A benchmark for generic, large-scale shuffle operations on continuous stream of data, implemented with state-of-the-art stream processing frameworks.
Currently, we provide implementations for the following frameworks:
- Apache Flink
- Apache Spark (Structured Streaming)
- Apache Kafka Streams
- Hazlecast (with its Jet engine)
Additionally, a load generator and a tool for measuring and exporting the latency are provided.
The most straightforward way to run experiments with ShuffleBench is to use the Theodolite benchmarking framework. This allows you to run experiments on Kubernetes clusters in a fully automated, reproducible way including setting up stream processing application, starting the load generator, measuring performance metrics, and collecting the results.
Theodolite benchmark specifications for ShuffleBench can be found in kubernetes. There, you can also find detailed instructions on how to run the benchmarks.
To engage at a lower level, you can also run the benchmark implementations and the load generator manually using the Kubernetes manifests in kubernetes or run the provided container images or the Java applications directly.
Gradle is used to build, test, and package the benchmark implementations, the load generator, and the latency exporter tool. To build all subprojects, run:
./gradlew buildExcept the Shufflebench implementations for Spark, all implementations can be packaged as container images and pushed to a registry using Jib by running:
ORG_GRADLE_PROJECT_imageRepository=<your.registry.com>/shufflebench ./gradlew jibFor Spark, we have to build and push the image manually (e.g., using the Docker deamon):
docker build -t <your.registry.com>/shufflebench/shufflebench-spark shuffle-spark/
docker push <your.registry.com>/shufflebench/shufflebench-sparkIf you use ShuffleBench in your research, please cite:
Sören Henning, Adriano Vogel, Michael Leichtfried, Otmar Ertl, and Rick Rabiser. 2024. ShuffleBench: A Benchmark for Large-Scale Data Shuffling Operations with Distributed Stream Processing Frameworks. In Proceedings of the 15th ACM/SPEC International Conference on Performance Engineering (ICPE '24). Association for Computing Machinery, New York, NY, USA, 2–13. DOI: 10.1145/3629526.3645036