The IP over ATM Mailing List Archive by date

[Jim Forster <forster@cisco.com>]

>    Also, there is an implied view (at least on my part) that if I attach
>    my host (or your router :-) directly to the ATM network, then I need to
>    pump _out_ a lot of data, which of course will be echoed back to me
>    using MCS. Hence, if my node has to also filter out echos at the IP layer, 
>    then I'm wasting CPU cycles (and worse, suffering from memory access 
>    latency) in dropping IP packets.
>    
>    I thought routing vendors developed rashes whenever memory has to be 
>    accessed :-)

Not nearly so bad as if you change encapsulations.  We have over a thousand
ATM interfaces that we've shipped, and we thought the encapsulation stuff
had been put to bed.  It would be unfortunate if vendors that had nothing
to lose were not sympathetic to the fact that there *is* an installed base.

> Indeed, if you push the problem into the IP layer, cycles (memory &
> cpu) that could have been used to forward legitimate packets will be
> spent doing this filtering.  Your mileage will vary based on how much
> multicast you send and the filtering implementation.

I don't see any difference in cycles whatsoever between doing at the IP
layer or below.  In either case it's one comparison statement.  There might
be some savings in memory cycles by not lengthening the encapsulation, but
all this is second or third order compared to changing an installed base.

  -- Jim