======================= Message Queue Test Plan ======================= :status: ready :version: 0 :Abstract: This document describes a test plan for quantifying the performance of message queues usually used as a message bug between OpenStack services. Test Plan ========= Test Environment ---------------- This section describes the setup for message queue testing. It can be either a single (all-in-one) or a multi-node installation. A single-node setup requires just one node to be up and running. It has both compute and controller roles and all OpenStack services run on this node. This setup does not support hardware scaling or workload distribution tests. A basic multi-node setup with RabbitMQ or ActiveMQ comprises 5 physical nodes: * One node for a compute node. This node simulates activity which is typical for OpenStack compute components. * One node for a controller node. This node simulates activity which is typical for OpenStack control plane services. * Three nodes are allocated for the MQ cluster. When using ZeroMQ, the basic multi-node setup can be reduced to two physical nodes. * One node for a compute node as above. * One node for a controller node. This node also acts as a Redis host for match making purposes. Preparation ^^^^^^^^^^^ **RabbitMQ Installation and Configuration** * Install RabbitMQ server package: ``sudo apt-get install rabbitmq-server`` * Configure RabbitMQ on each node ``/etc/rabbitmq/rabbitmq.config``: .. literalinclude:: rabbitmq.config :language: erlang .. * Stop RabbitMQ on nodes 2 and 3: ``sudo service rabbitmq-server stop`` * Make Erlang cookies on nodes 2 and 3 the same as on node 1: ``/var/lib/rabbitmq/.erlang.cookie`` * Start RabbitMQ server: ``sudo service rabbitmq-server start`` * Stop RabbitMQ services, but leave Erlang: ``sudo rabbitmqctl stop_app`` * Join nodes 2 and 3 nodes to node 1: ``rabbitmqctl join_cluster rabbit@node-1`` * Start app on nodes 2 and 3: ``sudo rabbitmqctl start_app`` * Add needed user: ``sudo rabbitmqctl add_user stackrabbit password`` ``sudo rabbitmqctl set_permissions stackrabbit ".*" ".*" ".*"`` **ActiveMQ Installation and Configuration** This section describes installation and configuration steps for an ActiveMQ message queue implementation. ActiveMQ is based on Java technologies so it requires a Java runtime. Actual performance will depend on the Java version as well as the hardware specification. The following steps should be performed for an ActiveMQ installation: * Install Java on nodes node-1, node-2 and node-3: ``sudo apt-get install default-jre`` * Download the latest ActiveMQ binary: ``wget http://www.eu.apache.org/dist/activemq/5.12.0/apache-activemq-5.12.0-bin.tar.gz`` * Unzip the archive: ``tar zxvf apache-activemq-5.12.0-bin.tar.gz`` * Install everything needed for ZooKeeper: * download ZK binaries: ``wget http://www.eu.apache.org/dist/zookeeper/zookeeper-3.4.6/zookeeper-3.4.6.tar.gz`` * unzip the archive: ``tar zxvf zookeeper-3.4.6.tar.gz`` * create ``/home/ubuntu/zookeeper-3.4.6/conf/zoo.cfg`` file: .. literalinclude:: zoo.cfg :language: ini .. note:: Here 10.4.1.x are the IP addresses of the ZooKeeper nodes where ZK is installed. ZK will be run in cluster mode with majority voting, so at least 3 nodes are required. .. code-block:: none tickTime=2000 dataDir=/home/ubuntu/zookeeper-3.4.6/data dataLogDir=/home/ubuntu/zookeeper-3.4.6/logs clientPort=2181 initLimit=10 syncLimit=5 server.1=10.4.1.107:2888:3888 server.2=10.4.1.119:2888:3888 server.3=10.4.1.111:2888:3888 * create dataDir and dataLogDir directories * for each MQ node create a myid file in dataDir with the id of the server and nothing else. For node-1 the file will contain one line with 1, node-2 with 2, and node-3 with 3. * start ZooKeeper (on each node): \textbf{./zkServer.sh start} * check ZK status with: \textbf{./zkServer.sh status} * Configure ActiveMQ (apache-activemq-5.12.0/conf/activemq.xml file - set the hostname parameter to the node address) .. code-block:: none ... ... After ActiveMQ is installed and configured it can be started with the command: :command:./activemq start or ``./activemq console`` for a foreground process. **Oslo.messaging ActiveMQ Driver** All OpenStack changes (in the oslo.messaging library) to support ActiveMQ are already merged to the upstream repository. The relevant changes can be found in the amqp10-driver-implementation topic. To run ActiveMQ even on the most basic all-in-one topology deployment the following requirements need to be satisfied: * Java JRE must be installed in the system. The Java version can be checked with the command ``java -version``. If java is not installed an error message will appear. Java can be installed with the following command: ``sudo apt-get install default-jre`` * ActiveMQ binaries should be installed in the system. See http://activemq.apache.org/getting-started.html for installation instructions. The latest stable version is currently http://apache-mirror.rbc.ru/pub/apache/activemq/5.12.0/apache-activemq-5.12.0-bin.tar.gz. * To use the OpenStack oslo.messaging amqp 1.0 driver, the following Python libraries need to be installed: ``pip install "pyngus$>=$1.0.0,$<$2.0.0"`` ``pip install python-qpid-proton`` * All OpenStack projects configuration files containing the line ``rpc_backend = rabbit`` need to be modified to replace this line with ``rpc_backend = amqp``, and then all the services need to be restarted. **ZeroMQ Installation** This section describes installation steps for ZeroMQ. ZeroMQ (also ZMQ or 0MQ) is an embeddable networking library but acts like a concurrency framework. Unlike other AMQP-based drivers, such as RabbitMQ, ZeroMQ doesn't have any central brokers in oslo.messaging. Instead, each host (running OpenStack services) is both a ZeroMQ client and a server. As a result, each host needs to listen to a certain TCP port for incoming connections and directly connect to other hosts simultaneously. To set up ZeroMQ, only one step needs to be performed. * Install python bindings for ZeroMQ. All necessary packages will be installed as dependencies: ``sudo apt-get install python-zmq`` .. note:: python-zmq version should be at least 14.0.1 .. code-block:: none python-zmq Depends: python Depends: python Depends: python Depends: libc6 Depends: libzmq3 **Oslo.messaging ZeroMQ Driver** All OpenStack changes (in the oslo.messaging library) to support ZeroMQ are already merged to the upstream repository. You can find the relevant changes in the zmq-patterns-usage topic. To run ZeroMQ on the most basic all-in-one topology deployment the following requirements need to be satisfied: * Python ZeroMQ bindings must be installed in the system. * Redis binaries should be installed in the system. See http://redis.io/download for instructions and details. .. note:: The following changes need to be applied to all OpenStack project configuration files. * To enable the driver, in the section [DEFAULT] of each configuration file, the 'rpc_backend' flag must be set to 'zmq' and the 'rpc_zmq_host' flag must be set to the hostname of the node. .. code-block:: none [DEFAULT] rpc_backend = zmq rpc_zmq_host = myopenstackserver.example.com * Set Redis as a match making service. .. code-block:: none [DEFAULT] rpc_zmq_matchmaker = redis [matchmaker_redis] host = 127.0.0.1 port = 6379 password = None **Running ZeroMQ on a multi-node setup** The process of setting up oslo.messaging with ZeroMQ on a multi-node environment is very similar to the all-in-one installation. * On each node ``rpc_zmq_host`` should be set to its FQDN. * Redis-server should be up and running on a controller node or a separate host. Redis can be used with master-slave replication enabled, but currently the oslo.messaging ZeroMQ driver does not support Redis Sentinel, so it is not yet possible to achieve high availability, automatic failover, and fault tolerance. The ``host`` parameter in section ``[matchmaker_redis]`` should be set to the IP address of a host which runs a master Redis instance, e.g. .. code-block:: none [matchmaker_redis] host = 10.0.0.3 port = 6379 password = None Environment description ^^^^^^^^^^^^^^^^^^^^^^^ Test results must include used environment description. This includes: * Hardware used (servers, switches, storage, etc.) * Network scheme * Messaging bus specification and OpenStack version deployed (if any). Test Case 1: Message Queue Throughput Test ------------------------------------------ Description ^^^^^^^^^^^ This test measures the aggregate throughput of a MQ layer by using the oslo.messaging simulator tool. Either RabbitMQ, ActiveMQ, or ZeroMQ can be used as the MQ layer. Throughput is calculated as the sum over the MQ clients of the throughput for each client. For each test the number of clients/threads is configured to one of the specific values defined in the test case parameters section. The full set of tests will cover all the "Threads count" values shown, plus additional values as needed to quantify the dependence of MQ throughput on load, and to find the maximum throughput. Parameters ^^^^^^^^^^ ======================= =========== Parameter name Value ======================= =========== oslo.messaging version 2.5.0 simulator.py version 1.0 Threads count 50, 70, 100 ======================= =========== List of performance metrics ^^^^^^^^^^^^^^^^^^^^^^^^^^^ ======== ========== ================ =================================== Priority Value Measurment Units Description ======== ========== ================ =================================== 1 Throughput msg/sec Directly measured by simulator tool ======== ========== ================ =================================== Result Type ^^^^^^^^^^^ ================ ======================= ========================= Result type Measurement Units Description ================ ======================= ========================= Throughput Value msg/sec Table of numerical values Throughput Graph msg/sec vs # of threads Graph ================ ======================= ========================= Additional Measurements ^^^^^^^^^^^^^^^^^^^^^^^ =========== ======= ============================= Measurement Units Description =========== ======= ============================= Variance msg/sec Throughput variance over time =========== ======= ============================= Test Case 2: OMGBenchmark Rally test ------------------------------------ Description ^^^^^^^^^^^ OMGBenchmark is a rally plugin for benchmarking oslo.messaging. The plugin and installation instructions are available on github: https://github.com/Yulya/omgbenchmark Parameters ^^^^^^^^^^ ================================= =============== =============== Parameter name Rally name Value ================================= =============== =============== oslo.messaging version 2.5.0 Number of iterations times 50, 100, 500 Threads count concurrency 40, 70, 100 Number of RPC servers num_servers 10 Number of RPC clients num_clients 10 Number of topics num_topics 5 Number of messages per iteration num_messages 100 Message size msg_length_file 900-12000 bytes ================================= =============== =============== List of performance metrics ^^^^^^^^^^^^^^^^^^^^^^^^^^^ ======= ================= ========================================== Name Measurement Units Description ======= ================= ========================================== min sec Minimal execution time of one iteration median sec Median execution time 90%ile sec 90th percentile execution time 95%ile sec 95th percentile execution time max sec Maximal execution time of one iteration avg sec Average execution time success none Number of successfully finished iterations count none Number of executed iterations ======= ================= ========================================== Result Type ^^^^^^^^^^^ ================= ======================= ========================= Result type Measurement Units Description ================= ======================= ========================= Throughput Graph msg size vs median Graph Concurrency Graph concurrency vs median Graph ================= ======================= =========================