NameNode benchmark using mapred is insufficient

The current namenode benchmark (org.apache.hadoop.examples.NNBench) uses map/reduce to distribute a load on the namenode. For the purposes of loading the namenode, this model gives insufficient control over job start and failure recovery. I propose the namenode benchmark be re-written to use slaves.sh directly to execute the namenode benchmark. The benchmark should also give finer control over the operations executed and the timings reported.