While efficient execution of the data pipeline is prerogative of the task scheduler, which is part of the Spark driver, sometimes Spark needs hints. Spark scheduling is primarily driven by the two parameters: CPU and memory. Other resources, such as disk and network I/O, of course, play an important part in Spark performance as well, but neither Spark, Mesos or YARN can currently do anything to actively manage them.
The first parameter to watch is the number of RDD partitions, which can be specified explicitly when reading the RDD from a file. Spark usually errs on the side of too many partitions as it provides more parallelism, and it does work in many cases as the task setup/teardown times are relatively small. However, one might experiment with decreasing the number of partitions, especially if one does aggregations.
The default number of partitions per RDD and the level of parallelism is determined by the spark.default.parallelism parameter, defined in the $SPARK_HOME...