Skip to main content
NetApp Solutions

Performance overview and validation in AWS

Contributors nkarthik kevin-hoke
PDFs

A Kafka cluster with the storage layer mounted on NetApp NFS was benchmarked for performance in the AWS cloud. The benchmarking examples are described in the following sections.

Kafka in AWS cloud with NetApp Cloud Volumes ONTAP (high-availability pair and single node)

A Kafka cluster with NetApp Cloud Volumes ONTAP (HA pair) was benchmarked for performance in the AWS cloud. This benchmarking is described in the following sections.

Architectural setup

The following table shows the environmental configuration for a Kafka cluster using NAS.

Platform component Environment configuration

Kafka 3.2.3

  • 3 x zookeepers – t2.small

  • 3 x broker servers – i3en.2xlarge

  • 1 x Grafana – c5n.2xlarge

  • 4 x producer/consumer — c5n.2xlarge
    *

Operating system on all nodes

RHEL8.6

NetApp Cloud Volumes ONTAP instance

HA pair instance – m5dn.12xLarge x 2node
Single Node Instance - m5dn.12xLarge x 1 node

NetApp cluster volume ONTAP setup

  1. For the Cloud Volumes ONTAP HA pair, we created two aggregates with three volumes on each aggregate on each storage controller. For the single Cloud Volumes ONTAP node, we create six volumes in an aggregate.

    This image depicts the properties of aggr3 and aggr22.

    This image depicts the properties of aggr2.

  2. To achieve better network performance, we enabled high speed networking for both the HA pair and the single node.

    This images shows how to enable high-speed networking.

  3. We noticed that the ONTAP NVRAM had more IOPS so we changed the IOPS to 2350 for the Cloud Volumes ONTAP root volume. The root volume disk in Cloud Volumes ONTAP was 47GB in size. The following ONTAP command is for the HA pair, and the same step is applicable for the single node.

    statistics start -object vnvram -instance vnvram -counter backing_store_iops -sample-id sample_555
kafka_nfs_cvo_ha1::*> statistics show -sample-id sample_555
Object: vnvram
Instance: vnvram
Start-time: 1/18/2023 18:03:11
End-time: 1/18/2023 18:03:13
Elapsed-time: 2s
Scope: kafka_nfs_cvo_ha1-01
    Counter                                                     Value
    -------------------------------- --------------------------------
    backing_store_iops                                           1479
Object: vnvram
Instance: vnvram
Start-time: 1/18/2023 18:03:11
End-time: 1/18/2023 18:03:13
Elapsed-time: 2s
Scope: kafka_nfs_cvo_ha1-02
    Counter                                                     Value
    -------------------------------- --------------------------------
    backing_store_iops                                           1210
2 entries were displayed.
kafka_nfs_cvo_ha1::*>

    This images shows how to modify volume properties.

The following figure depicts the architecture of an NAS-based Kafka cluster.

  • Compute. We used a three-node Kafka cluster with a three-node zookeeper ensemble running on dedicated servers. Each broker had two NFS mount points to a single volume on the Cloud Volumes ONTAP instance through a dedicated LIF.

  • Monitoring. We used two nodes for a Prometheus-Grafana combination. For generating workloads, we used a separate three-node cluster that could produce and consume to this Kafka cluster.

  • Storage. We used an HA-pair Cloud volumes ONTAP instance with one 6TB GP3 AWS-EBS volume mounted on the instance. The volume was then exported to the Kafka broker with an NFS mount.

This figure depicts the architecture of an NAS-based Kafka cluster.

OpenMessage Benchmarking configurations

  1. For better NFS performance, we need more network connections between the NFS server and the NFS client, which can be created using nconnect. Mount the NFS volumes on the broker nodes with the nconnect option by running the following command:

    [root@ip-172-30-0-121 ~]# cat /etc/fstab
UUID=eaa1f38e-de0f-4ed5-a5b5-2fa9db43bb38/xfsdefaults00
/dev/nvme1n1 /mnt/data-1 xfs defaults,noatime,nodiscard 0 0
/dev/nvme2n1 /mnt/data-2 xfs defaults,noatime,nodiscard 0 0
172.30.0.233:/kafka_aggr3_vol1 /kafka_aggr3_vol1 nfs defaults,nconnect=16 0 0
172.30.0.233:/kafka_aggr3_vol2 /kafka_aggr3_vol2 nfs defaults,nconnect=16 0 0
172.30.0.233:/kafka_aggr3_vol3 /kafka_aggr3_vol3 nfs defaults,nconnect=16 0 0
172.30.0.242:/kafka_aggr22_vol1 /kafka_aggr22_vol1 nfs defaults,nconnect=16 0 0
172.30.0.242:/kafka_aggr22_vol2 /kafka_aggr22_vol2 nfs defaults,nconnect=16 0 0
172.30.0.242:/kafka_aggr22_vol3 /kafka_aggr22_vol3 nfs defaults,nconnect=16 0 0
[root@ip-172-30-0-121 ~]# mount -a
[root@ip-172-30-0-121 ~]# df -h
Filesystem                       Size  Used Avail Use% Mounted on
devtmpfs                          31G     0   31G   0% /dev
tmpfs                             31G  249M   31G   1% /run
tmpfs                             31G     0   31G   0% /sys/fs/cgroup
/dev/nvme0n1p2                    10G  2.8G  7.2G  28% /
/dev/nvme1n1                     2.3T  248G  2.1T  11% /mnt/data-1
/dev/nvme2n1                     2.3T  245G  2.1T  11% /mnt/data-2
172.30.0.233:/kafka_aggr3_vol1   1.0T   12G 1013G   2% /kafka_aggr3_vol1
172.30.0.233:/kafka_aggr3_vol2   1.0T  5.5G 1019G   1% /kafka_aggr3_vol2
172.30.0.233:/kafka_aggr3_vol3   1.0T  8.9G 1016G   1% /kafka_aggr3_vol3
172.30.0.242:/kafka_aggr22_vol1  1.0T  7.3G 1017G   1% /kafka_aggr22_vol1
172.30.0.242:/kafka_aggr22_vol2  1.0T  6.9G 1018G   1% /kafka_aggr22_vol2
172.30.0.242:/kafka_aggr22_vol3  1.0T  5.9G 1019G   1% /kafka_aggr22_vol3
tmpfs                            6.2G     0  6.2G   0% /run/user/1000
[root@ip-172-30-0-121 ~]#

  2. Check the network connections in Cloud Volumes ONTAP. The following ONTAP command is used from the single Cloud Volumes ONTAP node. The same step is applicable to the Cloud Volumes ONTAP HA pair.

    Last login time: 1/20/2023 00:16:29
kafka_nfs_cvo_sn::> network connections active show -service nfs* -fields remote-host
node                cid        vserver              remote-host
------------------- ---------- -------------------- ------------
kafka_nfs_cvo_sn-01 2315762628 svm_kafka_nfs_cvo_sn 172.30.0.121
kafka_nfs_cvo_sn-01 2315762629 svm_kafka_nfs_cvo_sn 172.30.0.121
kafka_nfs_cvo_sn-01 2315762630 svm_kafka_nfs_cvo_sn 172.30.0.121
kafka_nfs_cvo_sn-01 2315762631 svm_kafka_nfs_cvo_sn 172.30.0.121
kafka_nfs_cvo_sn-01 2315762632 svm_kafka_nfs_cvo_sn 172.30.0.121
kafka_nfs_cvo_sn-01 2315762633 svm_kafka_nfs_cvo_sn 172.30.0.121
kafka_nfs_cvo_sn-01 2315762634 svm_kafka_nfs_cvo_sn 172.30.0.121
kafka_nfs_cvo_sn-01 2315762635 svm_kafka_nfs_cvo_sn 172.30.0.121
kafka_nfs_cvo_sn-01 2315762636 svm_kafka_nfs_cvo_sn 172.30.0.121
kafka_nfs_cvo_sn-01 2315762637 svm_kafka_nfs_cvo_sn 172.30.0.121
kafka_nfs_cvo_sn-01 2315762639 svm_kafka_nfs_cvo_sn 172.30.0.72
kafka_nfs_cvo_sn-01 2315762640 svm_kafka_nfs_cvo_sn 172.30.0.72
kafka_nfs_cvo_sn-01 2315762641 svm_kafka_nfs_cvo_sn 172.30.0.72
kafka_nfs_cvo_sn-01 2315762642 svm_kafka_nfs_cvo_sn 172.30.0.72
kafka_nfs_cvo_sn-01 2315762643 svm_kafka_nfs_cvo_sn 172.30.0.72
kafka_nfs_cvo_sn-01 2315762644 svm_kafka_nfs_cvo_sn 172.30.0.72
kafka_nfs_cvo_sn-01 2315762645 svm_kafka_nfs_cvo_sn 172.30.0.72
kafka_nfs_cvo_sn-01 2315762646 svm_kafka_nfs_cvo_sn 172.30.0.72
kafka_nfs_cvo_sn-01 2315762647 svm_kafka_nfs_cvo_sn 172.30.0.72
kafka_nfs_cvo_sn-01 2315762648 svm_kafka_nfs_cvo_sn 172.30.0.72
kafka_nfs_cvo_sn-01 2315762649 svm_kafka_nfs_cvo_sn 172.30.0.121
kafka_nfs_cvo_sn-01 2315762650 svm_kafka_nfs_cvo_sn 172.30.0.121
kafka_nfs_cvo_sn-01 2315762651 svm_kafka_nfs_cvo_sn 172.30.0.121
kafka_nfs_cvo_sn-01 2315762652 svm_kafka_nfs_cvo_sn 172.30.0.121
kafka_nfs_cvo_sn-01 2315762653 svm_kafka_nfs_cvo_sn 172.30.0.121
kafka_nfs_cvo_sn-01 2315762656 svm_kafka_nfs_cvo_sn 172.30.0.223
kafka_nfs_cvo_sn-01 2315762657 svm_kafka_nfs_cvo_sn 172.30.0.223
kafka_nfs_cvo_sn-01 2315762658 svm_kafka_nfs_cvo_sn 172.30.0.223
kafka_nfs_cvo_sn-01 2315762659 svm_kafka_nfs_cvo_sn 172.30.0.223
kafka_nfs_cvo_sn-01 2315762660 svm_kafka_nfs_cvo_sn 172.30.0.223
kafka_nfs_cvo_sn-01 2315762661 svm_kafka_nfs_cvo_sn 172.30.0.223
kafka_nfs_cvo_sn-01 2315762662 svm_kafka_nfs_cvo_sn 172.30.0.223
kafka_nfs_cvo_sn-01 2315762663 svm_kafka_nfs_cvo_sn 172.30.0.223
kafka_nfs_cvo_sn-01 2315762664 svm_kafka_nfs_cvo_sn 172.30.0.223
kafka_nfs_cvo_sn-01 2315762665 svm_kafka_nfs_cvo_sn 172.30.0.223
kafka_nfs_cvo_sn-01 2315762666 svm_kafka_nfs_cvo_sn 172.30.0.223
kafka_nfs_cvo_sn-01 2315762667 svm_kafka_nfs_cvo_sn 172.30.0.72
kafka_nfs_cvo_sn-01 2315762668 svm_kafka_nfs_cvo_sn 172.30.0.72
kafka_nfs_cvo_sn-01 2315762669 svm_kafka_nfs_cvo_sn 172.30.0.72
kafka_nfs_cvo_sn-01 2315762670 svm_kafka_nfs_cvo_sn 172.30.0.72
kafka_nfs_cvo_sn-01 2315762671 svm_kafka_nfs_cvo_sn 172.30.0.72
kafka_nfs_cvo_sn-01 2315762672 svm_kafka_nfs_cvo_sn 172.30.0.72
kafka_nfs_cvo_sn-01 2315762673 svm_kafka_nfs_cvo_sn 172.30.0.223
kafka_nfs_cvo_sn-01 2315762674 svm_kafka_nfs_cvo_sn 172.30.0.223
kafka_nfs_cvo_sn-01 2315762676 svm_kafka_nfs_cvo_sn 172.30.0.121
kafka_nfs_cvo_sn-01 2315762677 svm_kafka_nfs_cvo_sn 172.30.0.223
kafka_nfs_cvo_sn-01 2315762678 svm_kafka_nfs_cvo_sn 172.30.0.223
kafka_nfs_cvo_sn-01 2315762679 svm_kafka_nfs_cvo_sn 172.30.0.223
48 entries were displayed.
 
kafka_nfs_cvo_sn::>

  3. We use the following Kafka server.properties in all Kafka brokers for the Cloud Volumes ONTAP HA pair. The log.dirs property is different for each broker, and the remaining properties are common for brokers. For broker1, the log.dirs value is as follows:

    [root@ip-172-30-0-121 ~]# cat /opt/kafka/config/server.properties
broker.id=0
advertised.listeners=PLAINTEXT://172.30.0.121:9092
#log.dirs=/mnt/data-1/d1,/mnt/data-1/d2,/mnt/data-1/d3,/mnt/data-2/d1,/mnt/data-2/d2,/mnt/data-2/d3
log.dirs=/kafka_aggr3_vol1/broker1,/kafka_aggr3_vol2/broker1,/kafka_aggr3_vol3/broker1,/kafka_aggr22_vol1/broker1,/kafka_aggr22_vol2/broker1,/kafka_aggr22_vol3/broker1
zookeeper.connect=172.30.0.12:2181,172.30.0.30:2181,172.30.0.178:2181
num.network.threads=64
num.io.threads=64
socket.send.buffer.bytes=102400
socket.receive.buffer.bytes=102400
socket.request.max.bytes=104857600
num.partitions=1
num.recovery.threads.per.data.dir=1
offsets.topic.replication.factor=1
transaction.state.log.replication.factor=1
transaction.state.log.min.isr=1
replica.fetch.max.bytes=524288000
background.threads=20
num.replica.alter.log.dirs.threads=40
num.replica.fetchers=20
[root@ip-172-30-0-121 ~]#

    • For broker2, the log.dirs property value is as follows:

      log.dirs=/kafka_aggr3_vol1/broker2,/kafka_aggr3_vol2/broker2,/kafka_aggr3_vol3/broker2,/kafka_aggr22_vol1/broker2,/kafka_aggr22_vol2/broker2,/kafka_aggr22_vol3/broker2

    • For broker3, the log.dirs property value is as follows:

      log.dirs=/kafka_aggr3_vol1/broker3,/kafka_aggr3_vol2/broker3,/kafka_aggr3_vol3/broker3,/kafka_aggr22_vol1/broker3,/kafka_aggr22_vol2/broker3,/kafka_aggr22_vol3/broker3

  4. For the single Cloud Volumes ONTAP node, The Kafka servers.properties is the same as for the Cloud Volumes ONTAP HA pair except for the log.dirs property.

    • For broker1, the log.dirs value is as follows:

      log.dirs=/kafka_aggr2_vol1/broker1,/kafka_aggr2_vol2/broker1,/kafka_aggr2_vol3/broker1,/kafka_aggr2_vol4/broker1,/kafka_aggr2_vol5/broker1,/kafka_aggr2_vol6/broker1

    • For broker2, the log.dirs value is as follows:

      log.dirs=/kafka_aggr2_vol1/broker2,/kafka_aggr2_vol2/broker2,/kafka_aggr2_vol3/broker2,/kafka_aggr2_vol4/broker2,/kafka_aggr2_vol5/broker2,/kafka_aggr2_vol6/broker2

    • For broker3, the log.dirs property value is as follows:

      log.dirs=/kafka_aggr2_vol1/broker3,/kafka_aggr2_vol2/broker3,/kafka_aggr2_vol3/broker3,/kafka_aggr2_vol4/broker3,/kafka_aggr2_vol5/broker3,/kafka_aggr2_vol6/broker3

  5. The workload in the OMB is configured with the following properties: (/opt/benchmark/workloads/1-topic-100-partitions-1kb.yaml).

    topics: 4
partitionsPerTopic: 100
messageSize: 32768
useRandomizedPayloads: true
randomBytesRatio: 0.5
randomizedPayloadPoolSize: 100
subscriptionsPerTopic: 1
consumerPerSubscription: 80
producersPerTopic: 40
producerRate: 1000000
consumerBacklogSizeGB: 0
testDurationMinutes: 5

    The messageSize can vary for each use case. In our performance test, we used 3K.

    We used two different drivers, Sync or Throughput, from OMB to generate the workload on the Kafka cluster.

    • The yaml file used for Sync driver properties is as follows (/opt/benchmark/driver- kafka/kafka-sync.yaml):

      name: Kafka
driverClass: io.openmessaging.benchmark.driver.kafka.KafkaBenchmarkDriver
# Kafka client-specific configuration
replicationFactor: 3
topicConfig: |
  min.insync.replicas=2
  flush.messages=1
  flush.ms=0
commonConfig: |
  bootstrap.servers=172.30.0.121:9092,172.30.0.72:9092,172.30.0.223:9092
producerConfig: |
  acks=all
  linger.ms=1
  batch.size=1048576
consumerConfig: |
  auto.offset.reset=earliest
  enable.auto.commit=false
  max.partition.fetch.bytes=10485760

    • The yaml file used for the Throughput driver properties is as follows (/opt/benchmark/driver- kafka/kafka-throughput.yaml):

      name: Kafka
driverClass: io.openmessaging.benchmark.driver.kafka.KafkaBenchmarkDriver
# Kafka client-specific configuration
replicationFactor: 3
topicConfig: |
  min.insync.replicas=2
commonConfig: |
  bootstrap.servers=172.30.0.121:9092,172.30.0.72:9092,172.30.0.223:9092
  default.api.timeout.ms=1200000
  request.timeout.ms=1200000
producerConfig: |
  acks=all
  linger.ms=1
  batch.size=1048576
consumerConfig: |
  auto.offset.reset=earliest
  enable.auto.commit=false
  max.partition.fetch.bytes=10485760

Methodology of testing

  1. A Kafka cluster was provisioned as per the specification described above using Terraform and Ansible. Terraform is used to build the infrastructure using AWS instances for the Kafka cluster and Ansible builds the Kafka cluster on them.

  2. An OMB workload was triggered with the workload configuration described above and the Sync driver.

    Sudo bin/benchmark –drivers driver-kafka/kafka- sync.yaml workloads/1-topic-100-partitions-1kb.yaml

  3. Another workload was triggered with the Throughput driver with same workload configuration.

    sudo bin/benchmark –drivers driver-kafka/kafka-throughput.yaml workloads/1-topic-100-partitions-1kb.yaml

Observation

Two different types of drivers were used to generate workloads to benchmark the performance of a Kafka instance running on NFS. The difference between the drivers is the log flush property.

For a Cloud Volumes ONTAP HA pair:

  • Total throughput generated consistently by the Sync driver: ~1236 MBps.

  • Total throughput generated for the Throughput driver: peak ~1412 MBps.

For a single Cloud Volumes ONTAP node:

  • Total throughput generated consistently by the Sync driver: ~ 1962MBps.

  • Total throughput generated by the Throughput driver: peak ~1660MBps

The Sync driver can generate consistent throughput as logs are flushed to the disk instantly, whereas the Throughput driver generates bursts of throughput as logs are committed to disk in bulk.

These throughput numbers are generated for the given AWS configuration. For higher performance requirements, the instance types can be scaled up and tuned further for better throughput numbers. The total throughput or total rate is the combination of both producer and consumer rate.

Four different graphs are presented here. CVO-HA Pair throughput driver. CVO-HA pair Sync driver. CVO-single node throughput driver. CVO-single node Sync driver.

Be sure to check the storage throughput when performing throughput or sync driver benchmarking.

This graph shows the performance in latency, IOPS, and Throughput.