[nsp] 100% CPU load on nrp-1 due to DHCPD Receive process ?

[Miquel van Smoorenburg]

I wonder if anyone has seen this before and knows a solution
or a workaround.

We terminate our DSL  clients on a 6400 with NRP-1s. About 1000
clients per nrp-blade. Encapsulation is 1483/RBE, with a seperate
subinterface per client. We use "ip helper-address <dhcp-server>"
on each interface and "ip dhcp relay information option" in
global mode. IOS 12.2(15)T7.

Yesterday our DHCP server went down. After 2 hours, all customer
leases (lease time is 2 hours) were ended and most of them
started to ask for a new lease.

This caused the "DHCPD Receive" process to consume all CPU. On
ATM 0/0/0, we saw lots of dropped input packets - presumably
because the DHCPD Receive process couldn't process them fast
enough.

We revived the DHCP server but that didn't help much.

But surely an NRP-1 must be able to DHCP-relay a few hundred
packets per second, right ?

It took 4 hours for the NRPs in the 6400 to recover .. in the mean
time the NRP was so cpu-starved that it dropped lots of packets
and sometimes completely locked up for 10-20 seconds.

Any insights appreciated.

Mike.