Characterizing Web Response Time

Transcription

1 Characterizing Web Response Time Binzhang Liu Thesis submitted to the Faculty of the Virginia Polytechnic Institute and State University in partial fulfillment of the requirements for the degree of Master of Science in Computer Science Edward A. Fox, Chair Marc Abrams Roger W. Ehrich April 22, 1998 Blacksburg, Virginia Keywords: World Wide Web, Latency, Proxy Caching Copyright 1998, Binzhang Liu

2 Characterizing Web Response Time Binzhang Liu (ABSTRACT) It is critical to understand WWW latency in order to design better HTTP protocols. In this study we characterize Web response time and examine the effects of proxy caching, network bandwidth, traffic load, persistent connections for a page, and periodicity. Based on studies with four workloads, we show that at least a quarter of the total elapsed time is spent on establishing TCP connections with HTTP/1.0. The distributions of connection time and elapsed time can be modeled using Pearson, Weibul, or Log-logistic distributions. We also characterize the effect of a user s network bandwidth on response time. Average connection time from a client via a 33.6 K modem is two times longer than that from a client via switched Ethernet. We estimate the elapsed time savings from using persistent connections for a page to vary from about a quarter to a half. Response times display strong daily and weekly patterns. This study finds that a proxy caching server is sensitive to traffic loads. Contrary to the typical thought about Web proxy caching, this study also finds that a single stand-alone squid proxy cache does not always reduce response time for our workloads. Implications of these results to future versions of the HTTP protocol and to Web application design also are discussed.

3 Acknowledgments I thank Professor Edward A. Fox, my advisor, for his advice, help and support during my thesis and throughout my graduate program in the Department of Computer Science at Virginia Tech. I also thank the members of my advisory committee Professor Marc Abrams and Professor Roger W. Enrich for their comments and suggestions. Help and suggestions from other members of the Network Research Group (NRG) at Virginia Tech including Ghaleb Abdulla and Tommy Johnson also are appreciated. Grants from IBM and the National Science Foundation (NCR ) partially supported my graduate study. IBM donated the equipment used to collect the traffic log files, conduct experiments and analyze the data. Finally, many thanks go to Wen Wang, my wife and Andy Liu, my son for their understanding and support during the course of my graduate studies. iii

4 Contents 1 Introduction Introduction CharacteristicsofWebLatency ProblemStatement OrganizationoftheThesis Related Works WWWCharacterizations Client Server ProxyServer Document Network HTTPPerformanceandWebLatency HTTPperformanceproblems Approaches in improving HTTP performance and latency iv

5 3 Experimental Design WorkloadsUsedIntheStudy SoftwareUsedIntheExperiments Webjammasoftware Squidsoftware Experiments Experiment one: re-play four workloads without proxy via switched Ethernet Experiment two: re-play a subset of VT campus workload without a proxyviaa33.6kmodem Experiment three: re-play VT log with a proxy via switched Ethernet Experiment four: re-play a subset of VT campus workload with a proxyviaa33.6kmodem Experiment five: response time of a proxy and loads Experiment six: response time without a proxy via switched Ethernet under persistent connections Experiment seven: response time without a proxy via a 33.6 K modem under persistent connections Experiment eight: response time with a proxy via switched Ethernet under persistent connections Experiment nine: response time with a proxy via a 33.6 K modem under persistent connections Experiment ten: periodicity of response time v

6 Summary WWW Response Time Without a Proxy ResponseTimeViaSwitchedEthernet Averageconnectionandelapsedtime Distribution of connection time and elapsed time Distribution of ratio of connection time to elapsed time ResponseTimeViaa33.6KModem Averageconnectionandelapsedtime Distribution of connection time and elapsed time Summary Web Response Time with a Proxy AverageConnectionTimeandElapsedTime Distribution of Connection Time and Elapsed Time Comparison of response time of local accesses and remote accesses via 10baseTswitchedEthernet ResponseTimeandProxyTrafficLoads Summary WWW Response Time With Persistent Connections DistributionofNumberofEmbeddedImages Average Elapsed Time Via 10baseT Switched Ethernet vi

7 6.3 Average Connection Time and Elapsed Time Via a 33.6 K Modem ElapsedTimeSavingFromPersistentConnection Percentage of Elapsed Time Saving From Persistent Connection Ratio of Connection Time and Elapsed Time From Persistent Connection Summary Periodicity of WWW Response Time PlotoftheResponseTimeSeries SpectralAnalysis DailyandWeeklyPeriod CorrelationofResponseTimeandWebTraffic Summary Conclusions and Limitations Conclusions LimitationsandFurtherResearchNeeds VITA 66 vii

8 List of Figures 4.1 Cumulativedistributionofconnectiontime Cumulativedistributionofelapsedtime Cumulative distribution of the ratio of connection time to elapsed time Cumulative distribution of response time via a 33.6 K modem Cumulative distribution of connection time and elapsed time Cumulative distribution of connection time for local and remote access Response curve of response time to number of Webjamma processes Response curve of response time to proxy traffic load Histogram of number of unique embedded images in a page Distribution of number of unique embedded images in a page Timeplotofresponsetimeseries Time plot of standard deviation of response time series Plotofperiodogrambyperiod Responsetimebythehourofaday viii

9 7.5 Responsetimebythedayofaweek AccessestotheVTcampusproxy Accesses to the VT campus proxy by the hour of a day Accessestotheei.cs.vt.eduserver ix

10 List of Tables 3.1 Workloadsusedinthestudy Localaccessesvs.remoteaccesses Number of re-played requests, rates of successful requests and failed requests Experiment conducted between Oct. and Nov Problemdescriptionoftheexperiments Detaileddescriptionofthetenexperiments Average connection time, elapsed time and ratio of connection time to elapsed time Responsetimeforlocalandremoteaccesses Parameter estimates of distribution of connection time Parameterestimatesofdistributionofelapsedtime Connectiontimeforvariouscumulativefrequencies Responsetimeviaa33.6Kmodem Connection time of modem users under various cumulative frequencies Parameter estimates of distribution of connection time and elapsed time.. 28 x

11 5.1 Response time of proxy caching using VT campus workload Connection time of VT workload with a proxy under various cumulative frequencies Models and parameter estimates of distribution of response time Response time of proxy caching for local and remote accesses Connection time of VT campus proxy workload for local and remote accesses under various cumulative frequencies Regression analysis of the connection time and the number of processes Regression analysis of the elapsed time and the number of processes Parameter estimates of distribution of number of embedded images in a page Response time vs. connection type using 10baseT switched Ethernet Response time vs. connection type using a 33.6 K modem Elapsed time saving from persistent connection Percentage of elapsed time saving from persistent connection Ratio of connection time to elapsed time from persistent connection Regression analysis of the connection time and hourly accesses to the VT campusproxy Regression analysis of the connection time and hourly accesses to the ei.cs.vt.edu Webserver xi

12 Chapter 1 Introduction 1.1 Introduction In the past several years the World Wide Web has experienced tremendous growth, and became the dominant source of Internet traffic. Now millions are using the Internet for WWW traffic. The Web has become an important tool to access information. The Graphics, Visualization, & Usability Center s (GVU) 8th WWW User Survey reported that 84% of the users report that they considered access to the Web indispensable, nearly the same percentage as those who feel is indispensable, and 85% of the users use it daily. [1] HTTP/1.0 protocol is a simple request/response protocol, not designed for heavy use. In order to accommodate continually increasing WWW uses, HTTP needs to be effective and efficient. Based on studies of persistent connections and pipelining, HTTP/1.1 was designed to improve the performance. However, HTTP/1.1 has deficiencies of complexity, poor extensibility, lack of generality and poor scalability [2]. To solve these deficiencies, in July 1997, the World Wide Web Consortium (W3C) started the HTTP-NG project to design the next generation of the HTTP protocol that fulfills these requirements and resolves deficiencies of HTTP/1.1. The W3C group initiated a wide range of Web characterization studies. The 1

13 Binzhang Liu Chapter 1. Introduction 2 HTTP-NG activity statement indicates It is important to understand the actual system and how it is being used before attempting to optimize it. [3] In the past, though many studies have been characterizing Web traffic, little is known about the characteristics of Web latency. Web proxy caching is widely used in the Web system, but little is known about the effectiveness of proxy caching in improving Web latency. Research on these two issues should enhance the understanding of the overall Web system. 1.2 Characteristics of Web Latency Improving Web latency is a major research area. Many studies report that speed continues to be the number one concern of Web users. Some people even have called WWW the World-Wide Wait. [4] Web users do not like to wait for a Web page that takes a long time to retrieve. Web latency comes from many sources, involving the HTTP protocol, Web server implementations, network bandwidth, characteristics of Web documents, characteristics of Web clients and network topology. The W3C Web Characterization Group emphasizes characterizing the kinds of tasks actually performed using HTTP, and the kinds of documents that are retrieved. [3] It also is important to understand the nature of WWW latency in order to properly design, implement, and improve the WWW system. Long latency is observed from some Web sites. But the causes of long latency are not known yet. A study by Manley and Seltzer concluded that the latencies can not be explained by server over-loading and suggests that the bottleneck lies in the network [5]. Another study found that bandwidth-related delay may not account for much of the perceived latency [6]. Many techniques have been studied and used in reducing Web latency. These approaches include protocol change [7], Web proxy caching [8], prefetching [9], compression [10], and CCS style sheets [11].

14 Binzhang Liu Chapter 1. Introduction Problem Statement A majority of works on characterizing the Web were devoted to characterize Web traffic, Web usage patterns and Web documents. These studies are very important to understand the Web system. However, there is another important aspect of the Web system - Web latency. Characterization of Web response time will help to understand the nature of WWW latency and provide an empirical basis for improving Web scalability. It can give guidance to designers of Web system applications and HTTP-NG. Web proxy caching is aimed at saving network bandwidth and improving Web response time. Many studies have investigated the caching performance effects of different replacement algorithms [8]. Only a few studies have investigated the impacts of proxy caching on Web response time [19]. One objective of this study is to characterize Web response time. Then the usefulness of Web proxy caching in reducing Web latency will be analyzed. Specifically this study will answer the following questions: What kind of distribution does response time follow? Does proxy caching improve response time? What is the effect of network bandwidth on response time? How does response time change with different levels of traffic? What kind of distribution is followed by the count of embedded images in a page? How much elapsed time can be saved by persistent connections? Are there any periodic patterns for response time?

15 Binzhang Liu Chapter 1. Introduction Organization of the Thesis The remainder of this thesis is divided into seven chapters. Chapter 2 is a literature review. Chapter 3 describes the experimental design. Chapter 4 presents WWW response time without a proxy. Chapter 5 discusses Web response time with a proxy. Chapter 6 examines the effect of persistent connections on elapsed time. Chapter 7 studies the periodicity of response time. Chapter 8 concludes the thesis.

16 Chapter 2 Related Works 2.1 WWW Characterizations Many studies have characterized various aspects of the World Wide Web. Ghaleb Abdulla et al. identified five main Web objects: client, document, network, server, and proxy server [12]. The Web characterization efforts are mainly focused on these five aspects. We summarize some of the key findings for each of these aspects Client The first study of Web user behavior was performed by Catledge and Pitkow [13]. Based on empirical analysis of user event log files the paper showed interface usage data for XMosaic and characterization of user navigation patterns as serendipitous browsing, general browsing or searching. Cunha et al. [14] analyzed the characteristics of WWW client-based traces. Georgia Tech s Graphic, Visualization, & Usability Center (GVU) has conducted 8 WWW User Surveys since Jan These surveys provided valuable information about user characteristics. In the fourth survey conducted in Oct. 1995, it was reported that the main problem with using the Web was that it takes too long to view/download pages. In 5

17 Binzhang Liu Chapter 2. Related Work 6 the recent survey of April 1997, it was reported that the most common Web activity is to gather information (86%), followed by searching (63%), browsing (61%), work (54%), education (52%), communication (47%), and entertainment (45%) [1]. A study at Virginia Tech identified client sessions based on some reasonable heuristics and quantified the amount of searching and browsing per session [15] Server Mogul performed a comprehensive analysis of a WWW server in 1995 [16]. His study found that the majority of traffic was consumed by small images, the cumulative distribution of the number of requests was log-log, and the inter-arrival time of requests did not follow a Poisson process. Almeida et al. performed analysis on server logs from NCSA, SDSC, EPA and BU and showed that spatial reference locality can be characterized using the notion of self-similarity [17]. A study by Arlitt and Williamson identified ten invariants using six different data sets [18] Proxy Server Most research on proxy servers has focused on improving caching, especially the replacement algorithm [19]. Using a Digital Equipment Corporation proxy log, Glassman found that document access frequency followed a Zipf distribution [20]. Abdulla et al. have done a comprehensive analysis of proxy logs and identified nine invariants [21] Document Using 1.5 million textual objects retrieved from the WWW, in November 1995, Bray of Open Text Corporation found the number of unique URLs was 11,366,121. At that time, the size of an average page was between 6K and 7K bytes, and a large majority of all pages

18 Binzhang Liu Chapter 2. Related Work 7 contained at least one URL [22]. Woodruff et al. did a similar analysis using over 2.6 million HTML documents collected by the Inktomi Web crawler [23]. Douglis and Krishnamurthy used live traces from two large corporations to evaluate the rate and nature of change of Web resources and found that Web resources have high rate of change [24] Network In their WWW traffic study, Crovella and Bestavros demonstrated and explained the selfsimilar nature of WWW traffic [25]. Abdulla et al. modeled proxy traffic using Fourier analysis and identified periodic patterns in WWW traffic [21]. 2.2 HTTP Performance and Web Latency The performance of the Web is determined in large part by problems with the HTTP protocol. We summarize some of these problems as well as the broad classes of approaches to solve them HTTP performance problems HTTP/1.0 is the protocol used by the majority of current Web clients and servers. It is a very simple request/response protocol. It opens a separate TCP connection for each file requested and closes the connections after servicing the request. Typical Web pages contain an HTML document plus several embedded images. Each of these embedded images is a file, and is retrieved separately. Since TCP uses slow-start and congestion avoidance to manage congestion, there are certain overheads associated with each connection set up [26]. HTTP/1.0 is widely known as inefficient since it requires a separate connection for each file request. There are several known approaches to improve HTTP performance, discussed below.

19 Binzhang Liu Chapter 2. Related Work Approaches in improving HTTP performance and latency Persistent connections Pandmanabhan and Mogul were among the first to propose persistent connections [27]. Their study showed a significant performance improvement can be achieved by using persistent connections. This analysis formed the basic data and justification behind HTTP/1.1 s persistent connections design. However, one study by Touch, Heidemann and Obraczka indicates that the persistent connection optimizations do not substantially affect Web access for the vast majority of users [28]. Caching and pre-fetching Caching is based on locality of references of Web objects. It is assumed that significant numbers of Web accesses involve repetitive access to previously accessed objects. By moving objects from original servers to machines closer to the clients, it is supposed one can save network bandwidth and reduce latency. Client cache: Most Web browsers implement a client cache. During a user browsing session, a browser stores retrieved Web files in the disk cache. When a user makes a request, the browser will check if the requested document exists in the disk cache. If so, a cache hit occurs, and the document is retrieved from the local disk. The requested document does not need to be retrieved from the remote server. Server cache: When a server receives a few million hits a day, its performance will be degraded. Server caching is mainly used in mirroring or replicating documents on different servers to lessen this kind of problem. The best documented example of a server cache is the NCSA server [29]. Proxy caching and prefetching: A proxy cache is an intermediary between clients browsers and Web servers on the Internet. When a user requests a Web document, the

20 Binzhang Liu Chapter 2. Related Work 9 request goes through a proxy. If the document is in the proxy cache, the proxy returns the document to the client. If the document is not in the proxy cache, it will retrieve the document from the Web server, and store it in the cache. The benefits of proxy caching are supposed to reduce network traffic and reduce average latency. Research on proxy caching is very active. A study by the NRG group at Virginia Tech has shown that hit rates of 30% to 50% can be achieved by proxy caching [8][30]. However, one study found that Web resources change frequently and suggests that a simple cache may be of only limited utility [24]. Prefetching is an attempt to predict future accesses and pre-load documents into the cache. One study showed that prefetching could significantly improve network performance [31]. Compression One study showed that delta encoding can provide remarkable improvement in response size and response delay for an important subset of HTTP content types and found that for the proxy trace 77% of the delta-eligible response-body bytes and 22% of all response-body bytes could have been saved; at least 37% of the transfer time for delta-eligible responses and 11% of the total transfer time could have been avoided. [32] Other approaches In a study by W3C, Nielsen et al. investigated the effect of pipelining and changing Web content [3]. They found a pipelined HTTP/1.1 implementation outperformed HTTP/1.0, and the savings were at least a factor of two, and sometimes as much as a factor of ten, in terms of packets transmitted. They also found that use of style sheets would result in the greatest observed improvements in downloading new web pages.

21 Chapter 3 Experimental Design To find solutions to the questions raised in Chapter 1, we identify a set of workloads to study, gather and adapt a suitable set of software, and design a range of experiments that are further discussed in subsequent chapters. 3.1 Workloads Used In the Study Four proxy log files are used in the study. The America Online workload is about 40 minutes worth of proxy log file from America Online s proxy server. The Boston University workload is a proxy log file in the Computer Science Department at Boston University. The VT Campus workload is a proxy log file from the Virginia Tech campus-wide proxy server. VT Library is a proxy log file from the Newman Library proxy server at Virginia Tech. Table 3.1 describes these workloads. For additional details, see [33]. We split the workloads according to local accesses and remote accesses. For the VT Campus and VT Library workloads, any requests with.vt.edu in the host name are considered as local accesses, otherwise they are considered as remote accesses. For the Boston University workload, any requests with.bu.edu in the host name are considered as local accesses, 10

22 Binzhang Liu Chapter 3. Experimental Design 11 Table 3.1: Workloads used in the study WorkLoads Periods Total Accesses America Online Dec. 1, ,602 Boston University Jan. 27 to Feb. 8, ,928 VT Campus Sep. 28 to Oct. 5, ,975 VT Library Sep. 28 to Oct. 5, ,014,875 otherwise they are considered as remote accesses. For the America Online workload, any requests with.aol.com in the host name are considered as local accesses, otherwise they are considered as remote accesses. Table 3.1 lists the ratios of local accesses and remote accesses to the total accesses. Table 3.2: Local accesses vs. remote accesses WorkLoads Total Accesses Local accesses Remote accesses America Online 825, % 87.3% Boston University 522, % 79.0% VT Campus 696, % 93.0% VT Library 1,014, % 85.9% The above workloads were re-played using Webjamma. Table 3.1 gives ratios of successful requests and failed requests to the total requests. Successful requests are when responses are returned by referenced Web servers. Failed requests are when no responses are returned by referenced Web servers. For a failed request, the document either no longer exists, or the Web server was down when Webjamma tried to fetch the document.

23 Binzhang Liu Chapter 3. Experimental Design 12 Table 3.3: Number of re-played requests, rates of successful requests and failed requests WorkLoads Re-played Requests Successful Failed America Online 825, % 9.0% Boston University 522, % 76.4% VT Campus 696, % 3.1% VT Library 1,014, % 4.7% 3.2 Software Used In the Experiments Webjamma software Webjamma is a URL re-player program originally developed by Tommy Johnson [34]. Previous versions of Webjamma could only send HTTP queries to a proxy server. This author modified Webjamma so that it can send queries directly to Web servers. Webjamma is a multi-process program. It now has both proxy and no proxy options. The proxy option allows requests to be sent to a proxy server. The no proxy option allows requests to be sent to a Web server directly. The number of parallel Webjamma processes is configurable. The name of the log file, option, proxy server, maximum number of child processes, repetition number, and timeout period can be specified in the configuration file. A log file is processed by the parent process, and each request is sent to a child process through a pipe. A child process reads a request from the pipe, and then sends the request to the Web server. Webjamma replays a workload by reading a log file of URLs, sending HTTP queries, and timing the transfer. Since Webjamma just discards the transferred data, the only delay is from the transfer. Connection time is defined as the time between when a Webjamma process tries to establish the socket connection to a Web server or proxy server and the first byte is received by the Webjamma process. Transfer time is the time between when the Webjamma process receives the first byte and the last byte is received by the process. Elapsed time is equal to connection time plus transfer time.

24 Binzhang Liu Chapter 3. Experimental Design Squid software We use a modified version of squid in the experiments [37]. The squid software is installed in a IBM RS/6000, Model 132 with 98 Mbytes memory and 4 Gbytes disk space. The machine is connected to the network via 10baseT switched Ethernet. It is a stand alone proxy server, and is configured not to send or receive ICP queries to parent and neighbor caches. Cache memory is configured as 40 Mbytes, cache disk space is configured as 1 Gbyte. We use the default maximum object size of 4 Mbytes. 3.3 Experiments Ten experiments allow explanation of the problems listed in Chapter Experiment one: re-play four workloads without proxy via switched Ethernet To characterize response time without a proxy, four studies were conducted between Oct. 15 and Nov. 24, These investigations re-played four workloads using Webjamma, run on an IBM SMP machine with 6 processors and a half Gbyte of RAM in the Computing Center. Table 3.4 lists runs made between Oct. and Nov Table 3.4: Experiment conducted between Oct. and Nov Runs Workloads Option 1 America Online no-proxy 2 Boston University no-proxy 3 VT Campus no-proxy 4 VT Library no-proxy Note: Maximum number of child processes is 20.

25 Binzhang Liu Chapter 3. Experimental Design Experiment two: re-play a subset of VT campus workload without a proxy via a 33.6 K modem To compare the effects of end user network connections on the response time, experiment two was conducted, two runs at the same time. In the first, Webjamma was run from a IBM Power PC station via switched Ethernet connection. In the second, Webjamma was run from a home Pentium II machine with 33.6 K PPP modem connection. The log file used in the study was randomly selected URLs from the VT campus proxy log file. All the requests were sent to Web servers directly by Webjamma. The results were recorded and used in the analysis. The experiment was conducted from 17:40pm to 23:18pm on Feb. 17, Experiment three: re-play VT log with a proxy via switched Ethernet To investigate the effect of proxy caching on response time via switched Ethernet, this experiment was conducted from Nov. 30 to Dec. 1, The experiment re-played the VT campus workload to a proxy server installed on one of the IBM RS/6000 machines in the Computer Science Department at Virginia Tech. The proxy server we are running is the squid distribution modified locally by Tommy Johnson. Webjamma was run on the IBM SMP machine with 6 processors in the Computing Center. The maximum number of processes was configured to 20. The proxy option was on Experiment four: re-play a subset of VT campus workload with a proxy via a 33.6 K modem To examine the effect of proxy caching on response time via a 33.6 K modem, this experiment was conducted on Feb. 17, URLs were randomly selected from the VT

26 Binzhang Liu Chapter 3. Experimental Design 15 campus workload and re-played using Webjamma. Requests were sent to a proxy server and Webjamma was run from a home Pentium II machine via 33.6 K PPP modem connection. The number of Webjamma processes was configured to one Experiment five: response time of a proxy and loads Experiment five was designed to investigate how response time changes with different levels of traffic. In this experiment, the number of child processes in Webjamma was configured at the 1, 10, 20, 30, and 90 process levels. After each run, the cache was cleaned, so the hit rate would be the same for each run. By varying the number of child processes in the Webjamma, different load levels to a proxy server were simulated. The log file used in this study was a set of randomly selected URLs from the VT campus proxy log file. The experiment was conducted from 7:30 pm on Jan. 14, to 3:05 am on Jan. 15, The proxy server was squid. Webjamma was run from the IBM SMP machine with 6 processors in the Computing Center at Virginia Tech Experiment six: response time without a proxy via switched Ethernet under persistent connections This experiment was designed to study the performance effects of persistent connections without a proxy via switched Ethernet. We used the America Online and the VT campus workloads with 1000 randomly selected URLs of HTML documents from each workload. Then the log file subset was re-played using the program developed in this study to fetch all the embedded image files in each page. A log file also was generated to list all the files including the HTML file and all the embedded image files in a page. This log file was replayed using Webjamma to record response times. Then result files from Webjamma were used in the later analysis. The experiment was conducted between Mar. 12 and Mar. 16, 1998.

27 Binzhang Liu Chapter 3. Experimental Design Experiment seven: response time without a proxy via a 33.6 K modem under persistent connections This experiment was designed to study the performance effects of persistent connections without a proxy via a 33.6 K modem. Webjamma was run from a Pentium II machine with a 33.6 K modem connection. We used the same log files as in experiment six Experiment eight: response time with a proxy via switched Ethernet under persistent connections This experiment was designed to study the performance effects of persistent connections with a proxy via switched Ethernet. The proxy server was squid. Webjamma was run from the IBM SMP machine with 6 processors in the Computing Center at Virginia Tech. We used the same log files as used in experiment six Experiment nine: response time with a proxy via a 33.6 K modem under persistent connections This experiment was designed to study the performance effects of persistent connections with a proxy via a 33.6 K modem. The proxy server was squid. Webjamma was run from a Pentium II machine with a 33.6 K modem connection. We used the same log files as in experiment six Experiment ten: periodicity of response time This experiment was designed to check periodicity of Web response time. First, ten thousand unique static URLs were sampled randomly from the VT campus proxy log file. A Unix cron job was scheduled to run the Webjamma using this sample log file each hour from Jan. 15

28 Binzhang Liu Chapter 3. Experimental Design 17 to Jan. 29, The number of Webjamma processes was configured to 20. A run took about 15 min.; 340 runs were carried out. The result files from running Webjamma were used in the analysis Summary In this chapter, we described the ten major experiments conducted in this study. The following table summarizes the problem being studied and the chapter in which the result is discussed. Table 3.5: Problem description of the experiments Experiment Problem Description Chapter 1 Response time without a proxy via switched Ethernet Response time without a proxy via a 33.6 K modem Response time with a proxy via switched Ethernet 5.1, Response time with a proxy via a 33.6 K modem 5.1, Response time of a proxy under various traffic loads Response time without a proxy via switched Ethernet 6 under persistent connections 7 Response time without a proxy via a 33.6 K modem 6 under persistent connections 8 Response time with a proxy via switched Ethernet 6 under persistent connections 9 Response time with a proxy via a 33.6 K modem 6 under persistent connections 10 Periodicity of response time 7 Table 3.6 lists detailed information for each experiment including workload used, sample

29 Binzhang Liu Chapter 3. Experimental Design 18 size, computer used, network connection, number of Webjamma processes and proxy option. Table 3.6: Detailed description of the ten experiments Exper- Workload Sample Computer Network No. of Proxy iment Size Used Connection Webjamma Option Processes 1 AOL, Boston, Full log RS/6000 Ethernet 20 No VT, VTLIB SMP2 2 VT 10K RS/6000, Pentium II Modem 1 No 3 VT Full log RS/6000, SMP2 Ethernet 1, 20 Yes 4 VT 10K RS/6000, Pentium II Modem 1 Yes 5 VT 10K RS/6000 Ethernet 1, 10, 20, 30 Yes SMP2 50, 70, 90 6 AOL, VT 1K RS/6000, SMP2 Ethernet 1 No 7 AOL, VT 1K RS/6000, Pentium II Modem 1 No 8 AOL, VT 1K RS/6000, SMP2 Ethernet 1 Yes 9 AOL, VT 1K RS/6000, Pentium II Ethernet 1 Yes 10 VT 10K RS/6000 Ethernet 20 No KEY: AOL: America Online, Boston: Boston University, VT: VT Campus, VTLIB: VT Library, SMP2: IBM SMP machine with 6 processors and a half Gbyte of RAM, RS/6000: One is IBM RS/6000, Model 3CT, machine with 64 Mbytes RAM and 17 Gbytes

30 Binzhang Liu Chapter 3. Experimental Design 19 disk space. The other is IBM RS/6000, Model 132, machine with 98 Mbytes memory and 4 Gbytes disk space, Pentium II: Pentium II machine with 32 Mbytes RAM and 3.2 Gbytes disk space, Ethernet: 10baseT switched Ethernet, Modem: 33.6 K modem.

31 Chapter 4 WWW Response Time Without a Proxy In this chapter, the results from experiments one and two are presented. 4.1 Response Time Via Switched Ethernet Average connection and elapsed time From the data resulting from experiment one, average connection and elapsed time were calculated using the PERL scripts developed during this study. Table 4.1 gives average connection time, average elapsed time and ratio of connection time to elapsed time for the four workloads without a proxy. Table 4.1 shows that average connection time ranged from a low of 0.27 to a high of 0.54 seconds, and average elapsed time ranged from a low of 0.57 to a high of 2.0 seconds. The ratio of average connection time to average elapsed time ranged from a low of 0.27 to a high of In all workloads, this ratio is higher than It indicates that at least a quarter 20

32 Binzhang Liu Chapter 4. WWW Response Time Without a Proxy 21 Table 4.1: Average connection time, elapsed time and ratio of connection time to elapsed time Workloads Connection Time Elapsed Time Ratio American Online Boston University VT Campus VT Library Note: All time is in seconds. of the total elapsed time was spent in setting up the connection. To compare the response time of local and remote accesses, the results from the VT Campus and VT Library workloads were split into local accesses (.vt.edu domain) and remote accesses (domain other than.vt.edu). Table 4.2 lists the average connection time and average elapsed time for local and remote accesses. Both average connection time and average elapsed time for local accesses are shorter than for remote accesses. Table 4.2: Response time for local and remote accesses Workload Average Connection Time Average Elapsed Time Local Remote Local Remote VT Campus VT Library Note: All time is in seconds Distribution of connection time and elapsed time Connection time and elapsed time are useful parameters in modeling Web latency. Connection time measures the duration between a request sent by a client and the first byte received from the server. Elapsed time measures the duration between a request sent by a client and

33 Binzhang Liu Chapter 4. WWW Response Time Without a Proxy 22 the last byte received from the server. Figure 4.1 and 4.2 are the cumulative distributions of connection time and elapsed time for four workloads aol Boston VT VTlib Cumulative Frequency Connection Time Figure 4.1: Cumulative distribution of connection time Figure 4.1 shows that over 80% of the time the connection time is less than one second. Figure 4.2 shows that about 80% of the time the elapsed time is less than one second. The cumulative distributions of connection time and elapsed time follow Pearson distributions except the cumulative distribution of connection time of the VT campus workload follows a Weibul distribution. Tables 4.3 and 4.4 list the parameters of distributions for connection time and elapsed time. See [38] for a detailed description of various distribution functions.

34 Binzhang Liu Chapter 4. WWW Response Time Without a Proxy aol Boston VT VTlib Cumulative Frequency Elapsed Time Figure 4.2: Cumulative distribution of elapsed time Distribution of ratio of connection time to elapsed time The ratio of connection time to elapsed time is a very important parameter of the Web system. Figure 4.3 shows the cumulative distribution of the ratio of the connection time to elapsed time. A higher ratio indicates that most of the response time is spent on establishing network connections, and setting a network connection is the bottleneck of the system. A low ratio indicates that more time is spent on transferring the files, and network bandwidth

35 Binzhang Liu Chapter 4. WWW Response Time Without a Proxy 24 Table 4.3: Parameter estimates of distribution of connection time Workloads Distribution α 1 α 2 β American Online Pearson Boston University Pearson VT Campus Weibul 0.81 N/A 0.33 VT Library Pearson Table 4.4: Parameter estimates of distribution of elapsed time Workloads Distribution α 1 α 2 β American Online Pearson Boston University Pearson VT Campus Pearson VT Library Pearson is the bottleneck. The cumulative distribution varies a lot among the four workloads. For the Boston University workload, for about 65% of the accesses the ratio is greater than 0.80, indicating that connection time is the bottleneck. For the AOL, VT Campus and VT Library workloads, for only about 35% of the accesses the ratio is greater than 0.80, indicating that transfer time is the bottleneck. One possible explanation is that the Boston University workload is relatively old, so that many documents are broken links. When the Webjamma requests these documents, the referenced Web server only returns a header. Table 4.5 lists the connection time under various cumulative frequencies. For all workloads, 99% of the time, the connection time is less than 10 seconds. This results suggests that a Web client default timeout value should not be longer than 10 seconds.

36 Binzhang Liu Chapter 4. WWW Response Time Without a Proxy aol Boston VT VTlib Cumulative Frequency Ratio of Connection Time to Elapsed Time Figure 4.3: Cumulative distribution of the ratio of connection time to elapsed time 4.2 Response Time Via a 33.6 K Modem To examine the effects of network connection on response time, we chose a subset of the VT Campus proxy workload and re-played it using Webjamma via a 33.6 K modem connection. The results are presented here Average connection and elapsed time

37 Binzhang Liu Chapter 4. WWW Response Time Without a Proxy 26 Table 4.5: Connection time for various cumulative frequencies Workloads 90% 99% 99.9% American Online Boston University VT Campus VT Library Note: All time is in seconds. Table4.6:Responsetimeviaa33.6Kmodem Connection Time Elapsed Time Ratio Note: All time is in seconds. Although Tables 4.1 and 4.6 show that end user network connection speed has a significant effect on response time, we expected the difference in performance to be higher. However, the data show that there is a fixed delay introduced in both cases from servers and network load. The average connection time from a client via 33.6 K modem network connection is 2.2 times longer than that from a client via switched Ethernet network connection. Average elapsed time from a client via a 33.6 K modem connection is 3.19 times longer than that via switched Ethernet connection Distribution of connection time and elapsed time Table 4.7 lists the connection and elapsed time under various cumulative frequencies. Table 4.7 shows that 99% of the time, connection time via a 33.6 K modem is less than 4.5 seconds. Figure 4.4 shows the cumulative distribution of response time via a 33.6 K modem connection. Figures 4.1 and 4.4 show that the distributions with switched Ethernet and modem

38 Binzhang Liu Chapter 4. WWW Response Time Without a Proxy connection time elapsed time Cumulative Frequency Response Time Figure 4.4: Cumulative distribution of response time via a 33.6 K modem connections are different. It is found that connection time via a 33.6 K modem follows a Log-logistic distribution and elapsed time via a 33.6 K modem follows a Pearson distribution. Table 4.8 lists the parameter estimates of distributions of connection time and elapsed time via a 33.6 K modem. See [38] for a detailed description of various distribution functions.

39 Binzhang Liu Chapter 4. WWW Response Time Without a Proxy 28 Table 4.7: Connection time of modem users under various cumulative frequencies Workloads 90% 99% 99.9% VT Notes: All time is in seconds Table 4.8: Parameter estimates of distribution of connection time and elapsed time Response Time Distribution α 1 α 2 β γ Connection Time Log-Logistics 2.51 N/A Elapsed Time Pearson Summary In this chapter, we characterized the WWW response time for HTTP/1.0 without a proxy. Major findings for our workloads are: Average connection time ranged from a low of 0.27 to a high of 0.54 seconds. In all workloads, the ratio of average connection time to elapsed time is higher than 0.25, indicating at least a quarter of the total elapsed time was spent in establishing the connection under HTTP/1.0. Distribution of connection time and elapsed time follow heavy-tailed distributions (Pearson and Weibull), implying very high variability of connection time and elapsed time. Network connection has a significant effect on response time. Average connection time from a client via a 33.6 K modem is 2.2 times longer than that from a client via switched Ethernet. Average elapsed time from a client via a 33.6 K modem is 3.19 times longer than that via switched Ethernet.

40 Binzhang Liu Chapter 4. WWW Response Time Without a Proxy 29 For both modem and switched Ethernet users, 99% of the time, connection time is less than 10 seconds. Based on this result, we suggest that clients default timeout values should not be longer than 10 seconds.

41 Chapter 5 Web Response Time with a Proxy To study the impact of proxy caching on response time, experiments three, four and five were conducted. The results from these three experiments are presented in this chapter. 5.1 Average Connection Time and Elapsed Time In experiment three and four, the VT campus workload was re-played to a proxy server in the Computer Science Department. The results are summarized in Table 5.1, where Table 5.1: Response time of proxy caching using VT campus workload Number of Network Average Average Ratio1 Ratio2 Ratio3 Processes Connection Connection Elapsed Time Time 1 Ethernet Ethernet Modem Notes: All time is in seconds. 30

42 Binzhang Liu Chapter 5. Web Response Time with a Proxy 31 ratio1 is the ratio of the connection time with a proxy to the connection time without a proxy. Ratio2 is the ratio of the elapsed time with a proxy to the elapsed time without a proxy. Ratio 3 is the ratio of the average connection time to the average elapsed time with a proxy. Table 5.1 shows that for Ethernet users, if proxy traffic load is low (e.g., in the one process case) average connection time is about 1.76 times longer and the elapsed time is about 1.18 times longer than that without a proxy. If proxy traffic load is heavy (e.g., in the 20 process case), average connection time is about three times longer and the average elapsed time is about two times longer than that without a proxy. For a 33.6 K modem user, with a proxy the average connection time is 1.48 times longer, and average elapsed time is almost the same as the average elapsed time without a proxy. These results indicate that proxy caching increases both average connection time and elapsed time. It also indicates that increased traffic loads degrade the performance of a proxy caching server. 5.2 Distribution of Connection Time and Elapsed Time Table 5.2 lists the connection time with a proxy under various cumulative frequencies. Table 5.2: Connection time of VT workload with a proxy under various cumulative frequencies Option 90% 99% 99.9% proxy Note: All time is in seconds. By using Expert Fit software, we find that the cumulative distributions of connection time and elapsed time via 10baseT switched Ethernet follow Inverse-Gaussian distributions, the cumulative distribution of connection time via a 33.6 K modem follows a Gamma distribution and the cumulative distribution of elapsed time via 33.6 K modem follows a Weibul

43 Binzhang Liu Chapter 5. Web Response Time with a Proxy connection time elapse time Cumulative Frequency Seconds Figure 5.1: Cumulative distribution of connection time and elapsed time distribution. Table 5.3 lists models and parameter estimates of distributions for the connection time and the elapsed time. See [38] for a detailed description of various distribution functions. Table 5.3 shows that all the distributions are heavy-tailed, indicating that generally, time is short for connection time or elapsed time. These distributions can be used in modeling and predicting Web server or proxy server performance.

44 Binzhang Liu Chapter 5. Web Response Time with a Proxy 33 Table 5.3: Models and parameter estimates of distribution of response time Parameters 10 baset Switched Ethernet 33.6 K modem and Models Connection Elapsed Connection Elapsed Time Time Time Time Models Inverse Inverse Gamma Weibul Gaussian Gaussian α β Comparison of response time of local accesses and remote accesses via 10baseT switched Ethernet To compare the response time of local and remote accesses, the results from experiment three were split into local accesses (vt.edu domain) and remote accesses (domains other than vt.edu domain). Table 5.4 lists the average connection and elapsed time for local and remote accesses. Average connection times for local accesses and remote accesses are almost Table 5.4: Response time of proxy caching for local and remote accesses Experiments Domain Average Connection Average Elapsed Time Time 1 local access remote access Note: All time is in seconds. the same. Average elapsed time of local accesses are shorter than for remote accesses. Table 5.5 lists the connection time of VT campus proxy traffic for local and remote accesses under various cumulative frequencies. Figure 5.2 shows the cumulative distribution of connection time of the VT campus proxy

45 Binzhang Liu Chapter 5. Web Response Time with a Proxy 34 Table 5.5: Connection time of VT campus proxy workload for local and remote accesses under various cumulative frequencies Option 90% 99% 99.9% local remote Note: All time is in seconds. workload for local and remote access. Contrary to typical thought, Figure 5.2 shows that 90% of the time, the connection time for local accesses is longer than for remote accesses. For the upper 10%, the connection time for local accesses is shorter than for remote accesses. 5.3 Response Time and Proxy Traffic Loads In experiment five, the number of parallel Webjamma processes ranges from 1 to 90, the corresponding completed requests per second range from a low of 0.65 to a high of The results are presented here. Figure 5.3 shows that response time increases with an increase in the number of parallel Webjamma processes (equivalent to an increase in the request arrival rate). Slothouber [39] studied Web server performance using a queuing model and found that before a server reached its theoretical upper bound load, the response time was almost constant (a mere fraction of a second). When the server approached full utilization, response time grew asymptotically toward infinity. Contrary to his findings, our result shows that proxy server performance is generally sensitive to traffic load. Even at very low request arrival rates, when the number of processes increases from 1 to 10 (equivalent to a range of 56,471 accesses/day to 436,363 accesses/day), the average connection time increases by 37%, and the average elapsed time increases by 38%. In Figure 5.4, Ratio represents the ratio of the average connection time