[c-nsp] Lot of input errors on a NPE-G1 interface

[Matlock, Kenneth L]

Think about the buffering on a chassis like a 7200 like this.

You have 2 buffers on input, the RX ring (a hardware buffer), and the
input queue (a software buffer)

A packet comes in on the wire, and goes into the RX ring. That generates
a CPU interrupt. The CPU needs to finish its current task then go
address the interrupt. It takes the packet off the RX ring and puts it
into the input queue.  The CPU then takes the packet from the input
queue, applies ACLs/NAT/etc to it before deciding if/how to forward it.

Now, keep in mind that a 7200 only has an RX ring of *128*.  Worst-case
at 64 byte packets that's only 8192 bytes of hardware buffer space. 

At 1gb/sec (say a small burst) that only gives the CPU about 60
*Microseconds* to finish what it's doing and grab that first packet off
the RX ring before the queue fills up and you get an overrun. 

The only real fixes to this are 

1) Even out the traffic to remove the bursts (traffic shaping upstream)
2) Decrease the CPU to let it better handle the bursts
3) Get a bigger box that does the hardware-software transfers via
hardware, not on the CPU.

You can see how much of the CPU is being taken up by a 'show proc cpu'
(the %x/%y portion) %x shows the CPU utilization, the %y shows how much
is interrupt traffic. 

Ken Matlock
Network Analyst
303-467-4671
matlockk at exempla.org

*** Exempla Confidentiality Notice *** The information contained in this message may be privileged and confidential and protected from disclosure. If the reader of this message is not the intended recipient, or an employee or agent responsible for delivering this message to the intended recipient, you are hereby notified that any other dissemination, distribution or copying of this communication is strictly prohibited. If you have received this communication in error, please notify me immediately by replying to the message and deleting it from your computer. Thank you. *** Exempla Confidentiality Notice ***