Apache Kafka benchmarks

This tutorial shows you how to run OpenMessaging benchmarks for Apache Kafka. You can currently deploy to the following platforms:

Initial setup

To being, you’ll need to clone the benchmark repo from the openmessaging organization on GitHub:

$ git clone https://github.com/openmessaging/openmessaging-benchmark
$ cd openmessaging-benchmark

You’ll also need to have Maven installed.

Create local artifacts

The current Kafka client version is 2.8.1. In order to downgrade it edit the ./driver-kafka/pom.xml file.

Once you have the repo cloned locally, you can create all the artifacts necessary to run the benchmarks with a single Maven command:

$ mvn install

Deploy a Kafka cluster on Amazon Web Services

You can deploy a Kafka cluster on AWS (for benchmarking purposes) using Terraform and Ansible. You’ll need to have both of those tools installed as well as the terraform-inventory plugin for Terraform.

In addition, you’ll need to:

SSH keys

Once you’re all set up with AWS and have the necessary tools installed locally, you’ll need to create both a public and a private SSH key at ~/.ssh/kafka_aws (private) and ~/.ssh/kafka_aws.pub (public), respectively.

$ ssh-keygen -f ~/.ssh/kafka_aws

When prompted to enter a passphrase, simply hit Enter twice. Then, make sure that the keys have been created:

$ ls ~/.ssh/kafka_aws*

Create resources using Terraform

With SSH keys in place, you can create the necessary AWS resources using just a few Terraform commands:

$ cd driver-kafka/deploy
$ terraform init
$ terraform apply

This will install the following EC2 instances (plus some other resources, such as a Virtual Private Cloud (VPC)):

Resource Description Count
Kafka instances The VMs on which a Kafka broker will run 3
ZooKeeper instances The VMs on which a ZooKeeper node will run 3
Client instance The VM from which the benchmarking suite itself will be run 4

When you run terraform apply, you will be prompted to type yes. Type yes to continue with the installation or anything else to quit.

Once the installation is complete, you will see a confirmation message listing the resources that have been installed.

Variables

There’s a handful of configurable parameters related to the Terraform deployment that you can alter by modifying the defaults in the terraform.tfvars file.

Variable Description Default
region The AWS region in which the Kafka cluster will be deployed us-west-2
public_key_path The path to the SSH public key that you’ve generated ~/.ssh/kafka_aws.pub
ami The Amazon Machine Image (AWI) to be used by the cluster’s machines ami-9fa343e7
instance_types The EC2 instance types used by the various components i3.4xlarge (Kafka brokers), t2.small (ZooKeeper), c4.8xlarge (benchmarking client)

If you modify the public_key_path, make sure that you point to the appropriate SSH key path when running the Ansible playbook.

Running the Ansible playbook

With the appropriate infrastructure in place, you can install and start the Kafka cluster using Ansible with just one command:

$ ansible-playbook \
  --user ec2-user \
  --inventory `which terraform-inventory` \
  deploy.yaml

If you’re using an SSH private key path different from ~/.ssh/kafka_aws, you can specify that path using the --private-key flag, for example --private-key=~/.ssh/my_key.

SSHing into the client host

In the output produced by Terraform, there’s a client_ssh_host variable that provides the IP address for the client EC2 host from which benchmarks can be run. You can SSH into that host using this command:

$ ssh -i ~/.ssh/kafka_aws ec2-user@$(terraform output client_ssh_host)

Running the benchmarks from the client hosts

The benchmark scripts can be run from the /opt/benchmark working directory.

Once you’ve successfully SSHed into the client host, you can run any of the existing benchmarking workloads by specifying the YAML file for that workload when running the benchmark executable. All workloads are in the workloads folder. Here’s an example:

$ sudo bin/benchmark \
  --drivers driver-kafka/kafka.yaml \
  workloads/1-topic-16-partitions-1kb.yaml

Although benchmarks are run from a specific client host, the benchmarks are run in distributed mode, across multiple client hosts.

There are multiple Kafka “modes” for which you can run benchmarks. Each mode has its own YAML configuration file in the driver-kafka folder.

Mode Description Config file
Standard Kafka with message idempotence disabled (at-least-once semantics) kafka.yaml
Exactly once Kafka with message idempotence enabled (“exactly-once” semantics) kafka-exactly-once.yaml
Sync Kafka with durability enabled (all published messages synced to disk) kafka-sync.yaml

The example used the “standard” mode as configured in driver-kafka/kafka.yaml. Here are some examples for the other modes:

# Exactly once
$ sudo bin/benchmark \
  --drivers driver-kafka/kafka-exactly-once.yaml \
  workloads/1-topic-16-partitions-1kb.yaml

# Sync
$ sudo bin/benchmark \
  --drivers driver-kafka/kafka-sync.yaml \
  workloads/1-topic-16-partitions-1kb.yaml

Specify client hosts

By default, benchmarks will be run from the set of hosts created by Terraform. You can also specify a comma-separated list of client hosts using the --workers flag (or -w for short):

$ sudo bin/benchmark \
  --drivers driver-kafka/kafka-exactly-once.yaml \
  --workers 1.2.3.4:8080,4.5.6.7:8080 \ # or -w 1.2.3.4:8080,4.5.6.7:8080
  workloads/1-topic-16-partitions-1kb.yaml

Downloading your benchmarking results

The OpenMessaging benchmarking suite stores results in JSON files in the /opt/benchmark folder on the client host from which the benchmarks are run. You can download those results files onto your local machine using scp. You can download all generated JSON results files using this command:

$ scp -i ~/.ssh/kafka_aws ec2-user@$(terraform output client_ssh_host):/opt/benchmark/*.json .

Tearing down your benchmarking infrastructure

Once you’re finished running your benchmarks, you should tear down the AWS infrastructure you deployed for the sake of saving costs. You can do that with one command:

$ terraform destroy -force

Make sure to let the process run to completion (it could take several minutes). Once the tear down is complete, all AWS resources that you created for the Kafka benchmarking suite will have been removed.