Two thoughts.

We need to be more precise about our timeframes for all relevant
factors.

And we may need to split the standard into two for enterprise (the
primary objective historically?) and for operator.

Take an enterprise WAN; 2Mbit/s links, link propagation delay O(10mS),
I would expect most outages to be layer 2 failures in the public
network, would expect notifications to come from L2 and not Hellos,
would ignore a L2 outage under 20mS, think of subsecond as being
O(0.3s) (because ChampCar drivers excepted, that is as fast as humans
respond) and regard convergence within 5s as ok although 1s might be
nice to have. And when I saw in one case that SPF has run over 4000
times in the past month, I heave a sigh of relief that we have a 5s
delay.

An operator with 2G4bit/s links may see the world as 10 or 100 times
faster, but we should still start with a view of what outages there
are, how fast they are notified, how they are notified, etc etc

Tom Petch
nwnetworksdial.pipex.com

-----Original Message-----
From: Manav Bhatia <manavSAMSUNG.COM>
To: OSPFDISCUSS.MICROSOFT.COM <OSPFDISCUSS.MICROSOFT.COM>
Date: 26 February 2003 10:40
Subject: Re: Subsecond hello and dead rtr timer support in v2/v3

I believe this thread was initiated in an effort to realize sub-second
convergence or milli-secs convergence for OSPF (this could apply
easily to
IS-IS also!). Current standards call for implementations to support
these
timers in O(secs). With the ISPs requirements for SLAs increasing
aggressively and the current paranoia people have for faster
convergence a
lot of people think that there should be something done at the IGP
level.

MinLSUpdate/MinLSArrival/etc timers restrict your convergence timings
and
people are basically looking for different ways to over come these
problems. One is of reducing these to O(ms). Everybody is aware of the
problems this could bring in (i will not rehash them here in the
list!) and
people are trying to look for solutions. Obviously when you bring down
these timers to O(ms) we cant just leave the Hello interval hanging in
mid
air with granularity in secs. This too comes down to O(ms).

Now we need to look into how we can mitigate the ill effects of doing
this.

Nobody is saying that you just bring down your HELLO to O(ms). What we
need
to remember is that we are trying to bring down all the timers to this
fine
granularity! How we need to schedule SPF is a different issue though!

Am open to flames! :-)

Regards,
Manav

P.S.
I personally feel that there is a lot more that can be done in
hardware to
bring down our convergence timings before we start tinkering around
with
the IGPs. But that shouldn't put somebody on hold if he/she wants to
do
something at L3!!

• Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]