🚀 go-pugleaf

In article <xjeldc.39.716291157@lustorfs.ldc.lu.se>, xjeldc@lustorfs.ldc.lu.se (Jan Engvald LDC) writes:
> 
> I was testing FDDI using a small ring with 4 nodes on it. When I rebooted
> one node, the traffic between two other nodes was interrupted for
> almost a second. This surprised me. I thought that with the high data rate,
> short distances and few nodes it would be able to recreate a new token
> in some tens of milliseconds (most of which would be to detect that it
> is lost).
> 
> I've heard that the process in reality takes double the time you would
> expect because first the ring is felt broken and will reconfigure itself 
> into U-shape and start token regeneration, then immediately it will 
> reconfigure itself as a ring and again start a new token regeneration.
> So each regeneration was probably less than half a second, but that is
> still a lot.
> 
> I would like to speed up the token regeneration process in a small
> FDDI ring at least 10 times. Is there some parameters you can configure to
> do that?


Something is broken.  (I hope you're not using any SGI stations.)
    -"Ring recovery" is a matter of milliseconds.
    -Noticing that a neighbor is shut down (i.e. sending QLS or 
	"dark fiber") is a matter if microseconds
    -if you have optical bypass switches, they are required to switch
	in or out in less then dozens of milliseconds.

Obvious ways things might be broken:
    -"rebooting one node" does not reset the FDDI hardware in that
	station, and the FDDI hardware "went off into to the weeds"
    -two, not one station was rebooted, and on opposite sides of the ring.

A non-standard kludge I'm insufferably proud of in SGI products, that
I think everyone should implement, is in a dual-attached station to
arrange to make your PHY's into a repeater when your station is
"rebooting" or otherwise kaput but powered up.  If the station does not
have an optical bypass, this will keep the ring healthier.  On the
other hand, it's easy to make such a "repeater" that does more harm
than good (the polite will not ask how I know) and it seems the obvious
way to wire up at least a couple of the commercial chip sets requires a
real, often expensive "node processor" to defeat "repeat filters" in
MACs and PHYs.


Vernon Schryver,  vjs@sgi.com