grpc/tools/run_tests/performance

Overview of performance test suite

For design of the tests, see https://grpc.io/docs/guides/benchmarking.

This document contains documentation of on how to run gRPC end-to-end benchmarks using the gRPC OSS benchmarks framework (recommended) or how to run them manually (for experts only).

Approach 1: Use gRPC OSS benchmarks framework (Recommended)

gRPC OSS benchmarks

The scripts in this section generate LoadTest configurations for the GKE-based gRPC OSS benchmarks framework. This framework is stored in a separate repository, grpc/test-infra.

These scripts, together with tools defined in grpc/test-infra, are used in the continuous integration setup defined in grpc_e2e_performance_gke.sh and grpc_e2e_performance_gke_experiment.sh.

Generating scenarios

The benchmarks framework uses the same test scenarios as the legacy one. The script scenario_config_exporter.py can be used to export these scenarios to files, and also to count and analyze existing scenarios.

The language(s) and category of the scenarios are of particular importance to the tests. Continuous runs will typically run tests in the scalable category.

The following example counts scenarios in the scalable category:

$ ./tools/run_tests/performance/scenario_config_exporter.py --count_scenarios --category=scalable
Scenario count for all languages (category: scalable):
Count  Language         Client   Server   Categories
   77  c++                                scalable
   19  python_asyncio                     scalable
   16  java                               scalable
   12  go                                 scalable
   12  node                      node     scalable
   12  node_purejs               node     scalable
    9  csharp                             scalable
    7  python                             scalable
    5  ruby                               scalable
    4  csharp                    c++      scalable
    4  php7                      c++      scalable
    4  php7_protobuf_c           c++      scalable
    3  python_asyncio            c++      scalable
    2  ruby                      c++      scalable
    2  python                    c++      scalable
    1  csharp           c++               scalable

  189  total scenarios (category: scalable)

Client and server languages are only set for cross-language scenarios, where the client or server language do not match the scenario language.

Generating load test configurations

The benchmarks framework uses LoadTest resources configured by YAML files. Each LoadTest resource specifies a driver, a server, and one or more clients to run the test. Each test runs one scenario. The scenario configuration is embedded in the LoadTest configuration. Example configurations for various languages can be found here:

https://github.com/grpc/test-infra/tree/master/config/samples

The script loadtest_config.py generates LoadTest configurations for tests running a set of scenarios. The configurations are written in multipart YAML format, either to a file or to stdout. Each configuration contains a single embedded scenario.

The LoadTest configurations are generated from a template. Any configuration can be used as a template, as long as it contains the languages required by the set of scenarios we intend to run (for instance, if we are generating configurations to run go scenarios, the template must contain a go client and a go server; if we are generating configurations for cross-language scenarios that need a go client and a C++ server, the template must also contain a C++ server; and the same for all other languages).

The LoadTests specified in the script output all have unique names and can be run by applying the test to a cluster running the LoadTest controller with kubectl apply:

$ kubectl apply -f loadtest_config.yaml

Note: The most common way of running tests generated by this script is to use a test runner. For details, see running tests.

A basic template for generating tests in various languages can be found here: loadtest_template_basic_all_languages.yaml. The following example generates configurations for C# and Java tests using this template, including tests against C++ clients and servers, and running each test twice:

$ ./tools/run_tests/performance/loadtest_config.py -l go -l java \
    -t ./tools/run_tests/performance/templates/loadtest_template_basic_all_languages.yaml \
    -s client_pool=workers-8core -s driver_pool=drivers \
    -s server_pool=workers-8core \
    -s big_query_table=e2e_benchmarks.experimental_results \
    -s timeout_seconds=3600 --category=scalable \
    -d --allow_client_language=c++ --allow_server_language=c++ \
    --runs_per_test=2 -o ./loadtest.yaml

The script loadtest_config.py takes the following options:

• -l, --language
  Language to benchmark. May be repeated.
• -t, --template
  Template file. A template is a configuration file that may contain multiple client and server configuration, and may also include substitution keys.
• -s, --substitution Substitution keys, in the format key=value. These keys are substituted while processing the template. Environment variables that are set by the load test controller at runtime are ignored by default (DRIVER_PORT, KILL_AFTER, POD_TIMEOUT). The user can override this behavior by specifying these variables as keys.
• -p, --prefix
  Test names consist of a prefix_joined with a uuid with a dash. Test names are stored in metadata.name. The prefix is also added as the prefix label in metadata.labels. The prefix defaults to the user name if not set.
• -u, --uniquifier_element
  Uniquifier elements may be passed to the test to make the test name unique. This option may be repeated to add multiple elements. The uniquifier elements (plus a date string and a run index, if applicable) are joined with a dash to form a uniquifier. The test name uuid is derived from the scenario name and the uniquifier. The uniquifier is also added as the uniquifier annotation in metadata.annotations.
• -d
  This option is a shorthand for the addition of a date string as a uniquifier element.
• -a, --annotation
  Metadata annotation to be stored in metadata.annotations, in the form key=value. May be repeated.
• -r, --regex
  Regex to select scenarios to run. Each scenario is embedded in a LoadTest configuration containing a client and server of the language(s) required for the test. Defaults to .*, i.e., select all scenarios.
• --category
  Select scenarios of a specified category, or of all categories. Defaults to all. Continuous runs typically run tests in the scalable category.
• --allow_client_language
  Allows cross-language scenarios where the client is of a specified language, different from the scenario language. This is typically c++. This flag may be repeated.
• --allow_server_language
  Allows cross-language scenarios where the server is of a specified language, different from the scenario language. This is typically node or c++. This flag may be repeated.
• --instances_per_client
  This option generates multiple instances of the clients for each test. The instances are named with the name of the client combined with an index (or only an index, if no name is specified). If the template specifies more than one client for a given language, it must also specify unique names for each client. In the most common case, the template contains only one unnamed client for each language, and the instances will be named 0, 1, ...
• --runs_per_test
  This option specifies that each test should be repeated n times, where n is the value of the flag. If n > 1, the index of each test run is added as a uniquifier element for that run.
• -o, --output
  Output file name. The LoadTest configurations are added to this file, in multipart YAML format. Output is streamed to sys.stdout if not set.

The script adds labels and annotations to the metadata of each LoadTest configuration:

The following labels are added to metadata.labels:

• language
  The language of the LoadTest scenario.
• prefix
  The prefix used in metadata.name.

The following annotations are added to metadata.annotations:

• scenario
  The name of the LoadTest scenario.
• uniquifier
  The uniquifier used to generate the LoadTest name, including the run index if applicable.

Labels can be used in selectors in resource queries. Adding the prefix, in particular, allows the user (or an automation script) to select the resources started from a given run of the config generator.

Annotations contain additional information that is available to the user (or an automation script) but is not indexed and cannot be used to select objects. Scenario name and uniquifier are added to provide the elements of the LoadTest name uuid in human-readable form. Additional annotations may be added later for automation.

Concatenating load test configurations

The LoadTest configuration generator can process multiple languages at a time, assuming that they are supported by the template. The convenience script loadtest_concat_yaml.py is provided to concatenate several YAML files into one, so configurations generated by multiple generator invocations can be concatenated into one and run with a single command. The script can be invoked as follows:

$ loadtest_concat_yaml.py -i infile1.yaml infile2.yaml -o outfile.yaml

Generating load test examples

The script loadtest_examples.sh is provided to generate example load test configurations in all supported languages. This script takes only one argument, which is the output directory where the configurations will be created. The script produces a set of basic configurations, as well as a set of template configurations intended to be used with prebuilt images.

The examples in the repository grpc/test-infra are generated by this script.

Generating configuration templates

The script loadtest_template.py generates a load test configuration template from a set of load test configurations. The source files may be load test configurations or load test configuration templates. The generated template supports all languages supported in any of the input configurations or templates.

The example template in loadtest_template_basic_template_all_languages.yaml was generated from the example configurations in grpc/test-infra by the following command:

$ ./tools/run_tests/performance/loadtest_template.py \
    -i ../test-infra/config/samples/*_example_loadtest.yaml \
    --inject_client_pool --inject_server_pool \
    --inject_big_query_table --inject_timeout_seconds \
    -o ./tools/run_tests/performance/templates/loadtest_template_basic_all_languages.yaml \
    --name basic_all_languages

The example template with prebuilt images in loadtest_template_prebuilt_all_languages.yaml was generated by the following command:

$ ./tools/run_tests/performance/loadtest_template.py \
    -i ../test-infra/config/samples/templates/*_example_loadtest_with_prebuilt_workers.yaml \
    --inject_client_pool --inject_driver_image --inject_driver_pool \
    --inject_server_pool --inject_big_query_table --inject_timeout_seconds \
    -o ./tools/run_tests/performance/templates/loadtest_template_prebuilt_all_languages.yaml \
    --name prebuilt_all_languages

The script loadtest_template.py takes the following options:

• -i, --inputs
  Space-separated list of the names of input files containing LoadTest configurations. May be repeated.
• -o, --output
  Output file name. Outputs to sys.stdout if not set.
• --inject_client_pool
  If this option is set, the pool attribute of all clients in spec.clients is set to ${client_pool}, for later substitution.
• --inject_driver_image
  If this option is set, the image attribute of the driver(s) in spec.drivers is set to ${driver_image}, for later substitution.
• --inject_driver_pool
  If this attribute is set, the pool attribute of the driver(s) is set to ${driver_pool}, for later substitution.
• --inject_server_pool
  If this option is set, the pool attribute of all servers in spec.servers is set to ${server_pool}, for later substitution.
• --inject_big_query_table
  If this option is set, spec.results.bigQueryTable is set to ${big_query_table}.
• --inject_timeout_seconds
  If this option is set, spec.timeoutSeconds is set to ${timeout_seconds}.
• --inject_ttl_seconds
  If this option is set, spec.ttlSeconds is set to ${ttl_seconds}.
• -n, --name
  Name to be set in metadata.name.
• -a, --annotation
  Metadata annotation to be stored in metadata.annotations, in the form key=value. May be repeated.

The options that inject substitution keys are the most useful for template reuse. When running tests on different node pools, it becomes necessary to set the pool, and usually also to store the data on a different table. When running as part of a larger collection of tests, it may also be necessary to adjust test timeout and time-to-live, to ensure that all tests have time to complete.

The template name is replaced again by loadtest_config.py, and so is set only as a human-readable memo.

Annotations, on the other hand, are passed on to the test configurations, and may be set to values or to substitution keys in themselves, allowing future automation scripts to process the tests generated from these configurations in different ways.

Running tests

Collections of tests generated by loadtest_config.py are intended to be run with a test runner. The code for the test runner is stored in a separate repository, grpc/test-infra.

The test runner applies the tests to the cluster, and monitors the tests for completion while they are running. The test runner can also be set up to run collections of tests in parallel on separate node pools, and to limit the number of tests running in parallel on each pool.

For more information, see the tools README in grpc/test-infra.

For usage examples, see the continuous integration setup defined in grpc_e2e_performance_gke.sh and grpc_e2e_performance_gke_experiment.sh.

Approach 2: Running benchmarks locally via legacy tooling (still useful sometimes)

This approach is much more involved than using the gRPC OSS benchmarks framework (see above), but can still be useful for hands-on low-level experiments (especially when you know what you are doing).

Prerequisites for running benchmarks manually:

In general the benchmark workers and driver build scripts expect linux_performance_worker_init.sh to have been ran already.

To run benchmarks locally:

• From the grpc repo root, start the run_performance_tests.py runner script.

On remote machines, to start the driver and workers manually:

The run_performance_test.py top-level runner script can also be used with remote machines, but for e.g., profiling the server, it might be useful to run workers manually.

1. You'll need a "driver" and separate "worker" machines. For example, you might use one GCE "driver" machine and 3 other GCE "worker" machines that are in the same zone.

2. Connect to each worker machine and start up a benchmark worker with a "driver_port".

• For example, to start the grpc-go benchmark worker: grpc-go worker main.go --driver_port <driver_port>

Commands to start workers in different languages:

• Note that these commands are what the top-level run_performance_test.py script uses to build and run different workers through the build_performance.sh script and "run worker" scripts (such as the run_worker_java.sh).

Running benchmark workers for C-core wrapped languages (C++, Python, C#, Node, Ruby):

• These are more simple since they all live in the main grpc repo.

$ cd <grpc_repo_root>
$ tools/run_tests/performance/build_performance.sh
$ tools/run_tests/performance/run_worker_<language>.sh

• Note that there is one "run_worker" script per language, e.g., run_worker_csharp.sh for c#.

Running benchmark workers for gRPC-Java:

• You'll need the grpc-java repo.

$ cd <grpc-java-repo>
$ ./gradlew -PskipCodegen=true -PskipAndroid=true :grpc-benchmarks:installDist
$ benchmarks/build/install/grpc-benchmarks/bin/benchmark_worker --driver_port <driver_port>

Running benchmark workers for gRPC-Go:

• You'll need the grpc-go repo

$ cd <grpc-go-repo>/benchmark/worker && go install
$ # if profiling, it might be helpful to turn off inlining by building with "-gcflags=-l"
$ $GOPATH/bin/worker --driver_port <driver_port>

Build the driver:

• Connect to the driver machine (if using a remote driver) and from the grpc repo root:

$ tools/run_tests/performance/build_performance.sh

Run the driver:

1. Get the 'scenario_json' relevant for the scenario to run. Note that "scenario json" configs are generated from scenario_config.py. The driver takes a list of these configs as a json string of the form: {scenario: <json_list_of_scenarios> } in its --scenarios_json command argument. One quick way to get a valid json string to pass to the driver is by running the run_performance_tests.py locally and copying the logged scenario json command arg.

2. From the grpc repo root:

• Set QPS_WORKERS environment variable to a comma separated list of worker machines. Note that the driver will start the "benchmark server" on the first entry in the list, and the rest will be told to run as clients against the benchmark server.

Example running and profiling of go benchmark server:

$ export QPS_WORKERS=<host1>:<10000>,<host2>,10000,<host3>:10000
$ bins/opt/qps_json_driver --scenario_json='<scenario_json_scenario_config_string>'

Example profiling commands

While running the benchmark, a profiler can be attached to the server.

Example to count syscalls in grpc-go server during a benchmark:

• Connect to server machine and run:

$ netstat -tulpn | grep <driver_port> # to get pid of worker
$ perf stat -p <worker_pid> -e syscalls:sys_enter_write # stop after test complete

Example memory profile of grpc-go server, with go tools pprof:

• After a run is done on the server, see its alloc profile with:

$ go tool pprof --text --alloc_space http://localhost:<pprof_port>/debug/heap

Configuration environment variables:

• QPS_WORKER_CHANNEL_CONNECT_TIMEOUT

  Consuming process: qps_worker

  Type: integer (number of seconds)

  This can be used to configure the amount of time that benchmark clients wait for channels to the benchmark server to become ready. This is useful in certain benchmark environments in which the server can take a long time to become ready. Note: if setting this to a high value, then the scenario config under test should probably also have a large "warmup_seconds".

• QPS_WORKERS

  Consuming process: qps_json_driver

  Type: comma separated list of host:port

  Set this to a comma separated list of QPS worker processes/machines. Each scenario in a scenario config has specifies a certain number of servers, num_servers, and the driver will start "benchmark servers"'s on the first num_server host:port pairs in the comma separated list. The rest will be told to run as clients against the benchmark server.