USF Libraries
USF Digital Collections

Evaluation of available bandwidth estimation tools (abets) and their application in improving tcp performance

MISSING IMAGE

Material Information

Title:
Evaluation of available bandwidth estimation tools (abets) and their application in improving tcp performance
Physical Description:
Book
Language:
English
Creator:
Easwaran, Yegyalakshmi
Publisher:
University of South Florida
Place of Publication:
Tampa, Fla.
Publication Date:

Subjects

Subjects / Keywords:
Bandwidth measurement
Igi
Pathload
Pathchirp
Congestion control
Tcp
Available bandwidth
Dissertations, Academic -- Computer Science and Engineering -- Masters -- USF   ( lcsh )
Genre:
government publication (state, provincial, terriorial, dependent)   ( marcgt )
bibliography   ( marcgt )
theses   ( marcgt )
non-fiction   ( marcgt )

Notes

Summary:
ABSTRACT: Available bandwidth is a time-dependant variable that defines the spare bandwidth in an end-to-end network path. Currently, there is significant focus in the research community on the design and development of Available Bandwidth Estimation Tools (ABETs), and a few tools have resulted from this research. However, there is no comprehensive evaluation of these tools and the research work in this thesis attempts to fill that gap. A performance evaluation of important ABETs like Pathload, IGI and pathChirp in terms of their accuracy, convergence time and intrusiveness is conducted in several scenarios. A 2k factorial design is carried out to analyze the importance of the size of probe packets, number of probe packets per train, number of trains, and frequency of runs in these performance metrics. ABETs are very important because of their potential in solving many network research problems.For example, ABETs can be used in congestion control in transport layer protocols, network management tools, route selection and configuration in overlay networks, SLA verification, topology building in peer to peer networks, call admission control, dynamic encoding rate modification in streaming applications, traffic engineering, capacity planning, intelligent routing systems, etc. This thesis looks at applying ABETs in the congestion control of transmission control protocol (TCP).Current implementations of TCP in the Internet perform reasonably well in terms of containing congestion, but their sending rate adjustment algorithm is unaware of the accurate network conditions and available resources. TCP's Additive Increase Multiplicative Decrease (AIMD) congestion control algorithm cannot efficiently utilize the available bandwidth to the full potential and this is especially true in high bandwidth networks.
Thesis:
Thesis (M.S.C.S.)--University of South Florida, 2005.
Bibliography:
Includes bibliographical references.
System Details:
System requirements: World Wide Web browser and PDF reader.
System Details:
Mode of access: World Wide Web.
Statement of Responsibility:
by Yegyalakshmi Easwaran.
General Note:
Title from PDF of title page.
General Note:
Document formatted into pages; contains 56 pages.

Record Information

Source Institution:
University of South Florida Library
Holding Location:
University of South Florida
Rights Management:
All applicable rights reserved by the source institution and holding location.
Resource Identifier:
aleph - 001681070
oclc - 62744903
usfldc doi - E14-SFE0001048
usfldc handle - e14.1048
System ID:
SFS0025369:00001


This item is only available as the following downloads:


Full Text
xml version 1.0 encoding UTF-8 standalone no
record xmlns http:www.loc.govMARC21slim xmlns:xsi http:www.w3.org2001XMLSchema-instance xsi:schemaLocation http:www.loc.govstandardsmarcxmlschemaMARC21slim.xsd
leader nam Ka
controlfield tag 001 001681070
003 fts
005 20060215071209.0
006 m||||e|||d||||||||
007 cr mnu|||uuuuu
008 051227s2005 flu sbm s000 0 eng d
datafield ind1 8 ind2 024
subfield code a E14-SFE0001048
035
(OCoLC)62744903
SFE0001048
040
FHM
c FHM
049
FHMM
090
QA76 (Online)
1 100
Easwaran, Yegyalakshmi.
0 245
Evaluation of available bandwidth estimation tools (abets) and their application in improving tcp performance
h [electronic resource] /
by Yegyalakshmi Easwaran.
260
[Tampa, Fla.] :
b University of South Florida,
2005.
502
Thesis (M.S.C.S.)--University of South Florida, 2005.
504
Includes bibliographical references.
516
Text (Electronic thesis) in PDF format.
538
System requirements: World Wide Web browser and PDF reader.
Mode of access: World Wide Web.
500
Title from PDF of title page.
Document formatted into pages; contains 56 pages.
520
ABSTRACT: Available bandwidth is a time-dependant variable that defines the spare bandwidth in an end-to-end network path. Currently, there is significant focus in the research community on the design and development of Available Bandwidth Estimation Tools (ABETs), and a few tools have resulted from this research. However, there is no comprehensive evaluation of these tools and the research work in this thesis attempts to fill that gap. A performance evaluation of important ABETs like Pathload, IGI and pathChirp in terms of their accuracy, convergence time and intrusiveness is conducted in several scenarios. A 2k factorial design is carried out to analyze the importance of the size of probe packets, number of probe packets per train, number of trains, and frequency of runs in these performance metrics. ABETs are very important because of their potential in solving many network research problems.For example, ABETs can be used in congestion control in transport layer protocols, network management tools, route selection and configuration in overlay networks, SLA verification, topology building in peer to peer networks, call admission control, dynamic encoding rate modification in streaming applications, traffic engineering, capacity planning, intelligent routing systems, etc. This thesis looks at applying ABETs in the congestion control of transmission control protocol (TCP).Current implementations of TCP in the Internet perform reasonably well in terms of containing congestion, but their sending rate adjustment algorithm is unaware of the accurate network conditions and available resources. TCP's Additive Increase Multiplicative Decrease (AIMD) congestion control algorithm cannot efficiently utilize the available bandwidth to the full potential and this is especially true in high bandwidth networks.
590
Adviser: Dr. Miguel Labrador.
Co-adviser:
653
Bandwidth measurement.
Igi.
Pathload.
Pathchirp.
Congestion control.
Tcp.
Available bandwidth.
690
Dissertations, Academic
z USF
x Computer Science and Engineering
Masters.
773
t USF Electronic Theses and Dissertations.
4 856
u http://digital.lib.usf.edu/?e14.1048



PAGE 1

Ev aluation of A v ailable Bandwidth Estimation T ools (ABETs) and Their Application in Impro ving TCP Performance by Y e gyalakshmi Easw aran A thesis submitted in partial fulllment of the requirements for the de gree of Master of Science in Computer Science Department of Computer Science and Engineering Colle ge of Engineering Uni v ersity of South Florida Major Professor: Miguel A. Labrador Ph.D. K en Christensen Ph.D. N. Ranganathan, Ph.D. Date of Appro v al: March 04, 2005 K e yw ords: Bandwidth Measurement, IGI, P athload, pathChirp, Congestion control c r Cop yright 2005, Y e gyalakshmi Easw aran

PAGE 2

DEDICA TION T o my w onderful husband Bala, f amily and friends

PAGE 3

A CKNO WLEDGEMENTS I w ould lik e to e xtend my most sincere thanks to Dr Miguel Labrador for his guidance and support throughout this research w ork. It has been an enjo yable and re w arding e xperience w orking with him and learning a great deal in the process. He has been a source of inspiration for me and has helped me bring out my best. I tak e this opportunity to thank Dr K en Christensen and Dr N. Ranganathan for taking the time to re vie w and e xamine this thesis. I w ould also lik e to thank Dr Kaushal Chari at the ISDS department in the Colle ge of Business for gi ving me the opportunity to assist him in his research w ork and gain in v aluable e xperience. I cannot thank my husband Bala enough for encouraging and supporting me and belie ving in me. Finally I w ould lik e to thank my parents and my f amily for all their lo v e and support in all of my endea v ors.

PAGE 4

T ABLE OF CONTENTS LIST OF T ABLES iii LIST OF FIGURES i v ABSTRA CT v CHAPTER 1 INTR ODUCTION 1 1.1 Background and moti v ation 1 1.2 Problem statement 2 1.3 Contrib utions of this thesis 2 1.4 Or ganization of this document 3 CHAPTER 2 B A CKGR OUND AND LITERA TURE REVIEW 4 2.1 ABETs 7 2.1.1 IGI (Initial Gap Increasing) 7 2.1.2 P athload 10 2.1.3 pathChirp 12 2.2 TCP 13 2.2.1 TCP Reno 13 2.2.2 TCP Sack 14 CHAPTER 3 METHODOLOGY 16 3.1 Simulation design 16 3.1.1 Case 1: Lo w bandwidth scenario 16 3.1.2 Case 2: High bandwidth scenario 17 3.1.3 Case 3: W ith CBR staircase traf c 18 3.2 Design methodology 19 CHAPTER 4 PERFORMANCE EV ALU A TION OF ABETS 21 4.1 Case 1: Lo w bandwidth scenario 22 4.1.1 Analysis of IGI 22 4.1.2 Analysis of P athload 24 4.1.3 Analysis of pathChirp 25 4.2 Case 2: High bandwidth scenario 28 4.2.1 Analysis of IGI 28 4.2.2 Analysis of P athload 30 4.2.3 Analysis of pathChirp 30 4.3 Analysis of ABETs with CBR staircase 33 4.3.1 Conclusions from the analysis of IGI, P athload and pathChirp tools 34 i

PAGE 5

CHAPTER 5 APPLICA TION OF ABETS IN END T O END CONGESTION CONTR OL 35 5.1 TCP congestion control 35 5.1.1 Selection of ABET for TCP congestion control application 36 5.1.2 Algorithm for ABET -based TCP Sack 36 5.1.2.1 TCP W estw ood 37 5.1.2.2 Simulations in lo w and high bandwidth netw orks 38 5.1.3 Areas for future research 43 CHAPTER 6 CONCLUSION AND FUTURE W ORK 45 REFERENCES 46 ii

PAGE 6

LIST OF T ABLES T able 4.1 Case 1: 2 k f actorial design using IGI in lo w bandwidth netw orks 23 T able 4.2 Case 1: 2 k f actorial design using P athload in lo w bandwidth netw orks 24 T able 4.3 Case 1: 2 k f actorial design using pathChirp in lo w bandwidth netw orks 26 T able 4.4 F actors ef fects using IGI in lo w bandwidth netw orks 26 T able 4.5 F actors ef fects using P athload in lo w bandwidth netw orks 27 T able 4.6 F actors ef fects using pathChirp in lo w bandwidth netw orks 27 T able 4.7 Case 2: 2 k f actorial design using IGI in high bandwidth netw orks 28 T able 4.8 Case 2: 2 k f actorial design using P athload in high bandwidth netw orks 29 T able 4.9 Case 2: 2 k f actorial design using pathChirp in high bandwidth netw orks 31 T able 4.10 F actors ef fects using IGI in high bandwidth netw orks 31 T able 4.11 F actors ef fects using P athload in high bandwidth netw orks 32 T able 4.12 F actors ef fects using pathChirp in high bandwidth netw orks 32 T able 4.13 Accurac y measurements for ABETs with CBR staircase cross traf c 33 T able 4.14 Con v er gence time (Secs) measurements for ABETs with CBR stair case cross traf c 33 T able 4.15 Intrusi v eness measurements for ABETs with CBR staircase cross traf c 33 iii

PAGE 7

LIST OF FIGURES Figure 2.1 P ack et pair probe gap model [1] 7 Figure 2.2 P athload mechanism [2 ] 10 Figure 2.3 pathChirp queueing delay signature [3 ] 12 Figure 3.1 Lo w bandwidth netw ork topology 17 Figure 3.2 High bandwidth netw ork topology 17 Figure 3.3 Simulation with CBR stair cross traf c 18 Figure 5.1 Simulation topology for application of ABET in lo w bandwidth netw orks 39 Figure 5.2 Lo w bandwidth netw ork Instantaneous and a v erage throughput of TCP Sack 39 Figure 5.3 Lo w bandwidth netw ork Instantaneous and a v erage throughput of ABET -based TCP using IGI 40 Figure 5.4 Simulation topology for application of ABET in high bandwidth netw orks 40 Figure 5.5 High bandwidth netw ork Instantaneous and a v erage throughput of TCP Sack 41 Figure 5.6 High bandwidth netw ork Instantaneous and a v erage throughput TCP SA CK ABET -based TCP using IGI 42 Figure 5.7 Instantaneous and a v erage throughput of ABET -based TCP using P athload 43 i v

PAGE 8

EV ALU A TION OF A V AILABLE B AND WIDTH ESTIMA TION T OOLS (ABETS) AND THEIR APPLICA TION IN IMPR O VING TCP PERFORMANCE Y egyalakshmi Easwaran ABSTRA CT A v ailable bandwidth is a time-dependant v ariable that denes the spare bandwidth in an end-toend netw ork path. Currently there is signicant focus in the research community on the design and de v elopment of A v ailable Bandwidth Estimation T ools (ABETs), and a fe w tools ha v e resulted from this research. Ho we v er there is no comprehensi v e e v aluation of these tools and the research w ork in this thesis attempts to ll that gap. A performance e v aluation of important ABETs lik e P athload, IGI and pathChirp in terms of their accurac y con v er gence time and intrusi v eness is conducted in se v eral scenarios. A 2 k f actorial design is carried out to analyze the importance of the size of probe pack ets, number of probe pack ets per train, number of trains, and frequenc y of runs in these performance metrics. ABETs are v ery important because of their potential in solving man y netw ork research problems. F or e xample, ABETs can be used in congestion control in transport layer protocols, netw ork management tools, route selection and conguration in o v erlay netw orks, SLA v erication, topology b uilding in peer to peer netw orks, call admission control, dynamic encoding rate modication in streaming applications, traf c engineering, capacity planning, intelligent routing systems, etc. This thesis looks at applying ABETs in the congestion control of transmission control protocol (TCP). Current implementations of TCP in the Internet perform reasonably well in terms of containing congestion, b ut their sending rate adjustment algorithm is una w are of the accurate netw ork conditions and a v ailable resources. TCP' s Additi v e Increase Multiplicati v e Decrease (AIMD) congestion control algorithm cannot ef ciently utilize the a v ailable bandwidth to the full potential and this is especially true in high bandwidth netw orks. Based on the results of the comparati v e e v aluation, the most appropriate ABET for TCP congestion control is embedded in a modied v ersion of TCP Sack v

PAGE 9

to solv e the ”blindness” of TCP in changing its congestion windo w and threshold v alues. It is sho wn that using the a v ailable bandwidth estimates pro vided by IGI instead of the ”by half” reduction rule of TCP the throughput of the proposed ABET -based TCP v ersion is impro v ed compared to re gular TCP Sack. vi

PAGE 10

CHAPTER 1 INTR ODUCTION 1.1 Backgr ound and moti v ation ”A v ailable Bandwidth”, or the maximum unused bandwidth of a link or path is a v ery impor tant metric. Se v eral tools and techniques lik e IGI [1 ], P athload [2], pathChirp [3 ], spruce [4] and T OPP [5 ] ha v e been proposed and de v eloped to estimate the a v ailable bandwidth in an end-to-end path. Although the potential applicability of a v ailable bandwidth estimation tools is v ery high, there are tw o main problems that still limit their usage. These tools are still under in v estigation and we don' t kno w ho w well these tools perform and the main f actors af fecting their performance. Also, there is little kno wledge about w ays to use these tools in dif ferent scenarios or to inte grate these tools into other applications and possibly address netw orking problems. In man y cases, these techniques might need to be adapted to the specic application or problem at hand. These tw o f acts were recently recognized in [6], where the authors say that more research is needed to not only impro v e these techniques, b ut also to determine w ays to use these techniques in other applications. A v ailable bandwidth in an end-to-end netw ork path is hard to obtain for end users and only a v ailable to netw ork administrators who ha v e access to intermediate routers and switches. Lack of administrati v e access to netw ork management protocols limits the end users from ha ving infor mation about the a v ailable bandwidth of an end-to-end path. A v ailable bandwidth is a metric that dynamically changes o v er time and so the measurement has to be an a v erage of se v eral instantaneous v alues o v er a period of time and the measurement has to be done quickly and used immediately so as not to lose its v alidity A v ailable Bandwidth Estimation T ools (ABETs) need to be studied and understood better to be able to utilize them suitably A scenario where their application will ha v e a signicant impact is in the predominant transport layer in the Internet the T ransmission Control Protocol (TCP). 1

PAGE 11

1.2 Pr oblem statement The recent years ha v e seen considerable research ef forts going into the de v elopment of a fe w a v ailable bandwidth estimation tools (ABETs) [1 ] [5 ]. Ho we v er there e xists no comprehensi v e performance e v aluation of these tools and techniques. Furthermore, it is not kno wn as to ho w the v arious tool settings/parameters af fect the tool performance in terms of accurac y con v er gence time and intrusi v eness. IGI/PTR, P athload and pathChirp are the three ABETs included in the e v aluation as the y are among the better tools and their ns-2 [7 ] modules were a v ailable to perform an e v aluation. The e v aluation done in this thesis will serv e as a guide in understanding and picking suitable ABETs to use in a gi v en scenario or application. The ABETs ha v e se v eral potential applications b ut there is limited kno wledge about utilizing the ABETs in applications to solv e netw orking issues. The most common transport layer protocol in the Internet is TCP and the congestion control algorithm that it uses has some dra wbacks that cause TCP to under -perform. A kno wledge of the a v ailable bandwidth estimate could potentially impro v e the ef cienc y of TCP in ne xt generation high speed netw orks where it has been demonstrated that the ”blindness” of TCP causes serious performance problems [8 ]. The suitability of usage of ABETs in an application depends on the performance characteristics of the ABETs and the needs of the application and matching the tw o optimally 1.3 Contrib utions of this thesis This thesis mak es contrib utions in the areas of a v ailable bandwidth estimation tools and TCP A performance e v aluation of the a v ailable bandwidth estimation tools and a one to one comparison between the tools is carried out in lo w and high bandwidth netw ork scenarios. Experiments to measure the accurac y con v er gence time and intrusi v ess metrics are carried out for IGI/PTR, P athload and pathChirp in the presence of 1) TCP traf c alone; 2)TCP and continuous CBR traf c; 3)TCP and ON-OFF CBR traf c and 4) CBR Staircase traf c. The abo v e e v aluations are done in both lo w and high bandwidth netw orks and a 2 k f actorial design is done to determine the important f actors af fecting these ABETs. An analysis of the results from the e v aluation is used to determine the most appropriate ABET to aid in TCP congestion control. An ABET -based TCP that demonstrates superior performance is 2

PAGE 12

designed by inte grating the most appropriate ABET identied by the e v aluation with TCP Sack. It is also sho wn that although P athload and IGI are f airly good bandwidth estimation tools, IGI is better suited to impro ving TCP performance as compared to P athload. In f act, P athload when incorporated in TCP w orsens its performance. 1.4 Or ganization of this document The rest of the thesis is or ganized as follo ws. Chapter 2 describes the dif ferent types of bandwidth-related netw ork measurements and the tools and techniques that are a v ailable to measure them. It also contains a description of IGI/PTR, P athload and pathChirp, which are the ABETs considered in this thesis for e v aluation. Chapter 3 e xplains the e xperimental setup and topology for the simulations and the design methodology used. In Chapter 4, a performance e v aluation of the three ABETs is included in lo w and high bandwidth scenarios using a 2 k f actorial design. In Chapter 5, an ABET -based TCP is proposed and its performance is compared with re gular TCP Sack. Lastly Chapter 6 concludes the thesis with future directions for research in this area. 3

PAGE 13

CHAPTER 2 B A CKGR OUND AND LITERA TURE REVIEW There are a number of bandwidth related netw ork measurements. A fe w important ones are end-to-end capacity b ulk transfer capacity and a v ailable bandwidth [6]. The end-to-end capacity of a netw ork path is the maximum rate that can be achie v ed at the IP layer when there is no other traf c in the path. In other w ords, the capacity of a path is the minimum of the capacities of all the links in the path. The link in the path that has the minimum capacity is called the narro w link. The link that has the minimum a v ailable bandwidth is called the tight link. The capacity of the narro w link determines the capacity of the path and the link with the maximum free bandwidth determines the a v ailable bandwidth in the path. The Bulk T ransfer Capacity (BTC) of a path is the throughput of a b ulk TCP transfer A v ailable bandwidth in a netw ork is the maximum free unused throughput a v ailable to a o w without af fecting the throughput of an y traf c that is currently in the path. P athchar [9 ] and pchar [10 ] measure the link bandwidth of all the links in the path and not just the bottleneck bandwidth link. These tools use a common underlying measurement methodology called v ariable pack et size (VPS) probing. The k e y assumption in VPS is that each hop of a path increases the one-w ay delay of a pack et by a serialization latenc y gi v en by the ratio of the pack et size o v er that hop' s capacity If this is true, the VPS technique can estimate the capacity of a hop i based on the relation between the Round T rip T ime (R TT) up to hop i and the probing pack et size. The R TTs for dif ferent pack et sizes can be measured using T ime Exceeded ICMP messages as done by traceroute. Another technique of measuring the link bandwidth is the pack et pair method used by tools lik e nettimer [11 ] and bprobe [12 ]. P ack et pair methods send pairs of back-to-back pack ets and measure the dif ference in the pack et gaps at the source and destination. The bottleneck link determines the interpack et spacing and the links with higher bandwidth preserv e the interpack et spacing. 4

PAGE 14

A v ailable bandwidth depends on both the bottleneck link bandwidth and the cross traf c in a netw ork path. A v ailable bandwidth is a time-dependant metric representing the unused capacity in a link during a specic period. A v ailable bandwidth measurements use pack et trains which are streams of pack ets, instead of pack et pairs so as to obtain an a v eraged v alue o v er a period of time. The rst fe w tools that measured the a v ailable bandwidth are cprobe [13 ] and pipechar [14 ]. These tools send a pair of pack ets and measure the pack et dispersion, which is the interspacing between the pack ets at the destination and compute the a v ailable bandwidth as the ratio of pack et length to the pack et dispersion. It w as later identied in [15 ] that this is actually the asymptotic pack et dispersion rate (ADR) and not the a v ailable bandwidth. T OPP [5 ] also measures the a v ailable bandwidth by sending man y pack et pairs at increasing rates. When the pack et sending rate is less than the endto-end a v ailable bandwidth, the pack ets will arri v e at the recei v er at the same rate the y were sent, whereas if the pack et rate is more than the a v ailable bandwidth, then the measured rate at the recei v er will be less than at the sender The methodology used is v ery similar to the Self Loading Periodic Streams (SLoPS) with the stream rate increase follo wing a linear f ashion instead of a binary search. The disadv antage with T OPP is that in paths with multiple links, the estimates tend to be erroneous as the results start to depend on the order of the links in the path. Mathematically the a v ailable bandwidth in link i is gi v en by A i = C i (1 U i ) (2.1) where C i is the capacity of link i and U i is the utilization of link i Consequently the end to end a v ailable bandwidth is gi v en by A = N min i =1 A i (2.2) Current techniques de v eloped to estimate the a v ailable bandwidth use either the Pr obe Gap Model (PGM) or the Pr obe Rate Model (PRM) T echniques using the Probe Gap Model obtain the a v ailable bandwidth estimate from the time dif ference that e xists between the probe pack ets when the y are sent at the source and the interarri v al time when the y reach the destination. T echniques using this model assume a single bottleneck link, a non-empty queue between the departure of the rst probe pack et and the departure of the second one, and apriori kno wledge of the capacity of the link. Spread P aiR Unused Capacity estimate (Spruce) [4] and Initial Gap Increasing (IGI) [1 ] are e xamples of tools using this approach. T echniques using the Probe Rate Model usually send 5

PAGE 15

a series of probe pack ets (a pack et train) to induce congestion in the tight link. These techniques check the sending and recei ving rates of the probe train looking for the point where the recei ving rate starts matching the sending rate. This juncture represents the rate where the probe sending rate equals the a v ailable bandwidth in the tight link.The PRM techniques usually need more time to con v er ge as the y use an iterati v e approach to nd the turning point, and usually need statistical techniques to v alidate that this is in f act the turning point. A v ailable tools taking the PRM approach are P athload [2 ], P ack et T ransmission Rate (PTR) [1 ], pathChirp [3] and T OPP [5]. Both, the PGM and the PRM use a train of probe pack ets to cope with the b urstiness of competing cross traf c. P athload [2] uses one-w ay delay trends of periodic streams to estimate the end-to-end a v ailable bandwidth. It is based on the principle of self induced congestion whereby the source sends long constant bit rate pack et trains until the rate of the train equals or e xceeds the rate of the a v ailable bandwidth indicated by a consistently increasing one-w ay delay trend. P athload adopts an adapti v e search in its a v ailable bandwidth estimation. The PTR and IGI algorithms also use trains of probe pack ets by scheduling departures of pack et trains from source with increasing initial gaps. It does this by monitoring the dif ference between the a v erage source and destination gaps. Whene v er this dif ference becomes zero, the a v erage rate of the train equals the a v ailable bandwidth in the bottleneck link. The tool pathChirp [3 ] utilizes e xponentially distrib uted probe pack ets called a Chirp to measure the a v ailable bandwidth. By e xponentially increasing the pack et spacing, pathChirp is supposed to reduce the number of probe pack ets introduced. Spruce [4 ] sends pairs of probe packets so that the second pack ets arri v e at the bottleneck link before the rst pack et lea v es the queue. This is the case of the Joint Queueing Re gion (JQR) identied in IGI/PTR that guarantees a better accurac y Similar to IGI, Spruce assumes that the tool kno ws the capacity of the tight link, which is usually found by applying other tools such as bprobe [12 ], pathrate [15 ], or nettimer [11 ]. The three important ABET metrics are: Accurac y The accurac y of an ABET is measured in terms of ho w close the tool estimate of a v ailable bandwidth is as compared to the actual a v ailable bandwidth. The higher the accurac y that a tool can achie v e, the better it is. Con v er gence T ime The con v er gence time of an ABET is the time it tak es for the tool to con v er ge to an estimate. The shorter the con v er gence time, the better it is. 6

PAGE 16

P2r P1r Cross Trafficr Routerr Bottleneck Linkr P2r P1r Gap_Inr Gap_Outr Cross Trafficr Cross Trafficr Figure 2.1 P ack et pair probe gap model [1 ] Intrusi v eness or Ov erhead The ABETs introduce probing traf c in the netw ork to mak e a v ailable bandwidth measurements. The lo wer the o v erhead an ABET introduces, the better it is. There are a number of f actors lik e pack et size, pack ets per train, number of pack et trains, etc that af fect the ABET metrics. It may be dif cult to ha v e one ABET with the absolute best in all the three metrics (i.e. highest accurac y least con v er gence time and least o v erhead). In reality most applications may not e v en need the best of all the three metrics as trade-of fs can be made between the metrics to meet the requirements of an application. The le v el of importance of the ABET metrics is dictated by the application that the ABET will be used in. 2.1 ABETs 2.1.1 IGI (Initial Gap Incr easing) A sequence of pack et trains is sent from the source to the destination with increasing initial gaps. The interpack et gap in a pack et pair is measured at the source and at the destination and the dif ference in the interpack et gaps is used to estimate the a v ailable bandwidth in the netw ork. The bottleneck link rate is a direct measure of the spacing between the pack ets and links with higher a v ailable bandwidths maintain the spacing between the pack ets. A v ailable bandwidth in a netw ork is a dynamic metric that can change instantaneously and so a mean of samples measured o v er a period of time will be representati v e of the true a v ailable bandwidth. So IGI uses a train of pack et pairs instead of just a pack et pair to estimate the a v ailable bandwidth. When pair of probing pack ets is sent, the competing cross traf c along the path introduces delays in the pack et spacing and so is proportional to the pack et spacing as seen in Figure 2.1. The abo v e statement is true only in the scenario where the second pack et arri v es in the queue before the rst pack et lea v es the queue (also called as Joint Queuing Re gion (JQR)). In this scenario the output gap is the sum of the time to 7

PAGE 17

process the rst pack et and the time to process the competing cross traf c at the bottleneck link. The initial gap between the probing pack ets has a signicant ef fect on the IGI/PTR algorithms. Increasing the initial gap increases the Disjoint Queuing Re gion (DQR) area. DQR v anishes if the initial gap is smaller than the bottleneck gap. But this will amount to ooding of the bottleneck link and hence an underestimate of the a v ailable bandwidth. Initially the input gap is less than the output gap, b ut as we slo wly increase the input gap up to and be yond the bottleneck gap, at some point the output gap equals the input gap. This point is where the competing cross traf c pack ets interlea v e nicely with the probe pack ets and the probe pack et rate represents the a v ailable bandwidth rate. This point is called the turning point and corresponds to the smallest initial gap v alue without ooding the bottleneck link and at the same time k eeping the queue full. W ith respect to PTR, the initial gap at the turning point corresponds to the pack et transmission rate where the pack et trains consume all the a v ailable bandwidth without interfering with the competing cross traf c. The pack et train beha v es lik e an aggressi v e b ut well-beha v ed application o w and so its rate is a good estimation of the a v ailable bandwidth. In the IGI algorithm, the a v ailable bandwidth is obtained by subtracting the estimated competing traf c throughput from the bottleneck capacity So there are tw o parts to the bandwidth estimation. 1) Bottleneck capacity measurement A simplied v ersion of the nettimer [11 ] algorithm is used to compute the bottleneck capacity This measurement is done by sending a train of pack ets continuously from the source, and measuring the pack et gaps at the destination. The pack et gap v alue that repeats most often in the set of gap v alues for the rst train of pack ets measured corresponds to the bottleneck capacity 2) Competing throughput measurement T rain of pack ets are sent at increasing gap v alues from the source and the gap v alues at the destination are measured. The increased gap v alues at the destination as compared to the bottleneck gap represent the competing traf c, while decreased or no change in gap v alues means absence of or minimal cross traf c. The ratio of the sum of the increased gap v alues to the sum of all the gap v alues is a measure of the competing traf c throughput. In the DQR, the output gap is gi v en by the equation as sho wn in Eqn. 2.3 [1 ] g O = g B + B C g i B O (2.3) 8

PAGE 18

where g O is the output gap, g i is the input gap, g B is the bottleneck gap, B C and B O are competing traf c and bottleneck throughputs respecti v ely The right hand side of Eqn. 2.3 has tw o time components and the rst corresponds to the time it tak es to process P ack et P1 and the second is the time to process the competing traf c that arri v es between the tw o probing pack ets. Assume a probing train in which there are M probes with increasing gap v alues, K probes whose gap v alues are unchanged and N probes whose gap v alues are decreased. Applying the abo v e formula, we get the follo wing estimate for the competing traf c load as gi v en in Eqn. 2.4 [1 ] B O P Mi =1 ( g i + g B ) P Mi =1 g i + + P Ki =1 g i = + P Ni =1 g i (2.4) where, g i + g i = g i represent gaps that are increased, unchanged and decreased respecti v ely The numerator denotes the amount of competing traf c arri ving during the probing period. The denominator represents the total probing time. After the competing traf c is computed, the a v ailable bandwidth is computed as the dif ference between bottleneck capacity and the competing load. The f actors af fecting IGI/PTR probing methodology are: 1. Probing pack et size. It has been pointed out by the authors in [1] that small probing pack et sizes can result in errors in a v ailable bandwidth estimation. One reason for this is that small probe v alues lead to small gap v alues and so the gap dif ference does not con v er ge as well as it w ould with lar ger probing pack ets. It is also indicated that that smaller pack et sizes may underestimate the a v ailable bandwidth and lar ger pack ets run the risk of o v er estimating the a v ailable bandwidth 2. Length of the pack et train. A v ery small pack et train may tak e a shorter time b ut may lead to inaccurate results, whereas a longer pack et train may tak e longer b ut estimate the a v ailable bandwidth more accurately The tool can tak e se v eral phases of sending pack et trains to con v er ge to the best initial gap v alue. A short pack et train corresponds to a shorter sampling interv al and so the a v ailable bandwidth estimates are b ursty In conclusion, the IGI algorithm uses the pack et gap v alues of pack ets in a train to estimate the bandwidth of competing cross traf c in the path. 9

PAGE 19

1r 2r 3r 4r 1r 2r 3r 4r T=L/Rr 1r 2r 3r 4r K=4r At Senderr At Recieverr R>Ar At Receiverr R
PAGE 20

R max = 8><>: R i if R i > A R max O ther w ise R min = 8><>: R i if R i A R min O ther w ise R i +1 = R min + R max 2 where R max and R min are the outer and inner limits for the a v ailable bandwidth after i streams. Initially R min is set to 0 and R max is set to a suf ciently high v alue such that R max0 > A. The algorithm terminates when R max R min < resol where resol is an user -dened bandwidth estimation resolution. The transmission rate R is a function of pack et size L and interpack etgap T T is usually set to T min which is the smallest possible interpack et gap that can be achie v ed at the end computers. L can be computed as RT L should al w ays be smaller than the path MTU to a v oid fragmentation. P athload determines if R > A by sending a eet of N streams of K pack ets instead of a single stream of N by K pack ets. This allo ws the tool to see if R > A N successi v e times. All streams in a eet ha v e the same rate R The use of multiple streams pro vides an idle period that allo ws for the queues in the netw ork to drain. P athload uses statistical metrics lik e the P airwise Comparison T est (PCT) and the P airwise Dif ference T est (PDT) of the streams to detect the increasing O WD trend. The PCT measures the fraction of consecuti v e O WD pairs that are increasing. PDT quanties ho w strong is the start-to-end O WD v ariation, relati v e to the O WD absolute v ariations during the stream. The f actors af fecting P athload probing methodology are: 1. P ack ets per train. W ith more pack ets per train, there are comparisons between R and A o v er a longer time range and the v ariablility of the estimate reduces because the a v eraging time scale increases. But this will ha v e an impact on the measurement latenc y 11

PAGE 21

Chirpr Rater Queueingr Delayr available bandwidthr Figure 2.3 pathChirp queueing delay signature [3 ] 2. Fleet length or number of trains. As the number of trains increases, the v ariations of the measured a v ailable bandwidth across se v eral runs decrease. 2.1.3 pathChir p pathChirp uses the concept of self-induced congestion and sends an e xponential ”chirp” of probe pack ets to compute the a v ailable bandwidth. An e xponentially spaced stream of probe pack ets is sent from the source and a statistical analysis is done at the destination. Figure 2.3 illustrates the queueing delay signature of a chirp train. The relati v e queuing delays between the probe pack ets are monitored to calculate the a v ailable bandwidth. T ypically a signature consists of e xcursions from the zero axis corresponding to increasing queuing delays caused by increasing cross traf c. When the chirp rate is less than the actual a v ailable bandwidth, the queues relax and the queuing delay goes to zero. When the queuing delays mak e an e xcursion from the zero axis and continue to trend northw ard with no sign of returning to zero delays, it signies that the actual a v ailable bandwidth is less than the P ack et Chirp rate. pathChirp uses the shape of the signature to mak e an estimate of the per -pack et a v ailable bandwidth and then tak es a weighted a v erage of the per pack et estimate to estimate the per -chirp a v ailable bandwidth. The f actors af fecting pathChirp probing methodology are: 1. P ack et size. The size of the pack et is proportionate to the number of bytes transmitted. F or the same probing rates a smaller pack et means smaller probe interv als which means that the cross trac arri ving during this short interv al will be v ery b ursty This will lead to erratic signatures and hence inaccurate bandwidth estimates. So a lar ger pack et size is better for the bandwidth estimation. 12

PAGE 22

2. Spread f actor The Spread f actor controls the spectrum of probing rates in a chirp. A smaller Spread F actor leads to a dense spectrum of rates leading to more accurate estimates. 2.2 TCP TCP is a connection-oriente d, point-to-point, reliable transport layer protocol. TCP utilizes a 3-w ay handshak e procedure while establishing an end to end connection and e xchanges protocol specic information before initiating an y data transfer The connection termination is also done gracefully once all data pack ets are deli v ered successfully The TCP protocol at the transport layer breaks the application data into se gments and passes it do wn to the netw ork layer where the y are encapsulated into netw ork layer IP datagrams. At the recei v er end, the se gment' s data is passed up by the netw ork layer and is placed in the recei v e b uf fer of the TCP Connection. Once the se gment is sent, a timer is set, and it w aits for the recei v er to ackno wledge (A CK) the se gment. If an A CK is not recei v ed before the timer e xpires, TCP retransmits the se gment. The A CK mechanism of TCP is an important part of what mak es the protocol reliable. Reliability means that data from the sender is not lost, duplicated or recei v ed out of order TCP guarantees these properties e v en though the underlying communication medium may lose, duplicate or reorder pack ets. The other features of TCP are its o w control and congestion control. Flo w control checks for b uf fer space at the end systems and slo ws do wn the data transfer rate if the b uf fer capacity is reached. Hence TCP limits the sender to send only as much data as the recei v er can handle. Flo w control only a v oids the sender from o v ero wing the recei v er' s b uf fer b ut it does not tak e netw ork congestion into account. T o solv e the problem of o v erloading the netw ork nodes between end systems, TCP implements congestion control. While o w control is an end system issue, congestion control is a netw ork issue. There ha v e been a number of changes made to the original TCP congestion control algorithm de v eloped by V Jacobson [16 ] and ne w TCP a v ors lik e TCP Reno, TCP Ne wreno, TCP V e gas,etc ha v e been implemented. 2.2.1 TCP Reno TCP Reno is a v ery popular v ariant of TCP and models lik e TCP Sack and TCP Ne wreno are enhancements to TCP Reno. TCP Reno vie ws pack et losses as a signal of netw ork congestion. 13

PAGE 23

While there are no pack et losses, TCP Reno continues to increase its windo w size by one during each round trip time. When it e xperiences a pack et loss, it reduces its windo w size to one half of the current windo w size. This is called additi v e increase and multiplicati v e decrease. It uses a sliding windo w mechanism for its o w control, ackno wledgment and retransmission polic y The basic idea is that there is a windo w of size n that determines ho w man y successi v e se gments of data can be sent in the absence of a ne w ackno wledgment. The size of n is dynamic depending on a v ailable b uf fer space at the recei v er and congestion in the netw ork. Each se gment of data is sequentially numbered, so the sender is not allo wed to send se gment i +n before se gment i has been ackno wledged. Thus, if i is the sequence number of the se gment most recently ackno wledged by the recei v er there is a windo w of data numbered i + 1 to k which the sender can transmit. As successi v ely higher numbered ackno wledgments are recei v ed, the windo w slides forw ard. The ackno wledgment mechanism is cumulati v e in that if the recei v er ackno wledges se gment k where k > i + 1, it means it has successfully recei v ed all se gments up to k. Se gment k is ackno wledged by sending a request for se gment k + 1. In TCP data that is transmitted is k ept on a retransmission b uf fer until it has been ackno wledged. Thus, when k is ackno wledged, se gments with sequence number less than or equal to k are remo v ed from the retransmission b uf fer If k < n + i the sender may retransmit se gments k + 1 to n + i from the retransmission b uf fer The decision to retransmit these se gments depends on the receipt of duplicate ackno wledgments and on timeouts. TCP Reno also uses F ast Retransmit/F ast Reco v ery where when the source recei v es i duplicate ackno wledgments (e.g., i = 3) for the same se gment (say k ), it determines that se gment k w as lost. The source chooses to retransmit se gment k right a w ay rather than w ait for a retransmission timer to e xpire. The congestion control mechanism reduces the sender' s transmission windo w by half on retransmission of a pack et based on duplicate ackno wledgments. 2.2.2 TCP Sack The Congestion Control mechanism in TCP Sack [17 ] is an impro v ement o v er that of TCP Reno. The limitation of the cumulati v e ackno wledgment strate gy that is a v ailable in TCP Reno is that it can only indicate that e v ery se gment up to k has been recei v ed and k + 1 has not been recei v ed. When multiple pack ets are lost from a windo w the throughput of TCP can suf fer greatly When 14

PAGE 24

multiple pack ets in a windo w are dropped, the sender is constrained by the f ast retransmit polic y to retransmit only one pack et per roundtrip in TCP Reno. Ho we v er TCP Sack allo ws the recei v er to ackno wledge noncontiguous and isolated blocks of data that ha v e been recei v ed and queued, using a selecti v e ackno wledgment in addition to the cumulati v e ackno wledgment of contiguous data, in the e v ent of multiple pack et losses. TCP SA CK introduces the possibility for earlier reco v ery from loss of multiple pack ets resulting in higher throughput because it a v oids drastic congestion reco v ery mechanisms lik e slashing windo w size to one. TCP Sack is used in the simulations in this thesis for demonstrating the application of ABETs in TCP because half of the serv ers in the Internet today use TCP Sack and man y others use Ne wReno and almost all bro wsers use TCP Sack [18 ]. 15

PAGE 25

CHAPTER 3 METHODOLOGY IGI, P athload and pathChirp are e v aluated using simulations and statistical techniques. Exper iments are designed to study the ef fects of ABET f actors lik e size of pack ets, number of pack ets per train and number of pack et trains on the accurac y con v er gence time and the intrusi v eness of the tools under lo w bandwidth and high bandwidth scenarios. The tools were e v aluated using the ns-2 simulator [7 ] and the modules pro vided by the corresponding authors. Three main cases were studied. In the rst case, the tools are e v aluated in a lo w bandwidth en vironment. The second case e v aluates the tools in high bandwidth netw ork conditions. The last case e v aluates the tools under v ariable load conditions. 3.1 Simulation design 3.1.1 Case 1: Lo w band width scenario The topology sho wn in Figure 3.1 w as used to run four sets in a lo w bandwidth scenario. The topology is a 4-node 3-hop netw ork in which the second hop is the bottleneck link. The bottleneck bandwidth is set to 10 Mbps. In all the scenarios, the ABETs are attached to the end nodes and so are the TCP Sack source and sink. In the rst scenario, the estimation technique runs with only one TCP Sack connection. The second scenario is similar with the addition of a CBR source transmitting at a x ed rate of 5 Mbps to simulate cross traf c. The third scenario is lik e the second one e xcept that the CBR source goes ON and OFF between 5 Mbps and 0 Mbps instead of being x ed at one v alue. It stays ON for 100 seconds and goes OFF for 50 seconds. In the fourth case, CBR traf c is in the form of a staircase and changes in steps of 2Mbps o v er the range 0 Mbps and 10Mbps e v ery 50 seconds. The abo v e simulations were carried out for 500 seconds. 16

PAGE 26

n1r ABETr TCPr Sourcer n4r ABETr TCPr Sinkr n2r n3r UDPr Sourcer UDPr Sinkr 1000 Mbpsr 10 Mbpsr 1000 Mbpsr 2 msr 40 msr 2 msr Figure 3.1 Lo w bandwidth netw ork topology n1r ABETr TCPr Sourcer n4r ABETr TCPr Sinkr n2r n3r UDPr Sourcer UDPr Sinkr 4000 Mbpsr 1000 Mbpsr 4000 Mbpsr 2 msr 40 msr 2 msr Figure 3.2 High bandwidth netw ork topology 3.1.2 Case 2: High band width scenario The topology sho wn in Figure 3.2 w as used to run four sets of scenarios using high bandwidth links. Again, the topology consists of four nodes and three hops and the second hop is the bottleneck bandwidth that is set to 1000 Mbps. As described in Case 1, the ABET tools run from the same end points of a TCP Sack connection. In the rst scenario, there is only one TCP Sack connection transmitting pack ets. The second scenario is lik e the rst one b ut there is also a CBR source transmitting at a x ed rate of 500 Mbps. The third scenario has the CBR source goes ON and OFF It stays ON for 100 seconds and goes OFF for 50 seconds. Finally the CBR traf c is induced in the form of a stair w a v eform going from 0 to 1000 Mbps in steps of 200 Mbps, stepping e v ery 50 seconds. The abo v e simulations were carried out for 500 seconds. 17

PAGE 27

n1r ABETr TCPr Sourcer n4r ABETr TCPr Sinkr n2r n3r UDPr Sourcer UDPr Sinkr 4000 Mbpsr 10/100/300/500/700/1000 Mbpsr 4000 Mbpsr 2 msr 40 msr 2 msr CBR stairr waveformr Figure 3.3 Simulation with CBR stair cross traf c 3.1.3 Case 3: W ith CBR stair case trafc The ABET tools were e v aluated in the presence of v ariable load created by the use of CBR traf c in staircase form. The CBR staircase traf c v aried from 0 Mbps in steps of (bottleneck bandwidth/5) until the CBR traf c equaled the bottleneck bandwidth and again stepping do wn the cross traf c from bottleneck bandwidth all the w ay do wn to zero. The topology sho wn in Figure 3.3 w as used with bottleneck bandwidths v alues of 10 Mbps, 100 Mbps, 300 Mbps, 500 Mbps, 700 Mbps and 1000 Mbps. This simulation co v ers the spectrum of lo w medium and high bandwidth netw orks. The set of parameters that yielded the best results in Case 1 and Case 2 w as used to carry out the e xperiments in Case 3. The abo v e simulations were carried out for 450 seconds. 18

PAGE 28

3.2 Design methodology The f actors that are considered in the e v aluation of ABETs are probe pack et size, number of pack ets per train, number of trains, frequenc y of measurements, and spread f actor The ef fect of these f actors on the response of the tools is measured. The response is measured in terms of accurac y con v er gence time and intrusi v eness. The design metholodogy used is a 2 k f actorial design as described in [19 ]. Each f actor can be at tw o le v els (lo w and high) and the v alues are set from a reasonable range that it can tak e. e.g. the tw o le v els for the pack et size f actor is 500 bytes and 1500 bytes. The f actors may interact with one another i.e. the ef fect of one f actor may depend on the le v els of other f actors. The ef fect of the f actors and also the ef fect of the interactions between the f actors on the responses can be measured by carrying out the 2 k f actorial design where k is the number of f actors af fecting an ABET and 2 is the number of le v els. So for e xample in case of IGI, we ha v e 3 f actors pack et size, pack ets per train and frequenc y of runs. Assuming tw o le v els for each of these f actors we ha v e 8 responses ( 2 3 ). R 1 R 2 R 3 ... R 8 are the response v alues. Then the ef fect of f actor 1 is gi v en e 1 = ( R 2 R 1 ) + ( R 4 R 3 ) + ( R 6 R 5 ) + ( R 8 R 7 ) 4 (3.1) The ef fect of f actor 2 is gi v en by e 2 = ( R 3 R 1 ) + ( R 4 R 2 ) + ( R 7 R 5 ) + ( R 8 R 6 ) 4 (3.2) The ef fect of f actor 3 is gi v en by e 3 = ( R 5 R 1 ) + ( R 6 R 2 ) + ( R 7 R 3 ) + ( R 8 R 4 ) 4 (3.3) The ef fect of f actor 1 could depend on the le v el of f actor 2 and so the de gree of interactions between the f actors is gi v en by equations 3.4 3.6. It is also important to note that the tw o-f actor interaction ef fects are completely symmetric i.e., e 12 = e 21 and e 23 = e 32 and so on. The de gree of interaction between the tw o f actors is dened as half the dif ference between the a v erage ef fect 19

PAGE 29

of f actor 1 when f actor 2 is at its ”lo w” le v el (all f actors e xcept 1 and 2 are held constant) and the a v erage ef fect of f actor 1 when f actor 2 is at its ”high” le v el. e 12 = 1 2 [ ( R 4 R 3 ) + ( R 8 R 7 ) 2 ( R 2 R 1 ) + ( R 6 R 5 ) 2 ] (3.4) The ef fect of f actor 2 is gi v en by e 13 = 1 2 [ ( R 6 R 5 ) + ( R 8 R 7 ) 2 ( R 2 R 1 ) + ( R 4 R 3 ) 2 ] (3.5) The ef fect of f actor 3 is gi v en by e 23 = 1 2 [ ( R 7 R 5 ) + ( R 8 R 6 ) 2 ( R 3 R 1 ) + ( R 4 R 2 ) 2 ] (3.6) Finally a three f actor interaction can be done using e 123 = 1 2 [ ( R 8 R 7 ) + ( R 6 R 5 ) 2 ( R 4 R 3 ) + ( R 2 R 1 ) 2 ] (3.7) 20

PAGE 30

CHAPTER 4 PERFORMANCE EV ALU A TION OF ABETS Simulations were run for 500 seconds to determine the accurac y con v er gence time, and intrusi v eness of IGI, P athload and pathChirp. In the case of IGI, eight sets of simulations were run follo wing the 2 3 f actorial design where the f actors are the size of the probe pack ets, the number of pack ets per train and the frequenc y of runs. In the case of P athload, again eight sets of simulations were done follo wing the 2 3 f actorial design where the f actors are the number of pack ets per train, number of trains and the frequenc y of runs. In the case of pathChirp, four simulations were run follo wing the 2 2 f actorial design where the main f actors are the size of the probe pack ets and the Spread f actor Frequenc y is the time between the runs. Continuous frequenc y means that the tool is run again immediately after it pro vides an estimate. A frequenc y of x means that the tool is scheduled to run e v ery x seconds. The con v er gence time of an ABET is the time it tak es for the tool to con v er ge to an estimate. The intrusi v eness of the ABET is calculated as the number of probe pack ets used by the tool to pro vide the estimate multiplied by the bytes per probe pack et. The accurac y of the ABETs is computed as follo ws. If the tool' s estimate is T and the actual a v ailable bandwidth is A then the accurac y is gi v en by ( T A ) = ( T + A ) Therefore, the v alue of the accurac y can range from 1 to +1 where a v alue of 0 indicates 100% accurac y A ne gati v e v alue indicates that the tool underestimates the a v ailable bandwidth, and a positi v e v alue says that the tool o v erestimates the amount of a v ailable bandwidth. During the duration of the simulation, v alues of T and A are recorded and the accurac y is computed. At the end of the simulation, the a v erage accurac y is computed as the absolute v alue of the a v erage of all the recorded v alues. This a v erage v alue of accurac y gi v es the measure of ho w much f arther from the actual a v aliable bandwidth v alue did the tool go in each run, irespecti v e of whether the tool o v er estimated or under estimated. It is important to understand this as in an absolute sense a tool may ha v e lo w accurac y b ut it could be under estimation most of the time or o v er estimation most of time or a combination of both. 21

PAGE 31

The a v erage v alue for accurac y is listed in all the f actorial design result tables in this Chapter The thresholds for the metrics really are dictated by the application that the ABETs will be applied to. 4.1 Case 1: Lo w band width scenario F or each tool, tw o tables list the results of the four scenarios. The rst table lists the results of the f actorial design and the second table lists the ef fect of the interactions between the f actors on the response v ariables. T able 4.1 sho ws the results of the f actorial design when IGI w as used. T able 4.4 sho ws the results with the ef fects on the performance metrics of the three f actors alone and their combination. T able 4.2 and T able 4.5 are the tw o tables containing results for P athload. T able 4.3 and T able 4.6 sho w the results for pathChirp. 4.1.1 Analysis of IGI From T able 4.1, it can be seen that the simulation using 1500 byte pack et sizes, 64 pack ets per train and continuous frequenc y pro vides the best accurac y possible b ut the w orst con v er gence time and o v erhead while the best con v er gence times are obtained using 8 pack ets per train and running the tool continuously No w looking at T ables 4.1 and 4.4 together we can see that 1) Increasing the pack et size from 500 to 1500 bytes impro v es the accurac y slightly to w ards underestimation b ut w orsens the con v er gence time slightly and the o v erhead considerably which is e xpected. Therefore, a rather small pack et size, such as 500, might be the most appropriate; 2) Increasing the number of pack ets per train impro v es the accurac y and w orsens the con v er gence time and o v erhead consider ably W e can hence deduce that a small number of pack ets per train, such as 8 is a good choice and 3) Reducing the frequenc y of measurement w orsens the accurac y increases the con v er gence time slightly and reduces o v erhead. So picking a high frequenc y (more frequent runs) or continuous frequenc y gi v es optimal results. Also, the cases of lo w accurac y detected in IGI measurements are mostly a result of o v er -estimation when the netw ork is o v erloaded. 22

PAGE 32

T able 4.1 Case 1: 2 k f actorial design using IGI in lo w bandwidth netw orks P ack et Pkts per Frequenc y T ests Accurac y Con v er gence Intrusi v eness Size T rain T ime (Sec) (MB) TCP Only 0.56 1.68 0.042 500 8 Continuous TCP+CBR 0.81 2.58 0.074 TCP+CBR ON/Of f 0.66 2.03 0.05 CBR ladder only 0.08 0.88 0.05 TCP Only 0.43 1.9 0.13 1500 8 Continuous TCP+CBR 0.56 2.69 0.22 TCP+CBR ON/Of f 0.58 2.36 0.17 CBR ladder only 0.07 0.93 0.17 TCP Only 0.33 4.58 0.53 500 64 Continuous TCP+CBR 0.62 4.51 0.63 TCP+CBR ON/Of f 0.41 4.69 0.57 CBR ladder only 0.07 2.33 0.57 TCP Only 0.21 4.98 1.22 1500 64 Continuous TCP+CBR 0.26 4.99 1.47 TCP+CBR ON/Of f 0.24 4.93 1.29 CBR ladder only 0.06 2.99 1.29 TCP Only 0.61 2.82 0.061 500 8 6 TCP+CBR 0.88 2.36 0.068 TCP+CBR ON/Of f 0.65 2.66 0.064 CBR ladder only 0.13 1.03 0.06 TCP Only 0.59 3.3 0.199 1500 8 6 TCP+CBR 0.73 2.69 0.238 TCP+CBR ON/Of f 0.6 3.14 0.21 CBR ladder only 0.12 1.14 0.21 TCP Only 0.49 4.49 0.5 500 64 6 TCP+CBR 0.76 4.7 0.59 TCP+CBR ON/Of f 0.56 4.7 0.54 CBR ladder only 0.1 2.56 0.54 TCP Only 0.45 5.9 1.41 1500 64 6 TCP+CBR 0.43 4.8 1.02 TCP+CBR ON/Of f 0.42 4.94 1.26 CBR ladder only 0.1 3.41 1.26 23

PAGE 33

T able 4.2 Case 1: 2 k f actorial design using P athload in lo w bandwidth netw orks # of Pkts per Frequenc y T ests Accurac y Con v er gence Intrusi v eness T rains T rain T ime (Sec) (MB) TCP Only 0.31 3.99 0.3 4 50 Continuous TCP+CBR 0.24 4.3 0.3 TCP+CBR ON/Of f 0.27 4.25 0.3 CBR ladder only 0.012 2.97 0.3 TCP Only 0.36 11.39 0.92 12 50 Continuous TCP+CBR 0.25 12.27 0.92 TCP+CBR ON/Of f 0.34 11.9 0.92 CBR ladder only 0.02 8.68 0.92 TCP Only 0.39 4.8 0.61 4 100 Continuous TCP+CBR 0.34 5.27 0.61 TCP+CBR ON/Of f 0.36 5.01 0.61 CBR ladder only 0.01 3.37 0.61 TCP Only 0.47 13.86 1.84 12 100 Continuous TCP+CBR 0.36 15.44 1.84 TCP+CBR ON/Of f 0.41 14.78 1.84 CBR ladder only 0.1 10.1 1.84 TCP Only 0.24 4.11 0.3 4 50 6 TCP+CBR 0.29 4.62 0.3 TCP+CBR ON/Of f 0.33 12.54 0.92 CBR ladder only 0.02 3 0.3 TCP Only 0.3 11.49 0.92 12 50 6 TCP+CBR 0.29 12.38 0.92 TCP+CBR ON/Of f 0.33 12.54 0.92 CBR ladder only 0.03 8.78 0.92 TCP Only 0.26 4.91 0.61 4 100 6 TCP+CBR 0.35 5.44 0.61 TCP+CBR ON/Of f 0.34 5.11 0.61 CBR ladder only 0.009 3.45 0.64 TCP Only 0.39 14.34 1.84 12 100 6 TCP+CBR 0.36 15.43 1.84 TCP+CBR ON/Of f 0.39 14.52 1.84 CBR ladder only 0.05 10.12 1.84 4.1.2 Analysis of P athload From T able 4.2, it can be seen that the best performance is obtained when using 4 trains, 50 pack ets per train and running the tool continuously Using those v alues, we can obtain the best accurac y con v er gence time and o v erhead. Looking at T ables 4.2 and 4.5 together we can see that 1) Increasing the number of trains from 4 to 12 w orsens the accurac y v ery slightly to w ards o v erestimation b ut increases the con v er gence time and the o v erhead considerably Therefore, it can be concluded that a small to medium number of trains such as 4 to 8 is appropriate; 2) Increasing the 24

PAGE 34

number of pack ets per train w orsens the accurac y slightly w orsens the con v er gence time slightly and w orsens the o v erhead considerably W e can hence infer that 50 is a good choice for number of pack ets per train; and 3) Reducing the frequenc y of measurement impro v es the accurac y by a v ery small mar gin, af fects the con v er gence time v ery slightly and has no ef fect on the o v erhead. So running the tool continuously will be a good choice. Also, the cases of lo w accurac y detected in P athload measurements are mostly a result of under -estimation when the netw ork is not loaded. 4.1.3 Analysis of pathChir p From T able 4.3, the best con v er gence time is obtained using 500 byte pack ets and 1.1 as the Spread F actor Ho we v er the best intrusi v eness is obtained in the 500 and 1.2 case. The case with 1500 and 1.2 seems to be the one that of fers a good combined performance in terms of accurac y and o v erhead b ut the con v er gence time is not good in particular for the rst scenario. Looking at T ables 4.3 and 4.6 together we can see that 1) Increasing the pack et size from 500 to 1500 impro v es the accurac y b ut increases the con v er gence time to a great e xtent. So it can be concluded that a small to medium probe pack et size is better; 2) Increasing the spread f actor impro v es the accurac y and reduces the con v er gence time a great deal while also impro ving the o v erhead. W e can hence infer that 1.2 is a good choice for the spread f actor A medium pack et size of say 1000 with a Spread F actor of 1.2 is the best choice.Also, the cases of lo w accurac y detected in pathChirp measurements are mostly a result of under -estimation when the netw ork is not loaded. 25

PAGE 35

T able 4.3 Case 1: 2 k f actorial design using pathChirp in lo w bandwidth netw orks P ack et Spread T ests Accurac y Con v er gence Intrusi v eness Size F actor T ime (sec) (MB) TCP Only 0.8 2.4 0.14 500 1.1 TCP+CBR 0.33 3.73 0.14 TCP+CBR ON/Of f 0.616 3.73 0.14 CBR ladder only 0.04 0.7 0.14 TCP Only 0.64 17.86 0.14 1500 1.1 TCP+CBR 0.42 4.18 0.14 TCP+CBR ON/Of f 0.47 5.41 0.14 CBR ladder only 0.07 2.45 0.42 TCP Only 0.8 7.52 0.07 500 1.2 TCP+CBR 0.33 2.08 0.07 TCP+CBR ON/Of f 0.61 3.26 0.07 CBR ladder only 0.05 0.54 0.1 TCP Only 0.69 15.03 0.07 1500 1.2 TCP+CBR 0.38 2.2 0.07 TCP+CBR ON/Of f 0.43 4.79 0.07 CBR ladder only 0.05 1.61 0.313 T able 4.4 F actors ef fects using IGI in lo w bandwidth netw orks Response T ests E1 E2 E3 E12 E13 E23 E123 (PSize) (Pkts/train) (Freq) TCP Only -0.07 -0.17 0.15 -0.01 0.046 0.042 -0.01 Accurac y TCP+CBR -0.27 -0.22 0.13 -0.07 0.03 0.01 -0.02 TCP+CBR ON/Of f -0.11 -0.21 0.08 -0.04 0.01 0.07 -0 CBR ladder only -0.007 -0.015 0.045 0 0.002 -0.005 0 TCP Only 0.62 2.56 0.84 0.27 0.31 -0.42 0.18 Con v er gence TCP+CBR -0.05 1.86 -0.36 -0.27 -0.35 -0.25 -0.46 T ime(Sec) TCP+CBR ON/Of f 0.32 2.26 0.35 -0.08 0.03 -0.34 -0.03 CBR ladder only 0.41 1.83 0.25 0.33 0.06 0.07 0.03 TCP Only 0.45 0.8 0.06 0.34 0.06 0.019 0.044 Intrusi v eness TCP+CBR 0.39 0.78 -0.11 0.24 -0.98 -0.12 -0.1 (MB) TCP+CBR ON/Of f 0.42 0.79 -0.59 0.29 0.007 -0.026 -0.007 CBR ladder only 0.42 0.79 0 0.29 0.007 -0.026 -0.007 26

PAGE 36

T able 4.5 F actors ef fects using P athload in lo w bandwidth netw orks Response T ests E1 E2 E3 E12 E13 E23 E123 (# T rains) (Pkts/train) (Freq) TCP Only 0.08 0.07 -0.08 0.02 0.02 -0.02 0.01 Accurac y TCP+CBR 0.01 0.08 0.02 0.003 -0.007 -0.018 0.001 TCP+CBR ON/Of f 0.04 0.05 0.001 0.01 -0.01 -0.02 0.01 CBR ladder only 0.03 0.02 -0.01 0.02 -0.01 -0.02 -0.01 TCP Only 8.31 1.73 0.2 0.92 0.08 0.09 0.09 Con v er gence TCP+CBR 8.97 2 0.14 1.1 -0.09 -0.06 0.007 time (Sec) TCP+CBR ON/Of f 8.78 1.62 0.12 0.81 0.07 -0.2 -0.25 CBR ladder only 6.22 0.9 0.05 0.47 0.002 -0.007 -0.03 TCP Only 0.92 0.31 0 0.3 0 0 0 Intrusi v eness TCP+CBR 0.92 0.61 0 0.3 0 0 0 (MB) TCP+CBR ON/Of f 0.92 0.614 0 0.3 0 0 0 CBR ladder only 0.92 0.61 0 0.3 0 0 0 T able 4.6 F actors ef fects using pathChirp in lo w bandwidth netw orks Response T ests E1 E2 E12 (PSize) (Spread F actor) TCP Only -0.13 0.02 0.03 Accurac y TCP+CBR 0.06 -0.01 -0.01 TCP+CBR ON/Of f -0.15 -0.02 -0.02 CBR ladder only 0.01 -0.008 -0.018 TCP Only 6.48 -3.85 1.02 Con v er gence TCP+CBR 0.28 -1.81 -0.16 time(Sec) TCP+CBR ON/Of f 1.6 -0.54 -0.07 CBR ladder only 1.4 -0.5 -0.34 TCP Only 0 -0.07 0 Intrusi v eness TCP+CBR 0.001 -0.07 0 (MB) TCP+CBR ON/Of f 0.001 -0.07 0 CBR ladder only 0.24 -0.07 -0.04 27

PAGE 37

4.2 Case 2: High band width scenario 4.2.1 Analysis of IGI T able 4.7 Case 2: 2 k f actorial design using IGI in high bandwidth netw orks P ack et Pkts per Frequenc y T ests Accurac y Con v er gence Intrusi v eness Size T rain T ime (Sec) (MB) TCP Only 0.27 0.56 0.02 500 8 Continuous TCP+CBR 0.33 0.69 0.03 TCP+CBR ON/Of f 0.13 0.59 0.02 CBR ladder only 0.06 0.65 0.02 TCP Only 0.26 0.62 0.07 1500 8 Continuous TCP+CBR 0.32 0.73 0.08 TCP+CBR ON/Of f 0.12 0.60 0.07 CBR ladder only 0.05 0.71 0.08 TCP Only 0.23 1.71 0.22 500 64 Continuous TCP+CBR 0.26 1.76 0.24 TCP+CBR ON/Of f 0.11 1.69 0.21 CBR ladder only 0.06 1.68 0.21 TCP Only 0.21 1.8 0.7 1500 64 Continuous TCP+CBR 0.25 1.87 0.82 TCP+CBR ON/Of f 0.09 1.71 0.65 CBR ladder only 0.05 1.74 0.68 TCP Only 0.30 0.56 0.08 500 8 6 TCP+CBR 0.34 0.78 0.03 TCP+CBR ON/Of f 0.15 0.57 0.02 CBR ladder only 0.09 1.85 0.75 TCP Only 0.29 0.86 0.199 1500 8 6 TCP+CBR 0.32 0.99 0.12 TCP+CBR ON/Of f 0.13 0.67 0.08 CBR ladder only 0.06 0.74 0.09 TCP Only 0.28 1.76 0.232 500 64 6 TCP+CBR 0.26 1.92 0.29 TCP+CBR ON/Of f 0.12 1.71 0.22 CBR ladder only 0.08 1.72 0.22 TCP Only 0.27 1.82 0.75 1500 64 6 TCP+CBR 0.24 2.02 0.97 TCP+CBR ON/Of f 0.10 1.8 0.60 CBR ladder only 0.06 1.76 0.7 The results observ ed in high bandwidth netw orks follo w closely with that observ ed in lo w bandwidth netw orks. An interesting point that can be observ ed from T able 4.7 is that the o v erall accurac y v alues, con v er gence time and intrusi v eness v alues are signicantly better than those measured in lo w bandwidth netw orks. F or e xample: On comparison of accurac y con v er gence time and intrusi v eness v alues (0.56, 1.68 and 0.04 respecti v ely) measured for the 500, 8, Continuous case with TCP only 28

PAGE 38

in lo w bandwidth netw orks (Ro w 1 from T able 4.1) to the corresponding v alues (0.27, 0.56 and 0.02 respecti v ely) measured in high bandwidth netw orks (Ro w 1 from T able 4.7), it can be seen that IGI performs better in high bandwidth netw orks than it does in lo w bandwidth netw orks. This suggests that the IGI ABET will be more accurate, less intrusi v e with better turnaround time in high bandwidth applications as compared to lo w bandwidth netw orks. T able 4.8 Case 2: 2 k f actorial design using P athload in high bandwidth netw orks # of Pkts per Frequenc y T ests Accurac y Con v er gence Intrusi v eness T rains T rain T ime (Sec) (MB) TCP Only 0.40 6.29 0.59 4 50 Continuous TCP+CBR 0.63 6.06 0.562 TCP+CBR ON/Of f 0.56 4.23 0.42 CBR ladder only 0.08 2.85 0.26 TCP Only 0.33 21.45 0.67 12 50 Continuous TCP+CBR 0.592 19.2 1.76 TCP+CBR ON/Of f 0.54 18.20 1.69 CBR ladder only 0.18 10.48 0.99 TCP Only 0.39 6.59 1.22 4 100 Continuous TCP+CBR 0.616 5.08 0.9 TCP+CBR ON/Of f 0.48 4.65 0.49 CBR ladder only 0.08 2.79 0.51 TCP Only 0.30 21.80 4.09 12 100 Continuous TCP+CBR 0.597 21.8 4.09 TCP+CBR ON/Of f 0.47 19.58 3.67 CBR ladder only 0.08 8.08 1.52 TCP Only 0.45 5.32 0.49 4 50 6 TCP+CBR 0.646 5.78 0.535 TCP+CBR ON/Of f 0.58 4.86 0.45 CBR ladder only 0.08 3.10 0.29 TCP Only 0.34 25.66 2.42 12 50 6 TCP+CBR 0.582 20.1 1.81 TCP+CBR ON/Of f 0.50 21.12 1.95 CBR ladder only 0.17 11.50 1.09 TCP Only 0.42 5.71 0.99 4 100 6 TCP+CBR 0.64 5.22 0.94 TCP+CBR ON/Of f 0.48 5.21 0.96 CBR ladder only 0.08 2.91 0.54 TCP Only 0.34 21.09 3.94 12 100 6 TCP+CBR 0.557 17.6 3.17 TCP+CBR ON/Of f 0.50 21.37 4.00 CBR ladder only 0.17 8.40 1.59 29

PAGE 39

4.2.2 Analysis of P athload From T able 4.8, it can be seen that the best performance is obtained when using 4 trains, 50 pack ets per train and running the tool continuously or by 4 trains, 100 pack ets per train and running the tool continuously Using either of those v alues the best accurac y con v er gence time and o v erhead is achie v ed. Looking at T ables 4.8 and 4.11 together we can see that 1) Increasing the number of trains from 4 to 12 impro v es the accurac y b ut w orsens the con v er gence time to a great e xtent. W e can conclude that a smaller number of trains such as 4 is appropriate; 2) Increasing the number of pack ets per train from 50 to 100 impro v es the accurac y slightly and w orsens the con v er gence time and o v erhead only v ery slightly Hence, it can be inferred that 100 is a good choice for number of pack ets per train. Finally reducing the frequenc y of measurement impro v es the accurac y by a small mar gin, af fects the con v er gence and o v erhead v ery ne gligibly So running the tool continuously is a good option. Comparing the measurements in lo w and high bandwidth netw orks from T ables 4.2 and 4.8, P athload seems to be producing better results in lo w bandwidth netw orks as compared to high bandwidth netw orks. 4.2.3 Analysis of pathChir p From table 4.9, we can see that accurac y wise, pathChirp does particularly w orse in the case where we ha v e 'TCP Only' traf c as compared to all other cases. The best con v er gence time is obtained with 500 pack ets with 1.2 spread f actor This combination also gi v es a good accurac y and con v er gence time. It can also be seen that the con v er gence times and intrusi v eness v alues achie v ed in high bandwidth netw orks are signicantly better than those in lo w bandwidth netw orks. This indicates that pathChirp may be better geared for applications in high bandwidth netw orks. 30

PAGE 40

T able 4.9 Case 2: 2 k f actorial design using pathChirp in high bandwidth netw orks P ack et Spread T ests Accurac y Con v er gence Intrusi v eness Size F actor T ime (sec) (MB) TCP Only 0.42 0.10 0.38 500 1.1 TCP+CBR 0.09 0.10 0.38 TCP+CBR ON/Of f 0.07 0.10 0.38 CBR ladder only 0.03 0.06 0.02 TCP Only 0.81 0.19 1.14 1500 1.1 TCP+CBR 0.12 0.19 1.14 TCP+CBR ON/Of f 0.09 0.19 0.38 CBR ladder only 0.04 0.06 0.06 TCP Only 0.59 0.08 0.38 500 1.2 TCP+CBR 0.06 0.09 0.38 TCP+CBR ON/Of f 0.05 0.08 0.38 CBR ladder only 0.03 0.06 0.01 TCP Only 0.62 0.13 1.14 1500 1.2 TCP+CBR 0.09 0.13 1.14 TCP+CBR ON/Of f 0.07 0.13 0.38 CBR ladder only 0.03 0.06 0.03 T able 4.10 F actors ef fects using IGI in high bandwidth netw orks Response T ests E1 E2 E3 E12 E13 E23 E123 (PSize) (Pkts/train) (Freq) TCP Only -0.01 -0.03 0.04 0.0 0.04 0.07 -0.03 Accurac y TCP+CBR -0.01 -0.07 0.01 0.0 0.01 0.00 0.00 TCP+CBR ON/Of f -0.02 -0.01 0.01 0.0 0.00 0.01 0.01 CBR ladder only -0.01 -0.02 0.01 0.01 0.02 0.01 -0.02 TCP Only 0.17 1.17 0.03 -0.10 0.10 0.00 -0.12 Con v er gence TCP+CBR 0.12 1.10 0.16 -0.01 0.04 -0.01 -0.05 time(Sec) TCP+CBR ON/Of f 0.06 1.25 -0.16 0.13 -0.16 0.15 0.13 CBR ladder only 0.23 1.20 -0.14 -0.18 0.17 0.17 -0.18 TCP Only 0.3 0.4 0.01 0.2 0 0 0 Intrusi v eness TCP+CBR 0.36 0.5 0.06 0.3 0 0 0 (MB) TCP+CBR ON/Of f 0.2 0.4 0 0.2 0 0 0 CBR ladder only 0.2 0.4 0 0.2 0 0 0 31

PAGE 41

T able 4.11 F actors ef fects using P athload in high bandwidth netw orks Response T ests E1 E2 E3 E12 E13 E23 E123 (# T rains) (Pkts/train) (Freq) TCP Only -0.09 -0.02 0.03 0.00 -0.01 0.00 0.01 Accurac y TCP+CBR -0.05 -0.01 0.00 0.00 -0.02 -0.01 -0.01 TCP+CBR ON/Of f -0.02 -0.06 0.00 0.03 0.00 0.01 0.02 CBR ladder only 0.02 0.02 0.06 0.02 0.07 -0.02 -0.02 TCP Only 16.52 -0.88 0.41 -1.23 1.34 -1.21 -1.25 Con v er gence TCP+CBR 14.14 -0.36 -0.86 0.41 -0.79 -1.17 -1.38 time (Sec) TCP+CBR ON/Of f 15.33 0.60 1.48 0.22 0.88 -0.30 -0.27 CBR ladder only 4.08 1.18 3.05 1.31 2.86 -2.83 -2.76 TCP Only 1.96 1.52 0.32 0.95 0.48 -0.51 -0.44 Intrusi v eness TCP+CBR 1.97 1.11 -0.21 0.74 -0.22 -0.23 -0.26 (MB) TCP+CBR ON/Of f 2.25 1.15 0.27 0.86 0.02 0.13 -0.09 CBR ladder only 0.65 0.63 0.30 0.38 0.28 -0.26 -0.25 T able 4.12 F actors ef fects using pathChirp in high bandwidth netw orks Response T ests E1 E2 E12 (PSize) (Spread F actor) TCP Only 0.21 -0.01 -0.18 Accurac y TCP+CBR 0.03 -0.03 0.00 TCP+CBR ON/Of f 0.02 -0.02 0.00 CBR ladder only 0.01 -0.01 -0.01 TCP Only 0.07 -0.04 -0.02 Con v er gence TCP+CBR 0.07 -0.04 -0.02 time(Sec) TCP+CBR ON/Of f 0.07 -0.04 -0.02 CBR ladder only 0.00 0.00 0.00 TCP Only 0.76 0.00 0.00 Intrusi v eness TCP+CBR 0.76 0.00 0.00 (MB) TCP+CBR ON/Of f 0.00 0.00 0.00 CBR ladder only 0.03 -0.02 -0.01 32

PAGE 42

4.3 Analysis of ABETs with CBR stair case From T able 4.13, we see that in the presence of only CBR traf c v arying in a staircase pattern, all tools e xhibit good accurac y W e can see that as bottleneck bandwidth increases from to 10 Mbps to 1000 Mbps, accurac y of IGI and pathChirp increases, whereas accurac y of P athload w orsens. From T able 4.14, IGI and pathChirp ha v e much lo wer con v er gence times as compared to P athload. As bottleneck bandwidth increases from to 10 Mbps to 1000 Mbps, con v er gence times of IGI and pathChirp impro v es, whereas con v er gence time of P athload remains relati v ely high. From T able 4.15, IGI and pathChirp ha v e much lo wer o v erhead as compared to P athload. As bottleneck bandwidth increases from to 10 Mbps to 1000 Mbps, o v erhead for IGI and pathChirp are minimal as compared to that of P athload. T able 4.13 Accurac y measurements for ABETs with CBR staircase cross traf c Bottleneck BW 10Mbps 100Mbps 500Mbps 1000Mbps A v erage IGI 0.08 0.03 0.05 0.05 0.05 P athload 0.01 0.06 0.02 0.08 0.05 pathchirp 0.05 0.04 0.03 0.03 0.04 T able 4.14 Con v er gence time (Secs) measurements for ABETs with CBR staircase cross traf c Bottleneck BW 10Mbps 100Mbps 500Mbps 1000Mbps A v erage IGI 0.88 0.62 0.61 0.65 0.69 P athload 2.97 2.36 3.71 2.84 2.97 pathchirp 0.55 0.06 0.06 0.06 0.18 T able 4.15 Intrusi v eness measurements for ABETs with CBR staircase cross traf c Bottleneck BW 10Mbps 100Mbps 500Mbps 1000Mbps A v erage IGI 0.01 0.09 0.09 0.10 0.07 P athload 0.30 0.21 0.34 0.26 0.28 pathchirp 0.10 0.02 0.02 0.03 0.04 33

PAGE 43

4.3.1 Conclusions fr om the analysis of IGI, P athload and pathChir p tools Based on the results from our e xperiments, it can be concluded that in terms of accur acy in lo w bandwidth netw orks, P athload gi v es the best follo wed by IGI and pathChirp. In high bandwidth netw orks, IGI gi v es the best accurac y follo wed by P athload and then pathChirp. pathChirp performs particularly w orse in the rst scenario where only TCP traf c is present. In terms of intrusiveness both in lo w and high bandwidth netw orks, P athload introduces more o v erhead as compared to IGI and pathChirp, ho we v er it is only a small fraction of the total traf c. The most telling dif ference is observ ed in the con ver g ence times v alues for the three tools. In lo w bandwidth netw orks, IGI has the best (smallest) con v er gence times follo wed by pathChirp and P athload has the w orst con v er gence times. Ho we v er in high bandwidth netw orks, both IGI and pathChirp perform v ery well in terms of con v er gence times and P athload does e v en w orse. In general, it is seen that all the ABETs perform v ery well in terms of accurac y when only CBR cross traf c is present. It can be further noted that pathChirp e xhibits v ery good con v er gence times in this scenario. It can also be seen that changing some of the f actors from lo w to high or otherwise has a direct positi v e impact on one response and a direct ne gati v e impact on another response. F or e xample in T able 4.1 for IGI in lo w bandwidth netw orks, we can see that increasing the pack et size from 500 to 1500 and pack ets per train from 8 to 64, impro v es the accurac y by 60% b ut w orsens the con v er gence time by 160%. But if pack et size is increased from 500 to 1500 and pack ets per train stays at 8, then accurac y is impro v ed by 23% and con v er gence time w orsens only by 13%. Also, T radeof fs e xist between accurac y con v er gence time and intrusi v eness. In man y cases better accurac y may call for longer con v er gence times and vice v ersa. Similarly better con v er gence times may entail more intrusi v eness. Another important observ ation that can be made is that the cases of lo w accurac y detected in IGI measurements are mostly a result of o v er -estimation when the netw ork is o v er loaded and instances of lo w accurac y detected in P athload and pathChirp measurements are mostly a result of under -estimation when the netw ork is not loaded. These results can serv e as a source of guidance for users while picking ABETs to use in particular applications and selecting the conguration for the ABETs. 34

PAGE 44

CHAPTER 5 APPLICA TION OF ABETS IN END T O END CONGESTION CONTR OL A v ailable bandwidth estimation can be used in man y applications such as congestion control in TCP QoS and SLA v erication, serv er selection, adjustment of encoding rates in streaming applications, optimal route selection in o v erlay netw orks, end-to-end admission control, etc. 5.1 TCP congestion contr ol The w ork in this thesis is focused on applying ABETs to TCP congestion control. TCP (T ransmission control protocol) is the most predominant transport layer protocol used in the Internet today TCP uses a closed-loop, end-to-end windo w-based congestion control algorithm. TCP uses an additi v e increase multiplicati v e decrease (AIMD) strate gy increasing its sending rate slo wly if pack ets are recei v ed correctly and decreasing its sending rate drastically if congestion is detected. The cur rent implementation of TCP Sack has tw o phases slo w start phase and congestion a v oidance phase. In slo w start phase, TCP doubles the congestion windo w (cwnd) for e v ery ackno wledgement pack et recei v ed and thereby follo ws an e xponential pattern of windo w gro wth until the slo w start threshold (ssthresh) is reached. At that point, TCP switches o v er to congestion a v oidance phase and cwnd gro ws linearly by one pack et per round trip time. On detection of congestion in the netw ork, TCP tak es an aggressi v e approach to control congestion. Upon detection of a pack et loss, the cwnd size is reduced by half (multiplicati v e decrease). When a timeout occurs, TCP slashes the congestion windo w to one and reenters the slo w start phase. This sending rate adjustment mechanism used by TCP is done re gardless of the actual netw ork bandwidth a v ailable to TCP and causes it to under perform. TCP can gain a lot by being more a w are of the netw ork bandwidth a v ailability instead of blindly adjusting its transmission rate. In this thesis, a tw o-step approach is adopted while using ABETs in TCP congestion control. The rst step is the determination of the ABET that w ould w ork best for use in TCP congestion 35

PAGE 45

control to impro v e TCP performance. The second step is de vising an algorithm to use the ABET bandwidth estimation in TCP congestion control. 5.1.1 Selection of ABET f or TCP congestion contr ol application Applications dictate the ABET metrics that impact most and those that are less critical relati v ely F or e xample, for applications such as QoS v erication, accurac y may be more critical than con v er gence time and o v erhead. From the e v aluation of ABETs in Chapter 4, it can be seen that there is a tradeof f between con v er gence time, accurac y and o v erhead. In order to use ABETs in TCP congestion control, con v er gence time and accurac y are the most important v ariables, in that order TCP may w ork ne with reasonably accurate estimates b ut it cannot af ford to ha v e old estimates. The load in the netw ork changes in a v ery dynamic manner and TCP needs to k eep track of these changes and react to them in a timely manner As f ar as con v er gence time is concerned, IGI ABET gi v es the best performance in both lo w and high bandwidth netw orks. pathChirp tak es a long time to con v er ge in lo w bandwidth netw orks b ut does con v er ge quickly in high bandwidth netw orks. P athload tak es the longest to con v er ge in both lo w and high bandwidth netw orks. Between IGI and pathChirp, IGI allo ws dynamic scheduling of the estimations at an y time (user or application triggered) whereas pathChirp runs continuously and controls the chirp interarri v al times itself. As a result, TCP will not be able to tell if pathChirp completed an estimation yet or not as the control ne v er comes back to TCP Hence, IGI emer ges as the most appropriate ABET to be applied in our TCP congestion control application. 5.1.2 Algorithm f or ABET -based TCP Sack The congestion control in TCP Sack is modied such that it uses the a v ailable bandwidth estimate from IGI that is running in parallel between the same end points. TCP Sack [17 ] w as considered as it is the most widely used TCP v ersion today [18 ]. The congestion control mechanism for the ABET -based TCP Sack v ersion is based on the congestion control mechanism of TCP W estw ood [20 ]. 36

PAGE 46

5.1.2.1 TCP W estw ood In TCP W estw ood the sender continuously computes the connection Bandwidth Estimate (BWE), which is dened as the share of the bottleneck bandwidth used by the connection. BWE is equal to the rate at which data is deli v ered to the TCP recei v er; it is calculated utilizing the rate at which A CKs are recei v ed and the minimum R TT e xperienced by the connection. In TCP W estw ood, the congestion windo w increments during slo w start are e xponential and during congestion a v oidance are linear just as in TCP Reno. Ho we v er after a pack et loss e v ent or a timeout, TCP W estw ood gets the congestion windo w and ssthreshold according to the BWE as follo ws: if (3 DUPACKs are received) ssthresh = (BWE RTTmin) / segsize; if (cwin > ssthresh) / congestion avoid. / cwin = ssthresh; endif endif In case a pack et loss is indicated by a timeout e xpiration, cwin and ssthresh are set as follo ws: if (coarse timeout expires) cwin = 1; ssthresh = (BWE RTTmin) / segsize; if (ssthresh < 2) ssthresh = 2; endif endif In the abo v e algorithm after a timeout, cwin and the ssthresh are set equal to 1 and BWE, respecti v ely The congestion control algorithm used by the ABET -based TCP Sack v ersion sets the congestion windo w and ssthreshold v alues according to the a v ailable bandwidth estimate pro vided by the ABET tool at those points in time, as follo ws: if (3 DUPACKs are received) ssthresh = (TE RTTmin) / segsize; / TE is the tool estimate / 37

PAGE 47

if (cwin > ssthresh) / congestion avoid. / cwin = ssthresh; endif endif In case a pack et loss is indicated by a timeout e xpiration, cwin and ssthresh are set as follo ws: if (coarse timeout expires) cwin = 1; ssthresh = (TE RTTmin) / segsize; / TE is the tool estimate / if (ssthresh < 2) ssthresh = 2; endif endif5.1.2.2 Simulations in lo w and high band width netw orks A simple dumbbell topology is used with a bottleneck bandwidth of 25 Mbps, with FIFO queue and Drop T ail, and a propagation delay of 40 ms. TCP Sack and IGI tool were attached to the tw o end nodes and the IGI tool w as congured to use a probing pack et size of 500 bytes, 8 pack ets per train and continuous frequenc y as sho wn in Figure 5.1. In Figure 5.2 and 5.3, the instantaneous and a v erage throughput of original TCP Sack and ABET -based TCP Sack are sho wn. As it can be seen, the ABET -based TCP v ersion impro v es TCP performance during the initial part of the connection achie ving higher throughput f aster Similarly during the steady state phase the ABET -based TCP achie v es a throughput closer to the capacity of the channel. 38

PAGE 48

n1r ABETr TCPr Sourcer n4r ABETr TCPr Sinkr n2r n3r 1000 Mbpsr 25 Mbpsr 1000 Mbpsr 2 msr 40 msr 2 msr Figure 5.1 Simulation topology for application of ABET in lo w bandwidth netw orks 0 5e+06 1e+07 1.5e+07 2e+07 2.5e+07 3e+07 0 50 100 150 200 250 300 Throughput (bps) Time (in seconds) ORIG TCP SACK INSTANTANEOUS THROUGHPUT ORIG TCP SACK CUMULATIVE THROUGHPUT Figure 5.2 Lo w bandwidth netw ork Instantaneous and a v erage throughput of TCP Sack 39

PAGE 49

0 5e+06 1e+07 1.5e+07 2e+07 2.5e+07 3e+07 0 50 100 150 200 250 300 Bandwidth in BPS Time (in seconds) MODIFIED TCP SACK INSTANTANEOUS THROUGHPUT MODIFIED TCP SACK CUMULATIVE THROUGHPUT IGI ESTIMATE Figure 5.3 Lo w bandwidth netw ork Instantaneous and a v erage throughput of ABET -based TCP using IGI n1r ABETr TCPr Sourcer n4r ABETr TCPr Sinkr n2r n3r 4000 Mbpsr 700 Mbpsr 4000 Mbpsr 2 msr 40 msr 2 msr Figure 5.4 Simulation topology for application of ABET in high bandwidth netw orks Another simulation w as carried out to compare TCP Sack and the proposed ABET -based TCP Sack with the topology sho wn in Figure 5.4 and using a bottleneck bandwidth of 700 Mbps, with FIFO queue and Drop T ail, and a propagation delay of 40 ms. Figure 5.5 and Figure 5.6 sho w the instantaneous and a v erage throughput of original TCP Sack and ABET -based TCP Sack. It can be seen that TCP Sack is unable to scale with higher bottleneck capacity because it prematurely e xits 40

PAGE 50

slo w start and enters congestion a v oidance. TCP congestion control algorithm is not aggressi v e enough in utilizing its share of the a v ailable bandwidth in big pipes. So it tak es TCP Sack a long time to ll the pipe of 700 Mps. The ABET -based TCP Sack quickly utilizes the lar ge a v ailable bandwidth and achie v es superior throughput. The impro v ement achie v ed by ABET -based TCP Sack is v ery signicant in high bandwidth netw orks. 0 2e+08 4e+08 6e+08 8e+08 1e+09 0 50 100 150 200 250 300 Bandwidth in BPS Time (in seconds) TCP SACK INSTANTANEOUS THROUGHPUT TCP SACK CUMULATIVE THROUGHPUT Figure 5.5 High bandwidth netw ork Instantaneous and a v erage throughput of TCP Sack 41

PAGE 51

0 2e+08 4e+08 6e+08 8e+08 1e+09 0 50 100 150 200 250 300 Bandwidth in BPS Time (in seconds) MODIFIED TCP SACK INSTANTANEOUS THROUGHPUT MODIFIED TCP SACK CUMULATIVE THROUGHPUT Figure 5.6 High bandwidth netw ork Instantaneous and a v erage throughput TCP SA CK ABET based TCP using IGI The follo wing e xperiment is included to demonstrate the importance of choosing the right ABET for the application at hand. The topology used w as as sho wn in Figure 5.1 with the ABET tool here being P athload with parameters of 4, 50 and Continuous. Here we compare re gular TCP Sack and (P athload) ABET -based TCP Sack. Figure 5.7 sho ws the instantaneous and a v erage throughput of ABET -based TCP v ersion using P athload. P athload does not react f ast enough and therefore supplies inaccurate v alues to set the ssthr esh and cwnd v alues in TCP Sack. It is important to notice that these v alues are inaccurate not because the estimations made by P athload are inaccurate b ut because P athload has a long con v er gence time and the v alues used by TCP are older estimates and hence do not correctly reect the a v ailable bandwidth. As it can be seen, choosing the wrong tool or e v en the wrong parameters can ha v e a detrimental ef fect on the performance. 42

PAGE 52

0 5e+06 1e+07 1.5e+07 2e+07 2.5e+07 3e+07 0 50 100 150 200 250 300 Bandwidth in BPS Time (in seconds) MODIFIED TCP SACK INSTANTANEOUS THROUGHPUT MODIFIED TCP SACK CUMULATIVE THROUGHPUT PATHLOAD AVERGAE ESTIMATE Figure 5.7 Instantaneous and a v erage throughput of ABET -based TCP using P athload 5.1.3 Ar eas f or futur e r esear ch W e sa w that all ABETs send a train of probe pack ets back and forth between end points in a connection across the netw ork and mak e a v ailable bandwidth estimations. This means that applications that use the ABETs ha v e to pay for it in the form of sharing the netw ork with the probing pack ets and hence losing a portion of their eligible bandwidth to the probe pack ets. In this chapter we sa w ABETs applied to TCP and the performance impro v ements that resulted. W e also concluded that ABETs can be adv antageous for use in applications after careful consideration of their perfor mance metrics and matching those with the application requirements appropriately ABET -based TCP causes e xtra netw ork o v erhead because of the ABET embedded within. It is well kno wn that the Internet traf c distrib ution is hea vy tailed. 80% of Internet traf c is actually carried by a small number of long-li v ed connections (long o ws) while the remaining lar ge number of short connections (short o ws) carry the rest of the traf c. The ratio of o v erhead bandwidth to actual data bandwidth for ABET -based TCP will be dif ferent for short and long o ws. W ith long o ws, the benets of prompt netw ork bandwidth utilization by ABET -based TCP may o v ershado w the o v erhead that has to be borne in the form of probe traf c. W ith short o ws, it seems lik e the percentage of netw ork o v erhead will be signicant and usage of ABET -based TCP 43

PAGE 53

may possibly not be benecial. Although we recognize that this is an important issue, the analysis of the ef fects of o v erhead w as outside the scope of this research as the main topic of this thesis w as performance e v aluation of ABETs. Another interesting aspect to research further w ould be to see ho w the e xtra o v erhead w ould impact the number of ABET -based TCP connections that a router can handle simultaneously as compared to plain TCP connections. The f airness and friendliness e v aluations of ABET -based TCP with itself and with other TCP v ariants w ould be another intersting future topic of study 44

PAGE 54

CHAPTER 6 CONCLUSION AND FUTURE W ORK This thesis includes a performance e v aluation of the most important A v ailable Bandwidth Estimation T ools (ABETs) P athload, IGI and pathChirp in terms of their accurac y con v er gence time and intrusi v eness in se v eral scenarios under a v ariety of load conditions. The ABETs operate by sending streams or pairs of pack et probes into the netw ork and gauging the a v ailable bandwidth in their o wn unique w ay The functioning of the ABETs is go v erned by f actors lik e probe pack et size, number of pack et trains, number of pack ets per train, etc. A 2 k f actorial e xperimental design is carried out using simulations to determine the importance of ABET f actors on the performance metrics. The results obtained can be used as a guide for users to select the most appropriate ABET and parameters for their application. It is concluded that IGI with its lo w v alues of con v er gence time, reasonable accurac y and small intrusi v eness is the ABET with the most suitable characteristics for application in TCP congestion control. An ABET -based v ersion of TCP Sack is de v eloped that uses the a v ailable bandwidth estimates of IGI to set the congestion windo w and ssthreshold v alues in the congestion control mechanism. The ABET -based TCP Sack sho ws signicant impro v ements o v er re gular TCP Sack in terms of achie v ed throughput in both lo w and high bandwidth netw orks. Ev aluation of the ABETs can be done in wireless scenarios to see their viability Such a study will also re v eal other f actors lik e contention, that may come into play in determining their perfor mance. It will be also interesting to run the ABETs in scenarios with dif ferent number of hops and see the impact on their performance. Ev aluation of the ABETs in real e xperiments carried out in the Internet or with a small netw ork test bed will thro w some more light on these ABETs and their suitability in netw orking applications. 45

PAGE 55

REFERENCES [1] N. Hu and P Steenkiste, “Ev aluation and Characterization of A v ailable Bandwidth Probing T echniques, ” IEEE J ournal on Selected Ar eas in Communications v ol. 21, No. 6, pp. 879– 894, August 2003. [2] M. Jain and C. Do vrolis, “End-to-End A v ailable Bandwidth: Measurement Methodology Dynamics, and Relation with TCP Throughput, ” IEEE/A CM T r ansactions on Networking v ol. 11, No. 4, pp. 537–549, August 2003. [3] V Ribeiro, R. H. Riedi, R. G. Baraniuk, J. Na vratil, and L. Cottrell, “pathChrip: Ef cient A v ailable Bandwidth Estimation for Netw ork P aths. A v ailable at http://citeseer .ist.psu.e du/ 56 48 36.h tml, ” [4] J. Strauss, D. Katabi, and F Kaashoek, “A Measurement Study of A v ailable Bandwidth Estimation T ools, ” in Pr oceedings of A CM SIGCOMM confer ence on Internet measur ement 2003, pp. 39–44. [5] B. Melander M. Bjorkman, and P Gunningber g, “A Ne w End-to-End Probing and Analysis Method for Estimating Bandwidth Bottlenecks, ” in Pr oceedings of IEEE Globecom 2000, pp. 415–420. [6] R. Prasad, C. Do vrolis, M. Murray and K. C. Claf fy “Bandwidth Estimation: Metrics, Measurement T echniques, and T ools, ” IEEE Network v ol. 17, No. 6, pp. 27–35, No v ember/December 2003. [7] Netw ork Simulator 2 (ns2), http://www .isi.edu/nsnam/ns/ [8] X. Jianxuan, M.Labrador and M.Guizani, “Performance Ev aluation of TCP o v er Optical Channels and Heterogeneous Netw orks, ” in Cluster Computing P aris, June 2004, pp. 1574– 1578. [9] V Jacobson, “pathchar: A T ool T o Infer Characteristics Of Internet P aths, ” in ftp://ftp.ee .lbl.go v/path c har / April 1997, pp. 905–914. [10] B. A. Mah, ”pc har: A T ool F or Measuring Internet P ath Char acteristics” http://www .emplo yees.or g/bmah/ Softw ar e/p cha r/, 2001. [11] K. Lai and M. Bak er “Nettimer: A T ool for Measuring Bottleneck Link Bandwidth, ” in Pr oceedings of USENIX 2001, pp. 123–134. [12] R. L. Carter and M. E. Cro v ella, “Measuring Bottleneck Link Speed in P ack et-switched Netw orks, ” P erformance Evaluation v ol. 27, pp. 297–318, 1996. [13] R. L. Carter and M. E. Cro v ella, Dynamic Server Selection Using Bandwidth Pr obing in W ide-Ar ea Networks. T echnical Report TR-96-007, 1996. 46

PAGE 56

[14] G. Jin, G. Y ang, B. Cro wle y and D. Agarw al, “Netw ork Characterization Service (NCS), ” in 10th IEEE Symposium on High P erformance Distrib uted Computing August 2001. [15] C. Do vrolis, P Ramanathan, and D. Moore, “What Do P ack et Dispersion T echniques Measure ?, ” in IEEE InfoComm April 2001, pp. 905–914. [16] V Jacobson, “Congestion A v oidance and Control, ” in Pr oceedings of SIGCOMM P alo Alto, CA,, Aug 1988, pp. 216–225. [17] S. Flo yd, J. Mahda vi, M. Mathis, and M. Podolsk y An Extension to the Selective Ac knowledg ement (SA CK) Option for TCP IETF RFC 2883, July 2000. [18] S. Flo yd, Thoughts on the Evolution of TCP in the Internet (ver sion 2) http://www .icir .or g/o yd/talks /ICIR-Mar1 7.p df. [19] A. Lo w and W K elton, Simulation Modeling and Analysis McGra w Hill, 1991. [20] R. W ang, M. V alla, M. Y Sanadidi, and M. Gerla, “Using Adapti v e Rate Estimation to Pro vide Enhanced and Rob ust T ransport o v er Heterogeneous Netw orks, ” in Pr oceedings of the 10th IEEE International Confer ence on Network Pr otocols (ICNP) P aris, France, No v 2002. 47