benchmarks

[Matthew Dillon]

:Thanks.
:Any suggestions from the Dfly community on other benchmarking tools
:that would run across the OSs I mention in my first email?

    The only benchmarks that are actually meaningful are the ones that
    simulate or test real services using reasonably accurate profiles.
    For example, testing a web server requires simulating 'users', including
    simulating client side caching (since all browsers do it), not
    just randomly pulling pages up over HTTP.

    Even then a benchmark is not going to be particularly useful unless
    you actually need to run a service full-out.  Services rarely break down
    due to performance issues but instead break down due to design,
    filesystem, or configuration issues.  A java applet server, for example,
    might become limited by the number of sessions it needs to keep alive
    rather then disk or network performance.  A mail transfer agent might
    become limited by disk performance, or by filesystem efficiency for
    certain types of operations (where the design of the service and the
    filesystem backing it don't take advantage of each other's strengths).

    There are lots of reasons... pure benchamrks are usually not at the head
    of the list when it comes to actually maintaining production systems.

    Benchmarks can be useful for testing various efficiency aspects of a
    design, but their use in this area is completely independant of their
    use in identifying the production performance of whole systems.  For
    example, timing the speed of a getuid() system call helps us determine
    if we need to pay some attention to the system call path, but doesn't
    really correlate with application performance from a system manager's
    perspective.

					-Matt
					Matthew Dillon 
					<dillon at backplane.com>