[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Diffserv] Model - queues
> It seems that the description you have is limited to buffer management
> schemes that make acceptance decisions only at the time of packet
Not quite. If a discarding element is placed after a queue (as the model
permits) it can discard one or more packets from the head of the queue.
This clearly represent the bulk of existing schemes, but there
> are other options that i don't think we want to preclude in the future.
Sure. I generally favor head dropping for real time traffic, although it
tends to annoy hardware designers.
> In general, you can have a buffer management scheme that continuously
> makes state-dependent decisions on packets already in the buffer, i.e.,
> you have the option of changing your mind even after accepting the
> packet. One of the simplest example of such a scheme that has been used
> (with limited success though) in ATM switches is the pushout scheme, but
> there are other variations that are possible and I think we should try
> to have a structure capable of accommodating this.
Hmm. I hadn't heard of this. Do you have a reference?
> So I would say that a discarding element is an element which for every
> state change in the buffer, i.e., a packet leaving or arriving, is
> making decisions on which packets to keep/accept in the buffer based on
> a discarding discipline. This is not much different from what you have,
> but I thought it might be important to clarify things.
What you're pointing out is that the description doesn't capture the notion of
a trigger, which could either be fired by a packet arrival at the discarding
element, some state change in the FIFO element, scheduler, or meter, or a
timer. That's useful.
> One maybe trickier implication is wrt the FIFO structure as more complex
> discard decisions may require additional information. For example, a
> pushout scheme may require multiple linked lists, one for each pushout
> priority. I don't think that we want to change the FIFO assumption but
> it seems that there may be a need for additional "control" structures
> together with it. I don't think this drastically affects your outline,
> but this is again something I felt might be useful to keep in mind.
I was being deliberately vague... er... inclusive here.
diffserv mailing list