Experimental Setup Polus Framework Base-level reasoning

Meta-level operations

Input: A set of candidate actions, and $\phi$generated by the problem determination module
Output: The actual set of actions to be invoked
Approach:
The aim is to select a candidate action, using an optimization function based on the following parameters:

• The transient overhead associated with the action invocation. For example, invoking replication has more overhead compared to invoking prefetching.
• The behavior side-effects of the action (i.e., an action, in addition to improving the violated goals, can possibly have a negative impact on other behavior dimensions). For example, invoking data backup improves availability, but it has a negative side-effect on throughput and latency.
• The confidence level associated with the details of the action object. As mentioned earlier, the confidence level is associated with the learning function error, measured as a difference between the predicted and the actual observed values.

An important requirement for this optimization is that it should not be based on short term goals (e.g., invoking replication has high overheads, but it might be beneficial in the long-run compared to invoking prefetching multiple times). In Polus, the optimization algorithm is using n-step look ahead [20], which is extensively used in game theory. In n-step look ahead, the system simulates the impact of pursuing different options. The simulation is repeated $n$times, and the end result of the simulation is used to decide on the option to be selected in the current state. The simulation is based on the implication and precondition information of the action object. The n-step look ahead is just a rough estimate for the optimization, because external factors such as changes in workload or resources may render the predictions inaccurate. But it definitely helps in detecting instabilities arising due to cycles in state transitions (i.e. the system ping-pongs between two states, invoking the same set of actions repeatedly). It also helps in avoiding choices where a single action leads to a series of actions being invoked due to the cumulative side effects of actions. Finally, in addition to considering the candidate actions independently, it is possible to reason with combinations of actions. For example, instead of considering the choices of invoking either prefetching or replication, it is possible to consider a combination of prefetching and replication as an additional candidate choice. Composite actions can be handled using vector arithmetic addition techniques but the description of these techniques is beyond the scope of this paper.

Experimental Setup Polus Framework Base-level reasoning