A more general solution is to have the operating system profile the applications online and use the locally
collected information to make resource decisions. While
the operating system could potentially run the applications over a set of configurations when the application is
first installed to build the models, it may be more practical to simply start with a generic model which is refined
over time with on-line training methods.
The source of the data used to create the models is
somewhat orthogonal to the type of models used. Adopting a particular model type in turn influences the specific
method used for searching over all models to find the
best resource distribution across all Cells. We are currently interested in two types of models. Models of the
first type are based on a discrete, sparse set of predetermined operational points