1. Labs of The World, Unite!!! Walfredo Cirne walfredo@dsc.ufcg.edu.br Universidade Federal de Campina Grande

2. eScience Computers are changing scientific research −Enabling collaboration −As investigation tools (simulations, data mining, etc...) As a result, many research labs around the world are now computation hungry Buying more computers is just part of answer Better using existing resources is the other

3. Solution 1: Globus Grids promise “plug on the wall and solve your problem” Globus is the closest realization of such vision −Deployed for dozens of sites But it requires highly-specialized skills and complex off-line negotiation Good solution for large labs that work in collaboration with other large labs −CERN’s LCG is a good example of state-of-art

4. Solution 2: Voluntary Computing SETI@home, FightAIDS@home, Folding@home, YouNameIt@home have been a great success, harnessing the power of millions of computers However, to use this solution, you must −have a very high visibility project −be in a well-known institution −invest a good deal of effort in “advertising” −have a very good computer science team to run “the server”

5. And what about the thousands of small and middle research labs throughout the world which also need lots of compute power?

6. Solution 3: OurGrid OurGrid is easy to install and automatically configures itself Labs can freely join the system without any human intervention OurGrid is a peer-to-peer grid −Each lab correspond to a peer in the system To keep it doable, we focus on Bag-of-Tasks application

7. Bag-of-Tasks Applications Data mining Massive search (as search for crypto keys) Parameter sweeps Monte Carlo simulations Fractals (such as Mandelbrot) Image manipulation (such as tomography) And many others…

8. OurGrid Components OurGrid: A peer-to-peer network that performs fair resource sharing among unknown peers MyGrid: A broker that schedules BoT applications SWAN: A Xen-based sandbox that makes it safe running a computation for an unknown peer

9. OurGrid Architecture User Interface Application Scheduling User Interface Application Scheduling Site Manager Grid-wide Resourse Sharing Site Manager Grid-wide Resourse Sharing SWAN Sandboxing MyGrid 1 SWAN... n

10. An Example: Factoring with MyGrid task: init: put./Fat.class $PLAYPEN remote: java Fat 3 18655 34789789799 output-$TASK final: get $PLAYPEN/output-$TASK results task: init: put./Fat.class $PLAYPEN remote: java Fat 18656 37307 34789789799 output-$TASK final: get $PLAYPEN/output-$TASK results task: init: put./Fat.class $PLAYPEN remote: java Fat 37308 55968 34789789799 output-$TASK final: get $PLAYPEN/output-$TASK results....

11. MyGrid GUI

12. Network of Favors OurGrid forms a peer-to-peer community in which peers are free to join It’s important to encourage collaboration within OurGrid (i.e., resource sharing) −In file-sharing, most users freeride OurGrid uses the Network of Favor −All peers maintain a local balance for all known peers −Peers with greater balances have priority −The emergent behavior of the system is that by donating more, you get more resources −No additional infrastructure is needed

13. Network of Favors NoF is a local reputation system based on past history of service among peers. Peer A keeps track of work it has done for others v(A,B) and work done for it v(B,A). Reputation calculated based on past service provided. Users who are the best contributors of past processing power get higher priority access to cycles.

14. Network of Favors T A (B) : reputation of peer B at node A Function of v(A,B) and v(B,A) Example 1: T A (B) = v(B,A) - v(A,B) Vulnerable to identity changing attack. Example 2: T A (B) = max(0, v(B,A) - v(A,B)) Robust to identity changing attack, but cannot distinguish between malicious ID-changing node who never donates and collaborating peer who has. Example 3: T A (B) = max(0,v(B,A) - v(A,B) + log(v(B,A)) Add history term that reflects for peer A the history of its donations from B.

15. A B C D E NoF at Work [1] ConsumerFavor ProviderFavorReport * * * = no idle resources now broker B60 D45

16. NoF at Work [2] A B C D E B60 D45 E 0 ConsumerQuery ProviderWorkRequest * * = no idle resources now * broker

17. Free-rider Consumption Epsilon is the fraction of resources consumed by free-riders

18. Equity Among Collaborators

19. Scheduling with No Information Grid scheduling typically depends on information about the grid (e.g. machine speed and load) and the application (e.g. task size) However, getting good information is hard Can we schedule without information and deploy the system now? −This would make the system much easier to deploy and simple to use

20. Work-queue with Replication Tasks are sent to idle processors When there are no more tasks, running tasks are replicated on idle processors The first replica to finish is the official execution Other replicas are cancelled

21. Evaluation 8000 experiments Experiments varied in −grid heterogeneity −application heterogeneity −application granularity Performance summary:

22. WQR Overhead Obviously, the drawback in WQR is cycles wasted by the cancelled replicas Wasted cycles:

23. Data Aware Scheduling WQR achieves good performance for CPU- intensive BoT applications However, many important BoT applications are data-intensive These applications frequently reuse data −During the same execution −Between two successive executions

24. Storage Affinity Storage Affinity uses replication and just a bit of static information to achieve good scheduling for data intensive applications Storage Affinity uses information on which data servers have already stored a data item

25. Storage Affinity Results 3000 experiments Experiments varied in −grid heterogeneity −application heterogeneity −application granularity Performance summary: Storage AffinityX-SuffrageWQR Average (seconds)57.04659.523150.270 Standard Deviation39.60530.213119.200

26. SWAN: OurGrid Security Running an unknown application that comes from an unknown peer is a clear security threat We leverage the fact that Bag-of-Tasks applications only communicate to receive input and return the output Input/output is done by OurGrid itself The remote task runs inside a Xen virtual machine, with no network access, and disk access only to a designated partition

27. SWAN Architecture Guest OS Grid OS Grid Middleware Grid Application Guest OS Grid OS Grid Middleware Grid Application Guest OS Grid OS Grid Middleware Grid Application

28. Adding a second line of defense We can also reboot to add a second layer of protection to the user data and resources This has the extra advantage of enabling us to use an OS different from that chosen by the user −That is, even if the interactive user prefers Windows, we can still have Linux Booting back to the user OS can be done fast by using hibernation

29. SWAN + Mode Switcher

30. Making it Work for Real...

31. OurGrid Status OurGrid free-to-join grid is in production since December 2004 −We currently have 300 machines spread around 20 sites −See status.ourgrid.org for the current status of the system OurGrid is open source (GPL) and is available at www.ourgrid.org −We’ve had external contributions

32. http://status.ourgrid.org

33. HIV research with OurGrid B,c,FB,c,F HIV-2 HIV-1 M O ABCDFGHJKABCDFGHJK N ? prevalent in Europe and Americas prevalent in Africa majority in the world 18% in Brazil

34. HIV protease + Ritonavir Subtype B RMSD Subtype F

35. Performance Results for the HIV Application 55 machines in 6 administrative domains in the US and Brazil Task = 3.3 MB input, 1 MB output, 4 to 33 minutes of dedicated execution Ran 60 tasks in 38 minutes Speed-up is 29.2 for 55 machines −Considering an 18.5-minute average machine

36. Conclusions We have an free-to-join grid solution for Bag- of-Tasks applications working today Real users provide invaluable feedback for systems research Delivering results to real users is really cool! :-)

37. Next Steps Improve virtualization −to allow site-wide file sharing, and communication of tasks within a site −to enable communication among tasks spread in the grid −to make it possible for jobs to provide services outside the grid Move beyond Bag-of-Tasks applications Better present our components in a more diverse SOA setting Investigate how to make use of information that is becoming available in the grid Investigate how to better manage failures in grids

38. Questions?

39. Thank you! Merci! Danke! Grazie! Gracias! Obrigado! More at www.ourgrid.org