Is it possible to set service curve for individual flows?

[litaihei1]

Question / Feature Request

I would like to know if it's possible to configure a service curve for individual flows at a specific server, rather than only having a total service curve for the server.

Current Understanding

From the codebase, I can see that:

• Server class has a service_curve field representing the total service curve
• Individual flow service curves are computed as leftover service curves: β_lo = β - α_cross
• There doesn't seem to be a way to directly specify a service curve for a specific flow at a specific server

Use Case

• Different flows need different service guarantees at the same server

Questions

1. Is there an existing way to achieve this that I might have missed?
2. If not, is there a plan to support per-flow service curve configuration?
3. Would this be a reasonable feature to add, or is there a design reason why it's not supported?

Thank you for your time and consideration!
Li Yunpeng

[sbondorf]

Hi Li Yunpeng,

Yes, put simple, the tool allows you to route multiple flows over one server, assuming that the given service curve needs to be converted to a left-over service curve for individual flows.

There is an easy way to provide your own per-flow service curves is to create the server graph this way, i.e., with one server per flow.
The modeling assumption behind this is that a server need not correspond to a physical output port of a switch. Instead, it can be the service that a round robin scheduler guarantees to an individual flow. Then, the server graph will have as many servers per output port as there are flows. Constructing the server graph like this, you need to manually take care of connecting the servers and routing the flows before handing it over to the tool. The tool can then automate the rest, i.e., deriving bounds (and doing so, will not compute a left-over service curve as there are no cross-flows at any server in the new server graph).

An example can be found here:
https://dl.acm.org/doi/abs/10.1145/3649477.3649495

• In Fig. 2(a), the output ports are depicted under the assumption that each port corresponds to one server with one service curve
• In Fig. 2(b) one of the ports will be modeled as two "buffers", which simply means there are two distinct service curves to separate flows (in this paper it is done to break a cyclic dependency of flows)