North American Network Operators Group

Date Prev | Date Next | Date Index | Thread Index | Author Index | Historical

Real World OpenBSD/OpenBGPd statistics

  • From: Phil Pierotti
  • Date: Fri Dec 26 22:40:07 2008

Recently on this list there's been many and various discussions about
high-throughput on *some random OpenSource platform* aka
software-routing-on-a-PC.

And it's been a very interesting discussion, ranging from noting "vendors
claims" and discussing why (or not) they work (and how that differs from
'the PC universe") to detailed OS Kernel/IP stack implementations and their
isues/benefits.

Some of the statistics quoted are dated (eg 2006/2007 timeframe).
Some of the statistics quoted are "in theory" (eg in a remote lab, isolated
from anything resembling the real world)

I would like to invite people to comment on their real-world aka "doing this
in production" results. Not just "I've tested it" , but "we use this as our
production routing every day". Yes, I am assuming there's more than just
talk going on out there.

Please Provide Hard Numbers:
- system description (cpus, bus, nics, memory, OS, Version)
- packets per second as well as throughput
- average packet size
- how much did you have to tweak to achieve that

Ideally we're talking about a "router doing BGP and forwarding packets"
scale of  implementation.
Pure and simple routing packets in a real-world usage scenario.

Admittedly being "just a PC" you'll want some base firewalling rules to
secure the box itself, maybe Anti-Spoofing rules.
But nothing along the lines of "a do-everything box, load-balancing,
fireall, nat, complex and long ruleset, 12 routing protocols" etc.

Consensus *seems* to be that filling "a few" 1Gbps nics even full duplex
should not be difficult on "decent server class hardware" (eg Dual 2.3GHz
Xeon with 4GB RAM and good nics), even for real-world traffic patterns.
There is still some dispute/uncertainty on how far you can push that into
10Gbps (or even multiples of that), or at least on what it takes to get
there.

Thanks,
Phil P