Mastering Spark with R

Chapter 9. Tuning

Chaos isn’t a pit. Chaos is a ladder.

—Petyr Baelish

In previous chapters, we’ve assumed that computation within a Spark cluster works efficiently. While this is true in some cases, it is often necessary to have some knowledge of the operations Spark runs internally to fine-tune configuration settings that will make computations run efficiently. This chapter explains how Spark computes data over large datasets and provides details on how to optimize its operations.

For instance, in this chapter you’ll learn how to request more compute nodes and increase the amount of memory, which, if you remember from Chapter 2, defaults to only 2 GB in local instances. You will learn how Spark unifies computation through partitioning, shuffling, and caching. As mentioned a few chapters back, this is the last chapter describing the internals of Spark; after you complete this chapter, we believe that you will have the intermediate Spark skills necessary to be productive at using Spark.

In Chapters 10–12 we explore exciting techniques to deal with specific modeling, scaling, and computation problems. However, we must first understand how Spark performs internal computations, what pieces we can control, and why.

Overview

Spark performs distributed computation by configuring, partitioning, executing, shuffling, caching, and serializing data, tasks, and resources across multiple machines:

• Configuring requests the cluster manager for resources: total machines, memory, and so on. ...