BOINC Berkeley Open Infrastructure for Network Computing

1
BOINCBerkeley Open Infrastructure for Network
Computing

• Danny Hughes
• Middleware Reading Group
• 7th November 2005

2
Context

• BOINC is a system for distributed processing over
  P2P.
• Geared towards volunteer projects / public
  resource computing.
• From the Berkeley space sciences lab, creators of
  Seti_at_Home.

3
Motivation

• Lots of spare CPU cycles out there!
• Seti_at_Home currently provides 70 TeraFLOPs
  (PetaFLOPs are possible).
• Despite success, very low uptake of PRC.
• Grid middleware is currently poorly matched with
  PRC.

4
The goals of BOINC

• Lower barriers to PRC.
• Share resources among projects.
• Support diverse applications.
• Reward participants.

5
Lowering Barriers to PRC

• Easy to install on a generic server
• Linux
• Apache
• PHP
• MySQL
• Python
• Designed to be lower in complexity than Grid
  middleware such as OGSA implementations.

6
Sharing Resources

• Each project runs its own server and is
  administratively separate.
• The BOINC client allows participation in multiple
  projects with the same client software.
• User can manage participation level in different
  projects.

7
Diverse Applications

• Existing applications written in C / C and
  Fortran can run.
• Application updates can be installed over BOINC
  without user intervention.

8
Rewarding Participants

• PRC depends heavily on perceived user rewards.
• BOINC includes a crediting system with protection
  from cheating.

9
Server Side
10
Scheduler

• Offers a range of scheduling options.
• Deadlines
• Initial instances of job
• Minimum WE size
• Upper bounds on erroneous, correct etc. from each
  work unit.

11
Bits Im glossing over

• User credit system.
• Database stuff.
• Performance optimisations.

12
Client Side

• Project executables can be written in several
  languages Fortran, C, C.
• Security maintained between client and server
  through signed packages.
• Results disseminated using XML.

13
Client Architecture

• Core Client
• Communicates with server.
• Scheduling between client processes.
• Cross Platform
• Client communicates with project executables via
  local XML/RPC.

14
Issues / Limitations.

• Very centralized architecture.
• Server HW becomes bottleneck.
• No autonomy of nodes.
• Not suitable for ad-hoc environments.
• Simplistic awareness.
• Only good for Seti-Like tasks.
• Decentralized alternatives such as GPU (Gnutella
  Processing Unit) exist, but are less widely used
  / more primitive.

15
Projects using BOINC

• SETI_at_home
• Predictor_at_home
• Folding_at_Home
• CERN