Switch latency difference between 4000, 2950 and 6509s

[csbowser]

I have four 6509s fully meshed and redundant. Redundantly attached to the 6509s are 5 Catalyst 4006s, running CatOS 7.1(2). Also attached are two Catalyst 2950s running 12.1(9)EA1. Spanning tree is enabled, and the operating correctly so that there is a single, loop free path from any node to the core.

When pinging from any 6509 to another core 6509, a maximum of 4 ms is returned. Same when pinging from a 6509 to a 2950 or 2950 to 2950. However, when pinging a 4000 from any other device, max times are reported between 8 and 32 ms. All 4000s and 2950s are using the same STP root, so there is no difference in costs and no routing is involved - strictly layer 2 activity. Round trip times do not vary with packet size. My question is:

What is the default switching method (each) for the 6509 and 2950s? Would store and forward on the 4000 and cut-through on the others cause this difference?

Are 32 ms round trip time acceptable? (connections are GigE, mm at 100 meters max, so distance and bandwidth are not impacting).

What troubleshooting advice can you give?

[navasil]

Pinging the device may not be a good testing tool in this case since device will choose slow switching path to response ICMP request. They do this because they want to avoid "pinging of dead". Unfortunately I still have no good tool to measure latency on LAN switching network also

Any opinion/idea??

[rfroom]

All the switches mentioned are store-and-forward. The latency is due soley to the nature of the software on the Cat4000. Only packets directed to the management IP address will have the occasionally additional latency.

When an ICMP echo request ("ping request") comes to a Cat4k switch, the packet is queued. Processing it immediately would probably be a bad

choice, other important process may be in the queue such as STP, etc.

As a results, pings are low priority and will be preempted by other more important frames in queue to the CPU. The nature of the hardware and software causes this behavior and does not exist on Cat6k and 2950. The original 2900XL and 3500XL experienced the same issue.

Don't use pingsto the management IP address to measure system performance.

[csbowser]

I appreciate your advice on this. I agree that pinging is not the best measurement tool, but right now it is the only tool I have and I wanted to identify any potential architecture issues prior to application loading.

Thank you.

[rfroom]

The best way to measure latency performance is through the switch, 2 devices connected to the switch pinging each other.