SORT BY:

LIST ORDER
THREAD
AUTHOR
SUBJECT


SEARCH

IPS HOME


    [Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

    RE: TCP "reliability"



    Jonathan,
    
    To be more exact about the differences between the various summations and
    CRC, CRC offers a substantial improvement if due to bus related errors.  The
    same could be true with respect to switch fabric cell re-ordering.  Your
    sample data indicated errors with a prevalence at word bit locations that
    could relate to either distinct locations or to an internal bus.
    Assumptions as to a random nature of these errors could lead to conclusions
    of equivalence but protections in this case are slight if due to an internal
    bus for summations.  These concerns are based on empirical data but I find
    this compelling to remain cautious about making such assertions.
    
    Doug
    
    > In message <277DD60FB639D511AC0400B0D068B71E0709F5@corpmx14.us.dg.com>,
    > Black_David@emc.com writes:
    >
    > >Careful - there are two meanings of the word reliability being confused
    > >here.
    > >
    > >(1) Data is delivered in the face of dropped packets.  TCP is definitely
    > >	reliable in this sense - all the data is delivered, in order, or the
    > >	connection closes.
    >
    > Yep.
    >
    > >(2) Data is correctly delivered in the face of corruption.  TCP's 16-bit
    > >	checksum falls short of a 32-bit CRC in its ability to detect
    > >	corruption, and hence TCP leaves something to be desired here.
    >
    > David,
    >
    > Can I quote that in my dissertation?
    >
    > stricly, what you say is unquestionably true.  But I think you are
    > conflating two quite distinct properties here: 32-bit error-detecting
    > codes (EDCs) and the kinds of errors which a (32-bit) CRC guarantee to
    > catch, versus the errors which to which a transport-level error chehck
    > is, empirically, subjected.
    >
    > It turns out that, on the best data we (I and Craig Partridge) have on
    > empirically-observed transport-level errors, CRCs are just not a whole
    > lot better than a 32-bit mod-M additive sum (where M ~= 2^32).
    >
    > This is work in progress, so i cannot give a proper citation (best is
    > my nearly-complete dissertation).  OTOH, Julo Satran saw much of the
    > raw data at the recent framing/error-control meeting; i'd be very
    > happy to let Julo make a more impartial comment and see where that
    > goes.
    >
    >
    >
    > >The word "integrity" is a better term for (2) to avoid confusion.
    >
    > No argument there. The term "reliable transport" shoul,d perhaps, be
    > understood in contrast to the contatenated-x25-virtual-circuit model
    > which telcos were offering in serious competition to IP/TCP, once upon
    > a time.  (I think even Bob Braden would buy that.)
    >
    
    


Home

Last updated: Tue Sep 04 01:04:19 2001
6315 messages in chronological order