From the MapReduce paper: "The MapReduce master takes the location information of the input files into account and attempts to schedule a map task on a machine

From the MapReduce paper: "The MapReduce master takes the location information of the input files into account and attempts to schedule a map task on a machine that contains a replica of the corresponding input data." One of two possible designs immediately come to mind: (1) The MapReduce master interacts with Borg directly on behalf of the programmer and places a requirement on the task to run on or near any node that has a replica of the data required or (2) for each Borglet, there is an accompanying MapReduce agent running to accept RPCs from the MapReduce master.