We used a total of 92 nodes (92 cores with 2GB of RAM,
scattered in several physical nodes): 80 mappers (16 different
map tasks replicated 5 times) and 12 reducers (4 different
reduce tasks replicated 3 times). To simulate Internet connection
bandwidths, all nodes had their download and upload
bandwidths limited to 100Mbps and 10Mbps respectively
(except for the experiment in Figure 8). All the files
(16 input files, 16 intermediate outputs and 4 final outputs)
were different with 64MB each (except for the experiments
in Figures 7 and 8).