Related Work Polus : Growing Storage Category 5: Negation of

Discussion

During the evaluation of the Polus framework, we made a few simplifying assumptions.

• The system state is assumed to remain constant during the time that an action is invoked. While it may be argued that this is an unrealistic assumption for actions that require significant time to complete, it is a reasonable one for most actions that involve small overhead.
• We do not consider the case where the specification provided by a administrator may be incomplete or contains incorrect information that may mislead the management software in its decision making.
• The experimental evaluation contains a single implementation of learning, reasoning and base invocation algorithms. These algorithms have not been fine tuned or optimized, and it is possible to plug-in more sophisticated algorithms for tree searching, incremental base invocation and learning. An evaluation by varying these algorithms is beyond the scope of this paper.

In real life deployments of Polus, system vendors can provide templates containing rules of thumb specifications and initial threshold values (obtained from training runs) for different workloads and system configurations. The adaptive behavior of the Polus framework will fine tune the knowledge base (i.e. threshold values) and tailor it to specific user environment and workload characteristics. The evaluation framework presented in the paper is a system with just four possible actions. Hence, it is important to note that our aim was to understand the possible weaknesses of the Polus approach, and to get a first cut estimate of the number of iterations required in converging towards a specified QoS goal for different system states. As shown in the experiment section, the initial results show that this approach is promising; thus, we are currently implementing Polus as part of a real storage QoS management system. Polus can also be initially deployed as an aid to system administrators that allows them to perform what if analysis with respect to whether a system can support different QoS goals.

Related Work Polus : Growing Storage Category 5: Negation of