Spark Benchmarks is a benchmarking suite specific for Apache Spark that helps to evaluate a Spark deployment in terms of speed, throughput and system resource utilization. It comprises a representative set of Spark workloads including DFSIO, TeraSort, etc.
There already exists other benchmarks suites in the community that helps to evaluate different big data frameworks. The more representative one is HiBench which contains a set of Hadoop, Spark and streaming workloads suited for benchmarking different use cases: sorting, machine learning algorithms, web searches, graphs and so on.
However, not all workloads are implemented using only Spark jobs and rely on Hadoop MapReduce framework assuming Spark is running on top of a YARN cluster. Concretely, DFSIO benchmark, that tests the throughput of a HDFS cluster by generating a large number of tasks performing writes and reads simultaneously, does not have a Spark corresponding implementation.
The purpose of this suite is to help users to stress different scenarios of Spark combined with a distributed file system (HDFS, Alluxio, etc), regardless of whether it runs on Mesos, YARN or Spark Standalone. Moreover, it enables an exhaustive study and comparision for different platform and hardware setups, sizing tuning and system optimizations, making easier the evaluation of their performance implications and the identification of bottlenecks.
Currently, there is only one workload available:
Please visit the documentation associated to the corresponding workload.
The followings are needed for building Spark Benchmarks
To build all modules in Spark Benchmarks, use the below command.
sbt clean assemblyIf you are only interested in a single workload you can build a single module. For example, the below command only builds the dfsio workload.
sbt dfsio/clean dfsio/assemblyContributions are very welcome, see CONTRIBUTING.md or skim existing tickets to see where you could help out.
Spark Benchmarks is Open Source and available under the Apache 2 License.