“Network management” in a large organization often involves—whether explicitly or implicitly—the responsibility for ensuring the availability and responsiveness of network resources attached to the network, such as servers and printers. Users often think of the services they rely on, such as web sites and email, as part of “the network”. Although tools exist for ensuring the availability of the servers running these services, ensuring their performance is a more difficult problem. In this dissertation, I introduce a novel approach to managing the performance of servers within a large network broadly and cheaply. I continuously monitor the border link of an enterprise network, building for each inbound connection an abstract model of the application-level dialog contained therein without affecting the operation of the server in any way. The model includes, for each request/response exchange, a measurement of the server response time, which is the fundamental unit of performance I use. I then aggregate the response times for a particular server into daily distributions. Over many days, I use these distributions to define a profile of the typical response time distribution for that server. New distributions of response times are then compared to the profile to determine whether they are anomalous. I applied this method to monitoring the performance of servers on the UNC campus. I tested three months of continuous measurements for anomalies, for over two hundred UNC servers. I found that most of the servers saw at least one anomaly, although for many servers the anomalies were minor. I found seven servers that had severe anomalies corresponding to real performance issues. These performance issues were found without any involvement of UNC network managers, although I verified two of the issues by speaking with network managers. I investigated each of these issues using the contextual and structural information measured at the border link in order to diagnose the cause of the issue. I found a variety of causes: overloaded servers from increased demand, heavy-hitting clients, authentication problems, changes in the nature of the request traffic, server misconfigurations, etc. Furthermore, information about the structure of connections proved valuable in diagnosing the issues.

—This work investigates the performance of a hybrid ad-hoc network extension to UMTS using a three-layer architecture with low cost relays through simulation. A mobile web user is considered; next to which several interfering mobiles are placed. Using a holistic view which considers all layers in the network stack, the following results are obtained: transmit power and therefore the number of errors experienced by the RLC layer are much smaller for the relay case. TCP delay is also smaller for the relay case, as are Web page response times. Also, Web page slowdown is better for the relay case.

— Empirical evaluations to study network performance, whether in a laboratory setting or on GENI testbeds, rely heavily on measurement-based modeling of round trip times (RTTs) to emulate realistic end-to-end delays of local and metropolitan area networks. For generating realistic traffic, we studied several models to emulate RTTs. In this paper, we performed experiments on real testbeds using synthetic TCP traffic generated from measurement data from a large university campus. As a result of our study, we present the Discrete-Approximation model for RTT (DA-RTT) emulation. Using three different metrics for performance evaluation, which include queue length at routers, connection response times, and connection durations, we demonstrate that the simple DA-RTT model closely represents the per-connection RTTs in the original traffic. While these experiments were performed in our laboratory, and not using GENI infrastructure, we present this as a possible model for adoption on GENI testbeds to emulate Round Trip Time Distributions for GENI experiments.

x Chapter

x Chapter

vii

stable object network paper A distributed web caching system can supply fast and stable transmission of information to the user avoiding a congested network section, storing and supplying the content that the user requested to cache by distributing and sharing a cache like a proxy server. It is located near the user. This paper proposes a simple client-based distributed web caching system (2HRCS) that can assign an object and control the load by using the direct connection of a client to the shared caches without the help of the Additional-DNS. In addition, this paper investigates the hash routing schemes such as CARP (Cache Array Routing Protocol) and GHS (Global Hosting System) which are shared web cache protocols. They need the Additional-DNS for the load balancing. The proposed system in this study makes the existing system simpler by removing the Additional-DNS and reducing the DNS query delay time for all objects. Moreover, the overhead of DNS can be lowered by reducing the general use of the DNS query through the direct object assignment for the client browser. A hash routing scheme system and the proposed system are compared in terms of the average delay time according to the object sizes.

of the Dissertation High-Fidelity Evaluation Framework and Application-Centric Performance Analysis of Vehicular Networks

ns c is LBNL’s Network Simulator [20]. The simulator is written in C++; it uses OTcl a s command and configuration interface.nsv2 has three substantial changes from nsv1: (1) the more complex objects in nsv1 have been decomposed into simpler components for greater flexibility and composabili ty; (2) the configuration interface is now OTcl, an object ori ented version of Tcl; and (3) the interface code to the OTcl interpr te is separate from the main simulator. Ns documentation is available in html, Postscript, and PDF f ormats. Seehttp://www.isi.edu/nsnam/ns/ns-documentation. html for pointers to these.

ns c is LBNL’s Network Simulator [14]. The simulator is written in C++; it uses OTcl as a command and configuration interface. ns v2 has three substantial changes from ns v1: (1) the more complex objects in ns v1 have been decomposed into simpler components for greater flexibility and composability; (2) the configuration interface is now OTcl, an object oriented version of Tcl; and (3) the interface code to the OTcl interpreter is separate from the main simulator.

many flows are short Web transfers and about 1% of the largest flows carry more than 50% of all bytes. Motivated by the high variability property of Internet traffic, we first analyze the performance of LAS under the M/G/1 queue, where G is a job size distribution with varying degree of variability. The analysis for LAS shows that the conditional mean response time highly depends on the variability of a job size distribution and that the percentage of large jobs that see penalty under LAS is negligible when a job size distribution shows the high variability property. While shortest remaining processing time (SRPT) is known to be the optimal policy in reducing the mean response time, we show that LAS offers a conditional mean response time close to SRPT. When compared to first-come first-serve (FCFS), LAS significantly outperforms FCFS for job size distributions with the high variability property. Moreover, as opposed to FCFS, we prove that LAS can service some jobs even at overload. We then use simulations to investigate the performance of LAS in packet networks for various network configurations. We also study the interaction of LAS with TCP. We find that LAS has quite interesting features that cause it to reduce the transfer time for short TCP flows without penalizing large flows a lot. In addition, we show through simulations that LAS allocates bandwidth equally among competing TCP connections in congested heterogeneous networks. The simulation results show that LAS penalizes long-lived TCP and UDP flows at high network load above 75%. To address this issue, we propose a family of new LAS-based scheduling policies that differentiate the service in order to improve the performance of priority flows. To evaluate these policies, we develop analytic models to compute the mean flow transfer time for TCP flows that share a single bottleneck link that uses LAS or one of new LAS-based policies . Over a wide range of settings , we show that the analytic models agree with the ns2 simulation of the policies in packet networks when loss rates are low. Evaluation of the proposed policies shows that they notably improve the performance of priority flows in terms of reducing their mean transfer time, average jitter, and loss rate while affecting ordinary flows in a minor way. Finally, we evaluate a hybrid policy called LAS-FCFS that services small jobs using LAS and large jobs using FCFS. This policy can be useful when the penalty for the largest jobs under LAS is not acceptable. Analytical results show that LAS-FCFS improves the performance of the largest jobs when a job size distribution exhibits the high variability property. We also analyze an evaluate variants of LAS-FCFS that offer service differentiation.

ions Abstraction is possible because most protocol architectures are layered such that one protocol makes use of another. In design or evaluation of a level-n protocol, we need information provided by level n-1 and below, but not necessarily (depending on the research questions) all the details of the lower level protocols. For instance, a multicast transport protocol may need multicast routing tables in order to forward multicast packets. The detailed exchange of messages to generate those routing tables is often not important. If we abstract away unnecessary details, the memory and time consumption that would have been used by these details can be conserved and used to simulate larger scale scenarios. Based on this reasoning, we proposed centralized computation and end-to-end packet delivery. These two techniques are aimed to abstract away, respectively, network layer and hop-by-hop transmission details that higher layer protocol studies may or may not require. Protocol complexity is only one dimension of the scaling problems. To be able to simulate very large networks and substantial amounts of traffic, we need to enhance those elements that construct network connectivity or generate traffic in the simulations. Currently, in most network simulators, the cost to construct network connectivity is proportional to the cost of creating nodes O(N), links O(L) and routing table O(N). For large topologies (N > 1000), routing table cost is the most significant. Hierarchical routing has the

enterprise file server, we propose an estimation method for relationship between file number and cumulative file size in descending order of file size based on a model for file size distribution. We develop the model by weighted summation of multiple log normal distribution based on AIC. File size data from technical and non-technical divisions of a company show that our model fits well with observed distribution, and that the estimated relationship can be utilized for cost-effective operational management of file server.

by Emil Sit S.B. (Math), Massachusetts Institute of Technology (1999) S.B. (EECS), Massachusetts Institute of Technology (1999) Submitted to the Department of Electrical Engineering and Computer Science in partial fulfillment of the requirements for the degree of Master of Engineering at the MASSACHUSETTS INSTITUTE OF TECHNOLOGY -May 2000 @ Emil Sit, MM. All rights reserved. The author hereby grants to MIT permission to reproduce and distribute publicly paper and electronic copies of this thesis document in whole or in part.

Zusammenfassung Mit wachsender Komplexität moderner Kommunikationsund Informationssysteme werden entsprechende Leistungsanalysen unter verschiedenen Lastszenarien für Anbieter dieser Systeme besonders wichtig. Der Bedarf nach geeigneten Werkzeugen zur Unterstützung des Experimentators bei der Lastmodellierung und -generierung steigt somit auch kontinuierlich. In diesem Beitrag werden ausgehend von dem Konzept eines verallgemeinerten Lastgenerators UniLoG ein spezieller Adapter UniLoG.IP zur Generierung von künstlichen IP-Verkehrslasten und seine Leistungscharakteristika vorgestellt. Die praktische Nutzung von UniLoG.IP wird im Rahmen einer umfangreichen QoS-Studie von Videostreaming in einem IEEE 802.11g WLAN unter verschiedenen IP-Hintergrundlasten demonstriert.

Zusammenfassung In der algorithmischen Spieltheorie modellieren wir das Verhalten von eigennützigen Agenten in großen Netzwerken als strategisches Spiel: Jeder Agent versucht, den eigenen Nutzen zu maximieren, ohne dabei auf mögliche Verschlechterungen für andere Agenten Rücksicht zu nehmen. Das dominierende Lösungskonzept in solchen Spielen ist das Nash-Gleichgewicht, ein Zustand, in dem kein Agent seine eigene Situation verbessern kann. Zwar ist es unstreitig, dass ein solcher Zustand stabil ist, aber es ist nicht unmittelbar klar, wie er erreicht werden kann, da eine Berechnung durch die Agenten auf Grund fehlenden Wissens und mangelnder Berechnungskraft nicht möglich ist. In dieser Arbeit betrachten wir daher ein dynamisches Modell, in dem die Agenten parallel und unabhängig voneinander einfache Verbesserungsschritte ausführen, die nur auf lokal verfügbarer Information basieren. Wir zeigen, dass eine Population von Agenten auf lange Sicht zu einem Nash-Gleichgewicht konvergiert. Ein Hauptproblem ist dabei das unkoordinierte parallele Vorgehen der Agenten, das zu einem Überschießen führen kann. Wir geben ein Kriterium an, das sicher stellt, dass dieser Effekt nicht auftritt, und zeigen, dass auch in sehr großen Netzwerken schnelle Konvergenz zum Gleichgewicht möglich ist.

World-wide web as well as proxy servers rely for their scheduling on services provided by the underlying operating system. In practice, this means that some form of first-come-first-served (FCFS) scheduling is utilized. Although FCFS is a reasonable scheduling strategy for job sequences that do not show much variance, for the world-wide web it has been shown that the requested-object sizes do exhibit heavy tails. Under these circumstances, job scheduling on the basis of shortest-job first (SJF) or shortest remaining processing time (SRPT) has been shown to minimize the total average waiting time. However, these methods have the disadvantage of potential job starvation. In order to avoid the problems of both FCFS and SJF we present in this paper a new scheduling approach called class-based interleaving weighted fair queueing (CI-WFQ). This scheduling approach exploits the specific characteristics of the job stream being served, that is, the distribution of the sizes of the objects being requested, to set its parameters such that good mean response times are obtained and starvation does not occur. In that sense, the new scheduling strategy can be made adaptive to the characteristics of the job stream being served. In this paper we compare the new scheduling approach (using trace-driven simulations) to FCFS, SJF and the recently introduced α-scheduling, and show that CI-WFQ combines very good performance (as far as mean and variance of response time and blocking probability are concerned) with a scheduling complexity almost as low as for FCFS (and hence, lower than for SJF and α-scheduling). The use of trace-driven simulation is essential, since the special properties of the arrival process makes analytical solutions very difficult to achieve. Copyright © 2005 John Wiley & Sons, Ltd.

World-Wide Web, workload characterization, performance, servers, caching, World Cup This paper presents a detailed workload characterization study of the 1998 World Cup Web site. Measurements from this site were collected over a three month period. During this time the site received 1.35 billion requests, making this the largest Web workload analyzed to date. By examining this extremely busy site and through comparison with existing characterization studies we are able to determine how Web server workloads are evolving. We find that improvements in the caching architecture of the World-Wide Web are changing the workloads of Web servers, but that major improvements to that architecture are still necessary. In particular, we uncover evidence that a better consistency mechanism is required for World-Wide Web caches.

World Wide Web is a media in constant growth, joining several components and services in an accelerated evolution. The new kinds of sites, such as, E-commerce, Web-publishing and demand video still uses more servers resources. In this context, adapting the new Web environment to performance evaluation, the present work accomplished a characterization study of several Apache Web servers traces, allowing collect important data that define the form as users and servers interact. With these data, four types of sites categories were analyzed: Default (composed of all trace), Academic, Web-publishing and Traditional. In this analysis, it was evaluated four aspects: the arrival time, the status code, the object class and the object size and at the end, mathematical models are proposed to represent those characteristics. Furthermore, a synthetic workload generator was also developed. With a graphical interface, the W4Gen (World Wide Web Workload Generator), as called, allows the users to generate new workloads based on mathematical models. Besides, it also allows to modify the four essential aspects preseted above to simulate new types of workloads. Finally, to validate the results, the Web server model with differentiated services was used, verifying the performance in overload situations.

World Wide Web (WWW) traffic will dominate network traffic for the foreseeable future. Accurate predictions of network performance can only be achieved if network models reflect WWW traffic statistics. Through analysis of usage logs at a range of caches it is shown that WWW traffic is not a Poisson arrival process, and that it displays significant levels of self-similarity. It is also shown for the first time that the self-similar variability extends to demand for individual pages, and is far more pervasive than previously thought. These measurements are used as the basis for a cache-modelling tool-kit. Using this software the impact of the variability on predictive planning is illustrated. The model predicts that optimisations based on predictive algorithms (such as least recently used discard) are likely to reduce performance very quickly. This means that, far from improving the efficiency of the network, conventional approaches to network planning and engineering will tend to reduce efficiency and increase costs.