Category 4: Recurrent action Experimental Analysis Category 2: Multiple actions

Category 3: More than one goal not met

Figure: Comparing Polus and ECA for category 3 (more than one goal not met). In the graph, the latency goal = 4.5 msec

Analysis: Rule-based systems will invoke a single action in each iteration without analyzing the combined impact of the actions. On the other hand, Polus considers different permutations to combine the actions (figure 11). This is beneficial when the two actions act on the same resources, such that the invocation of one action beyond a threshold can violate the pre-conditions of other actions. As shown in the graph, ECA does not meet the latency goal due to lack of memory resources. In its previous iteration Prefetching was invoked for throughput goals and the rules did not consider the combined state while deciding the value for prefetching. Learning refines the attributes of actions allowing better combination strategies. Insights: Higher-order operation are powerful in deriving permutations that cannot be possibly defined statically.

Category 4: Recurrent action Experimental Analysis Category 2: Multiple actions