Skip to content
forked from ibmcb/cbtool

Cloud Rapid Experimentation and Analysis Toolkit #cloud

License

Notifications You must be signed in to change notification settings

maugustosilva/cbtool

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Welcome to the Cloud Rapid Experimentation and Analysis Toolkit

Cloud Rapid Experimentation and Analysis Tool (aka CBTOOL) is a framework that automates IaaS cloud benchmarking through the running of controlled experiments.

Subscribe to our mailing list:

New! CloudBench is now released as a component of SPEC Cloud IaaS 2018

To get started:

  1. Learn more about the tool

  2. Perform the initial installation

  3. Run the tool for the first time

  4. Prepare your cloud to be driven by the tool

  5. Prepare a VM image to be used with the tool

  6. Deploy your first Virtual Application

  7. Run simple experiments

  8. If needed, debug the initial setup

  9. Configure CBTOOL to run outside of the cloud (or with multiple tenant networks)

    1. Detailed instructions on how to use OpenVPN support
  10. Look at some auto generated plots made from the data collected.

  11. Read our latest paper IC2E 2013.

  12. Try administrating the tool with the Graphical Environment

    1. Then try monitoring your experiments with the Graphical Environment

    2. Try to customize your dashboard monitoring data with filters in the Graphical Environment

    3. You can also try to use the Graphical Wizard for a first-time connection.

  13. Read the Frequently Asked Questions

Features:

  • Automatic deployment and controlled execution of multiple multi-tier applications. Each individual application can have a specific load profile using probability distributions.
  • Adapters for multiple clouds (EC2 and OpenStack, among others), with a plugin structure that allows new cloud models to be added incrementally.
  • Can orchestrate different arrival rates and lifetimes for VMs using probability distributions.
  • Collects application and system (OS) performance data from hosts and guests in real time.
  • It is designed from the ground up to be highly scalable and parallel.
  1. Amazon EC2
  2. OpenStack (and RackSpace)
  3. Google Compute Engine
  4. DigitalOcean
  5. Docker/Swarm
  6. LXD/LXC
  7. Kubernetes
  8. Libvirt+KVM
  9. VMWare vCloud (NOT actively maintained)
  10. CloudStack (NOT actively maintained)
  11. SoftLayer

Want to add support for a new Cloud? Take a look at our Frequently Asked Questions

To get the most current list, start CBTOOL and type typelist on the CLI. To get more information about a given workload, typeshow <workload name>:

(MYCLOUD) typelist
AIs with the following types can be attached to to this experiment (Cloud MYSIMCLOUD) :

synthetic:
  bonnie (default, full)
  btest (default)
  coremark (default)
  ddgen (default)
  filebench (fileserver, oltp_noism, varmail, videoserver, webproxy)
  fio (default)
  iperf (tcp, udp)
  mlg (default)
  netperf (tcp_stream, tcp_maerts, udp_stream, tcp_rr, tcp_cc, tcp_crr, udp_rr)
  nuttcp (tcp, udp)
  postmark (default)
  unixbench (arithmetic, dhrystone, whetstone, load, misc, speed, oldsystem, system, fs, shell, index)
  xping (icmp)

application-stress:
  memtier (default)
  oldisim (default)
  wrk (default)

scientific:
  hpcc (default)
  linpack (default)
  multichase (simple, work:N, t0-2, nta, movdqa, mvntdqa, parallel2-10, critword:N, critword2:N)
  parboil (histo, bfs, stencil, mri-q, mri-gridding, lbm, tpacf, sad, spmv, sgemm, cutcp)
  scimark (default)

transactional:
  cassandra_ycsb (workloada, workloadb, workloadc, workloadd, workloade, workloadf)
  ibm_daytrader (default)
  mongo_ycsb (workloada, workloadb, workloadc, workloadd, workloade, workloadf)
  open_daytrader (default)
  redis_ycsb (workloada, workloadb, workloadc, workloadd, workloade, workloadf)
  specjbb (preset, hbir)
  sysbench (simple, complex, nontrx, sp)

data-centric:
  giraph (pagerank, topkpagerank)
  hadoop (sort, wordcount, terasort, dfsioe, nutchindexing, pagerank, bayes, kmeans, hivebench)

fake:
  nullworkload (default)

Contacts:

Marcio Silva marcios@us.ibm.com Michael R. Hines mrhines@digitalocean.com

About

Cloud Rapid Experimentation and Analysis Toolkit #cloud

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages

  • Python 70.8%
  • Shell 18.6%
  • HTML 6.2%
  • R 2.5%
  • JavaScript 0.4%
  • Filebench WML 0.4%
  • Other 1.1%