http:^^www.cs.washington.edu^homes^romer^rocky^java-x86^java-x86.html
来自「This data set contains WWW-pages collect」· HTML 代码 · 共 87 行
HTML
87 行
Date: Tue, 10 Dec 1996 03:27:57 GMTServer: NCSA/1.4.2Content-type: text/htmlLast-modified: Sat, 28 Sep 1996 04:44:41 GMTContent-length: 3333<html><head><title>Java Measurements on x86</title></head><body><h1>Java Measurements on x86</h1>This page contains a set of measurements of Sun's Java 1.0 interpreterexecuting on a Pentium Pro running Windows NT. We had originallymeasured Java on the x86 platform because no Java interpreter was yetavailable for the alpha platform. For consistency in the ASPLOS VIIpaper, we switched to the alpha when a Java port became available.But since we have the x86 measurements, we would also like to makethem available to those who are interested. If you are interestedin the tool that we used to help collect these measurements on the x86,take a look at the Etch home page. <p>The tables and figures below roughly correspond to the tables andfigures that appeared in the final version of the ASPLOS VII paper.However, since they are the tables and figures that were in thesubmitted version of the paper, they are presented in a slightlydifferent format from the final version. <p>The system on which we performed our measurements was a 150 MHzPentium Pro workstation running Windows NT 3.51. The Cmicrobenchmarks and the Java interpreter were compiled using MicrosoftVisual C++ 4.0. Also note that, even though the Java measurementswere from an x86 system, the measurements for the other interpretersare still from the alpha system. <p><hr><img src="micro.gif" align="middle"> <p><b>Table 1: </b> Microbenchmark results. This table shows theslowdown of each microbenchmark relative to the equivalent operationimplemented in C. Each microbenchmark ran for at least five secondsper trial. Each number presented is the average of 20 runs. Standarddeviations were no more than 10\% and were usually under 5%. <p><hr><img src="bench-desc.gif" align="middle"> <p><b>Table 2: </b> This table shows baseline performance measurements ofthe interpreters running a set of representative programs, along withtwo benchmark programs written in C. The % Fetch/Decode column showsthe percentage of native instructions due to fetching and decodingvirtual commands. Native instructions per virtual command is simplythe total number of instructions executed divided by the number ofvirtual commands. It includes fetch/decode overhead, as well as anyother program overheads. <p><hr><img src="inst-per-vc1.gif"> <img src="inst-per-vc2.gif"> <p><b>Figures 1 and 2: </b> (These two figures from the final paper werecompiled from the data shown in these six graphs.) Java nativeinstruction count distributions for categories of Java virtualcommands. Each category represents the total number of nativeinstructions executed while interpreting virtual commands in thatcategory during the execution phase. Successive categories arecumulative with respect to previous categories. The percentage aftereach category name denotes the number of times that virtual commandsin that category were executed as a percentage of all virtual commandsexecuted.<hr><h3>Figures 3 and 4</h3>We do not have any corresponding results for Figure 3 because we don'thave a Pentium Pro simulator with which to perform the equivalentexperiment. Similarly, we did not perform any cache simulations onJava executing on the x86. <p></body><hr> <address>rocky@cs.washington.edu</address></html>⌨️ 快捷键说明
复制代码Ctrl + C
搜索代码Ctrl + F
全屏模式F11
增大字号Ctrl + =
减小字号Ctrl + -
显示快捷键?