• Algorithimical skeleton
• Machine learning
• Predictive model
• Macro-dataflow
• TreeESN
• Structured domain

The thesis tries to investigate on how a machine learning tool can be used to achieve performance prediction in the algorithmical skeleton context. In the dissertation, an extension of the Echo State Network (ESN) able to deal with tree structured data (TreeESN) is examined to build a predictive model.

In the thesis has been realized:
* A general library for the automatic learning with TreeESN developed in C++ using the BLAS/LAPACK libraries.
* Two different parallel implementations of the model selection process targeting the multicore architecture has been developed using the FastFlow framework. They rely on the streaming oriented parallelism using respectively the farm and the macro-dataflow parallel patterns.
* Some experimental results on the application of the TreeESN tool to achieve a performance prediction model using structured parallel program (skeleton trees). The predictive model has been created by carrying out the whole design cycle typical of the machine learning.