Description Karsten Thoms 2019-03-14 03:50:22 EDT

The performance of ExpressionMatchQuery#perform() can be improved when the iteration is parallelized by using Spliterator. The performance gain can be observed when the iteration occurs over a high number of elements.

The examined use case is the CBI Aggregator for simrel.aggr and performing the "Validate Aggregation" action. The perform() method is invoked for a collection of 1077 InstallableUnits in that case.

When executing the validation action, ~12s is consumed in ExpressionMatchQuery#perform(). This is called quite frequently and often with larger collections. In the concrete case the method is sometimes called on a collection of 1077 IUs.

I tried to use a Spliterator with parallel execution and this shows a performance gain. The time used in the method is ~1s less. More importantly, this time is spent on multiple threads. The calling method is SimplePlanner#getSolutionFor(), whose execution time drops by 6s from 70s to 64.5s.

The profile shows half of the time is cut off for some methods:
- Unary#evaluate:  8.1s -> 4.1s
- Equals#evaluate: 6.1s -> 3.6s
- Member$DynamicMember#invoke: 5.9s->2.8s