performance-docs/doc/source/methodologies/tools.rst

Tools

In distributed computing systems three basic components of the underlying telecommunications architecture can be found - control plane, data plane and management plane, that is commonly considered a subset of the control plane.

Classically the control plane is defined as the part of a network that carries signaling traffic and is responsible for routing. Control plane functions include system configuration and its management. The data plane in this definition is the part of a network that carries user traffic and, served by control plane, enables data transfer to and from system clients.

In case of very specific use case - testing of OpenStack clouds - we can treat these terms in a bit different light.

For cloud technologies and infrastructures (and OpenStack in particular) we can define control plane as a part of cloud architecture, that is responsible for cloud resources management, including their on demand availability, their scheduling, creation, modification and deletion. These operations are covered by various distributed cloud services, communicating via their unified APIs. Data plane in general can be considered as all data operations performed by the workloads on top of OpenStack infrastructure. In particular, the most common parts are network operations which are responsible for data transfer over network protocols between running on cloud VMs or between a VM and the external networks including Internet. Other parts of Data Plane layer are related to the storage operations which are used by the workloads to store data on a persistent storage.

All tools listed below are orienting on the cloud-specific definition of both control plane and data plane layers.

Control Plane (API) testing

To evaluate cloud control plane performance, it's vital to understand performance of resource management operations, that are served by cloud APIs. Therefore, in this section API testing tools will be listed.

JMeter

The leader of the pack in awareness is probably Apache JMeter. This is an open-source Java application whose key feature is a powerful and complete GUI which you use to create test plans. A test plan is composed of test components which define every piece of the test such as:

• Multiple threads to generate a load
• Parametrizing HTTP requests
• Flexible output results via listeneres interface

This tool is very well established and is probably one of the best tools for functional load testing. It allows to model complex flows using conditions and allows to create custom assertions to validate the behavior. It also allows to simulate non-trivial HTTP scenarios like logging in before the actual HTTP call to a specific URL or perform file uploads. JMeter has a wide and well established community which produces various plugins to modify and extend the built-in behaviors. JMeter allows to test not only HTTP based API but also supports various protocols including:

• Web - HTTP and HTTPS
• SOAP and REST API protocols (over HTTP)
• FTP
• Databases via JDBC Java DB interfaces
• LDAP
• Message oriented middleware MOM via JMS • Mail - SMTP, POP3 and IMAP
• MongoDB
• TCP over IP

And last, but not the least, JMeter is open source and free. As with every tool, it has its own limitations and problems. JMeter comes with GUI which has a steep learning curve. It is overloaded with options and concepts which one should learn before being able to use this tool efficiently. GUI consumes a lot of compute and memory resources, so, in order to reduce the performance impact, the GUI can be switched off and tests can be executed in non-GUI mode. It will require saving the test scenario in an XML formatted file and using CLI tool to start the test execution. The desired throughput of the requests is controlled by several parameters of the test scenario and requires fine-tuning before test execution.

Gatling

Gatling is a highly capable load testing tool. It is designed for ease of use, maintainability, and high performance.

Out of the box, Gatling comes with excellent support of the HTTP protocol that makes it a tool of choice for load testing of any HTTP server (including OpenStack controllers, as all cloud management processes running on them are using HTTP APIs to communicate). As the core engine is actually protocol agnostic, it is perfectly possible to implement support for other protocols. For example, Gatling currently also ships JMS support. Gatling’s architecture is asynchronous as long as the underlying protocol, such as HTTP, can be implemented in a non-blocking way. This kind of architecture lets us implement virtual users as messages instead of dedicated threads, making them very resource cheap. Thus, running thousands of concurrent virtual users is not an issue. Gatling uses its own DSL for the test scenarios.

Wrk and Apache AB

Wrk and Apache AB are command line tools to test HTTP based resources written in C. They only provide a command line interface but can generate higher HTTP load than tools written in Java or Python. Apache AB is single threaded and can be extended via plugins (currently plugins for Kafka and RabbitMQ tests). Wrk can be multi-threaded and is one of the most scalable HTTP traffic generator. Wrk2 is a version that has been corrected to generate very accurate latency measurement. A fork of wrk2 further provides compressed latency histograms that are ready to be sent over the network to any test orchestrator for aggregation.

Rally

Rally is a benchmarking tool that was designed specifically for OpenStack API testing. To make this possible, Rally automates and unifies multi-node OpenStack deployment, cloud verification, benchmarking & profiling. Rally does it in a generic way, making it possible to check whether OpenStack is going to work well on, say, a 1k-servers installation under high load. The actual Rally core consists of four main components, listed below in the order they go into action:

• Server Providers provide a unified interface for interaction with different virtualization technologies (LXS, Virsh etc.) and cloud suppliers (like Amazon): it does so via SSH access and in one L3 network
• Deploy Engines either deploy some OpenStack distribution (like DevStack, Fuel or others) or use the existing one before any benchmarking procedures take place, using servers retrieved from Server Providers
• Verification runs Tempest (or another subunit-based tool) against the deployed cloud to check that it works correctly, collects results and presents them in a human readable form
• Benchmark Engine allows to write parameterized benchmark tasks & run them against the cloud (and in fact, it is the most powerful part of this framework, that allows to simulate any kind of usual OpenStack control plane load)

Rally is written in Python language and can easily be extended with plugins written in Python.

Data Plane testing

For now there is no 100% ready data plane testing tool for OpenStack, that can cover all possible data workloads, although, there are still options to use and to improve, to be able to run complex testing data workloads against the cloud.

VMTP

VMTP is a data path performance measurement tool built specifically for OpenStack clouds. It was written to provide a quick, simple and automated way to get VM-level or host-level single-flow throughput and latency numbers from any OpenStack cloud and to take into account various Neutron topologies.

VMTP is a small Python application that will automatically perform ping connectivity, round trip time measurement (latency) and TCP/UDP throughput measurement for the following East/West flows on any OpenStack deployment:

• VM to VM same network (private fixed IP)
• VM to VM different network using fixed IP (same as intra-tenant L3 fixed IP)
• VM to VM different network using floating IP and NAT (same as floating IP inter-tenant L3)

Optionally, when an external Linux host is available for testing North/South flows:

• External host/VM download and upload throughput/latency (L3/floating IP)

For VM-related flows, VMTP will automatically create the necessary OpenStack resources (router, networks, subnets, key pairs, security groups, test VMs) using the public OpenStack API, install the test tools, then orchestrate them to gather the throughput measurements then cleanup all related resources before exiting.

VMTP also supports:

• throughput and latency for VMs using SR-IOV
• provider network

VMTP online documentation

KloudBuster

While VMTP is a single flow measurement tool, KloudBuster can load the data plane or storage infrastructure of any OpenStack cloud at massive scale and measure how well the cloud behaves under load. The performance data comes from where it matters: the VMs standpoint where cloud applications run.

KloudBuster HTTP traffic load:

• Can load the data plane with one OpenStack cloud (single-cloud operations for L3 East-West scale) or two OpenStack clouds (dual-cloud operations with one cloud hosting the HTTP servers and the other loading HTTP traffic for L3 North-South scale testing)
• Can stage any number of tenants, routers, networks, and HTTP servers (as many as your cloud can handle)
• Uses real HTTP servers (Nginx) running in real Linux images (Ubuntu 14.04)
• Uses high performance and highly scalable HTTP traffic generators (wrk2) to simulate huge number of HTTP users and TCP connections (hundreds of thousands to millions of concurrent and active connections)
• Overall throughput aggregation and loss-less millisecond-precision latency aggregation for every single HTTP request (typically millions per run)
• Traffic shaping to specify on which links traffic should flow
• Supports periodic reporting for real-time monitoring and results aggregation

KloudBuster Storage load:

• Can launch any number of VMs accessing storage using FIO running inside VMs
• Supports VM-level Cinder volume (block storage) or Ephemeral disk file I/O performance measurement
• Supports random and sequential file access mode and any mix of read/write, block size and queue depth
• Returns IOPs, bandwidth and loss-less sub millisecond-precision latency aggregations for every IO operation (typically millions per run)
• Supports user configurable workload sequence

KloudBuster supports automated progression runs on the VM count (e.g. measure load for 10 to 100 Vms, in increment of 10 VMs).

Automatic cleanup is performed upon termination to avoid stale resources leaving in the cloud under test.

KloudBuster is accessible to anybody with basic knowledge of OpenStack, installs in minutes and runs off the box in a fully automated way with sensible default workloads. Available to run from CLI, REST or Web User Interface... you pick what works best for you.

KloudBuster online documentation

Examples of KloudBuster results

Shaker

The Shaker tool is a tool used and developed by Mirantis to understand the Data Plane capabilities of an OpenStack deployment. Data Plane testing helps cloud administrators understand their deployment from the perspective of the applications that are using the environment. This tool can be used for deployment planning, environment verification, and troubleshooting.

Today, Shaker focuses on network based tests using iperf to drive load across the network. Shaker has future plans to roll out testing to evaluate I/O and CPU.

Shaker utilizes Heat (OpenStack Orchestration) templates to deploy and execute Data Plane tests. It deploys a number of agents/compute nodes that all report back to a centralized Shaker server.

The server is executed by shaker command and is responsible for deployment of instances, execution of tests as specified in the scenario, for results processing and report generation. The agent is light-weight and polls tasks from the server and replies with the results. Agents have connectivity to the server, but the server does not (so it is easy to keep agents behind NAT).

Shaker runs three types of network tests with many different options (including TCP and UDP). Below is a summary of the tests and their characteristics:

• type of the test:
  • VMs in the same network (L2)
  • VMs in a different network (L3 East/West)
  • VMs hitting external IP addresses (L3 North/South)
• communication: either floating IPs or SNAT/internal
• number of VMs: from 1 to N/2, where N is number of compute nodes available
• external hosts to use: static hard coded
• VM placement:
  • one VM per compute
  • two VMs per compute
  • two VMs per compute (different networks)

Shaker L2 Segment Topology

With VMs in the same network (L2 network test), Shaker deploys two VMs in the same network using Heat templates, and runs iperf between them, measuring the single stream network performance between.

Shaker L3 East-West Topology

With VMs in different networks (L3 east/west), Shaker deploys two VMs in different networks using Heat templates, and runs iperf between them, measuring the single stream network performance between them. This will involve routing and will test the performance of the deployed SDN overlay.

Shaker L3 North-South Topology

The last case is about VMs hitting external IP addresses (L3 north/south). Shaker deploys one of the VMs with an external (floating) IP address, and runs iperf between the some given external node and the VM.

Rally

Although right now Rally is used for control plane testing, there is the approved blueprint for it to support various workloads testing, that means that in future it will be possible to use Rally for all data plane testing as well.

Browbeat

Browbeat is a powerful combination of several upstream tools for performance testing, tuning, comparison and monitoring of an OpenStack cloud. It uses Ansible to automate deploying infrastructure such as:

• Graphite: To serve as a store for system performance metrics being pushed from the cloud by Collectd
• Grafana: To visualize the system performance data in Graphite and act as a means of actively monitoring the cloud real-time to troubleshoot performance bottlenecks
• ELK Stack:
  • Elasticsearch: To act as a storage backend for performance testing data, metadata about how the cloud is setup and log data
  • Logstash: To manage the logs
  • Kibana: To Visualize data from Elasticsearch and aid in performance comparison

The tool also sets up Collectd on the OpenStack environment so that metrics are reported to Graphite. All the user needs to do is fill in minimal variables such as IP address, username, password of the infrastructure hosts and Browbeat will setup the infrastructure hosts and transport system metrics form the OpenStack environment so that the user can actively monitor the cloud through Grafana. Dashboards that visualize OpenStack related data can be uploaded to Grafana by simply running a playbook Playbooks to deploy containerized Graphite/Grafana are also available. Metrics such as CPU, memory utilization, process/thread counts of OpenStack processes along with general system metrics are collected.

Other important features of Browbeat are:

• Check Overcloud and Undercloud for common configuration/tuning errors and print out a bug report.
• Run control/data plane workloads from upstream tools such as Rally, Shaker and Perfkit by simply passing a configuration file. Browbeat orchestrates the entire run to include multiple benchmarks at varying concurrencies/configurations and creates a result directory for all results
• Pushes the results to the Elasticsearch Instance configured, along with metadata that represents how the cloud is configured(Nova, Neutron, Ceilometer, etc.) so that run-to-run, build-to-build and cloud-to-cloud comparisons can be done