Andrew Fort writes:
> On 4/5/07, Monty Ree <[EMAIL PROTECTED]> wrote:
>> I would like to capture all flows as you said.
>> But the traffic is over 10Gbps, so I should use sampling to reduce 6509 CPU
>> load.

> Note that if I remember this all correctly, you won't lower SP CPU
> loading, only RP CPU load as a result of generating netflow packets
> for export.

It's the other way round - on newer IOS releases (I think 12.2SXD or
SXE and later), the SP does practically all of the work of sending out
NDE packets, and the RP is mostly unimpacted by the rate of flows to
be exported.

This is for the centralized case.  I think the RP is involved in
exporting packets from DFCs.  This is a pity because otherwise adding
DFCs would be a great way to scale NetFlow performance (it definitely
increases the effective hardware flow-table size).

> I'd try to avoid running Netflow on the Sup720 at all costs, but
> best of luck.

The Sup720 has the highest-performance (unsampled) Netflow
implementation I know of.  If you can tune your timeouts so that your
flows fit into the hardware table (128Ki entries on the PFC-3B, 256Ki
entries on the PFC-3BXL, then it will perform nicely.  If you have
more flow, NetFlow will not be able to account for all packets.  But
at least forwarding still works at line rate, and you even get
counters that tell you how many packets could not be put in the flow
table.
-- 
Simon.
_______________________________________________
Flow-tools mailing list
[EMAIL PROTECTED]
http://mailman.splintered.net/mailman/listinfo/flow-tools

Reply via email to