[ On Mon, November 23, 1998 at 14:30:42 (-0500), Perry E. Metzger wrote: ]
> Subject: Re: This PMTU thread 
>
> Demon Internet in the U.K. was running NetBSD for its routers at the
> big exchange point in London for a while as I recall. Lets assume that 
> each of those machines was handling fourty or fifty thousand
> simultaneous TCP connections transiting through. Thats, what, fifty
> megabytes of state? Plus all the CPU for processing each packet? Plus
> a giant cache miss for every single packet going by?

You're not paying attention now Perry.  How many ports with smaller than
normal MTU do each of those routers have?  And how many simultaneous
connections to hosts with broken PMTU-D will pass through each port????

OK, since some folks are having trouble with the math here, let me try.
If you can count the number of broken and un-fixable PMTU-D endpoints
that your users will try to connect to, or will try to connect to your
users, on the fingers of one hand, and even if every "router" had 32
such ports, we're probably talking well under 200KB of data, and it's
all just an LRU cache which will re-prime itself as necessary.  Any
router could get by with one or two entries in each table, at some
potential cost to efficiency and retransmissions.  That's why the
logging and/or instrumentation is necessary -- for tuning!

-- 
							Greg A. Woods

+1 416 218-0098      VE3TCP      <gwoods@acm.org>      <robohack!woods>
Planix, Inc. <woods@planix.com>; Secrets of the Weird <woods@weird.com>