# The Spark tunable that gave me 8X speedup

There are many configuration tunables in Spark. However, if you have time for only one, set this one. It made a streaming application we run process data 8X faster. That’s 800% improvement, no code change needed!

### The tunables

Here are the tunables:

spark.sql.shuffle.partitions


and its sister:

spark.default.parallelism


The first setting applies to dataframes. The second setting applies to RDDs, which show up when you do calls like map() on dataframes.

By default, both are set to 200. This can be a lot, unless it is a production cluster with 20 nodes and 10 cores each!

Here is how I set this value on my Spark cluster managed by YARN. You need to know:

• how many worker nodes you have in the cluster, say W
• how many cores you have on each worker node, say C

The available parallelism is W * C and that is the number you want to use. This is why I said “20” and “10” above. Some people make an “allowance” of 2X on top of this, but I’m not convinced about it yet.

You can set these either in spark-defaults.conf (preferred, helps everyone on the cluster), or within your application, while constructing SparkConf.

### What happens if you don’t set this value?

There are two problems.

1. Your Spark (batch) application will take longer to finish, because Spark can only run W * C tasks at any given time. The other tasks have to wait.

2. Each task may get only a very small chunk of data, and it takes longer to set it up and run it, than for it to do any data munching.

It is very easy to see this when you look at the timeline view of a job under Spark’s UI. You can also see this when Spark reports “200 tasks” in any stage.