Subject: RE: OT: Routers
From: Christopher E. Brown (cbrown@woods.net)
Date: Wed Mar 05 2003 - 11:27:23 AKST
Sounds like an interesting network upcoming. Mike is very right
about the Ciscos, you are looking at alot of un-needed expense.
As to the complexity of a Linux vs Cisco setup, for this type config
there is not alot of difference. The complexity is in the network
design and route layout, not the devices. One way you have to deal
with some cryptic syntax and the other way you have to ... wait for it
... deal with some cryptic syntax.
You *COULD* build a much more complicated setup with Linux, if you
decide to start playing with say netfilter DS marking, bandwidth
control and packet prio/reorder, on the fly packet header rewrites, or
other magic items. But, in this case the Cisco setup would only be
simpler because Cisco gear of the class you are talking about cannot
do any of it (and the higher end stuff only some of it). In any case
it doesn't sound like you have the need (or the staffing and time) to
play with such an overkill system.
Linux on recent x86 hardware will handle an insane amount of traffic,
400Mbit aggregate was an easy job on a (properly tuned and selected
hardware) K6-4 450Mhz machine, with 80% of the CPU left over for
filtering/traffic management/etc. Things like Athlon-XP 2200+ based
units with PC2700 memory and PCI-64/66Mhz buses will handle things up
into the Gbit/sec level (though one does have to watch packets/sec and
make proper hardware choice to prevent interrupt loading).
I am assuming you are referring to pairgain ethernet bridges, but at
what speeds? 256kbit, 512kbit, 768kbit, 1.544mbit?
A routed solution is the way to go, assuming windows clients will be
present, a multi-location network in a switched config would easily
consume the entire remote link bandwidth for broadcast traffic alone.
You need a router at each remote location, and one or more router(s)
at the core, otherwise you end up passing all layer 2 and layer 3
broadcast traffic everywhere and eating your links. One could do
something of a hybrid design, a full routed system, with a switched
inter-router area if one were careful about the router setup. This
would say alot of $$ at the core on router ethernet ports.
The remote routers are trivial, assuming < 10Mbit links back to the
core, and that the service requirements are running the link, and
providing DHCP for the local subnet. An AMD 486DX-133 with 32megs ram
a PCI bus and a couple good (tulip, eepro, etc) network cards would do
the job easily, we are only talking about 2,000/pps - 36,000/pps for a
fully loaded 1.544mbit link (bi-directional), given normal office
traffic types and packet sizes, prolly in the range 6,000/pps for a
fully loaded bi-dir link. Assuming a well built 486-133 the router
would run out of CPU at around the 28,000/pps mark, and the only
possible way to hit that would be using packet generators set for a
64byte (min length) packet size running both ways. Oh yes, the pps
figures you were listing for the Cisco gear are not comparable to the
ones I am listing. The Cisco figures are for direct interface to
interface forwarding where the routes are already in hardware cache
and the packets are min length. To get a safe figure for real world
load handling (mixed traffic, multiple source/dest, etc) for a Cisco
in normal config divide the provided best case by 3, with a baseline
2600 listed at ~ 40,000/pps this gives about 14,000/pps safe figure.
For the core a 500Mhz or better machine with more than 64megs ram
would be more than enough to handle the worst case loads (All remote
locations doing large file copies/etc at the same time). Given a
routed network with an inter router switched area and properly
controlled and configured routes overhead would be minimal, and remote
to remote traffice would not even pass via the core router.
Oh yeah (putting on several of my hats), for some reason I seem to end
up doing alot of network design work, and Linux/*BSD based router
design, so far on x86, StrongARM, PPC, Alpha, UltraSparc and MIPS
platforms, from 1mbit embeded units up through 4Gbit aggragate SMP
Alpha 21264 machines. If you need actual network
design/implementation services, hardware/etc feel free to contact me
off list about it. Most of the places I work with/for deal with this
kinda of thing, on a daily basis, and I have sources for all sorts of
interesting things routing wise.
-- I route, therefore you are.
---------
To unsubscribe, send email to <aklug-request@aklug.org>
with 'unsubscribe' in the message body.
This archive was generated by hypermail 2a23 : Wed Mar 05 2003 - 11:27:05 AKST