Based on our observation, majority of Spark workloads are not bottlenecked by I/O or network, but rather CPU and memory. This project focuses on 3 areas to improve the efficiency of memory and CPU for Spark applications, to push performance closer to the limits of the underlying hardware.
Memory Management and Binary Processing
- Avoiding non-transient Java objects (store them in binary format), which reduces GC overhead.
- Minimizing memory usage through denser in-memory data format, which means we spill less.
- Better memory accounting (size of bytes) rather than relying on heuristics
- For operators that understand data types (in the case of DataFrames and SQL), work directly against binary format in memory, i.e. have no serialization/deserialization
- Faster sorting and hashing for aggregations, joins, and shuffle
- Faster expression evaluation and DataFrame/SQL operators
- Faster serializer
Several parts of project Tungsten leverage the DataFrame model, which gives us more semantics about the application. We will also retrofit the improvements onto Spark’s RDD API whenever possible.
This epic tracks work items for Spark 1.5. More tickets can be found in:
SPARK-7075: Tungsten-related work in Spark 1.5
SPARK-9697: Tungsten-related work in Spark 1.6
- relates to
SPARK-8159 Improve expression function coverage (Spark 1.5)