Skip to content

Latest commit

 

History

History
100 lines (90 loc) · 5.44 KB

README.md

File metadata and controls

100 lines (90 loc) · 5.44 KB
Raw

COSTREAM Management Toolkit

This repository contains several Ansible Playbooks to set-up and control multiple clusters of nodes to run distributed frameworks such as Apache Storm, Apache Zookeeper, Apache Kafka, Hadoop and MongoDB.

Currently, OpenStack and Cloud-Lab hypervisors are supported. This toolkit is able to create instances, interconnect them and to set-up different big-data frameworks. Furthermore, it is used to generate training data for COSTREAM - a learned cost estimation model for Distributed Stream Processing Systems.

Overview

The folder playbooks contains several Ansible-Playbooks. In general, there are three folders of playbooks:

  • openstack: Openstack-specific playbooks to create and destroy instances
  • infrastructure: Setting up distributed services on a given cluster
  • costream: Scripts to use for COSTREAM training data generation The folder conf contains configuration files with metadata that is required to run the scrupts
  • The folder cluster will be created during operation. It contains the required SSH-config-files and ansible-files. These are required to connect to the cluster propertly.

General Use

To run a particular playbook, e.g. the deploy_storm playbook against set of nodes, run the manager-script with the following command:

python3 manager.py --config ./config.ini --operation deploy_storm

It has the following arguments:

  • --config: Path to configuration file with some metadata to use in the further executions. See template under conf for more information.
  • --operation: Name of the playbook to run, e.g. deploy_storm

Create instances with OpenStack

In case you have access to an OpenStack-cluster and no resources are given, it is possible to create a set of instances by using this toolkit.

  • Creating a cluster: 
    python3 manager.py --config ./config.ini --operation create_cluster
  • The new cluster is set-up in a private network with a public IP-address, provided as floating-IP. This serves as entrypoint for the ssh-connections. After creation, relevant config-files are created under the cluster-folder that are later used to connect to the cluster again. Make sure that you have enough quota to create your nodes and that your environment file is sourced.
  • Removing a cluster: 
    python3 manager.py --config ./config.ini --operation remove_cluster

Use instances from CloudLab

In case that you use and access instances from CloudLab, it is possible to extract a rspec-file from the ongoing experiment. Please open up your experiment and find the .rspec-file under the tab: Manifest This file can be converted by the script convert_rspec to create the necessary set of config-files to connect to the cluster. The usage is as following. Please replace {cluster_name} with a meaningful name of your choice.

python3 convert_rspec.py --input /path/to/rspec/cluster.rspec --output ./cluster/{cluster_name} --identifier {cluster_name}

Please note that the CloudLab-instances are publically available and this might be attacked. Please consider additional security measures.

Prepare a cluster

Given a cluster as a set of nodes from any hypervisor and the necessary config-files, this toolkits prepares the nodes to work as a cluster by the playbook:

python3 manager.py --config ./config.ini --operation prepare_instances

This playbook sets up NTP for all the nodes and makes them visible & SSH-accessible by each other.

Install distributed frameworks

As usually given in distributed settings, the services are generally set-up in a master-slave-architecture, where one instance works as master and the other as workers.

The following distributed frameworks are supported:

  • Apache Zookeeper: The master runs a Zookeeper service that is used for other frameworks as a basis. It runs under port 2181.
  • Apache Storm: The master runs the nimbus-service of Storm, and the workers are running the supervisor-services. In addition, there is a GUI provided that runs on the master to display the cluster status. It uses the ports 6627 (Storm) and 8081 (GUI). Storm requires a Zookeeper installation.
  • MongoDB: A MongoDB-Server is set up as servoce mongod on the master. The workers receive a client installation, to query the master via the command line. Note that a password authentication is afforded.
  • Apache Kafka: The master runs a Kafka Broker kafka that can be queried by any node. It runs under port 9092.

To install one or many frameworks, e.g. Apache Kafka and Apache Zookeeper, please call:

python3 manager.py --config ./config.ini --operation deploy_zookeeper deploy_kafka

Prepare COSTREAM training data generation

  • To prepare the cluster for the execution of DSPS queries and the observation of query costs, please call the prepare_experiments script. This script uploads the modified storm jars to the cluster and makes necessary configurations. It furthermore enables cgroups in case they aren't yet.
  • To only upload the storm-jars, call upload_jars
  • To reset the full cluster (e.g. when a query execution has crasehd) it is recommended to reset the cluster with reset_cluster. This clears the logs, restarts the services and clears the central MongoDB database.

Troubleshooting

  • Sometimes the playbook prepare_isntances hangs at the very initial apt-update-command. Stopping and restarting the playbook might help here.