Re: tcptrace [OFFTOPIC?] Looking for cheap timestamping...

From: Jeremy Elson (jelson@circlemud.org)
Date: 02/27/03

  • Next message: uaca@alumni.uv.es: "Re: tcptrace [OFFTOPIC?] Looking for cheap timestamping..."

    Message-Id: <200302280429.h1S4T6513658@cambot.circlemud.org>
    Subject: Re: tcptrace [OFFTOPIC?] Looking for cheap timestamping... 
    Date: Thu, 27 Feb 2003 20:29:06 -0800
    From: Jeremy Elson <jelson@circlemud.org>
    
    

    uaca@alumni.uv.es writes:
    >What about interrupt latency mostly caused by interrupts being disabled?

    We haven't had this problem, perhaps because we haven't run our
    machines at a particularly high interrupt load. The precision seems
    to be limited by the clock resolution more than anything else.

    As others have pointed out, some cards might do interrupt coalescing,
    which can render timestamps useless.

    Timestamping in the card can overcome this problem, although for it to
    be useful often you need to find a way to reconcile the card's clock
    with your CPU clock. Our software can do this translation for you, if
    the card's firmware has some form of sync support. I'd love to find
    an off-the-shelf NIC that supports this feature; for some parts of our
    testbed, we use NICs with programmable firmware and add the feature
    ourselves. (But, for Ethernet, we just timestamp the interrupts and
    it works fine.)

    Jer

    ----------------------------------------------------------------------------
    To unsubscribe, send a message with body containing "unsubscribe tcptrace" to
    majordomo@tcptrace.org.



    This archive was generated by hypermail 2b30 : 02/28/03 EST