Jump to: navigation, search

Difference between revisions of "Rally"

(How amqp_rpc_single_reply_queue affects performance)
(Interesting results of Rally work)
Line 43: Line 43:
  
  
= Interesting results of Rally work =
+
= Rally in action =
  
 
=== How amqp_rpc_single_reply_queue affects performance ===
 
=== How amqp_rpc_single_reply_queue affects performance ===

Revision as of 13:05, 23 October 2013

Introduction

Rally is a Benchmark-as-a-Service project for OpenStack.

Rally is intended for providing the community with a benchmarking tool that is capable of performing specific, complicated and reproducible test cases on real deployment scenarios.

Rally flow diagram.png

In the OpenStack ecosystem there are currently several tools that are helpful in carrying out the benchmarking process for an OpenStack deployment. To name a few, there are DevStack and FUEL which are intended for deploying and managing OpenStack clouds, the Tempest testing framework that validates OpenStack APIs, some tracing facilities like Tomograph with Zipkin, and so on. The challenge, however, is to compile all these tools together on a reproducible basis. That can be a rather difficult task since the number of compute nodes in a practical deployment can be really huge and also because one may be willing to use lots of different deployment strategies that pursue different goals (e.g., while benchmarking the Nova Scheduler, one usually does not care of virtualization details, but is more concerned with the infrastructure topologies; while in other specific cases it may be the virtualization technology that matters). Compiling a bunch of already existing benchmarking facilities into one project, making it flexible to user requirements and ensuring the reproducibility of test results, is exactly what Rally does.

 

Use Cases

  1. Investigate how different deployments affect OS performance
    • Find the set of good OpenStack deployment architectures
    • Populate tables about deployment specification in case of different load (amount of controllers, swift nodes,...).
  2. Find the set of best Hardware for OpenStack cloud
  3. Automate production cloud specification generation
    • Max load for base cloud operations (VM start & stop, Block Device create/destroy & various OpenStack API methods)
    • Performance of base cloud operations in case of different load
  4. Automatically get numbers & profiling info about how your changes influence on OS performance
  5. Using Rally profiler detect scale & performance issues. E.g. when we delete 3 VMs by one request they are deleted one by one because of DB lock on quotas table

 

Architecture

Rally is basically split into 4 main components:

  1. Deploy Engine, which is responsible for processing and deploying VM images (using DevStack or FUEL according to user’s preferences). The engine can do one of the following:
    • deploying an OS on already existing VMs;
    • starting VMs from a VM image with pre-installed OS and OpenStack;
    • deploying multiple VMs each of which has running OpenStack compute node based on a VM image.
  2. Server Provider, provides servers (virtual servers) to deploy OpenStack.
  3. Benchmarking Tool, which carries out the benchmarking process in several stages:
    • runs Tempest tests, reduced to 5-minute length (to save the usually expensive computing time);
    • runs a set of benchmark scenarios (using the Rally testing framework);
    • collects all the test results and processes them by Zipkin tracer;
    • puts together a benchmarking report and stores it on the machine Rally was lauched on.
  4. Orchestrator, which is the central component of the system. It uses the Deploy Engine to run control and compute nodes and to launch an OpenStack distribution and, after that, calls the Benchmarking Tool to start the benchmarking process.


To dive deeper into the architectural concepts of Rally, see Rally architecture for developers.


Rally in action

How amqp_rpc_single_reply_queue affects performance

To show Rally's capabilities and potential we used NovaServers.boot_and_destroy scenario to see how amqp_rpc_single_reply_queue option affects VM bootup time. Some time ago it was shown that cloud performance can be boosted by setting it on so naturally we decided to check this result. To make this test we issued requests for booting up and deleting VMs for different number of concurrent users rangin from one to 30 with and without this option set. For each group of users a total number of 200 requests was issued. Averaged time per request is shown below:

amqp_rpc_single_replya_queue

So apparently this option affects cloud performance, but not in the way it was thought before.

How To

  1. Rally installation
  2. How to use Rally
  3. Available Deploy engines
  4. Available Server providers
  5. Available Benchmark scenarios
  6. Extend Rally functionality
  7. Rally Road Map

   

Links

Source

https://github.com/stackforge/rally

Pending Code Reviews

https://review.openstack.org/#/q/status:open+rally,n,z

Project space

http://launchpad.net/rally

Blueprints

active:    http://blueprints.launchpad.net/rally

v1 base: https://blueprints.launchpad.net/rally/+spec/init

Bugs

https://bugs.launchpad.net/rally

IRC chat

server: freenode.net

chanel: #openstack-rally