An autonomous search utility for pervasive storage virtualization

Storage virtualization provides abstraction to pervasive storage graphs. Data management operations require pervasive search utilities to discovery entities and services. In search methods, algorithmic complexity, such as memory bound problems, error convergence issues and supervised training is prohibitive for large state and solution spaces or high dimensional state spaces. In addition, among popular search strategies, heuristic algorithms may not guarantee search optimality or are hard to approximate it without close formed utility functions. Furthermore, model-building algorithms require large computation per iteration since every update needs to compute sums over the entire state space. Hence, dynamic search function and control optimization are major primitives to construct search utilities for stochastic system processes to ensure converged resource accesses. This research focuses on optimization of general search solution methods and proposes a formal search utility framework and algorithms rooted from reinforcement learning (RL) and dynamic programming (DP) techniques. To reduce space complexity within large dimension search spaces, a memory-less Q learning is augmented with self-organized index structure and algorithms for exact state-action value function mapping to optimize search procedures for optimal state actions. Data parallelization is ensured with this page-based index value mapping function. Hence, time complexity is reduced with threaded search parallelism. Convergence analysis and error estimation are presented for numeric and information evaluation. Finally, simulation and learning results are presented and discussed.