MR-plus

The implementation of MR+ is derived from Hadoop MapReduce. However, unlike Hadoop, MR+ enables both map and reduce to interleave, as well as the same key to be reduced by different reduce workers in parallel – permitting multi- level reduces.

In the MR+ implementation, although the role of the JobTracker and Task- Tracker is identical to that in Hadoop, the details of task scheduling are completely different. To achieve interleaving of maps and reduces, MR+ tasks are allocated according to a configurable map to reduce schedule ratio parameter1. The first reduce task is scheduled as soon as a certain number of map tasks complete. The reduces form an inverted tree of reduces over various reduce levels, doing away with the traditional brick wall between map and reduce stages.