Document Title: errata.txt Subject: Known Problems in SPEC CPU2000 Last updated: 16 Dec 1999 3:00pm jh (To check for possible updates to this document, please see http://www.spec.org/osg/cpu2000 ) ------------------------------------------------------------------------ This document describes known problems in the SPEC CPU2000 suite. Note: When updating this document, please do not ever re-use problem numbers. ------------------------------------------------------------------------ Problem number: 1 Problem summary: Rate from speed sometimes loses reportable attribute. Details: The CPU2000 run rules specify that it is legal to generate a single-user rate result from a single-user speed run. See runrules.txt 4.3.2 and the section on rawformat in utility.txt. However, when the instructions in utility.txt are followed, the output files sometimes do not preserve the reportable attribute. That is, reports say "INVALID RUN". The problem has been seen only when attempting to format base-only runs; no problems are known for rawfiles that contain both base and peak metrics. Workaround: None at this time. To generate the single-user rate metrics, re-run your test with runspec --rate --users=1. ------------------------------------------------------------------------ Problem number: 2 Problem summary: --username does not work as documented Details: The CPU2000 suite provides limited support for multiple users, as described in runspec.txt. It is also documented there that the switch "--username" can be used to change the value that would otherwise be defaulted from the current process. Nevertheless, at this time the value is always taken from the process. That is, if a user is logged in as "michael", but says runspec --username=paula, the run directories will still be tagged as michael's. Workaround: None at this time. Do not rely on this feature. The run directories will be recorded with your correct process name rather than with your pseudonym. ------------------------------------------------------------------------ Problem number: 3 Problem summary: runspec -V does not find specmake etc. on NT Details: The command "runspec -V" provides detailed information about versions of tools in the SPEC CPU2000 suite. However, when used on NT, it does not correctly find specinvoke, specmake, specgzip, and specperl. Workaround: You can find out the versions of these tools by hand. Try saying: shrc specinvoke -v specmake -v specgzip -h specperl -v ------------------------------------------------------------------------ Problem number: 4 Problem summary: runspec crashes quickly on Alpha/NT, trying to find PDF Details: On Alpha/NT, the runspec command crashes soon after starting up. The command window shows: Identifying output formats...asc...html... and then a dialog box is raised that says specperl.exe - Application Error The instruction at "0x011a8e04" referenced memory at "0x00000628". The memory could not be "read". Workaround: Disable PDF and Postscript output by doing: cd %SPEC%\bin rename PSPDF.pm hideme_PSPDF.pm Runspec will then be able to initialize correctly (and other forms of output will work). ------------------------------------------------------------------------ Problem number: 5 Problem summary: --max_active_compares cannot exceed --users Details: If you issue a runspec command with --max_active_compares set to a value larger than the number of users being tested, the tools may incorrectly report that a benchmark got the wrong answer (miscompared). Workaround: Do not set --max_active_compares to a value larger than --users. ----------------------------------------------------------------- Copyright © 1999-2000 Standard Performance Evaluation Corporation