Apache Spark is gaining momentum with an in-memory architecture that can outperform the two-stage disk based MapReduce 100 times. However there are ways to significantly boost Hadoop services, specifically the MapReduce performance, by relatively simple means. What I have in mind is adding storage virtualization that removes a lot of the performance bottleneck by moving data to in-memory caches, effectively changing the MapReduce dynamics. DataCore Software provides one such solution. Check it out.