>OK, I need an answer to this once an for all. Can someone stake their
>reputation on answering the following question:
>
>Does a Cisco router using PA-A3 hardware (or NPM-ATM) shape vbr-nrt traffic
>based on the cell rate or the payload rate? I hear so many people saying
>that it is done based on payload, and they break out these little
>calculations to 'account' for the overhead. I cannot fathom that this is
>done before SAR. As I see it, the GCRA algorithm is cell-based, therefore,
>leaky bucket operations are most efficiently done at a cell boundary. Given
>varying lengths of CPCS PDU sizes, I would think that shaping would have to
>be done in software were it done based on payload, which obviously would be
>a huge performance hit. Unless the SAR chipset does this first? Call me
>crazy...
The answer is cells. I've been interested in the same issue and did
some testing over an OC3 in our lab this week, which has confirmed
that a PA-A3 shapes the cell rate.
I used the stats reported on an LS1010 (somewhat confusing since the
"show int" display says "packets/sec" but it is really reporting
cells/sec) and an old Network General ATM sniffer. They both agree
that when using VBR-NRT parameters on a PVC the cell rate is shaped to
the SCR value on a sustained flow.
Note that the output bit rate on a "show int atm10/0/0" on the 7513
ATM interface reports the packet bit rate being sent, not the
cell-based line rate.
Another item I checked is that the shaping interface (PA-A3 in VIP2/50
with 128/8 RAM) drops whole packets, not just cells. In other words,
when one or more cells are dropped out of a given AAL5 PDU, the
interface drops the whole PDU rather than send it down the link to
generate an error at the far end.
It was also interesting to check the overhead issues. For example,
sending a stream of 64-byte Ethernet frames at the max rate of a Fast
Ethernet through the router and into a PVC shaped to 25 Mbps. This
resulted in a bit rate of roughly 15.5 Mbps reported on the output
interface on the 7513 and on the destination Ethernet as measured by a
Shomiti analyzer.
Due to LLC/SNAP (8 bytes) and AAL5 encap (8 byte trailer) along with
ATM cell headers (5 bytes) and the resulting 48-byte payload of a
cell, it takes two cells (106 bytes) to transport a 44-byte IP/UDP
packet over the PVC.
-Charles
Charles E. Spurgeon / UT Austin Networking Services
c.spurgeon@mail.utexas.edu / 512.475.9265
This archive was generated by hypermail 2b29 : Sun Aug 04 2002 - 04:12:13 EDT