This project contains a benchmarking framework and benchmarks for Apache Geode. It is based on the yardstick framework, but with a java based test configuration and test execution framework.

These benchmarks are under development.

The benchmarks require machines with passwordless ssh enabled in order to run, so ensure that the authentication key pair for SSH does not have a passphrase. If you have previously created a key pair with a passphrase, you can create a different key pair which is of a different type than the previously created one. Be sure to backup your existing key pair before creating a new one. The public key needs to be in PEM format, but some newer OpenSSH versions default to a new format. PEM format can be forced by using -m PEM:

ssh-keygen -m PEM -t rsa

While running a test on a single machine (i.e. localhost) add the generated key to authorized_keys to authorize the user:

cat <your_public_key_file> >> ~/.ssh/authorized_keys

Test if you can ssh to localhost:

ssh localhost

As long as that works, we are good to go.

Get your local hosts name:

hostname

Edit /etc/hosts and add the local host name with and without domain to localhost entries.

127.0.0.1 localhost mycomputer mycomputer.mydomain
::1       localhost mycomputer mycomputer.mydomain

To run all benchmarks, run the benchmark task and pass in a list of hosts.

For example:

./gradlew benchmark -Phosts=localhost,localhost,localhost,localhost

The benchmarks can take configuration options. Some using Gradle's -P flag and other, which adjust benchmark behavior, via Java system properties using -D.

These options may apply to all benchmarks.

These options only apply to Geode benchmarks.

These options only apply to Redis benchmarks.

These options should not be used when measuring benchmarks.

This project includes some scripts to automate running benchmarks in AWS and analyzing the results produced (as well as the results produced from running locally). See the README.md in the infrastructure/aws directory.

The project is divided into two modules

• harness - Contains test framework code for running benchmarks. Does not depend on Apache Geode.
• geode-benchmarks - Individual benchmarks of Apache Geode.

Benchmarks are defined in declarative configuration classes. Each configuration class is run as a junit test which calls the configure method and passes it to the TestRunner, which executes the test on the provided TEST_HOSTS.

Benchmarks are composed of before tasks, after tasks, and workload tasks. Each seperate before and after task is run once. Workload tasks are run repeatedly and their execution time is measured and reported by the yardstick framework.

/**
* Benchmark configuration class, which defines the topology of the test and
* the initialization tasks and workload tasks for the test.
*/
public class PartitionedPutBenchmark extends AbstractPerformanceTest {

  @Test
  public void run() throws Exception {
    TestRunners.defaultRunner().runTest(this::configure);
  }
  
  /**
  * Declare the configuration of the test by calling methods
  * on TestConfig.
  */
  public TestConfig configure() {
    TestConfig testConfig = new TestConfig();
    int locatorPort = 10334;

    //This test has three roles, a geode locator, server, and client
    config.role("locator", 1);
    config.role("server", 2);
    config.role("client", 1);
    
    //Define how the locator,server and client are initialized
    config.before(new StartLocator(locatorPort), "locator");
    config.before(new StartServer(locatorPort), "server");
    config.before(new StartClient(locatorPort), "client");
    config.before(new CreatePartitionedRegion(), "server");
    config.before(new CreateClientProxyRegion(), "client");
    //Define the benchmarked workload, which runs in a client
    config.workload(new PutTask());
    
    return config;
  }
}

/**
* Workload task, which extends the yardstick BenchmarkDriverAdapter
* 
* Workload tasks should execute a single unit of work, and will be run repeatedly
* for the duration of the test.
*/
public class PutTask extends BenchmarkDriverAdapter implements Serializable {
  private Region<Object, Object> region;
  
  @Override
  public void setUp(BenchmarkConfiguration cfg) throws Exception {
    super.setUp(cfg);
    ClientCache cache = ClientCacheFactory.getAnyInstance();
    region = cache.getRegion("region");
  }

  
  @Override
  public boolean test(Map<Object, Object> ctx) throws Exception {
    region.put(1,2);
    return true;
  }
}

You can run any benchmark on a topology that routes all client-server communication through an SNI proxy.

The withSniProxy property accepts:

• HAProxy for HAProxy based SNI proxy (default).
• Envoy for Envoy based SNI proxy.
• Manual for providing your own SNI proxy and managing its lifecycle.

The withSniProxyImage property can be used to provide an alternative Docker image to one of the supported proxy implementations. The value should be set to a valid Docker image coordinate.

To run a test, e.g. PartitionedGetBenchmark, with default SNI Proxy:

./run_tests.sh -t anytagname -- -Pbenchmark.withSniProxy --tests=PartitionedGetBenchmark

Since SNI is a feature of TLS, running with the SNI topology incurs TLS overheads with implied -Pbenchmark.withSsl.

An alternative topology uses a router sitting in front of the SNI proxy to simulate off network access to the cluster, enabled with -Pbenchmark.withRouter.

Enabling the router implies -Pbenchmark.withSniProxy.

The withRouter property accepts:

• HAProxy for HAProxy based router (default).
• Manual for providing your own router and managing its lifecycle.

Example:

./run_tests.sh -t anytagname -- -Pbenchmark.withRouter --tests=PartitionedGetBenchmark

You can run benchmarks utilizing the Redis protocol with various clients and backends. All Redis benchmarks take the pattern Redis*Benchmark. They expect 3 shards with 1 replica per shard, which when combined with 1 Geode locator and the benchmarking client needs a total of 8 hosts (./launch_cluster ... -c 8).

The withRedisClient property accepts:

• Jedis for using the Jedis library (default).
• Lettuce for using the Lettuce library.

The withRedisCluster property accepts:

• Geode for using the Geode server backend (default). Builds a Geode cluster utilizing 7 hosts, 1 locator and 6 servers.
• Redis for using the Redis server backend. Builds a Redis cluster utilizing 6 hosts and the Bitnami Redis image.
• Manual for using a manually configured Redis server backend, like Elasticache. Use withRedisServers to specify the address(es) to the Redis server endpoints.

Examples:

• Runs the RedisGetBenchmark against a Geode cluster using the Jedis client.

  ./run_tests.sh -t anytagname -- --tests=RedisGetBenchmark

• Runs the RedisGetBenchmark against a Geode cluster using the Lettuce client.

  ./run_tests.sh -t anytagname -- -Pbenchmark.withRedisClient=lettuce --tests=RedisGetBenchmark

• Runs the RedisGetBenchmark against a Redis cluster using the Jedis client.

  ./run_tests.sh -t anytagname -- -Pbenchmark.withRedisCluster=redis --tests=RedisGetBenchmark

• Runs the RedisGetBenchmark against an Elasticache cluster using the Jedis client.

  ./run_tests.sh -t anytagname -- -Pbenchmark.withRedisCluster=manual -Pbenchmark.withRedisServers=my-cluster...clustercfg.usw2.cache.amazonaws.com:6379 --tests=RedisGetBenchmark