Volunteer Computing with BOINC Dr. David P. Anderson University of California, Berkeley SC10 Nov. 14, 2010
Goals Explain volunteer computing Teach how to create a volunteer computing project using BOINC Target audience: High-throughput computing users Technical skills: Basic Linux/Apache sysadmin, familiarity with PHP, SQL and XML, C/C++ (optional)
Outline Why use volunteer computing? Basic concepts of BOINC Developing BOINC applications (15 minute break) Deploying a BOINC server Deploying applications Submitting jobs Organizational issues
Part 1: Why use volunteer computing?
The Consumer Digital Infrastructure 1 billion PCs current GPUs: 1 TeraFLOPS (1,000 ExaFLOPS total) Storage: ~1,000 Exabytes Commodity Internet: 10-1,000 Mbps to home Consumers pay for hardware sysadmin network costs electricity
Volunteer computing PC owners donate computing resources to projects (e.g., computational science) Applications run at zero priority while PC in use, and/or while PC is not in use
Examples Project start where area peak #hosts GIMPS 1994 math 10,000 distributed.net 1995 cryptography 100,000 SETI@home I 1999 UCB SETI 600,000 Folding@home 1999 Stanford biology 200,000 United Devices 2002 commercial biomedicine 200,000 CPDN 2003 Oxford climate change 150,000 LHC@home 2004 CERN physics 60,000 Predictor@home 2004 Scripps biology 100,000 WCG 2004 commercial biomedicine 200,000 Einstein@home 2005 LIGO astrophysics 200,000 SETI@home II 2005 UCB SETI 850,000 Rosetta@home 2005 U. Wash biology 100,000 SIMAP 2005 T.U. Munich bioinformatics 10,000...............
Current status ~50 projects 500,000 vounteers 800,000 computers
# processors 1 100 1000 Grid multiple jobs cluster (batch) single job cluster (MPI) 10K-1M Volunteer computing Commercial cloud High-throughput computing supercomputer High-performance computing
Volunteer computing is different You don t buy resources; you ask for them Resources are: - heterogeneous - sporadically available and connected - untrusted and not private - behind firewalls/nats/proxies
Part 2: Basic concepts of BOINC
About BOINC Funded by NSF since 2002 Open-source (LGPL) Based at UC Berkeley Few staff, but lots of volunteers software testing translation documentation support (email lists, message boards, Skype)
Volunteers and projects volunteers projects CPDN LHC@home attachments WCG
BOINC software overview scheduler MySQL daemons data server HTTP project server GUI client screensaver apps volunteer host
BOINC scheduler applications - HW, SW description - existing workload - per resource type: # of instances requested # of seconds requested Win32 Win64 app versions Win32 + NVIDIA Win32 N-core Mac OS X - app version descriptions - job descriptions jobs instances
Job replication Job instances may fail or return wrong results Job replication: do 2, see if they agree - agree may be fuzzy Homogeneous replication - numerical equivalence of hosts Adaptive replication - reduce replication for hosts that seem trustworthy
The job pipeline work generator BOINC validator assimilator
The BOINC data model App versions, job inputs, job output can consist of arbitrarily many files Each file has a physical name (unique, immutable); each reference to a file has a logical name Files have various attributes (e.g., sticky) Each file can have one or more URLs, and are transferred via HTTP App version files are digitally signed
What kinds of jobs can BOINC handle? Pretty much anything you d run on a Grid Bag of tasks (but IPC support soon) Short/long jobs Data intensive, up to a point Geared towards - Few apps, many jobs (high startup cost per app) - Jobs with high slack time
Part 3: Application development for BOINC
The BOINC runtime environment processes files
Native BOINC applications boinc_init() - create runtime system thread boinc_finish() - write finish file boinc_resolve_filename(logical, physical) boinc_fraction_done(x)
Checkpointing bool boinc_time_to_checkpoint() - call when in checkpointable state boinc_checkpoint_done()
The BOINC wrapper Can use for legacy apps XML input file lists sub-jobs - executable, input files What it does: - interfaces to BOINC client - copies files to/from slot directory - runs executables - does checkpointing at sub-job level
Building app versions Linux - gcc Windows - Visual Studio - mingw (gcc) Mac OS X - xcode
Multithread apps boinc_init_parallel() Allows suspend/resume of all threads - Unix: fork/exec - Windows: direct thread control
GPU app versions Develop for NVIDIA or ATI, with CUDA, CAL, OpenCL, etc. (BOINC supplies samples) Each version has a plan class For each plan class, supply a function that determines - can app run on this host? hardware, driver version, etc. - what resources will it use? #CPUs, #GPUs, GPU RAM, etc.
VM apps Develop apps on your favorite OS Create a VirtualBox VM image App version consists of - VM wrapper (supplied by BOINC) - VM image - app executable
Part 4: Deploying a BOINC server
Hardware options Native Linux host - download/compile BOINC software BOINC server VM (VMware/Debian) BOINC Amazon EC2 image
Components of a project Master URL name MySQL database Directory hierarchy A set of daemon processes and cron jobs
Processes clients scheduler feeder work generator validator assimilator transitioner file deleter DB purger MySQL DB
Project directory hierarchy apps/ application files bin/ daemon programs cgi-bin/ BOINC scheduler and upload GCI config.xml configuration file download/ downloadable files html/ web site; master URL points here keys/ keys for code signing, upload auth log_(hostname) daemon log files project.xml list of platforms and apps upload/ uploaded files
BOINC database platform app app_version user host workunit result...
Creating a project make_project name creates - directory hierarchy - DB - mods for httpd.conf - crontab entry
Project configuration and control config.xml - scheduling and other options - list of daemons - list of periodic tasks project control - bin/start: start daemons, enable scheduler - bin/stop: stop daemons, disable scheduler - bin/status
Scaling a BOINC server Components can run on different machines sharing a file system Each component can be distributed MySQL server is typically the bottleneck 1 server machine can issue ~100K jobs/day; 4 machines can issue > 1 million
Part 5: Deploying applications
Adding an application edit project.xml <app> <name>multi_thread</name> <user_friendly_name>test multi-thread apps</user_friendly_name> </app> run bin/xadd
Adding an application version Create application version directory apps/ uppercase/ uppercase_6.14_windows_intelx86 cuda.exe/ uppercase_6.14_windows_intelx86 cuda.exe graphics_app=uppercase_graphics_6.14_windows_intelx86.exe logo.jpg Helvetica.txf Sign files on offline computer run bin/update_versions
Part 6: Submitting jobs
Describing job inputs Input template file <file_info> <number>0</number> </file_info> <workunit> <file_ref> <file_number>0</file_number> <open_name>in</open_name> </file_ref> <target_nresults>1</target_nresults> <min_quorum>1</min_quorum> <command_line>-cpu_time 60</command_line> <rsc_fpops_bound>446797000000000</rsc_fpops_bound> <rsc_fpops_est>279248000000000</rsc_fpops_est> </workunit>
Describing job outputs Output template file <file_info> <name><outfile_0/></name> <generated_locally/> <upload_when_present/> <max_nbytes>5000000</max_nbytes> <url><upload_url/></url> </file_info> <result> <file_ref> <file_name><outfile_0/></file_name> <open_name>out</open_name> </file_ref> </result>
Submitting a job Stage input files cp test_files/12ja04aa `bin/dir_hier_path 12ja04aa` Submit job create_work appname A wu_name B wu_template C result_template D
Part 7: Organizational issues
Single-scientist projects Need to: Port apps Get publicity interface with public maintain servers Not many research groups have the resources And it creates a lot of competing brands
Umbrella projects Project publicity web development sysadmin app porting Example: IBM World Community Grid
The Berkeley@home model A university has scientists a powerful brand PR resources IT infrastructure lots of alumni (UCB: 500,000)
Hubs nanohub: science portal for nanoscience social network + app store sharing of ideas, data, software computational portal HUBzero: generalization to other areas currently ~20 hubs Integration of BOINC with HUBzero each hub has a volunteer computing project