[c-nsp] IPv6 not Fast/CEF-switched on 3640 12.3T

[Daniel Roesen]

Hi,

FastEthernet0/0
...
    Protocol  IP
          Switching path    Pkts In   Chars In   Pkts Out  Chars Out
                 Process      14000    1183726      15308    1963244
            Cache misses        111          -          -          -
                    Fast      30121    5299973      20333    6336875
               Auton/SSE          0          0          0          0
...
    Protocol  IPv6
          Switching path    Pkts In   Chars In   Pkts Out  Chars Out
                 Process      42868    4272263     420507   54021071
            Cache misses          0          -          -          -
                    Fast          2        188          0          0
               Auton/SSE          0          0          0          0
...

So basically all IPv6 traffic out of this interface is being
process-switched. Why?

Config has:

ipv6 unicast-routing
ipv6 cef
ipv6 multicast-routing

interface FastEthernet0/0
 ip address ...
 no ip redirects
 no ip proxy-arp
 ip nat inside
 ip virtual-reassembly
 speed 100
 full-duplex
 pppoe enable
 ipv6 address ...
 ipv6 enable
 ipv6 nd ra-interval 3
 ipv6 cef
 pppoe-client dial-pool-number 2
end

Usually I see no problems, but when subscribing to an IPv6 multicast
source, CPU load goes up:

CPU utilization for five seconds: 26%/12%; one minute: 22%; five minutes: 12%
 PID Runtime(ms)   Invoked      uSecs   5Sec   1Min   5Min TTY Process
  60      261924    324971        805  8.00%  6.69%  3.71%   0 IP Input
 203      204696    312161        655  4.87%  4.35%  2.37%   0 IPv6 Input

The multicast traffic is received on GRE Tunnel102 and forwarded to the
FastEthernet0/0. Traffic level is low though:

Tunnel102 is up, line protocol is up
  Tunnel protocol/transport GRE/IP, key disabled, sequencing disabled
  5 minute input rate 122000 bits/sec, 111 packets/sec
  5 minute output rate 0 bits/sec, 0 packets/sec

Any unicast traffic transitting the box then has some 1.8-1.9 seconds
RTT rise and ~25% packet loss.

In a different thread I think I've read that the 3640 maxes out at
~200pps process switching. This sounds reasonable when looking at the
CPU load while switching the ~110pps multicast traffic.

But I wonder why all that is being switched process-level and not CEF?
Is IPv6 multicast still non-CEF on 3640 with 12.3T? IPv6 unicast traffic
received from the GRE tunnel is also sent to the FastE process-switched.

Sidenote: incoming IPv6 traffic isn't being counted in the switching
stats properly, is that normal?


Best regards,
Daniel

-- 
CLUE-RIPE -- Jabber: dr at cluenet.de -- dr at IRCnet -- PGP: 0xA85C8AA0