Overview

Note: The Workload Services Framework is a benchmarking framework and is not intended to be used for the deployment of workloads in production environments. It is recommended that users consider any adjustments which may be necessary for the deployment of these workloads in a production environment including those necessary for implementing software best practices for workload scalability and security.

Introduction

Apache Kafka is a framework implementation of a software bus using stream-processing. It is an open-source software platform developed by the Apache Software Foundation written in Scala and Java. The project aims to provide a unified, high-throughput, low-latency platform for handling real-time data feeds.

The example use case for Kafka is for operational-related tasks such as application logs collection, and event streaming from one system/framework/platform to another.

Test Case

This workload is measuring Apache Kafka's performance by utilizing the built-in application bundled with Apache Kafka. Currently, the test case measures Apache Kafka producer and consumer performance.

• run_test.sh - Run test
  1. This script collects environment variables and invokes start_test.py to run test cases

• start_test.py - Start test

  1. This script generates a set number of producers and consumers processes according to producer and consumer number.
  2. Each producer process invokes script kafka-producer-perf-test.sh which will generate 5 million messages to the Kafka server. (By default, throughput=-1, record-size=1000, num-records=5000000, compression.type=lz4, buffer.memory=33554432, batch.size=65536, linger.ms=100)
  3. Each consumer process invokes script kafka-consumer-perf-test.sh which will read 10 million messages from Kafka server. (By default, messages=10000000, timeout=600000)

• kafka_jdk*_1n_gated - Gated test case, for this test case, PARTITIONS, PRODUCERS and CONSUMERS will be set to 1, and cannot be changed.

• kafka_jdk*_1n - Used for single node testing, all pods will be deployed on one K8S worker node.
• kafka_jdk*_multinode - Used for multi node testing, need at least 3 K8S worker nodes for this test case.
• kafka_jdk*_multinode_pkm - Used for multi node testing, need at least 3 K8S worker nodes for this test case.
• support jdk8, jdk11, jdk17 different openjdk/corretto versions.

Quick start

• How to run.
• Step 1: Clone code
• Step 2: Switch to code root path and run command "mkdir build".
• Step 3: Go to path "../script/terraform" and edit file terraform-congig.static|aws|gcp|...|.tf to set your node information.
• Step 4: Go to path "../build" and Run below command to configure the ctest.
  "cmake -DPLATFORM=SPR -DBACKEND=terraform -DCUMULUS_SUT=static -DREGISTRY="your_registry_IP:port" -DRELEASE=:latest ..".
  According to your test bed, 'DPLATFORM' can be changed to ICX. 'DCUMULUS_SUT' can be changed to aws or azure.
• Step 5: Switch to path ../build/workload/Kafka and run "make" command.
• Step 6: Run below command to list all test cases.
  "./ctest.sh -N"
• Step 7: Run test case by below command.
  "./ctest.sh -R test_static_kafka_jdk17_multinode"
  "_1n" means running on single nodes. "_multinode" means running on multiple nodes.

Customize Test Configurations

Parameters for workload configure:

Kafka Settings

Parameters	Default	Description
BROKER_SERVER_NUM	1	Number of broker server.
REPLICATION_FACTOR	1	Number of backups per partition.
KAFKA_BENCHMARK_TOPIC	KAFKABENCHMARK	Kafka topic name.
MESSAGES	10000000	Number of messages to consume.
NUM_RECORDS	5000000	Number of records to produce.
THROUGHPUT	-1	Throughput (records/second) the producer should achieve.
RECORD_SIZE	1000	Size of each record.
COMPRESSION_TYPE	lz4	Compression type
CONSUMER_TIMEOUT	600000	Time the consumer will wait for messages (in milliseconds).
BUFFER_MEM	33554432	Amount of memory available to the producer for buffering.
BATCH_SIZE	65536	Producer batch size.
LINGER_MS	100	Time in milliseconds that producer will wait to send batch.
ACKS	1	Number of response.
FETCH_SIZE	1048576	Size of bytes fetched per request.
ENCRYPTION	false	Use SSL for encryption.
PAYLOAD_NUM	10000	Number of random payloads in message.
NUM_NETWORK_THREADS	0	Number of threads used to process the network.
SERVER_PROTECTION	true	Limit RECORDS,PRODUCERS based on memory,disk and network.
NUM_REPLICA_FETCHERS	1	Number of fetcher threads used to replicate records.
REPLICA_FETCH_MAX_BYTES	1048576	Number of bytes of messages to fetch for each partition.
REPLICA_SOCKET_RECEIVE_BUFFER_BYTES	65536	Socket receive buffer for network requests to the leader.
PARTITIONS	Number of CPUs	Number of partitions of Kafka Topic.
PRODUCERS	Number of CPUs	Number of Kafka producers.Set to 0 if value is not specified
CONSUMERS	Number of CPUs	Number of Kafka consumers.Set to 0 if value is not specified
SERVER_NUMACTL_OPTIONS		numactl setting "--physcpubind=0-63%20--localalloc"
CONSUMER_NUMACTL_OPTIONS		numactl setting "--physcpubind=0-63%20--localalloc"
PRODUCER_NUMACTL_OPTIONS		numactl setting "--physcpubind=0-63%20--localalloc"
KAFKA_HEAP_OPTS		Set kafka JVM heap size. "-Xmx4G_-Xms4G"
ENABLE_MUL_DISK	false	Cloud platform only. Support multiple disks attached.
MOUNT_DISK_COUNT		Number of disk. MOUNT_DISK_COUNT=CSP_DISK_SPEC_1_DISK_COUNT
PRODUCER_NODE_NUM	1	Number of producer node.Virtual producers will be equally
distributed between physical nodes if you set PRODUCERS.
TOPIC_NUM	1	Number of topic.

CSP Settings

Parameters	Description
'CSP'_CONTROLLER_INSTANCE_TYPE	Controller instance type
'CSP'_WORKER_INSTANCE_TYPE	Worker instance type. Worker instance is used for broker
'CSP'_CLIENT_INSTANCE_TYPE	Client instance type. Client instance is used for producer/consumer
'CSP'_ZONE
'CSP'_CONTROLLER_OS_DISK_SIZE	Controller OS disk size
'CSP'_CLIENT_OS_DISK_SIZE	Client OS disk size
'CSP'_WORKER_OS_DISK_SIZE	Worker OS disk size
'CSP'_CONTROLLER_OS_DISK_TYPE	Controller OS disk type
'CSP'_CLIENT_OS_DISK_TYPE	Client OS disk type
'CSP'_WORKER_OS_DISK_TYPE	Worker OS disk type
'CSP'_DISK_SPEC_1_DISK_COUNT	Number of disks (apart from OS disk) that will be attached to Worker
'CSP'_DISK_SPEC_1_DISK_SIZE	Disk size that will be attached to Worker
'CSP'_DISK_SPEC_1_DISK_TYPE	Disk type that will be attached to Worker

If multiple disks is required, set ENABLE_MUL_DISK to true, enable DISK_SPEC_1 and attach additional disks to workers; Default MOUNT_DIR is /mnt/disk{N}, N is the number of disks. Please keep MOUNT_DISK_COUNT the same value as <CSP>_DISK_SPEC_1_DISK_COUNT for cloud tests.

RUN ctest with test config file:

All Kafka Settings and CSP Settings can be specified by users through test config file.
And applied by " ./ctest.sh -R testcase_name --config path_to_test_config.yaml -VV " .
Using test config file is helpful for users to reproduce test results and keep track of performance.
Test config is like [test config file](test-config/test-config-aws-bkc.yaml).

Docker Image

The workload contains 3 docker images: kafka-zookeeper-server, producer and consumer. The container interact with each other using Kubernetes Service (ClusterIP). Due to this configuration, it is recommended to run this workload using Kubernetes instead of docker.

• kafka-zookeeper-server - Kafka and Zookeeper server container
  • Used to receive messages from producer and send messages to consumer, exposes port 2181, 9092 and 9093 using ClusterIP
• producer - Producer container
  • Used to generate and send messages to Kafka and Zookeeper server
• consumer - Consumer container
  • Used to get messages from Kafka and Zookeeper server

KPI

Run the kpi.sh script to parse the KPIs from the validation logs.

./kpi.sh
kafka_p95_latency (ms): *
number_of_producer: *
*Maximum Throughput (MB/s): *
Maximum Throughput for Latency SLA (MB/s): *
max_p95_tx_latency (ms): *
max_p95_latency (ms): *
min_p95_latency (ms): *
avg_p95_latency (ms): *
max_p99_latency (ms): *
min_p99_latency (ms): *
avg_p99_latency (ms): *

Index Info

• Name: Kafka
• Category: DataServices
• Platform: SPR, ICX
• Keywords:
• Permission:

See Also

• Kafka Configurations