NetNews Usenet Archive 1992 #26/NN_1992_26.iso/spool/comp/sys/sgi/16088

Text File | 1992-11-08 | 3.4 KB | 78 lines

Newsgroups: comp.sys.sgi
Path: sparky!uunet!think.com!ames!sgi!rhyolite!vjs
From: [e-mail] (Vernon Schryver)
Subject: Re: late collision
Message-ID: <[e-mail]>
Keywords: collision
Organization: Silicon Graphics, Inc.  Mountain View, CA
References: <[e-mail]>
Date: Fri, 6 Nov 1992 19:27:23 GMT
Lines: 66

In article <[e-mail]>, [e-mail] (bill o'donnell) writes:
> Our lab has 12 networked Iris workstations. Some of our machines see about
>20-30 late collisions/hour, a couple of our machines never see late collisions.
> Also, I don't remember experiencing late collisions under the old 3.3 OS. The
> lab has a mix of 4.0.1 and 4.0.5 now.
>
> The problem has gotten so bad I prefer to rlogin into the server to do my
> text editing (Writing a 1000 line file over the net can take several minutes
> otherwise.).
>
> [e-mail]


Changing the operating system did not change your cable plant.  Late
collisions are reported as they have always been reported.  (well, for
about 5 years) SGI is unusual only in reporting late collisions instead
of hiding them.  Late collisions do terrible things to NFS and TCP
performance.  A 1% late collision rate is more than enough to reduce
NFS performance by 90%.

A late collision is happens after the first "slot time".  They can only
happen to big packets.  Small packets are just mysteriously lost as far
as some stations are concerned.  Late collisions will happen at some
stations but not others because of the nature of the configuration
timing errors that cause them.

Ethernets allocate the bus using a "cocktail party" protocol.  Everyone
who wants to talk starts talking, and if two or more people start to
talk at once, they are both supposed to stop immediately, and randomly
wait.  This scheme works great, is almost as media efficent as token
passing, far simpler than tokens, and arguably as or more robust than token
passing (no tokens to lose).  The only advantage of token passing is
that a token based system will collapse smoothly into uselessness under
overload instead of abruptly.

However, the ethernet scheme requires that everyone follow the rules.
It does not work if some blabber-mouth insists on talking regardless
of other people.

In a cocktail party, the hard-of-hearing have more trouble, in part
because they have more trouble knowning when other people are talking.
Analogously, if two stations have trouble hearing each other, the
ethernet bus allocation scheme, "collisions", breaks down.  If the bus
is too long, both stations can have been talking for longer than the
minimum size of a packet before they hear the other guy.  That is a
"late collision."  Obviously, if they had only a mimimum sized packet,
they will never know that their bits got stepped on and trashed;
minimum sized packets are simply lost.

Late collisions are commonly caused by:
    (1) long AUI drop cables on 10baseT transceivers
        (a popular game on the SGI campus)
    (2) violations of other cabling rules:
    (a) some fiber extenders
    (b) more than 500 m of thick wire
    (c) cascading too many MUX's
    (d) excessively long drop cables (generalization of (1))
    (e) violation of the 4-repeater rule (or 2, if you use the
        old definitions)
    (3) defective stations
    (a) partially connected cables (e.g. stations that do not
        get the collison signal from the transciever)
    (b) other broken hardware.


Vernon Schryver,  [e-mail]