[lkml]   [2008]   [Sep]   [25]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
    SubjectRe: [RFC PATCH 1/3] Unified trace buffer

    * Linus Torvalds <> wrote:

    > That said, if people think they can do a good job of ns conversion,
    > I'll stop arguing. Quite frankly, I think people are wrong about that,
    > and quite frankly, I think that anybody who looks even for one second
    > at those "alternate" sched_clock() implementations should realize that
    > they aren't suitable, but whatever. I'm not writing the code, I can
    > only try to convince people to not add the insane call-chains we have
    > now.

    hm, i'd really hope hw makers see the light and actually make the hw do
    it all. Signs are that they are cozying up to these ideas.

    Good and fast timestamps are important, and it is _infinitely_ more easy
    to do it in hw than in sw.

    Firstly they need a low-frequency (10khz-100khz) shared clock line
    across all CPUs. A single line - and since it's low frequency it could
    be overlaid on some existing data line and filtered out. That works
    across NUMA nodes as well and physics allows it to be nanosec accurate
    up to dozens of meters or so. Then they need some really cheap way to
    realize what absolute value the clock counts, and read it out every now
    and then in the CPU, and approximate it inbetween, and have a secondary
    stage cheap few-transitors long-latency multiplicator that keeps passing
    on the nanosec-ish value to a register/MSR that can be read out by the

    This trivially works fine even if the CPU is turned off. It uses nary
    any power as it's low freq, and can be spread across larger system
    designs too. In fact it would be a totally exciting new capability for
    things like analysis of SMP events. PEBS/BTS could be extended to save
    this kind of timestamp, and suddenly one could see _very_ accurately
    what happens between CPUs, without expensive bus snooping kit.

    and CPUs wont go beyond the '~1nsec' event granularity for quite some
    time anyway - so nanoseconds is not a time scale that gets obsoleted

    [ i guess this proves it that everyone has his pipe dream ;-) ]


     \ /
      Last update: 2008-09-26 00:31    [W:0.023 / U:3.860 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site