[j-nsp] M20 SSB/IP2 CPU usage on IPv6 forwarding
Kevin Day
toasty at dragondata.com
Fri Feb 22 01:13:28 EST 2008
Hey, J-NSPers!
I'm nearly ready to launch what's probably going to be the largest
public use of IPv6 so far[*]. The end result is that we're going to be
shoving many gigabits of v6 traffic through our M20. In preparation
for the real deployment, I've been doing some internal testing and
simulations to make sure all of my infrastructure can handle this.
When pushing about 4gbps of v6 traffic through our router (between
several directly connected hosts, nothing targeting the RE or router
interfaces), the SSB starts using 100% CPU:
SSB status:
Slot 0 information:
State Master
Temperature 40 degrees C / 104 degrees F
CPU utilization 99 percent
Interrupt utilization 0 percent
Doing a "start shell pfe network ssb" took about 30 seconds for it to
actually log me in. I typed "show threads" to see if I could figure
out what it was doing. It took about a minute for that command to even
echo back to me. After another few minutes of waiting for a reply it
dumped me back to the RE without responding.
The only log messages I got were some:
ssb CM(0): received message error, subtype=11, op=169
and a few:
pfe_send_failed(index 2, type 3), err=32
And finally when I stopped everything, I got this before it all came
back up:
rdp keepalive expired, connection dropped - src 1:1020 dest 19:34816
pfe_listener_disconnect: conn dropped: listener idx=3, tnpaddr=0x13,
reason: reconnect timeout
Running the exact same test over v4 caused no problems at all. Does
anyone know what gets dumped to the SSB CPU while forwarding v6
packets? Is there any kind of document that shows what gets done on
the CPU on the v6 side? We also had an offer that we previously turned
down to borrow a M160 from someone - is there any reason to believe it
would be any better at v6 forwarding?
Because of the affect this had on our production v4 network I wasn't
able to experiment much with trying to narrow down the cause, so I
apologize for the vagueness of this post. :) I'll try to repeat this
if I have a better idea of what to be looking at while it's going wrong.
-- Kevin
[*] http://www.ipv6experiment.com if you're curious
More information about the juniper-nsp
mailing list