[c-nsp] RMON history collection caveats?

[Kevin Graham]

Is anyone doing large-scale RMON history collection presently? We're
considering it for capacity and error monitoring, as history buckets allow for
really trivial implementation of otherwise complex checks like "on any given
port, I don't care about a 1000 errors in a bucket, but I do care about 10
errors in each of 3 consecutive buckets".

Most of the CCO notes seem to date back to when RMON was new and cool and how
to turn it up on a cat5k.

An efficient implementation would suggest that the switch collects existing
interface counters every history bucket; presumably these will exhibit all of
the usual platform-specific counter quirks?

Is there any reason at all to be concerned about CPU/memory utilization?
Obviously on most platforms I wouldn't think twice, but for example a large
3750 stack serious skews the ratio of ports to cpu/mem.

We'd be looking at doing this indiscriminately across 6500's, 4500's and any
mix of small DSBU switches (from 3750E's to some 3500XL's that refuse to die).