Alan,
>
> In order to get nice traffic shaping, this should be pushed out to
> the application device. Note that this is generally handled quite
> nicely by TCP, available on most hosts :-)
Not much use if you aren't in control of the application device
at either end though :-)
> The ATM_Lite had only one queue as far as I can recall, w/ the ATM
> Deluxe having multiple queues.
The ATM solution I was thinking of was running something horrible
like LANE (*) through a BPX to give ABR VS/VD / Foresight traffic management.
One PVC per server (one VLAN per server if necessary).
That way I get exactly the control I want as I have verified to my
satisfaction a BPX can do this. But as I said, this sounds like
overkill...
(*) actually you could do it with policy routing, a BPX, one AIP, and
sensible encapsulation. Equally ugly.
> I don't think you are whining over nothing, but focus on the
> fact that the limiting of packet transmission rate will cause
> TCP to conform to that bandwidth without peaking too near the
> upper constraint.
>
> I must confess I'm unclear as to how buffering and CAR work
> together, except to know that there's a single queue. It would
> seem that anytime the 'drop' CAR functionality is employed,
> it is just dropped. Since there's no 'buffer' option, I'd suggest
> that the 'transmit' then throws it towards the queue. Sounds like
> I need to go play in the lab.
In a different document, Cisco wrote (whilst extoling the virtues
of limiting SCR oversubscription, and traffic management technologies
which allow minimization of cell loss in the ATM core):
Effect of Discards on TCP/IP Traffic
* A single cell discard causes TCP to react by halving its window size
* The Window is slowly reopened (slow start) until another discard
is experienced
* In practice multiple TCP sessions become 'synchronised' and start 'porpoising'
* TCP Throughput converges towards 50% of the trunk capacity
* User throughput drops WELL BELOW the SCR
That whole document reads like a thinly veiled criticism of Cisco's (ATM)
competitors whose traffic management is no more and no less dumb than,
erm, CAR.
As you say, time to play in the lab.
-- Alex Bligh GX Networks (formerly Xara Networks)
This archive was generated by hypermail 2b29 : Sun Aug 04 2002 - 04:13:13 EDT