Mark Miller, Wayne Pfeiffer, and Terri Schwartz

1. The CIPRES Science Gateway: Enabling High-Impact Science for PhylogeneticsResearchers with Limited Resources Mark Miller, Wayne Pfeiffer, and Terri Schwartz San Diego Supercomputer Center

2. Phylogeneticsis the study of the diversification of life on the planet Earth, both past and present, and the relationships among living things through time ?

3. Evolutionary relationships can be inferred from DNA sequence comparisons: 2. Infer evolutionary relationships based on some set of assumptions: 1. Align sequences to determine evolutionary equivalence:

4. Inferring Evolutionary relationships from DNA sequence comparisons is powerful: DNA sequences are determined by fully automated procedures. Sequence data can be gathered from many species at scales from gene to whole genome. The high speed and low cost of NexGen Sequencing means new levels of sensitivity and resolution can be obtained. The speed of sequencing is still increasing, while the cost of sequencing is decreasing.

5. Inferring Evolutionary relationships from DNA sequence comparisons is powerful, BUT: Current analyses often involve 1000’s of species and 1000’s of characters, creating very large matrices. Sequence alignment and Tree inference are NP hard;even with heuristics, computational power often limits the analyses (already). The length of tree search analysis scales exponentially with number of taxa and with number of characters with codes in current use. There are at least 107 species, each with 3000 - 30,000 genes, so the need for computational power and new approaches will continue to grow.

6. Inferring Evolutionary relationships from DNA sequence comparisons is powerful, BUT: Current analyses often involve 1000’s of species and 1000’s of characters, creating very large matrices. Sequence alignment and Tree inference are NP hard;even with heuristics, computational power often limits the analyses (already). The length of tree search analysis scales exponentially with number of taxa and with number of characters with codes in current use. There are at least 107 species, each with 3000 - 30,000 genes, so the need for computational power and new approaches will continue to grow.

7. Inferring Evolutionary relationships from DNA sequence comparisons is powerful, BUT: Current analyses often involve 1000’s of species and 1000’s of characters, creating very large matrices. Sequence alignment and Tree inference are NP hard;even with heuristics, computational power often limits the analyses (already). The length of tree search analysis scales exponentially with number of taxa and with number of characters with codes in current use. There are at least 107 species, each with 3000 - 30,000 genes, so the need for computational power and new approaches will continue to grow.

8. Inferring Evolutionary relationships from DNA sequence comparisons is powerful, BUT: Current analyses often involve 1000’s of species and 1000’s of characters, creating very large matrices. Sequence alignment and Tree inference are NP hard;even with heuristics, computational power often limits the analyses (already). The length of tree search analysis scales exponentially with number of taxa and with number of characters with codes in current use. There are at least 107 species, each with 3000 - 30,000 genes, so the need for computational power and new approaches will continue to grow.

9. Inferring Evolutionary relationships from DNA sequence comparisons is powerful, BUT: Current analyses often involve 1000’s of species and 1000’s of characters, creating very large matrices. Sequence alignment and Tree inference are NP hard;even with heuristics, computational power often limits the analyses (already). The length of tree search analysis scales exponentially with number of taxa and with number of characters with codes in current use. There are at least 107 species, each with 3000 - 30,000 genes, so the need for computational power and new approaches will continue to grow.

10. In this new, DNA sequence-rich world, laptops and desktops are no longer adequate for phylogenetic analysis….

11. The CIPRES Portal was created to allow users to analyze large sequence data sets using popular community codes on a significant computational resource. The CIPRES Portal provided: • Login-protected personal user space for storing results indefinitely. • Access to most/all native command line options for each code. • Support for adding new tools and new versions as needed.

12. The CIPRES Portal was created to allow users to analyze large sequence data sets using popular community codes on a significant computational resource. The CIPRES Portal provided: • Login-protected personal user space for storing results indefinitely. • Access to most/all native command line options for each code. • Support for adding new tools and new versions as needed.

13. The CIPRES Portal was created to allow users to analyze large sequence data sets using popular community codes on a significant computational resource. The CIPRES Portal provided: • Login-protected personal user space for storing results indefinitely. • Access to most/all native command line options for each code. • Support for adding new tools and new versions as needed.

14. The CIPRES Portal was created to allow users to analyze large sequence data sets using popular community codes on a significant computational resource. The CIPRES Portal provided: • Login-protected personal user space for storing results indefinitely. • Access to most/all native command line options for each code. • Support for adding new tools and new versions as needed.

15. Workflow for the CIPRES Portal: • CIPRES Portal Assemble Sequences Upload to Portal Run Alignment Store Run Tree Inference Post-Tree Analysis Download

16. Limitations of the original CIPRES Portal • all jobs were run serially (efficient, but no gain in wall time) • the cluster was modest (16 X 8-way dual core nodes) • runs were limited to 72 hours • the cluster was at the end of its useful lifetime • funding for the project was ending • demand for job runs was increasing

17. The solution: make parallel versions of community codes available on scalable, sustainable resources via the Science Gateway Program CIPRES Cluster Workbench Framework

18. The solution: make parallel versions of community codes available on scalable, sustainable resources via the Science Gateway Program TeraGrid/XSEDE Triton Parallel codes Workbench Framework Serial codes

19. Greater than 90% of all computational time was used for three tree inference codes: MrBayes, RAxML, and GARLI. Deploy parallel versions of these codes on TeraGrid Machines; initially using Globus/GRAM. Work with community developers to improve the speed-up available through the parallel codes offered by CSG. Add new parallel codes (e.g. MAFFT) as they appear in the community. Keep other serial codes on local SDSC resources that provide the project with fee-for-service cycles.

20. Parallel code profiles on Trestles 19 X 36 X

21. Use of the New Gateway exceeded all expectations…. ?!?

22. Use of the New Gateway exceeded all expectations…. ?!? Initial allocation

23. Use of the New Gateway exceeded all expectations…. The initial allocation was greater than the capacity of the original cluster! ?!? Initial allocation

24. Given the high level of consumption,how to make sure Resource Usage is efficient?

25. Job Attrition on the CIPRES Science Gateway* *March – August 2010

26. Error Impact analysis

27. Prevent Job Loss When Contact with the Resource is Lost To make the system robust to system outages, jobs in the DB are tracked in a “running jobs table.” When a job completes, results are transferred to the DB, and the job is marked complete. Every 24 hours, a daemon looks for results for jobs that are incomplete, any results found are transferred to the DB, and the job is marked as complete. If there is a service interruption while the job is running, the job results will still be automatically delivered to the user.

28. Monitor Submissions/Usage to Track Efficiency Usage 12/2009 – 4/2012

29. SU/job increases Usage 12/2009 – 4/2012

30. MrBayes jobs (20%) fail when Lustre file system is under load Long running jobs fail even when there is no other traffic on the system. Users immediately resubmit, driving SU use up. The failure is due to resource contention in the Lustre file system As a result, the code cannot write output files, and jobs fail Moving to a mounted ZFS system eliminated the problem

31. Move off Lustre Usage 12/2009 – 4/2012

32. With such high consumption, we must track usage by individuals* *Reporting Period: Sept, 2010 – May, 2011

33. With such high consumption, we must track usage by individuals* *Reporting Period: Sept, 2010 – May, 2011 We need to monitor individual users, because we want all XSEDE users to be subject to the same level of peer review.

34. Establish a Fair Use Policy • Anyone, anywhere can sign up for an account. An account is not required to submit a job. • Users at US institutions are permitted to use 50,000 SUs from the community allocation annually. • Users at institutions in other countries can use up to 30,000 SUs annually. • Users can apply for a personal XSEDE allocation if they require more SUs.

35. Tools required to implement the CIPRES SG Fair Use Policy: • ability to halt submissions from a given user account • ability to monitor usage by each account automatically • ability for users to track their SU consumption • ability to forecast SU cost of a job for users • ability to charge to a user’s personal XSEDE allocation

36. Tools required to implement the CIPRES SG Fair Use Policy: • ability to halt submissions from a given user account • ability to monitor usage by each account automatically • ability for users to track their SU consumption • ability to forecast SU cost of a job for users • ability to charge to a user’s personal XSEDE allocation

37. Tools required to implement the CIPRES SG Fair Use Policy: • ability to halt submissions from a given user account • ability to monitor usage by each account automatically • ability for users to track their SU consumption • ability to forecast SU cost of a job for users • ability to charge to a user’s personal XSEDE allocation

38. XDB CDB job id, SU charge job id, user name Nightly usage reports Post to user’s work area User/management email notifications

39. Help users track their resource consumption: Notify users of their usage level

40. Tools required to implement the CIPRES SG Fair Use Policy: • ability to halt submissions from a given user account • ability to monitor usage by each account automatically • ability for users to track their SU consumption • ability of users to forecast SU cost of a job • ability to charge to a user’s personal XSEDE allocation

41. Create a conditional “warning” element in the interface XML

42. Tools required to implement the CIPRES SG Fair Use Policy: • ability to halt submissions from a given user account • ability to monitor usage by each account automatically • ability for users to track their SU consumption • ability to forecast SU cost of a job for users • ability to charge to a user’s personal XSEDE allocation

43. Steps required to use a personal allocation: • User receives personal allocation from XRAC • PI of allocation adds “cipres” user to their account • CSG staff changes the user profile to charge to the personal allocation account id

44. Steps required to use a personal allocation: • User receives personal allocation from XRAC • PI of allocation adds “cipres” user to their account • CSG staff changes the user profile to charge to the personal allocation account id • 3 users have completed this process

45. Impact of Policy on Usage Dec 2009 – April 2012 When Lustre file system is not used, submissions and SU usage are linear. 29,000 more SUs requested each month. Projected use for 2012 - 2013 is 13.4 million SUs

46. Impact of Policy on Usage Dec 2009 – April 2012 12 more users submit 160 more jobs each month Growth in usage is driven by new users J Feb Apr Jun Aug Oct Dec Feb Apr

47. What works about Trestles: • The Trestles machine is managed to keep queue depth low. This is a key requirement for many of our users who run a lot of relatively short jobs, and for class instruction. • Run times of up to 334 hours are allowed. This is important because most of the CSG codes do not have restart capability and scalability is typically limited to 64 cores or less.

48. Impact on Scientific Productivity: Publications enabled by the CIPRES Science Gateway/CIPRES Portal: YearNumber 2012* 106 2011 130 2010 89 2009 59 2008 4 *As of June 1, 2012 Publications in the pipeline: StatusNumber In preparation 91 In review 25 :

49. Impact on Scientific Productivity: • In Q1 2012, 29% of all XSEDE users who ran jobs ran them from the CSG • 50% of users said they had no access to local resources, nor funds to purchaseaccess on cloud computing resources • Used for curriculum delivery by at least 68 instructors. • Jobs run for researchers in 23/29 EPSCOR states. • Routine submissions from Harvard, Berkeley, Stanford….. • 76% of users are in the US or have a collaborator in the US

50. Impact on Scientific Productivity : “It is hard for me to imagine how I could work at a reasonable pace without this resource, especially when things like MS or grant submission deadlines loom….”