberTan Technology Inc. wll develop WLAN models that adhere to both n and standards. D-Lnk Corp. has lkewse revealed plans to develop and launch n products wth enhanced user-frendly features [Tech 05]. Furthermore, several smaller ctes n US, such as Chaska, Mnn., have deployed ctywde W-F. Large ctes such as Phladelpha and San Francsco see wreless broadband technology as a low-cost soluton to provdng broadband access to low-ncome resdents. Phladelpha plans to have ts ctywde W-F network up and runnng by summer 2006 [Reardon 05]. In Adelade (Australa) the company Agle communcatons s also offerng WF throughout the cty for free [CtyLan]. Protocol Man Functon Protocol Man Functon verson verson a 5GHz OFDM PHY b 2.4GHz CCK PHY c brdgng d Internatonal roamng e QoS/effcency f Inter AP protocol enhancements g 2.4GHz OFDM PHY h 5GHz regulatory extensons Securty enhancements j Japan 5GHz band extensons k Rado resource measurement l Skpped (typographcally unsound) m Mantenance n Hgh throughput PHY Table 1.I WLAN standards versons Whereas wde area cellular networks are tradtonally speech-orented and WLANs are typcally nstalled as (extensons to wrelne) offce data networks, the evoluton of servces and the desre for seamless roamng underles the sgnfcant current nterest n the ntegraton of both network types. Swsscom [Swsscom 04] started to combne GPRS, UMTS and WLAN technologes, whch can offer users a natonwde broadband network. In addton, the moble termnal company DoCoMo [NTT 04] started to produce dual-mode 3G and wreless LAN phones. Ths s very useful for busness users wshng to work whle on the move (e.g., to access e-mals, ther company ntranet or the Internet) as well as for resdental users wth a need for bandwdth-ntensve applcatons such as the transfer of audo and vdeo fles. Moreover, a broadband wreless access technology WMAX (Worldwde Interoperablty for Mcrowave Access) s also developng, whch provdes hgh-throughput broadband connectons over long dstances. It s also called "802.16", whch seems to be a promsng extenson of WF, but wth a much broader coverage than WF (radus up to 50 km n rural areas), wth speeds up to 70 Mbps [McCullagh 05]. Whether t s 3G, WLAN or WMax, the customer s nterested n servces, not access technology. A key success factor n the commercal deployment of these servces s 3

12 performance, for example, n terms of throughput and total amount of retreved data. For nstance, GPRS can provde broad coverage and constant qualty of servce but the acheved bt rates are relatvely low. In WLAN on the other hand, the wreless channel has a hgher bt rate, but s shared among all users. Ths results n a hgh bt rate, f only a few users are served. Hence, for servce provders t s crucal to be able to predct the performance of ther cellular networks under varous scenaros, such as the number of subscrbers, the locatons of the base statons, and the moblty pattern of the users. Ths advocates the need for the development, valdaton and analyss of quanttatve models to descrbe and predct the performance experenced by the end user, for any gven set of system parameters. Examples of these parameters are: the locatons of the base statons, the number of subscrbers, the duratons of sessons, the download volume per sesson, the average load for each of the cells, etc. The IEEE standard s the de facto standard for wreless local area networks (WLANs). It has been deeply studed n the last few years, and a multplcty of papers has appeared n the lterature, coverng almost all aspects of the standard. A common problem n moble cellular networks s that the amount of transmsson capacty s lmted and shared among the dfferent users wthn the same cell. However, up to the knowledge of the authors, no satsfactory analytcal model of the nteracton between the user moblty and the network performance has been presented so far. 1.3 Research goal and approach The goal of ths project s to understand the mpact of the moblty of the moble termnal on performance of nternet access va cellular networks. To realze ths goal, the followng research questons are addressed. The key research questons can be stated as follows: How does the moblty of the user mpact the performance of hs moble nternet access? How can we use those valdated expressons to do the optmzatons? Although smulatons play an essental role n comparng competng desgn alternatves, smulatons of wreless network systems, especally wreless lnks, are sometmes unrelable. Ths may lead to ncorrect desgn choces. Generally, hgher accuracy n modelng requres more complexty of computaton. In ths case, t s desrable to have models wth low computaton complexty, whle mantanng the desred accuracy. If we can develop and study mathematcal models, t s not only fast but t can also gve us some nsght of the relatonshp between the factors n the network. Another advantage s that we can do what-f analyses,.e., we can predct the performance under any gven scenaro. Ths can never be done easly usng smulatons. To answer those research questons, quanttatve models are developed that predct the performance as experenced by the end users n ths doman (n terms of throughput and total amount of retreved data), for a gven set of desgn choces and realstc user-behavor scenaros. Smulatons are done to valdate those expressons of performance obtaned from the models. After that, system optmzaton can be done by usng those valdated approxmaton expressons for performance as a functon of system parameters. 4

13 1.5 Structure of ths thess The remander of ths thess s organzed n the followng way: The next chapter starts wth starts wth an ntroducton to the wreless networks. After ths, ntroductons of the IEEE protocol, the TCP protocol and the Network smulator are gven. After these ntroductons, the related work n ths area complements ths chapter. Chapter 3 frst derves the performance expresson for the sngle moble staton model, whch has been valdated by the smulatons. Based on the results of the prevous step, the multple statons n a sngle-class scenaro are analyzed. The expressons of the average throughput and the total amount of receved data of any number of statons are derved, whch also was thoroughly valdated by the smulatons. Chapter 4 focuses on the mult-class users n the sngle cell scenaro, that s, the users drve through one AP cell wth dfferent average veloctes. The approxmatng expressons of the performance are derved. Smulatons are run for valdatng the models. At the end of ths chapter, the practcal optmzaton usage of our model s analyzed. Chapter 5 focuses manly on nvestgated the performance of multple-class users across multple AP cells. We derved smple closed-form expressons for the average throughput and the total amount of receved data for any class of users. After that, we provded two examples on how our model can be used to analyze practcal problems. Chapter 6 presents the conclusons. Frst, a summary of the developed models are gven. Other topcs n ths chapter nclude a dscusson on what ths thess contrbutes to the research communty. The last paragraph focuses on future work. 5

14 Chapter 2 Background 2.1 Introducton Ths chapter starts wth an ntroducton to the wreless networks. Then, ntroductons of the IEEE protocols are gven n Secton 2.3. The related detals of the TCP protocol are present n Secton 2.4, and the ntroductons of Network smulator are gven n Secton 2.5. In Secton 2.6 related work n ths area complements ths chapter. 2.2 Wreless networks Wreless and network have become two commonly used words today. Many companes ether have products or present future products that are to be used n wreless network envronments. The concept of wreless networks ncorporates several dfferent networkng technologes, communcaton ranges and transmsson bandwdths. They range from local coverage networks (as IEEE ) to large wde area coverage networks, such as the thrd generaton moble telephony systems (e.g., GPRS, UMTS). The General Packet Rado Servce (GPRS) [ETSI 01] s part of the evoluton path towards thrd generaton (3G) moble systems. It s desgned to provde packet-based data servce n a cellular system based on GSM. It s an overlay to the crcut swtched GSMnetworks. GPRS ncreases the possble bandwdth for data transmsson by ntroducng packet swtchng. Inherently, packet-swtchng enables a more flexble and effcent utlzaton of the rado resources. In ths thess we concentrate on close range networks, whch are often called Wreless Local Area Networks (WLANs). Recently, hardware prces have dropped drastcally for nfrastructure equpment, and as a result of ths, WLANs are deployed almost everywhere. The most common standard for these networks today s the IEEE standard [IEEE 99]. There exst other standards such as HperLan/2 [HpGF], and HomeRF [HRF], but they are not so wdely used. WLAN usually operates on the lcense free ISM (Industry, Scence and Medcal) frequency band at 2.4GHz. And the extenson a s movng towards 5 GHz whch enables hgher transfer rates but decreases the coverage area. The transmsson bandwdth ranges from 1Mbt/s to approxmately 50Mbt/s and the possble communcaton dstance ranges from 30 to 1500m. 2.3 The IEEE standard Snce ths thess focuses on the TCP performance of the IEEE protocol, we only gve an explanaton of the general nformaton about standards that are related to our research. Further detals can be found n [IEEE 99] and [ b]. 6

15 The protocol archtecture The standard was frst standardzed by IEEE n 1997 and revsed n It defnes two layers, shown n Fgure 2.1. The frst layer s the Physcal layer (PHY), whch specfes the modulaton scheme used and sgnalng characterstcs for the transmsson through rado frequences. The second layer s the medum access control (MAC) layer. Ths layer determnes how the medum s used. Moblty of wreless statons may be the most mportant feature of a wreless LAN. A WLAN would not serve much purpose f statons were not able to move freely from locaton to locaton ether wthn a specfc WLAN or between dfferent WLAN segments. For compatblty purposes, the MAC must appear to the upper layers of the network as a standard 802 LAN. The MAC layer s forced to handle staton moblty n a fashon that s transparent to the upper layers of the 802 LAN stack. Ths forces functonalty nto the MAC layer that s typcally handled by upper layers. Fgure 2.1 The protocol model The archtecture s comprsed of several components and servces that nteract to provde staton moblty transparent to the hgher layers of the network stack, shown n Fgure 2.2. An network, n general, conssts of Basc Servce Sets (BSSs) that are nterconnected wth a Dstrbuton System (DS). Each BSS conssts of moble nodes (referred to as statons, Wreless LAN Staton (STA)). The staton s the most basc component of the wreless network. A staton s any devce that contans the functonalty of the protocol, e.g., a laptop PC, handheld devce, or an Access Pont. Statons n a BSS gan access to the DS and to statons n remote BSSs through an Access Pont (AP). The access ponts communcate amongst themselves to forward traffc from one BSS to another to facltate movement of statons between BSSs. The access pont performs ths communcaton through the dstrbuton system. The dstrbuton system s the backbone of the wreless LAN and may consst of ether a wred LAN or a wreless network. Before a staton can access the wreless medum t has to be assocated wth an AP. A staton can be assocated wth only one AP at any gven tme. A network of nterconnected BSSs, as n Fgure 2.2, n whch mobles can roam wthout loss n connectvty, s frequently referred to as an Extended Servce Set (ESS). The IEEE standard also specfes an addtonal ad-hoc archtecture for a WLAN. Ad-hoc WLANs are characterzed by lack of an AP, no functonalty to support moblty, and support for data transfer only between statons that belong to the same WLAN. Therefore, an ad hoc WLAN s smply an ndependent BSS where a staton communcates drectly wth one or more other statons. In our studes, we focus on the ESS as n Fgure

16 Fgure 2.2 Components of an WLAN system The MAC layer The MAC layer provdes functonalty to allow relable data delvery for the upper layers over the wreless PHY meda. The data delvery tself s based on an asynchronous, best-effort, connectonless delvery of MAC layer data. The standard provdes two modes of operaton. The basc access method n the MAC protocol s the Dstrbuted Coordnaton Functon (DCF) whch s best descrbed as the Carrer Sense Multple Access wth Collson Avodance (CSMA/CA) protocol [IEEE 99]. In addton to the DCF the also ncorporates an alternatve optonal access method known as the Pont Coordnaton Functon (PCF) an access method that s smlar to pollng and uses a pont coordnator (the AP) to determne whch staton has the rght to transmt. Ad-hoc WLANs support only the DCF snce they lack an AP. Most of the WLAN cards actually avalable on the market do not mplement PCF for complexty reasons. Snce the PCF s not the prmary concern of my study, ts operaton s not covered here. DCF offers two access methods; Basc Access (BA) and ready to send/clear to send (RTS/CTS). The basc access method When usng the DCF basc access mode, a staton, before ntatng a transmsson, senses the state of the channel to determne f another staton s transmttng. If the medum s determned to be dle for an nterval that exceeds the Dstrbuted InterFrame Spacng (DIFS), the staton proceeds wth ts transmsson. If the medum s busy, the staton defers untl after a DIFS s detected and then generates a random backoff perod for an addtonal deferral tme before transmttng. Ths mnmzes collsons durng contenton between multple statons. The backoff perod s used to ntalze the backoff tmer. The backoff tmer s decremented only when the medum s dle; t s frozen when the medum s busy. After a busy perod the decrementng of the backoff tmer resumes only after the medum has been free longer than DIFS. A staton ntates a transmsson when the backoff tmer reaches zero. The length of the backoff nterval, expressed n slots, s unformly chosen n the set {0,1,, CW-1}, where CW denotes the actual contenton wndow sze. At the begnnng, the contenton wndow takes an ntal value of CWmn for each frame queued for transmsson. To reduce the probablty of collsons, the contenton wndow of the statons nvolved n the collson s doubled and another transmsson attempt s made. The contenton wndow cannot grow ndefntely, but may reach a maxmum value of 2 γ CWmn = CWmax, and then the contenton wndow wll reman at CWmax for the remanng retres; 8

17 moreover, f a packet ncurs m collsons, where m γ, t s dropped. The set of CW values are 7 (CWmn), 15, 31, 63, 127, 255 (CWmax) [IEEE 99]. Snce no channel load sensng mechansm s provded,.e., a transmtter cannot determne f the data frame was successfully receved by lstenng to ts own transmsson as n wred LANs, an explct acknowledgement s necessary to nform the staton of the success/falure of ts transmsson. To do that, Immedate postve acknowledgements are employed to determne the successful recepton of each data frame. Ths s accomplshed by the recever ntatng the transmsson of an acknowledgement frame after a tme nterval Short InterFrame Spacng (SIFS), whch s less than DIFS, mmedately followng the recepton of the data frame. In case an acknowledgement s not receved the data frame s presumed lost and a retransmsson s scheduled (by the transmtter). Ths access method referred to as the Basc Access mechansm s summarzed n Fgure 2.3 [Ltjens 03_1]. Fgure 2.3 Basc access mechansm Fgure 2.4 InterFrame spacng n The standard encompasses also an Extended Interframe Spacng (EIFS), whch s used whenever the PHY ndcates to the MAC that a frame recepton was not successfully completed. EIFS s provded to let the transmttng staton retransmt the packet wthout nterference from the recevng staton. However, under the assumpton of deal channel condtons, EIFS does not have a great mpact on network performance. Ths s due to the fact that EIFS acts at the recever sde. As an example, consder a par of nodes exchangng data. In such a stuaton, collsons may clearly occur. However, snce a transmttng node cannot receve data, nodes wll become aware of the falure of a transmsson attempt only 9

18 by means of the ACK tmeout expraton. Thus, EIFS wll never be used. Hence, even f EIFS may actually be used when there are more than two nodes contendng for the channel, we wll, n the followng, neglect ts mpact on network performance. The dfferent Mac layer nterframe space tme s summarzed n Fgure 2.4 [IEEE 99]. RTS/CTS access method The DCF also provdes an alternatve way of transmttng data frames that nvolve transmsson of specal short Request To Send (RTS) and Clear To Send (CTS) frames pror to the transmsson of the actual data frame. Ths RTS/CTS mechansm s used to avod the well-known hdden termnal problem of CSMA-based MAC protocols. In the basc access method, a data frame could be corrupted by transmssons due to statons that are hdden from the transmtter any tme durng the transmsson of the data frame. The above leads to an ncreased probablty of collsons, and the transmsson tme wasted as a result of each collson. Ths wasted tme s sgnfcant snce the transmtter s made aware of a collson only after t tmes out watng for the correspondng acknowledgement. The RTS/CTS exchange attempts to perform a fast collson detecton and a transmsson path check. A successful exchange of RTS and CTS frames attempts to reserve the channel for the tme duraton needed to transfer the data frame under consderaton. The rules for the transmsson of an RTS frame are the same as those for a data frame under basc access,.e., the transmtter sends an RTS frame after the channel has been dle for a tme nterval exceedng DIFS. On recevng an RTS frame the recever responds wth a CTS frame (the CTS frame acknowledges the successful recepton of an RTS frame), whch can be transmtted after the channel has been dle for a tme nterval exceedng SIFS. After the successful exchange of RTS and CTS frames the data frame can be sent by the transmtter after watng for a tme nterval SIFS. In case a CTS frame s not receved wthn a predetermned tme nterval, the RTS s retransmtted followng the backoff rules as specfed n the basc access procedures outlned above. The channel access method usng RTS and CTS frames s summarzed n Fgure 2.5 [Ltjens 03_1]. Fgure 2.5 The RTS/CTS Access Mechansm The RTS and CTS frames contan a duraton feld that ndcates the perod the channel s to be reserved for transmsson of the actual data frame. Ths nformaton s used by statons that can hear ether the transmtter and/or the recever to update ther Net Allocaton Vectors (NAV) - a tmer that s always decreasng f ts value s non-zero. A staton s not allowed to ntate a transmsson f ts NAV s non-zero. The use of NAV to determne the busy/dle status of the channel s referred to as the Vrtual Carrer sense mechansm. Snce statons that can hear ether the transmtter or the recever resst from transmttng durng the transmsson of the data frame under consderaton the probablty of ts success s ncreased. However, ths ncrease n the probablty of successful delvery s acheved at the expense of the ncreased overhead nvolved wth the exchange of RTS and CTS frames, whch can be sgnfcant for short data frames. 10

19 The physcal layer The physcal layer (PHY) s the nterface between the MAC and the wreless meda where frames are transmtted and receved provdes three dfferent PHY defntons. Both Frequency Hoppng Spread Spectrum (FHSS) and Drect Sequence Spread Spectrum (DSSS) support 1 and 2 Mbps data rates. An extenson to the archtecture (802.11a) defnes dfferent multplexng technques that can acheve data rates up to 54 Mbps. Another extenson to the standard (802.11b) defnes 11 Mbps and 5.5 Mbps data rates (n addton to the 1 and 2Mbps rates) utlzng an extenson to DSSS called Hgh Rate-Drect Sequence Spread Spectrum (HR-DSSS), also known as Complementary Code Keyng (CCK). However, the MAC mechansms descrbed n the above secton are unchanged for IEEE b, so generally b can be treated as a legacy wth a hgher access rate [802.11b]. Currently b WLAN s wdely appled and has replaced most legacy WLANs. So we focus on the DSSS, and, furthermore, use the parameters descrbed for operatons n the 2.4 GHz ISM band, known as b. The PHY defnes a packet structure wth three major parts, the preamble, header, and payload. Two dfferent preambles and headers are defned: the mandatory Long Preamble and header and an optonal Short Preamble and header. Fgure 2.6 shows the format for the Long PHY protocol data unts (PPDU). Most b statons manly support only the Long Preamble and header mode. The Short Preamble and header mode s ntended for operatons where maxmum throughput between the statons s desred and nteroperablty wth legacy equpments s not a consderaton. Our research s manly based on the b standards and only the Long Preamble and header mode are consdered. Furthermore, dfferent parts are transferred wth dfferent data rates. For example, n the case of the Long Preamble and header a long preamble of 144 bts s used. Both preamble and header are transmtted wth a 1 Mbps data rate. The bts of the Payload (real data), however, use b, and can thus be transmtted wth 11Mbps. Fgure 2.6 Packetzaton on WLAN PHY layer wth Long Preamble and Header 11

20 2.4 The Transmsson Control Protocol Transport protocols have an mportant functon n data transmsson over the Internet. The protocols, whch are manly used, are the User Datagram Protocol (UDP) and the Transmsson Control Protocol (TCP). Snce wreless access to the Internet s the most common applcaton of ths knd of networks, t s reasonable to assume that the traffc wll be carred over the well-known TCP/IP protocol sute. Hence, we only explan the TCP protocols n more detals below. The transmsson control protocol [TCP] s the standard transmsson layer protocol for provdng a pont-to-pont full-duplex connecton-orented servce. In contrast to UDP, a TCP necesstates a connecton to be set up and kept alve between two hosts for sendng a certan amount of data. A three-way handshake s done to nform each other about allocated buffers (sender and recever) and ntal values of sequence numbers one for the sender and one for the recever. The sequence number s ncreased when each new segment sent and s used to keep track of the segments n the transmsson. The recever n turn, wll send socalled acknowledgement (ACK) packets. Wth the bult-n feedback mechansm, a TCP source can detect f a packet has been lost and has to be transmtted agan, thus provdng a relable transfer of data. The feedback mechansm also makes t possble for TCP to detect the network condtons, upon whch ts sendng rate s calculated. TCP adds more overhead to the network than UDP. Besdes the addtonal TCP ACK packets, TCP s header s also larger than UDP. To regulate the rate at whch segments are sent, TCP mplements two control mechansms, the flow control and the congeston control TCP flow control TCP s flow control s a technque whose prmary purpose s to properly match the transmsson rate of the sender to that of the recever and the network. TCP s flow control s end-to-end, whch means that the sender must not nject nto the network more than the recever can hold n ts buffer. It s mportant for the transmsson to be at a hgh enough rate to ensure good performance, but also to protect aganst overwhelmng the network or the recevng host. Ths control s done wth a wndow that lmts the number of packets the sender can nject nto the network and that sldes when the sender receves an acknowledgement sayng that the recever has correctly receved some data and handed them to the applcaton. A sendng staton s allowed to transmt more packets durng one RTT. The tme nterval between the moment a sgnal s sent and the moment a response s receved s defned as the Round-Trp Tme (RTT). The RTT contans two major components. The frst one s a fxed part that s caused by the propagaton delay. The second component s a varable part that s caused by queung delays, whch depend on the degree of congeston at the transmsson lnks. The maxmum number of outstandng packets, whch have been sent and are stll not acknowledged, s determned by a recever advertsed wndow, mantaned at the sender. The sze of the advertsed wndow s determned at the recever, by subtractng buffer space used from the buffer allocated. Ths wndow s advertsed by the recever at the connecton set-up perod and t s updated durng the connecton lfetme f the buffer space at the recever changes. Snce the sender knows the number of unacknowledged segments, t can compare ths wth the advertsed wndow to make sure that t does not overflow the recever buffer. 12

21 TCP congeston control TCP congeston control and Internet traffc management ssues n general are an actve area of research and expermentaton. Congeston control n data networks means that the dfferent sources must adapt ther transmsson rates as a functon of network load. The rate control n TCP s done by changng the sze of a congeston wndow that has been added to the protocol [Jacobson 88] and that also lmts the number of packets the source can transmt before the recept of an acknowledgement. Furthermore, ths wndow s upper bounded by the recever wndow for end-to-end flow control purposes [Stevens 97]. The congeston control mechansm of TCP conssts of four dfferent algorthms: slow start, congeston avodance, fast retransmt and fast recovery [Stevens 97]. They act together to try to avod overflowng ntermedate router buffers n the network between the sender and recever. Congeston control uses an addtonal congeston wndow, to regulate the senders transmsson rate. The amount of unacknowledged data must be less than or equal to the mnmum of congeston wndow and recever advertsed wndow.[jacobson 90]. Below we present a bref descrpton of these features. Slow start Slow Start s a mechansm used by the sender to control the transmsson rate, otherwse known as the sender-based flow control. Ths s accomplshed through the return rate of acknowledgements from the recever. In other words, the rate of acknowledgements returned by the recever determnes the rate at whch the sender can transmt data. As stated before, the sender n each TCP connecton mantans a wndow of the maxmum number of outstandng packets, the so-called congeston wndow (C wnd ). At the same tme there s also an advertsed wndow by the recever. When a new TCP connecton s establshed n a network, the Slow Start algorthm ntalzes the C wnd to one segment, whch s the maxmum segment sze (MSS) ntalzed by the recever durng the connecton establshment phase. When acknowledgements are returned by the recever, the congeston wndow ncreases by one segment for each acknowledgement returned. Thus, the sender can transmt the mnmum of the congeston wndow and the advertsed wndow of the recever, whch s smply called the transmsson wndow. Ths provdes an exponental ncrease of C wnd durng slow start n each RTT. The maxmum number of ongong packets s ncreased accordngly. Congeston avodance Durng the ntal data transfer phase of a TCP connecton the Slow Start algorthm s used. However, there may be a pont durng the Slow Start that the network s forced to drop one or more packets due to overload or congeston. If ths happens, Congeston Avodance s used to slow the transmsson rate. In the Congeston Avodance algorthm a retransmsson tmer exprng or the recepton of duplcate ACKs can mplctly sgnal the sender that a network congeston stuaton s occurrng. If the sender detects that the lnk s congested, TCP s Congeston Avodance algorthm apples. At the establshment of a TCP connecton the SlowStart Threshold (SSThresh) s set to a default value, whch s dependent on the TCP mplementaton. At the moment that the sender s aware of lnk congeston,.e., tme-outs or duplcated ACKs have occurred, t sets SSThresh to one-half of the current wndow sze C wnd (the mnmum of the congeston wndow and the recever's advertsed wndow sze) and sets C wnd to one. It means that the TCP source has to do slow start agan. At a gven moment C wnd wll be greater than SSThresh (half way to the prevous congeston pont), and then collson avodance wll take over. From now on C wnd wll ncrease by 1/ C wnd at each successful recepton of each ACK. Ths results n an ncrease of one C wnd n each RTT, as opposed to the exponental ncrease durng slow start. If congeston was ndcated by duplcate ACKs, the Fast Retransmt (and/or Fast Recovery) algorthms are nvoked (see below). 13

22 Fast retransmt When a duplcate ACK s receved, the sender does not know f t s because a TCP segment was lost or smply that a segment was delayed and receved out of order at the recever. If the recever can re-order segments, t should not be long before the recever sends the latest expected acknowledgement. Typcally no more than one or two duplcate ACKs should be receved when smple out of order condtons exst. If however more than two duplcate ACKs are receved by the sender, t s a strong ndcaton that at least one segment has been lost. In ths case, the sender does not even wat for a retransmsson tmer to expre before retransmttng the segment (as ndcated by the poston of the duplcate ACK n the byte stream). Ths process s called the Fast Retransmt algorthm and was frst defned n [Jacobson 90]. In Fgure 2.7 the slow start, collson avodance and fast retransmt mechansm of a TCP connecton are llustrated. Fgure 2.7 TCP congeston control mechansms 2.5 Network smulator 2 (NS-2) In most studes on moble ad hoc networks, smulatons are used for the evaluaton of network protocols and devces under certan specfc condtons. Of the dfferent academc and commercal network smulaton tools, n ths project, we wll focus on one such academc network smulaton tool called the Network Smulator or mostly referred to as NS-2. The Network Smulator 2 (NS-2) s commonly used and wdely accepted n the networkng research communty as the basc smulatng tool for network evaluatons, mostly because t s open-source [Fall 97], [NS-2]. NS-2 s a dscrete event smulator targeted at networkng research. NS-2 provdes substantal support for smulaton of TCP, routng, and multcast protocols over wred and wreless (local and satellte) networks. Because of ts good TCP mplementatons, t has been often used to smulate and evaluate TCP performance as well. 14

23 Berkeley released the ntal code that made wreless network smulatons possble n NS-2. That code provded some support to model wreless LANs, but was farly lmted. As a result of the Monarch project at Carnege Mellon Unversty [CMU] the smulator was extended wth support for node moblty, a realstc physcal layer, rado network nterfaces and an mplementaton of the IEEE DCF MAC protocol. Ths work was presented as part of a larger study of performance for dfferent ad-hoc routng protocols. It was ths contrbuton that made t possble to perform real wreless smulatons [Broch 99] wth NS. NS-2 s an event based smulator, where a scheduler actvates events sequentally. The dea of a dscrete event scheduler s that actons may only be started as a result of an event. At present, there are four dfferent schedulers whch use dfferent data structures mplemented n NS. Usually the scheduler selects the next event from the event queue and fnshes t. After that the next event s taken and so on. There are no parallel operatons n NS. If two or more events have same startng tme, the scheduler selects the event that was frst put nto the event queue. NS-2 s a hybrd of C++ and OTcl (Object Tcl) programmng languages. OTcl works manly as a confguraton and command nterface but s also a part of an mplementaton of NS. Ths 'two language'- dlemma s also the orgn of the some problems encountered n NS; whle C++ has to be compled, OTcl s nterpreted. One great beneft of ths s that there s no need to recomple the smulator between dfferent smulatons snce you are able to set up topology, lnk bandwdth, traffc sources, etc. from the OTcl scrpts. So, once you have mplemented the basc functonalty wthn the smulator (C++ code) you only have to change the OTcl scrpts to run varous smulatons. The underlyng channel model n NS-2 s qute smple. Each moble node has a poston and a velocty and moves around on a topography that s specfed usng ether a dgtal elevaton map or a flat grd. The poston of a moble node can be calculated as a functon of tme, and s used by the rado propagaton model to calculate the propagaton delay from one node to another and to determne the power level of a receved sgnal at each moble node. Each moble node has one or more wreless network nterfaces, wth all nterfaces of the same type (on all moble nodes) lnked together by a sngle physcal channel. When a network nterface transmts a packet, t passes the packet to the approprate physcal channel object. Ths object then computes the propagaton delay from the sender to every other nterface on the channel and schedules a packet recepton event for each. Ths event notfes the recevng nterface that the frst bt of a new packet has arrved. At ths tme, the smulator calculates the recevng power Pr for every transmsson between two nodes wth the chosen propagaton model. The power level s compared to two dfferent values: the carrer sense threshold CS Thresh and the receve threshold RX Thresh. The channel model dstngushes prmarly between three cases. In case Pr s greater than the recevng threshold RX Thresh, the transmsson has enough power to allow proper recepton at the recever sde. Other smultaneous transmssons wth reasonable transmsson powers may certanly nterfere wth ths transmsson and make a correct recepton mpossble. Then the packet s smply handed up to the MAC layer. If Pr s below RX Thresh but greater than the carrer sense threshold CS Thresh, the packet s marked as a packet n error before beng passed to the MAC layer and the recevng node must drop the packet. However, the recevng power of ths transmsson s stll strong enough to nterfere wth other smultaneous transmssons. Consequently, these nterfered packets are also nvald and nodes must drop them as well. Transmssons wth recevng powers Pr smaller than CS Thresh do not even obstruct other smultaneous transmssons at the same node. And the packet s dscarded as nose. 15

24 To allow reasonable smulatons wthn an acceptable amount of tme, propagaton models must smplfy calculatons and reduce the requred computaton to a mnmum. The network smulator NS-2 knows three dfferent propagaton models to smulate wreless ad hoc networks, the free space (FS) model, the two ray ground (TRG) model and the shadowng model. The Rcean and Raylegh model s extended by the ARC Group of Carnege Mellon Unversty [Punnoose 00]. In our research s smulaton, the TRG model s used. More detals are descrbed n the next chapter. Once the MAC layer receves a packet, t checks to nsure that ts receve state s presently dle. If the recever s not dle, one of the followng two thngs can happen. If the power level of the packet already beng receved s at least 10 db greater than the receved power level of the new packet, we assume that capture ths packet, dscard the new packet, and allow the recevng nterface to contnue wth ts current receve operaton. Otherwse, a collson occurs and both packets are dropped. If the MAC layer s dle when an ncomng packet s passed up from the network nterface, t smply computes the transmsson tme of the packet and schedules a packet recepton complete event for tself. When ths event occurs, the AC layer verfes that the packet s error-free, performs destnaton address flterng, and passes the packet up the protocol stack. The lnk layer mplements the complete IEEE standard [IEEE 99] Medum Access Control (MAC) protocol Dstrbuted Coordnaton Functon (DCF) n order to accurately model the contenton of nodes for the wreless medum. DCF s desgned to use both physcal carrer sense and vrtual carrer sense mechansms to reduce the probablty of collsons due to hdden termnals. The transmsson of each uncast packet s preceded by a Request-to-Send/Clear-to-Send (RTS/CTS) exchange that reserves the wreless channel for transmsson of a data packet. Each correctly receved uncast packet s followed by an Acknowledgment (ACK) to the sender, whch retransmts the packet a lmted number of tmes untl ths ACK s receved. Broadcast packets are sent only when vrtual and physcal carrer sense ndcates that the medum s clear, but they are not preceded by RTS/CTS and are not acknowledged by ther recpents. 2.6 Related work A lot of research s done n the feld of performance modelng of the IEEE protocol. Most of the earler work s focused on the smulaton of the effcency of the IEEE protocol, by nvestgatng the maxmum throughput that t can acheve under varous network confguratons, e.g., [Wenmller 97]. Banch developed and analyzed a detaled mathematcal performance model of DCF [Babch 00], whch has been mproved later by Wu et al [Wu 02]. Both papers have studed the saturaton throughput of IEEE s MAC layer, whch represents the maxmum load the system can carry n stable condtons. Next to ths, both assume a smplfed physcal layer model. They propose to model the MAC access mechansm wth a Markov Chan, neglectng only mnor dependences among the behavor of dfferent statons. Ths yelds an accurate approxmaton for the WLAN saturaton throughput. However, persstent UDP flows nstead of TCP are studed n ther work, whch means that the sources always have packets ready to send n ther queue. Based on ths Markov Chan model, the stuaton wth non-persstent traffc sources s studed n [Ltjens 03_2] and [Wnands 04]. The number of actve statons vares dynamcally n tme accordng to the ntaton and completon of fle transfers at random tme nstants. In ther research they have proposed an ntegrated packet/flow level modelng approach. On packet level the WLAN DCF performance can be modelled usng a dscrete 16