RE: Progressing MAR

["Francois Le Faucheur (flefauch)" <flefauch@cisco.com>]

Jerry,
 
>> 
>> Once bandwidth is allocated, there is *no assumption* that 
>> this will actually represent, or be close to, the actual 
>> traffic that arrives.  Extremely large swings and surges in 
>> traffic occur, massive overloads occur, failures occur.  MAR 
>> is expected to operate well under these conditions (and does 
>> in actual operation, see below).  All the BC models in fact, 
>> should operate well under these same conditions.  There 
>> should be no presumption of 'accuracy' in bandwidth 
>> allocations, etc. in any model, or that the actual load is 
>> necessarily always close to what you are allocating on all links.

I entirely agree with the above. 

That was precisely my point. I fail to see how MAR can "simultaneously
achieves bandwidth efficiency, bandwidth isolation, and protection
against QoS degradation without preemption" in situations where loads
for each CT are quite different from allocated loads:

Imagine that Bwalloc(CT0)=10, Bwalloc(CT1)=30, Bwalloc(CT2)=60
Imagine at a given time load(CT0)=100, load (CT1)=10, load(CT2)=10.
Then I understand that with MAR:
	- CT2 will be allowed to grab 10
	- CT1 will be allowed to grab 10
	- CT0 will be allowed to grab 80-delta= say 70
So far so good.
But then let's assume the demand for CT1 and CT2 starts increasing again
without CT0 decreasing its load (which we agreed above is very possible
since CT0 load can be differnet from its allocated bandwidth), so we
have:
	- load(CT0)=80, load(CT1)=30, load(CT2). 
then in the absence of preemption I understand we will end up with
something like:
	- CT0 keeps its 70
	- CT1 will increase to say 15
	- CT2 will increase to say 15.
Is this not what would happen, more or less, with MAR?
If yes, then CT0 is hogging 70 when its engineered for 10, while CT2
only has access to 15 when it is engineered for 60. This would not be a
great result in terms of class isolation.
Now you could make "delta" bigger, this would result in better
isolation, but would also result in loss of bandwidth efficiency. For
example if delta=30, then in above example, intially CT0 will be limited
to 50. And when CT1 and CT2 later increase their load, they will be able
to grab say 25 each. This is a little better in terms of isolation (but
still far from good). And we initially had to reject a load of 30 of CT0
so that the max efficiency was 70 out of 100.

Hence my perception that :
	- MAR can "simultaneously achieve bandwidth efficiency,
bandwidth isolation, and protection against QoS degradation without
preemption" IF AND ONLY IF we make some assumpitions on the load
patterns of each CT.
	- MAR can NOT ""simultaneously achieves bandwidth efficiency,
bandwidth isolation, and protection against QoS degradation without
preemption" if we do not make assumptions about load patterns of each
CT.

What am I missing?

Thanks for your help in understanding MAR properties.

Francois