From: Howard C. Berkowitz (hcb@gettcomm.com)
Date: Wed Jun 25 2003 - 21:27:27 GMT-3
At 8:11 PM -0300 6/25/03, Carlos G Mendioroz wrote:
>Howard,
>I don't see how what you say can happen.
>Would it be possible for you to come up with a simple example where
>a ratio assignment would generate high short term throughput
>variability,
>and a good matching with one calculated by normalize/divide/multiply
>stuff ?
Four classes, each to receive 25%.
Let's say that one of the classes is an old IBM Token Ring
application that still uses the 4K packet size. For ease in
computation, let the average packet size of the other queues be 1K.
If you set up the classes 1K:1K:1K:1K, the IBM queue will be serviced
only every 4th cycle.
If you set it up 4K:4K:4K:4K, and there is enough traffic to fill
each cycle, the traffic flow will be smooth (1 packet:4 packets:4
packets: 4 packets).
>
>Using big numbers may do, but using ratios it would be easier to
>actually use smaller numbers anyway.
>
>Thank you,
>-Carlos
>
>Howard C. Berkowitz wrote:
>>At 4:02 PM -0500 6/25/03, John Humphrey wrote:
>>
>>Well, yes, but there are warnings involved. Yes, it's true that
>>remembering the deficit will make the long-term statistical
>>allocation of bandwidth more accurate with less effort.
>>
>>The problem, however, is that if you'v not estimated well (i.e.,
>>used estimates that reasonably reflect the average packet size),
>>you will get short term peaks and valleys in the throughput per
>>class. That may be OK for non-interactive traffic, but bad for
>>interactive and catastrophic for real-time.
This archive was generated by hypermail 2.1.4 : Fri Jul 04 2003 - 11:11:09 GMT-3