Summary of Functional Platform Requirements arising from 5G Use Case

In view of multiple (still unresolved) comments, particularly from Cisco, on the Sub Case C ("Optimization"), my understanding is that we don't have consensus on the corresponding requirements. For example, "Optimization - Conflict Resolution" relates to non-realistic scenario in which the network operator launches several "individual SON functions" that may be conflicting with each other.

For the "Optimization - Onboarding" requirement, there is no basis in the Sub Case C where the term "onboarding" is not used.   

About optimization/conflict resolution:

Basically we need a policy driven approach to avoid conflicts in a SON multi-algorithm environment; much like an "arbiter".  However there is an implication for optimization here. Given an operating condition and a performance objective, what is the best combination of algorithms? And their "parameter" settings?

About onboarding:

ONAP-OF provides a framework in DCAE to implement optimization algorithms as micro services, in an efficient manner. The efficiency stems from two aspects: (1) model driven optimization and (2) model driven data preparation to support the optimization models. On-boarding means, implementation of optimization services in that environment making use of the recipes and libraries, as well as managing them using the DCAE controller. 

About optimization/conflict resolution:

Not every thing we need is possible.

The proposed scenario with several conflicting optimization MSs is not realistic. I presume, we are talking of optimization MSs that are independently designed, possibly by different vendors.  The conflict often considered in this context is when two optimization MSs, X and Y, are trying to simultaneously modify same configuration parameter. The example mentioned on the recent call was about X = mobility optimization and Y = load balancing. Then the “arbiter” function A detects the collision and resolves it in the way that X is allowed to go while the request from Y is blocked.

I’d like to challenge the whole idea of launching independently designed optimization MSs that may be conflicting. For the above example, Y should improve network performance (balance the cells load) in the scenario when the configuration parameters are changed by another independent process X running in parallel with Y. From the point of view of Y, these changes come in random times and in random directions. Is it possible to design Y in the way that would provide for reasonable performance in such hostile environment? I don’t think so. The problem exists even when X and Y never collide.

If X and Y collide, this creates additional problems for the design of Y: to survive the combination of any imaginable X with any imaginable arbitration rules, so that the actions requested by Y are executed or blocked in a random way.

Bottom line: scenario with conflicting independently designed optimization MSs is not realistic.Therefore, the requirement for conflict resolution is requesting implementation of something that cannot be used.

About onboarding:

The comment kindly provided by Sarat does not explain what exactly is onboarded. Also, I don't know what model driven optimization means. Example of this e.g. for RAN optimization will be very helpful.

In any case, onboarding is not mentioned in the Use Case. It should be if we want to follow right methodology.

I have a concern about the following statement

"Since optimization problems in general have   a common structure (viz. maximizing/minimizing an objective subject to   various constraints), a framework exploiting this commonality would render   significant efficiency in creating optimization functions in these two areas. "

I read it as all (many) optimization algorithms are similar / follow same patterns / .... My observation is that practically used optimization algorithms are all different. There are many reasons for that, one of them being wide use of heuristics. Another is that in live network the measured values and KPIs are permanently changing even with no configuration change. The cause of changes are not necessarily observable. It would be oversimplification to consider this as calculus examples where we maximize well defined function f(x,y,z) with the set of constraints g(x,y,z)=0 and the gradient descent method always applies. The statement of common structure is therefore not trivial at all, and needs a strong proof.

I think I did not articulate this well. Of course the algorithm can be different and problem dependent. The value of optimization framework is the formulation of problems in an effortless manner. Basically we create a model for the problem to be solved, and then create a data model to procure the data to solve the problem. Then you can use the best algorithm to solve the problem. Let me try to further explain this with an example. A variety of constrained optimization problems can be solved as LPs (linear programs). These problems can be vastly different and the algorithms to solve the problems can be also different (simplex, barrier functions, interior point – you name it). But these problems can be formulated in a similar fashion – a model description and data feeding the model. This is a common practice in OR community. The ONAP-OF is taking it a step further. Constraints are represented as policies and associated with a math model – and a data model supplements this; fetching data on the fly from appropriate sources defined by the model.

I have discussed the possibility of some  ONAP-OF community members presenting these (possibly with a demo as the basic functionality is getting ready – placement problems to start with). So let us discuss.

I think it is premature to completely remove the requirement for conflict coordination. I strongly propose to add this back and work out the various cases that can be coordinated in the next phase. In general, Amdocs' position is that there should be scope for multiple analytics that can be deployed as per the operator defined policies and conflicts resolved during closed loop function design time.

Hi Vimal Begwani, I have added Amdocs as a willing contributor to further develop optimization requirements.

Good discussion.  Since this group (5G use case group) intends to gather and eventually provide 5G specific optimization requirements to ONAP-OF group, it is good to have some understanding what types of 5G optimization use cases we know of.

In 4G and before, main use cases are "Radio Parameter Optimization" and "Transport Parameter Optimization".  In case of "Radio Parameter Optimization", based on the information central function gathers from eNB RLC/MAC/PHY layers across multiple eNBs, runs some optimization algorithms, gets the output from these algorithms and sets parameter values such as TX Power,  Antenna tilt,  Handover parameters,  RLC/MAC protocol parameters etc..   Same is the case with "Transport parameter optimization", which results in setting new transport parameter values such as backhaul QOS/PHB,  peer eNBs,  new S1/X2 sessions,  IPSec tunnels etc... 

It is true that optimization algorithms that are deployed differ from operator to operator.  So, providing flexibility of selecting optimization algorithms is good. 

On the conflict resolution:  This problem is not new in 4G too.  In 4G also,  optimization needs to keep in mind "Capacity" (How many UEs to be supported by eNB for example),  "Performance" (How much throughput to be provided to each UE), Cost and coverage (larger coverage) and the parameters values resulting for each type of optimization could be different and that is where conflict resolution is required.  

So far, one SW entity does all the job of optimizing parameters keeping all types of optimizations (Capacity,Performance,coverage, cost) in mind etc..   

I guess, now the discussion seems to be whether there is one high level optimization algorithm or provide flexibility to have multiple optimization algorithms (one for each optimization type such as capacity, cost, performance, coverage) and use rules to collect output from each discrete algorithm do any conflict resolution of parameter values.  I hope I got the problem statement.  

Essentially, the discussion is about one monolithic optimization program (could be a micro service) for each optimization use case (Radio, transport etc..) or multiple optimization programs for each use case for various types of optimization) and connect them together with some conflict resolution rules.

I think it would be good to know from experts on following:

What additional optimization use cases 5G require beyond "Radio" and "Transport"?

For each optimization use case, what are different types of optimizations needed (such as Capacity, Cost, Power, Performance, Coverage etc...)?

Are there any SW vendors who provide optimization algorithms for each type? or do vendors typically provide one monolithic software package?

What are the requirements from this group for each optimization use case?

Some food for thought.

Slice Management: In summary of requirement, there is PNF related information. Could it be copy and paste issue?

Slice (instance) may contain PNFs; e.g. "sliced" RAN necessarily contains PNFs.