directory by setting SPARK_HOME and locating the executor relative to
that. Entries on SPARK_CLASSPATH and SPARK_LIBRARY_PATH are also passed
along to worker nodes.

- Use collect instead of toArray
- Disable the "running on Mesos" test when MESOS_HOME is not set

CPU and memory requested per task.

for each split. This replaces the previous method of calling
split.toString, which would produce different results for the same split
each time it is deserialized (because the default implementation returns
the Java object's address).

ensure that the proper keys will be generated when they cached.

caching was not working because Split objects did not have a
consistent toString value

- Don't include trailing $'s in class names of Scala objects
- Report errors using logError instead of printStackTrace

Changed printlns to log statements and fixed a bug in run that was causing it to fail on a Mesos cluster

HTTP server instead of a shared (NFS) file system.

Added fork()/join() operations for SparkContext, as well as corresponding changes to MesosScheduler to support multiple ParallelOperations.