On each partition, after the entries have been passed to the aggregator, the accumulation runs in parallel. It means that each aggregator is cloned and receives a sub-set of the entries received from a partition. Then, it runs the accumulation phase in all of the cloned aggregators - at the end, the result is combined into a single accumulation result. It speeds up the processing by at least the factor of two - even in case of simple aggregations. If the accumulation logic is more "heavy", the speed-up may be more significant.
In order to switch the accumulation into a sequential mode just set the hazelcast.aggregation.accumulation.parallel.evaluation property to false (it's set to true by default).