Link Search Menu Expand Document Documentation Menu

OpenSearch Benchmark quickstart

This tutorial outlines how to quickly install OpenSearch Benchmark and run your first OpenSearch Benchmark workload.

Prerequisites

To perform the Quickstart steps, you’ll need to fulfill the following prerequisites:

  • A currently active OpenSearch cluster. For instructions on how to create an OpenSearch cluster, see Creating a cluster.
  • Git 2.3 or greater.
  • Python 3.8 or later

Set up an OpenSearch cluster

If you don’t already have an active OpenSearch cluster, you can launch a new OpenSearch cluster to use with OpenSearch Benchmark.

OpenSearch Benchmark has not been tested with the Window’s distribution of OpenSearch.

After installation, you can verify OpenSearch is running by going to localhost:9200. If you’re running your cluster with the Security plugin enabled, OpenSearch will expect SSL connections with the username “admin” and password “admin”. However, since the localhost address is not a unique public address, no certificate authority will issue an SSL certificate for it, so certificate checking will need to be disabled using the -k option.

Use the following command to verify OpenSearch is running with SSL certificate checks disabled:

curl -k -u admin:<custom-admin-password> https://localhost:9200			# the "-k" option skips SSL certificate checks

{
  "name" : "147ddae31bf8.opensearch.org",
  "cluster_name" : "opensearch",
  "cluster_uuid" : "n10q2RirTIuhEJCiKMkpzw",
  "version" : {
    "distribution" : "opensearch",
    "number" : "2.10.0",
    "build_type" : "tar",
    "build_hash" : "eee49cb340edc6c4d489bcd9324dda571fc8dc03",
    "build_date" : "2023-09-20T23:54:29.889267151Z",
    "build_snapshot" : false,
    "lucene_version" : "9.7.0",
    "minimum_wire_compatibility_version" : "7.10.0",
    "minimum_index_compatibility_version" : "7.0.0"
  },
  "tagline" : "The OpenSearch Project: https://opensearch.org/"
}

With your cluster running, you can now install OpenSearch Benchmark.

Installing OpenSearch Benchmark

To install OpenSearch Benchmark with Docker, see Installing OpenSearch Benchmark > Installing with Docker.

To install OpenSearch Benchmark from PyPi, enter the following pip command:

pip3 install opensearch-benchmark

After the installation completes, verify that OpenSearch Benchmark is running by entering the following command:

opensearch-benchmark --help

If successful, OpenSearch returns the following response:

$ opensearch-benchmark --help
usage: opensearch-benchmark [-h] [--version] {execute-test,list,info,create-workload,generate,compare,download,install,start,stop} ...

   ____                  _____                      __       ____                  __                         __
  / __ \____  ___  ____ / ___/___  ____ ___________/ /_     / __ )___  ____  _____/ /_  ____ ___  ____ ______/ /__
 / / / / __ \/ _ \/ __ \\__ \/ _ \/ __ `/ ___/ ___/ __ \   / __  / _ \/ __ \/ ___/ __ \/ __ `__ \/ __ `/ ___/ //_/
/ /_/ / /_/ /  __/ / / /__/ /  __/ /_/ / /  / /__/ / / /  / /_/ /  __/ / / / /__/ / / / / / / / / /_/ / /  / ,<
\____/ .___/\___/_/ /_/____/\___/\__,_/_/   \___/_/ /_/  /_____/\___/_/ /_/\___/_/ /_/_/ /_/ /_/\__,_/_/  /_/|_|
    /_/

 A benchmarking tool for OpenSearch

optional arguments:
  -h, --help            show this help message and exit
  --version             show program's version number and exit

subcommands:
  {execute-test,list,info,create-workload,generate,compare,download,install,start,stop}
    execute-test        Run a benchmark
    list                List configuration options
    info                Show info about a workload
    create-workload     Create a Benchmark workload from existing data
    generate            Generate artifacts
    compare             Compare two test_executions
    download            Downloads an artifact
    install             Installs an OpenSearch node locally
    start               Starts an OpenSearch node locally
    stop                Stops an OpenSearch node locally

Find out more about Benchmark at https://opensearch.org/docs

Running your first benchmark

You can now run your first benchmark. The following benchmark uses the percolator workload.

Understanding workload command flags

Benchmarks are run using the execute-test command with the following command flags:

For additional execute-test command flags, see the execute-test reference. Some commonly used options are --workload-params, --exclude-tasks, and --include-tasks.

  • --pipeline=benchmark-only : Informs OSB that users wants to provide their own OpenSearch cluster.
  • workload=percolator: The name of workload used by OpenSearch Benchmark.
  • --target-host="<OpenSearch Cluster Endpoint>": Indicates the target cluster or host that will be benchmarked. Enter the endpoint of your OpenSearch cluster here.
  • --client-options="basic_auth_user:'<Basic Auth Username>',basic_auth_password:'<Basic Auth Password>'": The username and password for your OpenSearch cluster.
  • --test-mode: Allows a user to run the workload without running it for the entire duration. When this flag is present, Benchmark runs the first thousand operations of each task in the workload. This is only meant for sanity checks—the metrics produced are meaningless.

The --distribution-version, which indicates which OpenSearch version Benchmark will use when provisioning. When run, the execute-test command will parse the correct distribution version when it connects to the OpenSearch cluster.

Running the workload

To run the percolator workload with OpenSearch Benchmark, use the following execute-test command:

opensearch-benchmark execute-test --pipeline=benchmark-only --workload=percolator --target-host=https://localhost:9200 --client-options=basic_auth_user:admin,basic_auth_password:admin,verify_certs:false --test-mode

When the execute-test command runs, all tasks and operations in the percolator workload run sequentially.

Validating the test

After an OpenSearch Benchmark test runs, take the following steps to verify that it has run properly:

  • Note the number of documents in the OpenSearch or OpenSearch Dashboards index that you plan to run the benchmark against.
  • In the results returned by OpenSearch Benchmark, compare the workload.json file for your specific workload and verify that the document count matches the number of documents. For example, based on the percolator workload.json file, you should expect to see 2000000 documents in your cluster.

Understanding the results

OpenSearch Benchmark returns the following response once the benchmark completes:

------------------------------------------------------
    _______             __   _____
   / ____(_)___  ____ _/ /  / ___/_________  ________
  / /_  / / __ \/ __ `/ /   \__ \/ ___/ __ \/ ___/ _ \
 / __/ / / / / / /_/ / /   ___/ / /__/ /_/ / /  /  __/
/_/   /_/_/ /_/\__,_/_/   /____/\___/\____/_/   \___/
------------------------------------------------------

|                                                         Metric |                                       Task |       Value |   Unit |
|---------------------------------------------------------------:|-------------------------------------------:|------------:|-------:|
|                     Cumulative indexing time of primary shards |                                            |     0.02655 |    min |
|             Min cumulative indexing time across primary shards |                                            |           0 |    min |
|          Median cumulative indexing time across primary shards |                                            |  0.00176667 |    min |
|             Max cumulative indexing time across primary shards |                                            |   0.0140333 |    min |
|            Cumulative indexing throttle time of primary shards |                                            |           0 |    min |
|    Min cumulative indexing throttle time across primary shards |                                            |           0 |    min |
| Median cumulative indexing throttle time across primary shards |                                            |           0 |    min |
|    Max cumulative indexing throttle time across primary shards |                                            |           0 |    min |
|                        Cumulative merge time of primary shards |                                            |   0.0102333 |    min |
|                       Cumulative merge count of primary shards |                                            |           3 |        |
|                Min cumulative merge time across primary shards |                                            |           0 |    min |
|             Median cumulative merge time across primary shards |                                            |           0 |    min |
|                Max cumulative merge time across primary shards |                                            |   0.0102333 |    min |
|               Cumulative merge throttle time of primary shards |                                            |           0 |    min |
|       Min cumulative merge throttle time across primary shards |                                            |           0 |    min |
|    Median cumulative merge throttle time across primary shards |                                            |           0 |    min |
|       Max cumulative merge throttle time across primary shards |                                            |           0 |    min |
|                      Cumulative refresh time of primary shards |                                            |   0.0709333 |    min |
|                     Cumulative refresh count of primary shards |                                            |         118 |        |
|              Min cumulative refresh time across primary shards |                                            |           0 |    min |
|           Median cumulative refresh time across primary shards |                                            |  0.00186667 |    min |
|              Max cumulative refresh time across primary shards |                                            |   0.0511667 |    min |
|                        Cumulative flush time of primary shards |                                            |  0.00963333 |    min |
|                       Cumulative flush count of primary shards |                                            |           4 |        |
|                Min cumulative flush time across primary shards |                                            |           0 |    min |
|             Median cumulative flush time across primary shards |                                            |           0 |    min |
|                Max cumulative flush time across primary shards |                                            |  0.00398333 |    min |
|                                        Total Young Gen GC time |                                            |           0 |      s |
|                                       Total Young Gen GC count |                                            |           0 |        |
|                                          Total Old Gen GC time |                                            |           0 |      s |
|                                         Total Old Gen GC count |                                            |           0 |        |
|                                                     Store size |                                            | 0.000485923 |     GB |
|                                                  Translog size |                                            | 2.01873e-05 |     GB |
|                                         Heap used for segments |                                            |           0 |     MB |
|                                       Heap used for doc values |                                            |           0 |     MB |
|                                            Heap used for terms |                                            |           0 |     MB |
|                                            Heap used for norms |                                            |           0 |     MB |
|                                           Heap used for points |                                            |           0 |     MB |
|                                    Heap used for stored fields |                                            |           0 |     MB |
|                                                  Segment count |                                            |          32 |        |
|                                                 Min Throughput |                                      index |     3008.97 | docs/s |
|                                                Mean Throughput |                                      index |     3008.97 | docs/s |
|                                              Median Throughput |                                      index |     3008.97 | docs/s |
|                                                 Max Throughput |                                      index |     3008.97 | docs/s |
|                                        50th percentile latency |                                      index |     351.059 |     ms |
|                                       100th percentile latency |                                      index |     365.058 |     ms |
|                                   50th percentile service time |                                      index |     351.059 |     ms |
|                                  100th percentile service time |                                      index |     365.058 |     ms |
|                                                     error rate |                                      index |           0 |      % |
|                                                 Min Throughput |                   wait-until-merges-finish |       28.41 |  ops/s |
|                                                Mean Throughput |                   wait-until-merges-finish |       28.41 |  ops/s |
|                                              Median Throughput |                   wait-until-merges-finish |       28.41 |  ops/s |
|                                                 Max Throughput |                   wait-until-merges-finish |       28.41 |  ops/s |
|                                       100th percentile latency |                   wait-until-merges-finish |     34.7088 |     ms |
|                                  100th percentile service time |                   wait-until-merges-finish |     34.7088 |     ms |
|                                                     error rate |                   wait-until-merges-finish |           0 |      % |
|                                                 Min Throughput |     percolator_with_content_president_bush |       36.09 |  ops/s |
|                                                Mean Throughput |     percolator_with_content_president_bush |       36.09 |  ops/s |
|                                              Median Throughput |     percolator_with_content_president_bush |       36.09 |  ops/s |
|                                                 Max Throughput |     percolator_with_content_president_bush |       36.09 |  ops/s |
|                                       100th percentile latency |     percolator_with_content_president_bush |     35.9822 |     ms |
|                                  100th percentile service time |     percolator_with_content_president_bush |     7.93048 |     ms |
|                                                     error rate |     percolator_with_content_president_bush |           0 |      % |

[...]

|                                                 Min Throughput |          percolator_with_content_ignore_me |        16.1 |  ops/s |
|                                                Mean Throughput |          percolator_with_content_ignore_me |        16.1 |  ops/s |
|                                              Median Throughput |          percolator_with_content_ignore_me |        16.1 |  ops/s |
|                                                 Max Throughput |          percolator_with_content_ignore_me |        16.1 |  ops/s |
|                                       100th percentile latency |          percolator_with_content_ignore_me |     131.798 |     ms |
|                                  100th percentile service time |          percolator_with_content_ignore_me |     69.5237 |     ms |
|                                                     error rate |          percolator_with_content_ignore_me |           0 |      % |
|                                                 Min Throughput | percolator_no_score_with_content_ignore_me |       29.37 |  ops/s |
|                                                Mean Throughput | percolator_no_score_with_content_ignore_me |       29.37 |  ops/s |
|                                              Median Throughput | percolator_no_score_with_content_ignore_me |       29.37 |  ops/s |
|                                                 Max Throughput | percolator_no_score_with_content_ignore_me |       29.37 |  ops/s |
|                                       100th percentile latency | percolator_no_score_with_content_ignore_me |     45.5703 |     ms |
|                                  100th percentile service time | percolator_no_score_with_content_ignore_me |      11.316 |     ms |
|                                                     error rate | percolator_no_score_with_content_ignore_me |           0 |      % |



--------------------------------
[INFO] SUCCESS (took 18 seconds)
--------------------------------

Each task run by the percolator workload represents a specific OpenSearch API operation—such as Bulk or Search—that was performed when the test was run. Each task in the output summary contains the following information:

  • Throughput: The number of successful OpenSearch operations per second.
  • Latency: The amount of time, including wait time, taken for the request and the response to be sent and received by Benchmark.
  • Service Time: The amount of time, excluding wait time, taken for the request and the response to be sent and received by Benchmark.
  • Error Rate: The percentage of operations run during the task that were not successful or returned a 200 error code.

For more details about how the summary report is generated, see Summary report.

Running OpenSearch Benchmark on your own cluster

Now that you’re familiar with running OpenSearch Benchmark on a cluster, you can run OpenSearch Benchmark on your own cluster, using the same execute-test command, replacing the following settings.

  • Replace https://localhost:9200 with your target cluster endpoint. This could be a URI like https://search.mydomain.com or a HOST:PORT specification.
  • If the cluster is configured with basic authentication, replace the username and password in the command line with the appropriate credentials.
  • Remove the verify_certs:false directive if you are not specifying localhost as your target cluster. This directive is needed only for clusters where SSL certificates are not set up.
  • If you are using a HOST:PORTspecification and plan to use SSL/TLS, either specify https://, or add the use_ssl:true directive to the --client-options string option.
  • Remove the --test-mode flag to run the full workload, rather than an abbreviated test.

You can copy the following command template to use in your own terminal:

opensearch-benchmark execute-test --pipeline=benchmark-only --workload=percolator --target-host=<OpenSearch Cluster Endpoint> --client-options=basic_auth_user:admin,basic_auth_password:admin

Next steps

See the following resources to learn more about OpenSearch Benchmark:

  • User guide: Dive deep into how OpenSearch Benchmark can you help you track the performance of your cluster.
  • Tutorials: Use step-by-step guides for more advanced Benchmarking configurations and functionality.