schweikh3/schweikh3.hint
Best Utility: <schweikh@noc.dfn.de> Jens Schweikhardt

    Jens Schweikhardt
    DFN Network Operation Center
    Schlartaeckerweg 3 (home address)
    D-71384 Weinstadt
    Germany

    http://www.uni-stuttgart.de/People/schweikhardt/home.html


Judges' comments:

    To use:
	make schweikh3

    WARNING!

    Do not run this program without reading this text down to the end.
    It may render your system unusable for a limited amount of time
    or force you to reboot using the Big Red Button!  This is not a joke.
    You have been warned.  

    The source fails to compile on compilers that recognize the inline
    keyword by default.  For gcc, you may need to use -ansi.

    See the file, schweikh3.info for example output from various systems.
    
    But still it is a useful utility.  To use try:

	./schweikh3

    and thanks for the virtual memories!  :-)

    NOTE: On some systems such as SunOS, one may need to compile with:

	${CC} ${CFLAGS} \
	  -I. -D_POSIX_SOURCE '-Ddifftime(a,b)=(double)(b-a)' \
	  schweikh3.c -o schweikh3


Selected notes from the author:

    Why I think my entry is obfuscated
    ----------------------------------

    There are some layout obfuscations that hinder readability, like
    some (but not all) one character tokens in column 1. Includes,
    defines and declarations of data and functions appear intermixed.
    iso646.h is even included in the middle of an expression, how
    disgusting. Besides, requiring iso646.h will render the program
    uncompilable on many systems that have not yet incorporated
    Amendment One to The Standard.  If you, dear reader, suffer from
    this disease, just read iso646.h.

    There are no `if' statements. They have been replaced by
    "switch () case 1:" or "switch () case 0:" etc. At least one
    switch has a declaration after the opening brace, booh!
    Identifiers are named after random programming languages, Basic,
    Cobol, F77, awk etc.

    You have to select maximum ANSI C conformance of your
    compiler. The source fails to compile under a C++ compiler due to
    the `new' identifier which is a C++ reserved word:

	$ g++ prog.c
	prog.c:9: parse error before `new'
	prog.c:16: parse error before `new'
	prog.c:46: parse error before `void'
	prog.c:65: parse error before `Lisp'
	prog.c:68: parse error before `;'
	[...]
	prog.c:91: confused by earlier errors, bailing out

    The source fails to compile on compilers that recognize the inline
    keyword by default, e.g. gcc:

	$ gcc prog.c
	prog.c:46: parse error before `void'
	prog.c: In function `main':
	prog.c:91: parse error before `inline'
	prog.c: At top level:
	prog.c:97: parse error before `void'

    The -ansi option to gcc fixes this.

    The source fails to compile on obsolete systems with obsolete
    memory models due to the identifier `far' (to be honest: this is
    untested :-).

    Does your indent cope with backslash/newline pairs inbetween
    keywords like in "re\<newline>turn"? Solaris' indent inserts spaces 
    and thereby renders the source uncompilable...

    The indented source will exhibit different semantics when indenting
    changes the number of lines (due to using the __LINE__ macro in an
    expression).

    And what about sizeof __TIME__?  This one's for the philosophers 
    and physicists among us programmers. What is it? sizeof (char *) -- no.
    Strlen ("hh:mm:ss")? -- no. It's the same as sizeof __DATE__ -
    sizeof "no" :-)

    The program is POSIX conformant and lints without warning under
    Solaris' Sunpro lint. Allocated memory is correctly freed on all
    possible execution paths.

    What this program does
    ----------------------

    This program tests your operating system's allocation honesty. It
    forks a child that will allocate as much as possible with byte
    granularity and then write to every 1024th byte, at the same time
    writing a status report on stdout. This way each page of the
    allocated virtual memory will eventually be referenced. Without
    any arguments, malloc() will be used to allocate memory. If the
    first arg starts with c, calloc() will be used instead.

    I distinguish three different behaviours:

    1) OS without memory overcommit:
       malloc does not lie -- you can use all memory the OS granted to you.

    2) OS with memory overcommit:
       malloc lies -- any program may get killed at random when the
       OS can not acquire a free page.

    3) OS with pseudo memory overcommit:
       malloc lies -- The OS puts processes to sleep for which it can
       not acquire a free page. Because other processes are also
       affected this way, the system slows down and quickly becomes
       totally unusable (also known as "thrashing"). Because more and
       more processes wait for memory returned by other exiting processes
       the system eventually enters a deadlock situation and effectively
       halts.

    I had the pleasure and pain to find all three behaviours on the
    platforms I use regularly. What follows are the outputs for
    Solaris, FreeBSD and Linux. The Linux case is interesting in that
    the behaviour changes when calloc is used instead of malloc.
    For FreeBSD, the last two lines are not messages stemming from
    my program but issued by the syslog daemon.

    Implementation remarks
    ----------------------

    How can we distinguish the three different OS types from the
    table above?

    Obviously, type 1), no overcommit, is easy.

    For type 2), random kill, we need to fork a child that can be killed
    and whose exit status the parent collects and decodes. Here's
    where POSIX enters the scene.

    Type 3), thrashing, was the hardest part. The first attempt was to
    measure the elapsed processor time it takes to write a single byte
    (using clock()). If a certain threshold is reached I assume the
    system is thrashing. When thrashing, however, the process does not
    consume processor time in a noticeable amount any longer and I lose.
    So I switched to measuring wall clock time using time(). On the one
    system that I observe thrashing (Linux) this works (although the
    value returned by time seems to 'hang' sometimes when the OS thrashes.
    Exceeding a threshold of 5 seconds may be detected only after half a
    minute or so.