Next: Category 3: More than
Up: Experimental Analysis
Previous: Category 1: Single action
|
Figure: Comparing Polus and ECA for category 2 (multiple candidate actions). In the graph, the throughput goal = 100 MBps |
Analysis: This category (figure 10) exposes the "weak-spot" in Polus. When the candidate actions are indistinguishable, Polus tries them one-by-one till it either leads to a negative impact on the observable values or the goals are met. As shown in the graph, Polus initially selects the wrong action (i.e. prefetching). After the value dips further, Polus tries the next candidate action (i.e. replication). Learning adds the threshold values (in this example at the pre-conditions level) and enables distinguishing between the actions. Insights: In systems with larger action-sets, it is quite possible that Polus never converges due to side-effects of trying wrong actions.
