Apache Geode Benchmarks

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.

Running the benchmarks

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

Options:

    -Phosts               : Hosts used by benchmarks on the order of client,locator,server,server (-Phosts=localhost,localhost,localhost,localhost)
    -PoutputDir           : Results output directory (-PoutputDir=/tmp/results)
    -PtestJVM             : Path to an alternative JVM for running the client, locator, and servers. If not specified JAVA_HOME will be used. Note all compilation tasks will still use JAVA_HOME.
    -PwithSsl             : Flag to run geode with SSL. A self-signed certificate will be generated at runtime.
    -PwithSecurityManager : Flag to start Geode with the example implementation of SecurityManager
    -PwithGc              : Select which GC to use. Valid values CMS (default), G1, Z.
    -PwithHeap            : Specify how large a heap the benchmark VMs should use, default "8g". Accepts any `-Xmx` value, like "32g".
    -PwithThreads         : Specify how many threads to use when executing the benchmark. Default varies by benchmark.
    -PwithWarmup          : Specify how long to warm up the benchmark in seconds. Default is 60 seconds.
    -PwithDuration        : Specify how long to measure the benchmark in seconds. Default is 300 seconds.
    --tests               : Specific benchmarks to run (--tests=PartitionedPutBenchmark)
    -d                    : Debug
    -i                    : Info

Scripts for running in aws and analyzing results

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.

Project structure

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.

Sample benchmark

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 implements PerformanceTest {

  @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;
  }
}

SNI

On AWS, you can run any benchmark on a topology that routes all client-server communication through an SNI proxy (HAproxy).

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

./run_tests.sh -t anytagname -- -PwithSniProxy '--tests=PartitionedGetBenchmark'

Since SNI is a feature of TLS, running with the SNI topology incurs TLS overheads.

TODO for SNI

• verify StartSniProxy runs on proxy node
• don't require operator to supply -PwithSSL/-DwithSSL=true when running SNI tests
• generate haproxy.cfg with client-visible SNI hostnames
• turn on SNI via setPoolSocketFactory in a new StartClientSNI task
• set --hostname-for-clients on locator and servers for SNI
• reinstate thread-per-core in PrePopulateRegion.run() and in PartitionedPutBenchmark[SNI] ya
• set keyRange back to 1e6 in PartitionedPutBenchmark[SNI] after client-server connections are healthy
• make topology orthogonal to tests so all tests can run with SNI; have a -PwithSniProxy/-DwithSniProxy=true flag
• Potential performance improvement: HAproxy as configured runs one process with the max threads-per-process of 64 threads, ostensibly using 64/72 cores (89%.) We might be able to improve performance by configuring HAproxy to run in daemon mode where we can run two processes, each multithreaded, to run more than 64 threads, thereby utilizing 100% of our cores.

TODO (General)

• add logic to clean up existing locator.dat files before running a locator on a node
• eliminate harness module dependency on Geode by moving Geode keystore/truststore setting out of harness module and up into geode-benchmarks i.e. set 'em in properties sent to Locator.startLocatorAndDS in StartLocator, StartServer
• move docker-compose.yml distribution out of harness module up into geode-benchmarks so it gets distributed whenever it changes (without requiring rebuilding AWS AMI and cluster on AWS)
• generate 2048-bit keys (instead of 1024-bit ones) for TLS; will slow TLS handshakes which may necessitate a new baseline
• make StartServer task use ServerLauncher (instead of CacheFactory) for symmetry with LocatorLauncher—also too: encapsulation!
• ./run_tests.sh sometimes seems to hang after benchmarks have completed, requiring operator to enter ^C to un-stick it
• make rsync: Git "scheme" work in run_tests.sh script for benchmark repo (not just for geode repo)