netcal / dnc_networkmodel_converter Goto Github PK

The DNC core is designed with the assumption that a service curve is shared by all the flows crossing a server — thus the left-over computations with some multiplexing assumption. Looking at this in a very generic way this can be interpreted as class-based scheduling with only a single class (‘and arbitrary or fifo multiplexing within this class). The other extreme of this generic interpretation is your idea: one class per flow, all classed being of a fixed priority and having a service curve.

I agree that an extension to a more flexible class-based model should go into the converter. Implementation-wise, the most non-invasive way for the DNC core is in my opinion to map one device graph entity (like output port) to multiple server graph servers. The converter should then

• keep track of this mapping and its interpretation (e.g., classes, priorities among classes, multiplexing within a class),
• compute the per-class service curve from the overall service curve of the device graph entity (e.g., an output port’s service curve can be allocated to classes by some scheduling like generalized processor sharing)
• route flows such that they only cross servers in their class (with 1 to n flows per class, n being all the flows in the network as it is currently done).

In particular the last point, routing, needs more information that available in the server graph. Thus, I think it is better to keep a map in the converter than having multiple service curves per server.