Simulation tools

It is sometimes necessary create an test environment and incur artificial load to observe how well load managers handle the load. The load simulation controller, the load simulation client, and the broker monitor were created as an effort to make create this load and observe the effects on the managers more easily.

Simulation Client

The simulation client is a machine which will create and subscribe to topics with configurable message rates and sizes. Because it is sometimes necessary in simulating large load to use multiple client machines, the user does not interact with the simulation client directly, but instead delegates their requests to the simulation controller, which will then send signals to clients to start incurring load. The client implementation is in the class org.apache.pulsar.testclient.LoadSimulationClient.

Usage

To Start a simulation client, use the pulsar-perf script with the command simulation-client as follows:

  1. pulsar-perf simulation-client --port <listen port> --service-url <pulsar service url>

The client will then be ready to receive controller commands.

Simulation Controller

The simulation controller send signals to the simulation clients, requesting them to create new topics, stop old topics, change the load incurred by topics, as well as several other tasks. It is implemented in the class org.apache.pulsar.testclient.LoadSimulationController and presents a shell to the user as an interface to send command with.

Usage

To start a simulation controller, use the pulsar-perf script with the command simulation-controller as follows:

  1. pulsar-perf simulation-controller --cluster <cluster to simulate on> --client-port <listen port for clients>
  2. --clients <comma-separated list of client host names>

The clients should already be started before the controller is started. You will then be presented with a simple prompt, where you can issue commands to simulation clients. Arguments often refer to tenant names, namespace names, and topic names. In all cases, the BASE name of the tenants, namespaces, and topics are used. For example, for the topic persistent://my_tenant/my_cluster/my_namespace/my_topic, the tenant name is my_tenant, the namespace name is my_namespace, and the topic name is my_topic. The controller can perform the following actions:

  • Create a topic with a producer and a consumer
    • trade <tenant> <namespace> <topic> [--rate <message rate per second>] [--rand-rate <lower bound>,<upper bound>] [--size <message size in bytes>]
  • Create a group of topics with a producer and a consumer
    • trade_group <tenant> <group> <num_namespaces> [--rate <message rate per second>] [--rand-rate <lower bound>,<upper bound>] [--separation <separation between creating topics in ms>] [--size <message size in bytes>] [--topics-per-namespace <number of topics to create per namespace>]
  • Change the configuration of an existing topic
    • change <tenant> <namespace> <topic> [--rate <message rate per second>] [--rand-rate <lower bound>,<upper bound>] [--size <message size in bytes>]
  • Change the configuration of a group of topics
    • change_group <tenant> <group> [--rate <message rate per second>] [--rand-rate <lower bound>,<upper bound>] [--size <message size in bytes>] [--topics-per-namespace <number of topics to create per namespace>]
  • Shutdown a previously created topic
    • stop <tenant> <namespace> <topic>
  • Shutdown a previously created group of topics
    • stop_group <tenant> <group>
  • Copy the historical data from one ZooKeeper to another and simulate based on the message rates and sizes in that history
    • copy <tenant> <source zookeeper> <target zookeeper> [--rate-multiplier value]
  • Simulate the load of the historical data on the current ZooKeeper (should be same ZooKeeper being simulated on)
    • simulate <tenant> <zookeeper> [--rate-multiplier value]
  • Stream the latest data from the given active ZooKeeper to simulate the real-time load of that ZooKeeper.
    • stream <tenant> <zookeeper> [--rate-multiplier value]

The “group” arguments in these commands allow the user to create or affect multiple topics at once. Groups are created when calling the trade_group command, and all topics from these groups may be subsequently modified or stopped with the change_group and stop_group commands respectively. All ZooKeeper arguments are of the form zookeeper_host:port.

Difference Between Copy, Simulate, and Stream

The commands copy, simulate, and stream are very similar but have significant differences. copy is used when you want to simulate the load of a static, external ZooKeeper on the ZooKeeper you are simulating on. Thus, source zookeeper should be the ZooKeeper you want to copy and target zookeeper should be the ZooKeeper you are simulating on, and then it will get the full benefit of the historical data of the source in both load manager implementations. simulate on the other hand takes in only one ZooKeeper, the one you are simulating on. It assumes that you are simulating on a ZooKeeper that has historical data for SimpleLoadManagerImpl and creates equivalent historical data for ModularLoadManagerImpl. Then, the load according to the historical data is simulated by the clients. Finally, stream takes in an active ZooKeeper different than the ZooKeeper being simulated on and streams load data from it and simulates the real-time load. In all cases, the optional rate-multiplier argument allows the user to simulate some proportion of the load. For instance, using --rate-multiplier 0.05 will cause messages to be sent at only 5% of the rate of the load that is being simulated.

Broker Monitor

To observe the behavior of the load manager in these simulations, one may utilize the broker monitor, which is implemented in org.apache.pulsar.testclient.BrokerMonitor. The broker monitor will print tabular load data to the console as it is updated using watchers.

Usage

To start a broker monitor, use the monitor-brokers command in the pulsar-perf script:

  1. pulsar-perf monitor-brokers --connect-string <zookeeper host:port>

The console will then continuously print load data until it is interrupted.