Nexmark is a benchmark suite for queries over continuous data streams. This project is inspired by the NEXMark research paper and Apache Beam Nexmark.

These are multiple queries over a three entities model representing on online auction system:

• Person represents a person submitting an item for auction and/or making a bid on an auction.
• Auction represents an item under auction.
• Bid represents a bid for an item under auction.

Note: q1 ~ q8 are from original NEXMark queries, q0 and q9 ~ q13 are from Apache Beam, others are extended to cover more scenarios.

For evaluating the performance, there are two performance measurement terms used in Nexmark that are throughput and cores.

Throughput is the number of events executed by the stream processing system per seconds. As the three entities (Person, Auction, Bid) are generated in a single generator, we measures the total throughput of all events. So that we have a consistent measurement for every queries. For Flink, we collect the <source_operator_name>.numRecordsOutPerSecond metric via Flink Monitoring REST API.

Cores is the CPU usage used by the stream processing system. Usually CPU allows preemption, not like memory can be limited. Therefore, how the stream processing system effectively use CPU resources, how much throughput is contributed per core, they are important aspect for streaming performance benchmark. For Flink, we deploy a CPU usage collector on every worker node and send the usage metric to the benchmark runner for summarizing. We don't use the Status.JVM.CPU.Load metric provided by Flink, because it is not accurate.

The Nexmark benchmark framework runs Flink queries on standalone cluster, see the Flink documentation for more detailed requirements and how to setup it.

The cluster should consist of one master node and one or more worker nodes. All of them should be Linux environment (the CPU monitor script requries to run on Linux). Please make sure you have the following software installed on each node:

• JDK 1.8.x or higher (Nexmark scripts uses some tools of JDK),
• ssh (sshd must be running to use the Flink and Nexmark scripts that manage remote components)

If your cluster does not fulfill these software requirements you will need to install/upgrade it.

Having passwordless SSH and the same directory structure on all your cluster nodes will allow you to use our scripts to control everything.

The following environment variable should be set on every node for the Flink and Nexmark scripts.

• JAVA_HOME: point to the directory of your JDK installation.
• FLINK_HOME: point to the directory of your Flink installation.

Before start to run the benchmark, you should build the Nexmark benchmark first to have a benchmark package. Please make sure you have installed maven in your build machine. And run the ./build.sh command under nexmark-flink directoy. Then you will get the nexmark-flink.tgz archive under the directory.

• Step 1: Download the latest Flink package from the download page. Say flink-<version>-bin-scala_2.11.tgz.
• Step2: Copy the archives (flink-<version>-bin-scala_2.11.tgz, nexmark-flink.tgz) to your master node and extract it.

  tar xzf flink-<version>-bin-scala_2.11.tgz; tar xzf nexmark-flink.tgz
  mv flink-<version> flink; mv nexmark-flink nexmark
  

• Step3: Copy the jars under nexmark/lib to flink/lib which contains the Nexmark source generator.
• Step4: Configure Flink.
  • Edit flink/conf/workers and enters the IP address of each worker node. Recommand to set 8 entries.
  • Replace flink/conf/sql-client-defaults.yaml by nexmark/conf/sql-client-defaults.yaml
  • Replace flink/conf/flink-conf.yaml by nexmark/conf/flink-conf.yaml. Remember to update the following configurations:
    • Set jobmanager.rpc.address to you master IP address
    • Set state.checkpoints.dir to your local file path (recommend to use SSD), e.g. file:///home/username/checkpoint.
    • Set state.backend.rocksdb.localdir to your local file path (recommend to use SSD), e.g. /home/username/rocksdb.
• Step5: Configure Nexmark benchmark.
  • Set nexmark.metric.reporter.host to your master IP address.
• Step6: Copy flink and nexmark to your worker nodes using scp.
• Step7: Setup the benchmark cluster by running nexmark/bin/setup_cluster.sh on the master node.
• Step8: Start Flink Cluster by running flink/bin/start-cluster.sh on the master node.

You can run the Nexmark benchmark by running nexmark/bin/run_query.sh all on the master node. It will run all the queries one by one, and collect benchmark metrics automatically. It will take hours (1h30m) to finish the benchmark by default (6 min for each query). At last, it will print the benchmark summary result (TPS and Cores for each query) on the console.

You can also run specific queries by running nexmark/bin/run_query.sh q1,q2.

You can also tune the workload of the queries by editing nexmark/conf/nexmark.yaml with the nexmark.workload.* prefix options.

• 3 worker node (ecs.i2g.2xlarge instances on Aliyun)
• Each machine has 1 Xeon 2.5 GHz CPU (8 vCores) and 32 GB RAM
• 800 GB SSD local disk
• 2 Gbps between compute nodes

Use the default configuration file flink-conf.yaml and sql-client-defaults.yaml defined in nexmark-flink/src/main/resources/conf/.

Some notable configurations including:

• 8 TaskManagers, each has only 1 slot
• 4GB for each TaskManager and JobManager
• Job parallelism: 8
• Checkpoint enabled with exactly once mode and 3 minutes interval
• Use RocksDB state backend with incremental checkpoint enabled
• MiniBatch optimization enabled with 2 seconds interval and 5000 rows
• Splitting distinct aggregation optimization is enabled

Flink version: manually build for release-1.11 branch on commit b9ca9bb.

Source generates 10M records per seconds. The percentage of 3 stream is Bid: 92%, Auction: 6%, Person: 2%. Each query will warm up for 3 minutes and then collect metrics for 3 minutes.

+-------------------+-------------------+-------------------+-------------------+
| Nexmark Query     | Throughput (r/s)  | Cores             | Throughput/Cores  |
+-------------------+-------------------+-------------------+-------------------+
|q0                 |1.9 M              |8.17               |235 K              |
|q1                 |1.8 M              |8.17               |228 K              |
|q2                 |2.1 M              |8.16               |258 K              |
|q3                 |1.9 M              |9.66               |198 K              |
|q4                 |305 K              |11.55              |26 K               |
|q5                 |311 K              |11.71              |26 K               |
|q7                 |153 K              |12.14              |12 K               |
|q8                 |1.8 M              |13.65              |135 K              |
|q9                 |170 K              |11.86              |14 K               |
|q10                |633 K              |8.23               |76 K               |
|q11                |428 K              |10.5               |40 K               |
|q12                |937 K              |12.35              |75 K               |
|q13                |1.4 M              |8.26               |179 K              |
|q14                |1.8 M              |8.28               |228 K              |
|q15                |729 K              |9.06               |80 K               |
+-------------------+-------------------+-------------------+-------------------+

1. Run Nexmark benchmark for more stream processing systems, such as Spark, KSQL. However, they don't have complete streaming SQL features. Therefore, not all of the queries can be ran in these systems. But we can implement the queries in programing way using Spark Streaming, Kafka Streams.
2. Support Latency metric for the benchmark. Latency measures the required time from a record entering the system to some results produced after some actions performed on the record. However, this is not easy to support for SQL queries unless we modify the queries.

• Pete Tucker, Kristin Tufte, Vassilis Papadimos, David Maier. NEXMark – A Benchmark for Queries over Data Streams. June 2010.