Volunteer Computing and Cloud Computing: Opportunities for Synergy

Transcription

1 Volunteer Computing and Cloud Computing: Opportunities for Synergy Derrick Kondo INRIA, France

2 Performance vs. Reliability vs. Costs high Cost Reliability high low low low Performance high

3 Performance vs. Reliability vs. Costs high Supercomputers Grids Clusters Cost Reliability high low low low Performance high

4 Performance vs. Reliability vs. Costs high Supercomputers Grids Clusters Cost Reliability high Clouds low low low Performance high

5 Performance vs. Reliability vs. Costs high Supercomputers Grids Clusters Cost Reliability high Clouds low low low Performance Volunteer Computing high

6 Performance vs. Reliability vs. Costs high Supercomputers Grids Clusters Cost Reliability high Clouds Clouds + Volunteer Computing low low low Performance Volunteer Computing high

7 Cost of 1 TFLOPS-year Cluster: $145K Computing hardware; power/ac infrastructure; network hardware; storage; power; sysadmin Cloud: $438K Volunteer: $1K - $10K Server hardware; sysadmin; web development

8 Performance Current 500K people, 1M computers 6.5 PetaFLOPS (3 from GPUs, 1.4 from PS3s) Potential 1 billion PCs today, 2 billion in 2015 GPU: approaching 1 TFLOPS How to get 1 ExaFLOPS: 4M GPUs * 0.25 availability How to get 1 Exabyte: 10M PC disks * 100 GB

9 History of volunteer computing now distributed.net, GIMPS s Middleware SETI@home, Folding@home Academic: Bayanihan, Javelin,... Climateprediction.net Predictor@home IBM World Community Grid Einstein@home Rosetta@home... Commercial: Entropia, United Devices,... BOINC

10 The BOINC computing volunteers ecosystem projects CPDN attachments WCG Directed and developed by David Anderson, UC Berkeley Projects compete for volunteers Volunteers make their contributions count Optimal equilibrium

11 What apps work well? Bags of tasks parameter sweeps simulations with perturbed initial conditions compute-intensive data analysis Job granularity: minutes to months Native, legacy, Java, GPU soon: VM-based

12 Example projects Climateprediction.net IBM World Community Grid Primegrid

13 Creating a volunteer computing project Server Set up a server Develop software for job submission and result handling System, DB admin (Linux, MySQL) Client Port applications, develop graphics Marketing Develop web site Publicity, volunteer communication

14 How many CPUs will you get? Depends on: PR efforts and success public appeal availability of internal resources 12 projects have > 10,000 active hosts 3 projects have > 100,000 active hosts

15 Organizational issues Creating a volunteer computing project has startup costs and requires diverse skills This limits its use by individual scientists and research groups Better model: umbrella projects Institutional Lattice Corporate IBM World Community Grid Community AlmereGrid

16 Summary Volunteer computing is an important paradigm for high-throughput computing price/performance performance potential Low technical barriers to entry (due to BOINC) Organizational structure is critical

17 Performance vs. Costs vs. Reliability high Supercomputers Grids Clusters Cost Reliability high Clouds Clouds + Volunteer Computing low low low Performance Volunteer Computing high

18 Cloud Background Vision Hide complexity of hardware and software management from a user by offering computing as a service Benefits Pay as you go Scale up or down dynamically No hardware management, less software management

19 Outline Performance tradeoffs Monetary tradeoffs Client hosting Server hosting

20 Apples to Apples Loosely-coupled, high-throughput, compute-intensive applications Tightly-coupled, data-intensive real-time applications low complexity high complexity VC s Clouds Comparison assuming embarrassingly parallel, compute-intensive applications

21 Method Use real performance measurements Exported BOINC project data Use real costs Large/small BOINC projects / XtremLab) Amazon Elastic Computing Cloud (EC2)

22 Stages of Project & Platform Construction Deployment Execution Completion

23 Platform Construction Deployment Execution Completion How long before I get X TeraFLOPS?

24 Platform Construction Deployment Execution Completion How long before I get X TeraFLOPS? Number of cloud nodes TeraFLOPS Months for registration

25 Platform Construction Deployment Execution Completion How long before I get X TeraFLOPS? Number of cloud nodes Can get over 20 TeraFLOPS within 6 months TeraFLOPS Months for registration

26 Platform Construction Deployment Execution Completion How long before I get X TeraFLOPS? Strategy: Add to BOINC project list Press releases Forum Announcements Google Ad Sense Respond to users (leverage volunteers) Number of cloud nodes Can get over 20 TeraFLOPS within 6 months Months for registration TeraFLOPS

27 Platform Construction Deployment Execution Completion How long to deploy my batch of tasks needing faster response time?

28 Platform Construction Deployment Execution Completion How long to deploy my batch of tasks needing faster response time? 7,82/9*,34.4:*,.;*:3)4,6<!" '!" &!"!!" ".!"!!.!"".4=6>6!""".4=6>6!"""".4=6>6!" #!" $!" % ()*+,-./0.1/2)34,,-.3/5,6

29 Platform Construction Deployment Execution Completion How long to deploy my batch of tasks needing faster response time? 7,82/9*,34.4:*,.;*:3)4,6<!" '!" &!"!!" " For 1000 tasks, ~10 minutes with 10 5 hosts.!"!!.!"".4=6>6!""".4=6>6!"""".4=6>6!" #!" $!" % ()*+,-./0.1/2)34,,-.3/5,6

30 Platform Construction Deployment Execution Completion How long to deploy my batch of tasks needing faster response time? Strategy: Specify lower latency bounds [Heien et al.] 7,82/9*,34.4:*,.;*:3)4,6<!" '!" &!"!!" "!"!!. For 1000 tasks, ~10 minutes with 10 5 hosts!"".4=6>6!""".4=6>6!"""".4=6>6!" #!" $!" % ()*+,-./0.1/2)34,,-.3/5,6.

31 Platform Construction Deployment Execution Completion

32 Platform Construction Deployment Execution Completion How many volunteer nodes are equivalent to 1 cloud node?

33 Platform Construction Deployment Execution Completion How many volunteer nodes are equivalent to 1 cloud node?

34 Platform Construction Deployment Execution Completion How many volunteer nodes are equivalent to 1 cloud node? 2.8 active volunteer hosts per 1 cloud node. (Total performance still orders of magnitude better)

35 Platform Construction Deployment Execution Completion How many volunteer nodes are equivalent to 1 cloud node? Strategy: Use statistical prediction of availability 2.8 active volunteer hosts per 1 cloud node. (Total performance still orders of magnitude better)

36 Platform Construction Deployment Execution Completion

37 Platform Construction Deployment Execution Completion How long should I wait for task completion?

38 Platform Construction Deployment Execution Completion How long should I wait for task completion?

39 Platform Construction Deployment Execution Completion How long should I wait for task completion? Median project latency bound: 9 days for 3.7 hour work unit (on 3GHz host). Ratio of lat. bound / exec time > 5. Good success rates: 96.1% of WCG tasks met out of 227,000 tasks

40 Platform Construction Deployment Execution Completion How long should I wait for task completion? Strategy: See BOINC Catalog for typical deadlines and compute/comm/mem ratios. Median project latency bound: 9 days for 3.7 hour work unit (on 3GHz host). Ratio of lat. bound / exec time > 5. Good success rates: 96.1% of WCG tasks met out of 227,000 tasks

41 Monetary Tradeoffs Client hosting on cloud Possible solution Not worth it and never will Server hosting on the cloud More details in HCW, 2009 paper...

42 Performance vs. Costs vs. Reliability high Supercomputers Grids Clusters Cost Reliability high Clouds Clouds + Volunteer Computing low low low Performance Volunteer Computing high

43 Motivation Non-dedicated resources are cheap, but unreliable VC are cheaper by at least order of magnitude Amazon EC2 Spot instances at least 50% compared to dedicated ones Clouds are more expensive but highly reliable Amazon SLA: 99.95% availability s have different levels of performance, cost, and reliability requirements

44 Goal & Approach Given reliability requirements, what is the cheapest combination of dedicated and undedicated resources Improve reliability of VC resources with prediction and reliability ranking Compute optimal mixture that satisfies reliability requirement while minimizing cost

45 Resource Provisioning using Predictions User specifies availability guarantee (ag) and # of desired hosts (n). System determines # of given hosts (N) and # of dedicated hosts (d) N = # given hosts n = # desired hosts d = # dedicated hosts working set for intervali selection of non-d. hosts initial data migration host replacement and data migration!!!!!!!!!!!!!!!!!! non-dedicated hosts dedicated hosts i-th prediction interval (length pil in hours) selection of replacements for failed non-d. hosts

46 Pareto Optimal for Availability >=.90 for 50 requested hosts 30 undedicated hosts out of 55 given minimizes monetary costs

47 Summary Prediction: use of ranking to predict availability of a collection of hosts Operational model: identify cost-optimal mixture of undedicated and dedicated hosts given reliability requirements

48 Overall Summary Volunteer Computing Good for high-throughput applications at low cost Cloud Computing At least order of magnitude more expensive, but higher reliability Volunteer and Cloud Computing Can support applications with diverse reliability requirements while minimizing costs

49 Thank you Acknowledgement: David P. Anderson s material on BOINC