1. Designing a Runtime System for Volunteer Computing David P
Designing a Runtime System for Volunteer Computing David P. Anderson – UC Berkeley Carl Christensen – Oxford University Bruce Allen – UW Milwaukee and Max Planck Institute

2. Volunteer computing Distributed computing using volunteered computers
Early projects: GIMPS, distributed.net,
What makes volunteer computing hard?
anonymous, untrusted
heterogeneous, intermittent, high churn
require incentives
What makes it interesting?
More/cheaper computing power than other paradigms
Leverages consumer product R&D
Education/outreach

3. (constantly changing)
BOINC
Software platform for volunteer computing
Open-source (LGPL), NSF-funded
Supports wide range of application characteristics
Encourages multi-project participation
Goal: enable an ecosystem, not a static set of projects
Projects
(constantly changing)
Volunteers

4. Some BOINC-based projects
Climateprediction.net
Oxford; global climate change study
LIGO; gravitational wave astronomy
U. Washington; protein study
U.C. Berkeley; SETI
CERN; accelerator and collision simulation
Swiss Tropical Inst., U. of Geneva; malaria epidemiology
Chess960
game study
...and about 30 others

5. Computing power and cost
~450,000 active computers
Hardware cost: < $1/GFLOP
Future: game consoles (e.g. Playstation 3)
ExaFLOP computing in 2007?

6. BOINC client structure
schedulers, data servers
screensaver
application
local TCP
core client
BOINC library
GUI
Runtime system

7. Goals and requirements
Volunteer-oriented features
preferences (compute while idle, time-of-day, CPU throttling, RAM usage limits, etc.)
credit accounting
autonomic
Support diverse apps
compound apps
FORTRAN, legacy apps
long jobs
non-CPU-intensive apps
Heterogeneity
Win 98+, Mac OS X, all Unix/Linux

8. Shared-memory message passing

9. Thread structure, simple apps

10. Compound applications

11. Runtime system functionality
Task management
scheduling: suspend/resume/quit/abort
orphaned and duplicate processes
reliable termination
Status and credit
CPU time, fraction done reporting
app-specific benchmarks, FLOP reporting
Checkpoint handshaking
Error handling and reporting
symbolic stack trace on app crashes
detection/abort if exceed CPU/mem/disk limits

12. Runtime system functionality, cont.
Application graphics
app supplies render() function
Long-running applications
trickle messages
intermediate file upload
Virtual/physical filename resolution

13. Conclusion Volunteer computing a new computing paradigm
leader in FLOPS, price/performance
involves public in science
BOINC: enabling technology
Ongoing/future work (technical)
sandboxing
distributed storage
low-latency computing
Bitorrent integration
Ongoing/future work (organizational)
Meta-projects: Virtual Campus Supercomputing Center, IBM World Community Grid, etc.
Account management systems