Elsevier

Computer Networks

Volume 50, Issue 12, 24 August 2006, Pages 1952-1973
Computer Networks

Design and evaluation of web proxies by leveraging self-similarity of web traffic

https://doi.org/10.1016/j.comnet.2005.10.003Get rights and content

Abstract

In this paper a new concept for the analysis of communication systems and their performance is presented. Our point of view is that insights in system workload can help in developing new methods for improving the perceived system performance. From measurements, we gained ideas about the typical request pattern. These obtained insights have been used in developing new methods for improving the system performance. To validate these new approaches, simulations have been used.

First, we present a fitting algorithm which directly deals with measurement data instead of an intermediate heavy-tailed distribution. This method provides good results for approximating the object-size distribution as well as the performance measures in an MG∣1 queue. The results of the fitting procedure allow a classification of the considered events; they provide a perfect classification of the space of data sizes in different classes.

Furthermore, we develop a new caching algorithm, class-based, least recently used (C-LRU), with the aim to obtain a good balance between small and large documents in the cache. Similarly, the new scheduling algorithm, class-based interleaving weighted fair queueing (CI-WFQ), exploits the distribution of the object sizes being requested to set its parameters such that good mean response times are obtained and starvation does not occur. We have found that both methods are suitable for the use in Web proxy servers, and present, in many cases, an improvement over the yet existing strategies. For the comparison of the methods, we have used trace-driven simulations. Both algorithms are parameterized using information on the requested object-size distribution. In this way, they can be seen as potentially adaptive to the considered workload.

Introduction

The appearance of modern high-speed networks conjoins new and different switching and transmission technologies with diverse mixtures of applications and services. The number of supported application types (telephone, e-mail, e-commerce), World-Wide Web increases permanently; consequently, the complexity of communication systems rises permanently. The transfer rates become higher to meet the increasing needs, so that, in a given tunnel, more and more data packets are in transit. To serve current and future users well, a reliable performance prediction for communication systems is crucial. Communication systems should have enough capacity to meet the expected demand from the users, thereby maintaining a reasonable quality of service. Before reconfiguring or adapting existing systems and before building new ones, performance evaluation can be used to forecast the implications of implementation or architectural modifications on the system performance.

Performance measurements are an useful aid in evaluating communication systems. They are based on the recording and analysis of quantities which reflect the authentic traffic in existing systems. Although measurements are exact, they are costly and time consuming to obtain. Furthermore, measurement data can only be obtained from existing systems, not for new, planned system. Frequently, discrete-event simulations are made to determine the performance of communication systems. One can also gain information on the performance of a communication system by using analytical or numerical methods. It has been shown that queueing systems constitute suitable models for communication systems [36]; traffic is fed to a queue and various performance measures, such as the queue length, the waiting and response times, the server utilization, can be computed.

However, many used models are not convenient for present and future communication systems; the new system characteristics lie outside the modeling capabilities of these models. Therefore, developing new performance evaluation approaches to adequately dimension new communication systems is a big challenge. The models to be developed should be realistic, so that all relevant system aspects are represented accurately. Thus, insights in the true traffic characteristics should be considered.

Historically, traffic modeling has its origin in conventional telephony [59]. Traditionally, Poisson processes have been used for modeling arrival processes in communication systems, or more generally, in queueing systems. Furthermore, Markovian assumptions about arrival patterns as well as about resource holding times have been made. However, recent measurement studies of network traffic from local-area networks [46], in multimedia and video traffic [6], [14], [27], signaling traffic [20], and high-speed networks [13], [25] have shown that the traffic has slowly decaying variances. In many studies, the presence of long-term correlations, burstiness, self-similarity, fractality and long-range dependencies in network traffic has been shown. The property of self-similarity has been associated with the heavy-tailedness of the involved distributions, e.g., for the object-size distributions or inter-arrival times. Furthermore, it has been realized that traffic with self-similar properties has serious implications on the system performance, and that ignoring these properties can lead to underestimation of performance measures [15], [46], [59]. Consequently, designing and dimensioning systems by only taking into account the mean load requirements is not acceptable, since it fails to consider the behavioral dynamics associated with burstiness, self-similarity and long-range dependency [47].

This very specific nature of network traffic necessitates a new paradigm in the modeling and dimensioning of communication systems. The challenge here is to develop models that can describe the observed behavior correctly and that are able to quantify the performance measures as well. In fact, various efforts have been pursued to develop appropriate models. Unfortunately, almost all of them are focussed on the models’ descriptive capabilities, the engineering implications and analyses have been largely ignored [59]. However, the success of models dealing with such issues as self-similarity should not only depend on how well they describe actual network traffic, but should also depend on the ability to use these models in network analysis and control.

Due to the importance of the above mentioned properties and their implications on the communication system performance, many researchers started developing traffic models dealing with these properties. A number of researchers attempted to develop analytically tractable models to characterize self-similarity; they used Markovian arrival processes or Markov modulated Poisson processes [1], [7], [8], [24], [31], [32], [33], [34], [53]. Many other studies focus on the analysis of the heavy-tailed distributions and their impact on the system performance [22], [37], [38], [43], [40], [41], [51], [52], [58]. Both study directions try to use Markovian models by their considerations.

This paper provides a new concept for the analysis of communication systems and their performance. A workload-driven design and evaluation of Web-based systems has been realized. Insights in system workload can help in developing new methods for improving the perceived system performance. From measurements, we have gained ideas about the typical request pattern. These obtained insights have been used in developing new methods for improving the system performance. To validate these new approaches, simulations have been used.

The rest of this paper is organized as follows. In Section 2, we present a direct fitting method which directly deals with measurements data instead of an intermediate HTD (Heavy-Tailed Distribution). We will show that our method provides better results for describing the object-size distribution as well as for performance results in an MG∣1 queue. Out of the considerations made by the analysis of Web server traffic workload, new caching and scheduling strategies for Web servers are developed. A class-based least recently used replacement strategy (C-LRU) is presented in Section 3 as well as a class-based interleaving weighted fair queueing scheduling algorithm (CI-WFQ) in Section 4. For the comparison of the new strategies, we have used trace-driven simulations. In Section 6, we conclude with a summary of the key results.

Section snippets

Fitting measurements data to distribution

Extensive statistical analyzes of high time-resolution traffic measurement, as reported, e.g., in [59], have provided convincing evidence that actual traffic data from packet networks in operation are consistent with statistical self-similar or fractal characteristics. Although bursty or fractal phenomena have been observed in many branches of science and engineering, and fractal models have been applied with some success in areas such as hydrology, financial economics and biophysics, they are

Caching: class-based least recently used

Today, the largest share of traffic in the Internet originates from WWW requests. The increasing use of WWW-based services has not only led to high frequented Web servers but also to heavily used components of the Internet. Fortunately, it is well known that there are popular and frequently requested sites, so that object caching can be employed to reduce Internet network traffic [28] and to decrease the perceived end-to-end delays. Caching has been recognized as one of the most important

Scheduling

Modern WWW servers have to serve hundreds of requests simultaneously, each of which requires enough bandwidth and hardware resources. The operating system and the server software are responsible for exploiting these resources efficiently. The operating system provides interfaces and services for the server which chooses the processing sequence for the jobs according to some strategy. It decides whether the requested files have to be read from main memory or from disk, and delivers the requested

CI-WFQ: conclusions

In this section we have investigated the role of scheduling algorithms on the performance of world-wide web proxy servers. Although FCFS scheduling is normally applied, this strategy is by no means optimal. Much better for performance is SJF. However, it might cause starvation of long jobs. We have introduced in this paper a new, non-preemptive scheduling approach that avoids the problem of starvation. The new scheduling approach is parameterised using information on the requested object-size

Conclusion and key results

In this paper, we have presented new analysis methods that help in improving the performance of Web-based communications systems. Specific for this work is the consideration of the observed workload characteristics to develop new algorithms for caching and scheduling for world-wide Web proxy servers.

Markovian traffic models cannot provide a perfect modeling of real traffic; this fact has been realized by considering their properties. However, Markovian models can deliver good approximation of

Rachid El Abdouni Khayari is working as an Assistant Professor/Research Scientist in the Institute for Computer Engineering, Department of Computer Science at the University of the Armed Forces Munich, Germany.

From 1998 to 2002 he worked as research assistant at the Department of Computer Science, Institute for Performance Evaluation and Distributed Systems at the Technical University of Aachen (RWTH) where he obtained his Ph.D. Nowadays, he works on the design and performance evaluation of

References (59)

  • Apache, The Apache Software Foundation. Available from:...
  • M.F. Arlitt et al.

    Internet web servers: workload characterization and performance implications

    IEEE/ACM Transactions on Networking

    (1997)
  • S. Asmussen, O. Nerman. Fitting phase-type distributions via the EM algorithm, in: Symposium i Anvendt Statistik, 1991,...
  • P. Barford et al.

    Changes in web client access patterns

    WWW Journal

    (1999)
  • J. Beran et al.

    Long-range dependence in variable-bit rate video traffic

    IEEE Transactions on Communications

    (1995)
  • S.C. Borst et al.

    Heavy tails: the effect of the service discipline

  • P. Cao, S. Irani, Cost-aware WWW Proxy Caching Algorithms, in: Proceedings of USENIX, Monterey, CA, December, 1997, pp....
  • L. Cherkasova

    Scheduling strategy to improve response time for web applications

  • T. Chiotis, F. Stanatelopoulos, B. Maglaris, Traffic source models for realistic ATM performance modelling, in: Proc....
  • T. Chiotis, C. Stathis, B. Maglaris, The impact of self-similarity on the statistical multiplexing of MPEG video data,...
  • M.E. Crovella et al.

    Self-similarity in world wide web traffic: evidence and possible causes

    IEEE/ACM Transactions on Networking

    (1997)
  • M.E. Crovella, R. Frangioso, M. Harchol-Balter, Connection Scheduling in web servers, in: Proceedings of the 2nd USENIX...
  • C.R, Cunha, A, Bestavros, M.E. Crovella, Characteristics of WWW client-based traces, Technical Report TR-95-010, Boston...
  • J. Dahmen, K. Beulen, H. Ney, A mixture density based approach for object recognition for image retrieval, 6th...
  • Digital Equipment Cooperation, Digital’s Web Proxy Traces....
  • D.E. Duffy et al.

    Statistical analysis of CCSN/SS7 traffic data from working subnetworks

    IEEE Journal on Selected Areas in Communications

    (1994)
  • A. Feldmann

    Characteristics of TCP connection arrivals

  • A. Feldmann et al.

    Fitting mixtures of exponentials to long-tail distributions to analyze network performance models

    Performance Evaluation

    (1998)
  • W. Fischer et al.

    The Markov-modulated Poisson process (MMPP) Cookbook

    Performance Evaluation

    (1992)
  • Cited by (0)

    Rachid El Abdouni Khayari is working as an Assistant Professor/Research Scientist in the Institute for Computer Engineering, Department of Computer Science at the University of the Armed Forces Munich, Germany.

    From 1998 to 2002 he worked as research assistant at the Department of Computer Science, Institute for Performance Evaluation and Distributed Systems at the Technical University of Aachen (RWTH) where he obtained his Ph.D. Nowadays, he works on the design and performance evaluation of distributed computer and communications systems, focusing on network traffic, caching and scheduling in World Wide Web-based systems.

    View full text