Vinsamlegast notið þetta auðkenni þegar þið vitnið til verksins eða tengið í það: http://hdl.handle.net/1946/33361
This thesis has two goals: The first is to design and implement software that is able to simulate (large) parallel server systems in order to analyze the performance of different scheduling and dispatching policy implementations taking place in eg. cloud computers. The second is to apply reinforcement learning practices to the dispatching problem in order to obtain (near) optimal policies for given systems. The first goal is achieved by using Python to develop modular simulation software. The software is written in an object-oriented way, such that each class in the system represents some natural entity in the (large) parallel server system, aside from a central class (named “Global”) that facilitates the processing of events. Each part can be implemented in different ways to allow for a ’plug-and-play’ implementation for policies and schedulers. The second goal is achieved by formulating the problem of optimal dispatching as a Markov decision process (MDP). In the MDP framework, we can apply, eg., policy iteration and reinforcement learning (RL) like Temporal difference learning.
|Yfirlýsing.pdf||398.42 kB||Lokaður||Yfirlýsing um meðferð verkefnis|