Optimal load allocation algorithm for large distributed systems

This master thesis presents, a family of algorithms for load allocation in a large scale computing infrastructures, using the concept of resource saturation in a multi-class environment. New request admission policies and routing algorithms have been developed to im- prove the performance metrics based on the concept of bottleneck in a network. The underlying idea is to selectively admit in the system a job of class that maximize the utilization of the resources unused by the job in execution. The algorithm have been coded and simulated, in JMT, 'Java Modelling Tool' a suite of applications developed by Politecnico di Milano and released under GPL license, to evaluate the gain achieved when the new admission policy is used. For modeling a large computing network, the technique of Queuing Network are used where a Queuing Station represents any given Computing Resource with a given service time, capable of serving user requests. Several simulation experiments were performed on the above model for proper calibration of the admission policy to per- form at optimum level. Results are in agreement with the theoretical predictions and show a significant im- provement over prevailing static routing algorithms. The work presented also has several implications for future studies of Green IT since it can also help solving the problem of energy consumption in large computing centers.

La tesi presenta un nuovo algoritmo per la ripartizione del carico in sistemi o reti di sistemi informatici avanti dimensioni molto elevate. Il concetto di base quello di valuatare i livelli di contesa delle risorse ed in base alla. Gli algorithmi sono stati codificati e simulati, in JMT, Java Modelling Tool, una suite di applicazioni sviluppato dal Politecnico di Milano e rilasciata sotto licenza GPL, per vantaggi, ottenuti, attraverso la nuova politica di ammissione. I risultati sono ulteriormente discussi nella tesi. I risultati mostrano un accordo parziale con i risultati teorici, ed un miglioramento significativo rispetto agli algoritmi di routing statico. Il lavoro qui presentato ha anche profonde implicazioni per studi futuri di Green IT e possono anche contribuire ad affrontare il problema del consumo di energia in grandi centri di calcolo.