[c-nsp] Older Cisco Routers - which one to go with?

[Tony]

--- On Wed, 23/2/11, Youssef El Fathi <youssef.el.fathi at gmail.com> wrote:

> Be careful with the router performance, because by
> experience when you want
> to implement QoS the performance is nearly divided by 2

I recently has reason to do some testing of this on a 2811 and think it might be worth sharing the results with the list. All of the testing was done in one direction only and from devices connected directly to the ethernet interfaces of the 2811.

The firsts tests were done using iperf with the following parameters:

Packet type: UDP
Packet length: 400 bytes
Test duration: 60 secs
Test bandwidth 20Mbps
Number of packets: 374995

1. With no QoS settings applied on the router there was zero packet loss and the CPU on the 2811 sat on about 11% (that is the CPU usage just to punt that many packets around).
2. With QoS matching packets inbound on the source interface (and setting DSCP), but no outbound queuing the CPU increased to 28% (still zero packet loss).
3. With a normal output QoS policy (not hierarchical) the CPU was at 34% with 0.0005% loss (which could just be statistical anomally).
4. With H-QoS applied to shape to 20M and prioritise VoIP traffic (with no VoIP traffic) CPU was 58% with 10% packet loss, which is due to the fact that 20Mbps of traffic from iperf is 20Mbps of payload, whereas the 20M shaping on the box includes headers.

I repeated the tests, but this time with these parameters:

Packet type: UDP
Packet length: 1400 bytes
Test duration: 60 secs
Test bandwidth 30Mbps
Number of packets: 160843

1.	5%, zero loss. 
2.	13%, zero loss.
3.	15%, zero loss
4.	32%, 3% loss (less loss, because less overhead per packet due to larger packet size)

The reason for less CPU across the board with this second test is that even though the bandwidth is greater (30M v’s 20M), the number of packets to be forwarded is less, so CPU doesn’t work as hard.

I then decided to try and simulate some VoIP traffic, using two iperf sessions, one for “data”, one for “voiP”

-=VOIP session=-
Packet type: UDP
Packet length: 100 bytes
Test duration: 60 secs
Test bandwidth 3Mbps
Number of packets: 225559

-=DATA session=-
Packet type: UDP
Packet length: 1400 bytes
Test duration: 60 secs
Test bandwidth 30Mbps
Number of packets: 160843

During this test the CPU on the box sat on about 72%. I’m not sure if this is because there was more traffic being chucked out (only an extra 3%), the extra number of packets being pushed through the box, or because it was having to put stuff into a priority queue ahead of other stuff. There were zero packets lost from the VOIP session and 17% lost from the DATA session (so H-QoS did it's job properly).


HTH & YMMV


regards,
Tony Miles.