Nexmark is a suite of pipelines inspired by the ‘continuous data stream’ queries in Nexmark research paper
These are multiple queries over a three entities model representing on online auction system:
The queries exercise many aspects of Beam model:
We have augmented the original queries with five more:
Here are some of the knobs of the benchmark workload (see NexmarkConfiguration.java).
These configuration items can be passed to the launch command line.
Here is an example output of the Nexmark benchmark run in streaming mode with the SMOKE suite on the (local) direct runner:
The Nexmark launcher accepts the --runner
argument as usual for programs that use Beam PipelineOptions to manage their command line arguments. In addition to this, the necessary dependencies must be configured.
When running via Gradle, the following two parameters control the execution:
-P nexmark.args The command line to pass to the Nexmark main program. -P nexmark.runner The Gradle project name of the runner, such as ":runners:direct-java" or ":runners:flink:1.8. The project names can be found in the root `settings.gradle`.
Test data is deterministically synthesized on demand. The test data may be synthesized in the same pipeline as the query itself, or may be published to Pub/Sub or Kafka.
The query results may be:
Decide if batch or streaming:
--streaming=true
Number of events generators:
--numEventGenerators=4
Queries can be run by their name or by their number (number is still there for backward compatibility, only the queries 0 to 12 have a number)
Run query N:
--query=N
Run query called PASSTHROUGH:
--query=PASSTHROUGH
The suite to run can be chosen using this configuration parameter:
--suite=SUITE
Available suites are:
--manageResources=false --monitorJobs=false
--manageResources=false --monitorJobs=true \ --enforceEncodability=false --enforceImmutability=false --project=<your project> \ --zone=<your zone> \ --workerMachineType=n1-highmem-8 \ --stagingLocation=gs://<a gs path for staging> \ --runner=DataflowRunner \ --tempLocation=gs://<a gs path for temporary files> \ --filesToStage=target/beam-sdks-java-nexmark-{{ site.release_latest }}.jar
--manageResources=false --monitorJobs=true \ --enforceEncodability=false --enforceImmutability=false
--manageResources=false --monitorJobs=true \ --flinkMaster=[local] --parallelism=#numcores
--manageResources=false --monitorJobs=true \ --sparkMaster=local \ -Dspark.ui.enabled=false -DSPARK_LOCAL_IP=localhost -Dsun.io.serialization.extendedDebugInfo=true
Set Kafka host/ip (for example, “localhost:9092”):
--bootstrapServers=<kafka host/ip>
Write results into Kafka topic:
--sinkType=KAFKA
Set topic name which will be used for benchmark results:
--kafkaResultsTopic=<topic name>
Write or/and read events into/from Kafka topic:
--sourceType=KAFKA
Set topic name which will be used for benchmark events:
--kafkaTopic=<topic name>
These tables contain statuses of the queries runs in the different runners. Google Cloud Dataflow and Apache Gearpump statuses are yet to come.
Yet to come
Yet to come
The DirectRunner is default, so it is not required to pass -Pnexmark.runner
. Here we do it for maximum clarity.
The direct runner does not have separate batch and streaming modes, but the Nexmark launch does.
These parameters leave on many of the DirectRunner's extra safety checks so the SMOKE suite can make sure there is nothing broken in the Nexmark suite.
Batch Mode:
./gradlew :sdks:java:testing:nexmark:run \ -Pnexmark.runner=":runners:direct-java" \ -Pnexmark.args=" --runner=DirectRunner --streaming=false --suite=SMOKE --manageResources=false --monitorJobs=true --enforceEncodability=true --enforceImmutability=true"
Streaming Mode:
./gradlew :sdks:java:testing:nexmark:run \ -Pnexmark.runner=":runners:direct-java" \ -Pnexmark.args=" --runner=DirectRunner --streaming=true --suite=SMOKE --manageResources=false --monitorJobs=true --enforceEncodability=true --enforceImmutability=true"
The SparkRunner is special-cased in the Nexmark gradle launch. The task will provide the version of Spark that the SparkRunner is built against, and configure logging.
Batch Mode:
./gradlew :sdks:java:testing:nexmark:run \ -Pnexmark.runner=":runners:spark" \ -Pnexmark.args=" --runner=SparkRunner --suite=SMOKE --streamTimeout=60 --streaming=false --manageResources=false --monitorJobs=true"
Streaming Mode:
./gradlew :sdks:java:testing:nexmark:run \ -Pnexmark.runner=":runners:spark" \ -Pnexmark.args=" --runner=SparkRunner --suite=SMOKE --streamTimeout=60 --streaming=true --manageResources=false --monitorJobs=true"
Batch Mode:
./gradlew :sdks:java:testing:nexmark:run \ -Pnexmark.runner=":runners:flink:1.8" \ -Pnexmark.args=" --runner=FlinkRunner --suite=SMOKE --streamTimeout=60 --streaming=false --manageResources=false --monitorJobs=true --flinkMaster=[local]"
Streaming Mode:
./gradlew :sdks:java:testing:nexmark:run \ -Pnexmark.runner=":runners:flink:1.8" \ -Pnexmark.args=" --runner=FlinkRunner --suite=SMOKE --streamTimeout=60 --streaming=true --manageResources=false --monitorJobs=true --flinkMaster=[local]"
Batch Mode:
./gradlew :sdks:java:testing:nexmark:run \ -Pnexmark.runner=":runners:apex" \ -Pnexmark.args=" --runner=ApexRunner --suite=SMOKE --streamTimeout=60 --streaming=false --manageResources=false --monitorJobs=true"
Streaming Mode:
./gradlew :sdks:java:testing:nexmark:run \ -Pnexmark.runner=":runners:apex" \ -Pnexmark.args=" --runner=ApexRunner --suite=SMOKE --streamTimeout=60 --streaming=true --manageResources=false --monitorJobs=true"
Set these up first so the below command is valid
PROJECT=<your project> ZONE=<your zone> STAGING_LOCATION=gs://<a GCS path for staging> PUBSUB_TOPCI=<existing pubsub topic>
Launch:
./gradlew :sdks:java:testing:nexmark:run \ -Pnexmark.runner=":runners:google-cloud-dataflow-java" \ -Pnexmark.args=" --runner=DataflowRunner --suite=SMOKE --streamTimeout=60 --streaming=true --manageResources=false --monitorJobs=true --project=${PROJECT} --zone=${ZONE} --workerMachineType=n1-highmem-8 --stagingLocation=${STAGING_LOCATION} --sourceType=PUBSUB --pubSubMode=PUBLISH_ONLY --pubsubTopic=${PUBSUB_TOPIC} --resourceNameMode=VERBATIM --manageResources=false --numEventGenerators=64 --numWorkers=16 --maxNumWorkers=16 --firstEventRate=100000 --nextEventRate=100000 --ratePeriodSec=3600 --isRateLimited=true --avgPersonByteSize=500 --avgAuctionByteSize=500 --avgBidByteSize=500 --probDelayedEvent=0.000001 --occasionalDelaySec=3600 --numEvents=0 --useWallclockEventTime=true --usePubsubPublishTime=true --experiments=enable_custom_pubsub_sink"
Building package:
./gradlew :sdks:java:testing:nexmark:assemble
Submit to the cluster:
spark-submit \ --class org.apache.beam.sdk.nexmark.Main \ --master yarn-client \ --driver-memory 512m \ --executor-memory 512m \ --executor-cores 1 \ sdks/java/testing/nexmark/build/libs/beam-sdks-java-nexmark-{{ site.release_latest }}-spark.jar \ --runner=SparkRunner \ --query=0 \ --streamTimeout=60 \ --streaming=false \ --manageResources=false \ --monitorJobs=true"
Below dashboards are used as a CI mechanism to detect no-regression on the Beam components. They are not supposed to be benchmark comparision of the runners or engines. Especially because:
At each commit on master, Nexmark suites are run and plots are created on the graphs.
There are 2 kinds of dashboards:
There are dashboards for these runners (others to come):
Each dashboard contains:
Nexmark performance direct runner
Nexmark performance flink runner
Nexmark performance spark runner
Nexmark performance dataflow runner
Nexmark output size direct runner
Nexmark output size flink runner