|
|
Microsoft Windows NT Working Set Tuner README File
April 15, 1998
1. Description:----------------
Windows NT (32-bit) Working Set Tuner (WST) for x86, Mips, Alpha, and PowerPC platforms.
The working set of a program is a collection of those pages in the program's virtual address space which have recently been referenced. As the working set size decreases, memory demand decreases. Since it is advantageous to minimize the memory demand, the working set tuner DLL is developed to measure code working set, and produce an ordering of the functions (procedures) within the code which will be small, if not minimal, in a paging environment. In any practical case there are probably many optimal assignments of functions to pages. Although this method does not claim optimality, it will produce results at least as good as a programmer could achieve by hand in a reasonable amount of time. This method consists of two parts:
1) Data Collection:
Data is collected in C programs by using the attributed profiling hook provided by the Microsoft C/C++ Compiler (compiling with the "-Gh" option). Data is not collected in assembler language programs. A hook routine in the working set tuner DLL, WST.DLL, is called before each C/C++ function is called. The hook examines its environment to determine the identity of the called function. A bit is set in a bitmap every time a function is called within the specified time segment. A new bitmap is used (new snapshot) for each new time segment: smaller time segments will cause higher frequency of snapshots which will produce better working set (with some diminishing point of return...). The collected WST data is dumped upon the program's termination or via WST's dump facility called WstDump.exe. Functions' symbols in the program's symbol table are used by the WST.dll to match function names to bits in the bitmap.
2) Data Reduction:
Two post-processing tools, WstCat.exe & WsTune.exe, process the collected data and produce an ordering of the functions (packing list) which, when placed into pages, will result in a smaller working set of pages in memory compared to its original working set size for the activity measured. The packing list, created using an ordering algorithm, can be used by the linker to perform the actual packing within the executable's object modules.
2. WST Files:--------------
o wst.ini -- WST initialization file. o wst.dll -- WST DLL. o wst.lib -- WST library file containing WST entry point. o wstdump.exe -- Dump utility for dumping/clearing WST data, and stopping data collection at any time. o wstune.exe -- Data analysis utility for creating the packing list from collected data. o wstcat.exe -- Concatenation utility for multiple data files. o penter.lib -- An empty library containing a dummy WST entry point. This can be used instead of WST.lib to link with an executable object module which was compiled with the -Gh option, so the application can run without recompiling and without any of the WST overhead.
3. Using the Working Set Tuner Dll:------------------------------------
o Create the WST directory in the root directory of C: drive. (c:\wst)
o Place WST.INI in the WST directory of C: drive and list all the DLL/EXE images to be tuned as follows:
NOTE:- WST.INI must exist for WST to run.
[EXES] Name(s) of module(s) to be tuned. Each name should be on a new line. A semicolon ';' in front of a name removes it from the list.
[PATCH IMPORTS] obsolete - no longer used
[TIME INTERVAL] Elapsed time in milliseconds between recording snapshots (if left blank, a default time interval of 1000 is used). This section may include an optional entry 'MaxSnaps=n', where n is the maximum number of snapshots to take (the default maximum is 3200).
o Sample WST.INI file:
[EXES] myapp.exe
[TIME INTERVAL] 1500 MaxSnaps=256
o Attach WST.DLL to the application process:
- Recompile your EXE/DLL using the "-Gh" and "-Zi" or "-Z7" compiler options. (NOTE - on Mips platform, use "-Od" to disable compiler optimization.)
- Link it with WST.LIB using the options "-debug", either "-debugtype:coff" or "-debugtype:both", and "-pdb:none" or "-pdb:<filename>".
- This method will cause all the C functions in the recompiled sources to be accessible to WST.DLL.
NOTE:- The win32.mak file in the Win32 SDK will set these options correctly for your application if you build it with the tune=1 switch. I.e. "nmake -all tune=1".
NOTE:- If symbols have been removed from your exe or dll, WST will try to locate the symbols files (DBG or PDB files) in the current directory, then in the path specified by the environment variables "_nt_symbol_path", "_nt_alt_symbol_path", and "systemroot", in that order. E.G. if the DBG file of your exe is in c:\symbols\exe, set _nt_symbol_path to c:\symbols.
o Place WST.DLL in your SYSTEM directory (i.e. ..\nt\system32) or in the same directory with the application being tuned.
o If this is the only run that is to be considered for tuning, clear any existing .TMI and .WSP files in the C:\WST directory.
o Run your application(s).
o Data will be collected for all the modules specified in the [EXES] section of the WST.INI.
o Exit the application to dump the WST data, or run WstDump.exe to dump the data. See section 4 "WST Data" for details.
o If you have multiple TMI & WSP files, run WstCat.exe to concatenate data from multiple runs. See section 5 "Using WstCat" for details.
o Run WsTune.exe to generate the packing list (.prf file) based on the collected WST data. See section 6 "Using WsTune" for details.
o Use the packing list to direct the linker to order the modules as specified in the packing list (.prf file). See section 8 "Linking with the Packing List" for details.
4. WST Data:-------------
o WST data can be captured in two ways:
1) Upon termination of the application WST data is dumped automatically.
2) Using the dump utility, WstDump.exe, WST data can be dumped at any time. See section 7 "Using WstDump" for details.
o Data files are created in C:\WST directory.
o WST data is dumped into data files using the application and its DLLs names with .TMI & .WSP extensions. No data is dumped for modules which lack symbol information in the COFF format.
o If TMI & WSP files exist in the c:\wst directory, new data files with .T?? and .W?? are created where ?? is a number between 0x01-0xff. The number indicates the run number. All the data files can be concatenated (using WstCat.exe) to generate a single packing list from the multiple run data files.
5. Using WstCat:-----------------
o WstCat.exe can be used to concatenate multiple .WSP and .TMI files (multiple files will be named *.T00 - *.Tff and *.W00 - *.Wff) into a single file for analysis. This allows the user to run multiple scenarios, over time, and analyze them as a single run. WST will also create multiple files if a scenario exceeds its snapshots limit.
Usage: WSTCAT moduleName
"moduleName" is the name of the module to combine. The module is the name of the DLL or EXE which is being tuned and data has been collected for.
NOTE:- The original TMI & WSP files will be renamed to TXX & WXX by WstCat.
6. Using WsTune:-----------------
o WsTune.exe is used to analyze the data created by WST. WsTune will also create the packing list, .PRF file, which is used by the linker to pack your DLL or EXE.
o The DATA files created by wstune are:
- .DT (Data Tuned) contains the analysis data/bitstrings for all api's called in the "packing list" order.
- .DN (Data Not tuned) contains the api's called in the "unpacked" order.
- .PRF contains the packing list to be used by the linker.
- .WSR is a worker index file created for internal use.
Usage: WSTUNE [/O] [/D] [/N] [?] moduleName1[.WSP] [...]
/O Dump analysis data to file (*.DT tuned, *.DN not tuned) /N Analyze bitstring data only, create .WSR and .PRF files (turns off /O) /D Dump analysis data only; use existing .WSR file (turns off /N) /P Display a progress indicator /? Display this usage message
where "moduleNameX" is the name(s) of the module file(s) to tune. Each module file will result in a separate set of tuning data based on the input switch settings. If neither /O nor /N are specified, tuning data (normally the contents of the .DT file) is dumped to the standard output device.
7. Using WstDump:------------------
o WstDump.exe can be used to stop WST data collection and clear/dump WST collected data for all the applications being tuned, at any time.
o The following options are available via WstDump.exe:
- Stop : Stops WST data collection (applications continue to run).
- Clear and Restart : Clears any existing WST data and restarts data collection.
- Dump and Stop : Dumps any existing WST data and stops the data collection (applications continue to run).
o Data is dumped to c:\wst using the names of the application and its DLLs with .TMI & .WSP extensions. No data is dumped for modules with no symbols.
o If TMI & WSP files exist in the c:\wst directory, new data files with .T?? and .W?? are created, where ?? is a number between 0x01-0xff. The number indicates the run number. All the data files can be concatenated (using WstCat.exe) to generate a single packing list from the multiple run data files.
8. Linking with the Packing List:----------------------------------
o Recompile your applications using the Microsoft C/C++ Compiler's "-Gy" option. (Remove the "-Gh" option to avoid extra overhead.)
o Relink your applications using the link "-order:@filename" option where filename is the name of your packing list (i.e. .prf file generated by WsTune.exe).
9. Caveats:------------
o If a symbol is not available in an EXE/DLL that is being tuned, ??? is displayed as the function name in data and analysis files (??? is not listed in the packing list, however.)
o WST data is *NOT* collected for functions in assembler language programs/objects.
o The linker currently will not reorder static functions. The Microsoft C/C++ Compiler automatically generates certain static functions (their names look like "_$Ennn", where nnn is a number the compiler assigns internally).
** END OF README **
|
