[c-nsp] 6500 service modules and process switching

[Ross Vandegrift]

Hey everyone,

Last week's converstaion on 6500 scalaibility numbers prompted me to
dig back into the performance issues I have on a pair of boxes.  I
discovered two issues, both related to the presence of CSMs in the
chassis.  An important note about my scenario - I have two basically
identical installations, one which performs brilliantly, one which
doesn't.

The first is a semi-documented bridged mode issue.  The CSM is not
enough of a bridge to bridge traffic that stays on the server VLAN.
It must be bridged onto the client VLAN, handled by the MSFC, and then
bridged back.  (I don't have the cisco.com reference for this, but
could find it.)  It seems that if the two endpoints are
directly-connected to the client SVI, that traffic is software
switched.  This only affects servers without a direct layer 2 path on
a downstream access switch.

The second issue is a complete mystery to me.  Many flows directed at
VIPs (every flow?) on the CSM have their first packet switched in
software.  This adds up to an enormous amount of traffic for us.
We also have a well-performing pair of 6500s that do *not* display
this behavior.

I always thought that service module internal interfaces behaved just
like normal interfaces, except that they traversed the bus and not
the fabric.  Clearly, this isn't the case - there's something else
going on.

Anyone know anything about process switching the first packet of flows
to service modules?

-- 
Ross Vandegrift
ross at kallisti.us

"If the fight gets hot, the songs get hotter.  If the going gets tough,
the songs get tougher."
	--Woody Guthrie