Conclusion Polus : Growing Storage Discussion

Related Work

Rome [25], Minerva [1], Hippodrome [2], and ``attribute-managed storage'' [13] projects from HP, SELF* project [9] from Carnegie Mellon, Storage Tank [21], SledRunner [6] projects from IBM, Control Centre product line from EMC, Storage Central product line from Veritas, and BrightStor product line from Computer Associates all aim to simplify storage management by automating different aspects of storage management. The Polus framework presented in this paper is complementary to these projects, since none of these projects specifically address the QoS goal transformation problem being addressed in this paper. Moreover, there is nothing inherent in the Polus framework that prevents its adoption by these different frameworks and products as part of their QoS solutions. The Polus framework was built using specification, learning and reasoning techniques from the artificial intelligence (AI) domain. These technologies have a proven track record as they have been successfully used to build expert systems in medical, system configuration, video games and speech/handwriting processing application domains. To the best of our knowledge, Polus is the only system of its kind (in the domain of storage performance management) that integrates a rules-of-thumb specification model, reasoning (including higher-order operations) and a self-refining learning engine to manage a storage system. Polus leverages concepts in AI and uses them as building blocks in its solution. Techniques for specification in expert systems are broadly classified as imperative (e.g. rule-based), declarative (e.g. logic programming) or mixed. Brittleness has been identified as the biggest drawback of imperative rule-based systems [7], whereas, logic based systems overcome this problem by using a reasoning engine to combine facts/beliefs in the knowledge base to draw conclusions. The Polus specification of action attributes is similar to the declarative approach. Further, reasoning in Polus is a combination of specification search algorithms and higher-order operations. Polus uses forward chaining to search the specifications, but it is possible to use other approaches such as backward chaining or heuristic-based searching. Other popular approaches for reasoning are: Model-based, Constraint-based, and Case-based reasoning [17]. As explained earlier, Polus uses CBR as part of the reasoning engine to tie in the knowledge acquired by learning in the decision-making. Finally, learning in Polus systematically refines the specifications. It leverages research in the domain of machine learning algorithms such as neural networks and reinforcement learning [11,15]. Currently, there are many competing policy specification standards [14,19]. Polus can leverage any one these existing standards for specifying the base rules-of-thumb. Furthermore, there are no in-built dependencies that prevent Polus from leveraging the canonical SNIA SMI-S storage device standard [23] as the representation for low level system actions. A Case-Based Reasoning approach, in which a system starts off with no specifications and uses the previously learnt cases to decide how a goal should be transformed, has been employed in the web-server configuration domain [24]. The bootstrapping behavior of that approach is not attractive in real-world scenarios where the reasonable number of cases that need to be learned a priori are zero (resource states, workload characteristics, goals, action set). In comparison, as shown in the experiment section, the combination of rule-of-thumb specification and a learning engine has a reasonable bootstrapping behavior. That is, the Polus approach is able to dynamically adapt even when it does not start from the most desired bootstrapped state. Mark, et al., [3] propose an approach to separate the goal from the base rule specification. They create a mapping between the rule and user-requirements, making it easy for validation and usage. The Polus approach is more sophisticated, in that it encodes the goal implications and uses them to automate the reasoning process. Another approach [8] uses genetic algorithms for self-tuning. In this approach each system parameter is tuned by an individual algorithm and the genetic algorithm decides the best combination of algorithms. Unlike Polus, this approach does not allow refinement of the decision-making based on learning. Zinky, et al., [27] present a general framework, called QuO, to implement QoS-enabled distributed object systems. The QoS adaptation is achieved by having multiple implementations. Each implementation is mapped to an environment and to a QoS region. The QuO approach is static, as it does not implement semantics for reasoning about the various possible configurations.

Conclusion Polus : Growing Storage Discussion