1. Transparently Gathering Provenance with Provenance Aware Condor Christine Reilly and Jeffrey Naughton Department of Computer Sciences University of Wisconsin - Madison TaPP ‘09 February 23, 2009

2. 2 Motivations for Our Work Scientific Computing Grid Computing Condor job scheduling system

3. 3 Motivation: Scientific Computing What input data was used to produce this output data? Example: –Blast sequence DB is updated periodically. –Did my last computation use the latest version of the Blast DB?

4. 4 Motivation: Scientific Computing Need to find the root of anomalous results. Example: –Multiple members of a research group can update the code for the simulation. –Did two researchers use the same version of the executable?

5. 5 Motivation: Grid Computing What grid resources did my computation use? Example: –A machine in the grid has a hardware problem (bad DIMM, corrupt disk). –Did any of my computations use the bad machine?

6. 6 Motivation: Condor Condor: –Distributed computing system. –Runs jobs from wide range of applications and fields of study. Quill: Captures operational information exposed by Condor. Could Quill be used for provenance? –Users would get provenance “for free”.

7. 7 Provenance System Goals Generic: Can be used by many different applications. Transparent: Users don’t need to alter their applications. Can we use Condor as the execution system for this type of provenance system? Can we do this with minimal impact on Condor developers?

8. 8 Our Model of Provenance File level granularity. The provenance of a file is: –All files involved in its creation –Job execution environment. input exeoutput Execution Environment

9. 9 Outline of Talk Motivation and Introduction PAC - Provenance Aware Condor Is PAC Practical? (Storage requirements and overhead) DBLife: Benefits and Limitations of PAC Conclusions

10. 10 Provenance Aware Condor Three parts: Condor, Quill, FileTrace Condor is the job execution system. Quill gathers operational information from Condor. FileTrace gathers information about files used by Condor jobs. Provenance is obtained by querying the Quill database.

11. 11 What is Condor? Provides a simple interface to a cluster of machines. –User gives computing jobs to Condor. –Condor finds available machine that meets job requirements. –Condor ensures that the job completes execution, even in the presence of failures. Used for many different applications: –Physics, Biology, Chemistry, Computer Science, Engineering applications. –Academic research, national labs, industry.

12. 12 What is Condor? (part 2) Research project at UW, but also resembles commercial software. –35 faculty, full-time staff, students. –Used at ~2000 locations world-wide. –Running on more than 250,000 machines. Annual user conference attracts hundreds of attendees.

13. 13 Submit Machine Execute Machine Central Manager User’s Machine User’s Local Files Shared File System Job Specs Exe, In Files Output Files Output Files Executable, Input Files Match Info Match Info Job Requirements Machine Resources Condor Overview

14. 14 What is Quill? Gathers operational data from Condor. –Improves performance of user’s inquiries about their jobs. –Information about Condor operations is stored in RDBMS. Created by group of database researchers (including us). Ships with Condor.

15. 15 Quill’s Design Requirements Expose Condor operational data for querying through a DBMS interface. Cannot affect how Condor operates: –No changes to the existing Condor code. –No change to how Condor is used. Failures in Quill cannot cause Condor to fail.

16. 16 Submit Machine Execute Machine Central Manager User’s Machine User’s Local Files Shared File System Job Specs Exe, In Files Output Files Output Files Executable, Input Files Job Requirements Match Info Match Info Machine Resources Quill Overview

17. 17 File Access in Condor Three methods of accessing files: –Remote system calls –Condor File Transfer –Shared file system Quill’s information about files: –Detailed information about File Transfers –Limited information about some other files –No guarantee that it detects all files Problem: Because Condor does not track all file access, Quill may miss file information. FileTrace gathers information about all files.

18. 18 FileTrace Modification of UNIX strace. Transparent to Condor. For each file open or close system call FileTrace gathers: –Condor Job Id. –File information: pathname, last modified time, checksum, size. –File access information: activity type, open flags, file pointer.

19. 19 Provenance in PAC Provenance of a file is: –All files involved in its creation: Executable, Input files, Libraries and system files. –Job execution environment: when, on what machine. PAC has no control over files in the user’s file system.

20. 20 Provenance Queries in PAC What files were used by this job? SELECT pathname, activity_type, open_flags, last_modified, size, checksum FROM filetrace WHERE globaljobid=‘my_job_id’;

21. 21 Provenance Queries in PAC More questions PAC can answer: Were any of my files created on the bad machine? Was this output created using the current versions of the input and executable files? Did any input files change between two runs of the application?

22. 22 Outline of Talk Motivation and Introduction PAC - Provenance Aware Condor Is PAC Practical? (Storage requirements and overhead) DBLife: Benefits and Limitations of PAC Conclusions

23. 23 Is PAC Practical? Users will not tolerate PAC if: –Data storage requirements are too onerous. –Computational overhead is significant.

24. 24 Estimate Table Growth Use the run time and number of files used by seven scientific applications (BLAST, IBIS, CMS, Nautilus, Messkit Hartree-Fock, AMANDA, SETI@home). Simulate running each application constantly for a year on a 1000 machine cluster. How big is the FileTrace table?

25. 25 1.15 TB 900 GB 9 GB 30 GB 41 GB 190 GB 540 GB

26. 26 Test Cluster for Overhead Machine Specs: –Pentium 2.40 GHz Core 2 Duo –4GB RAM –Two 250GB SATA-I hard disks. Condor Cluster: –Central Manager –Submit Machine –Database Machine –10 execute machines

27. 27 Overhead of PAC Use synthetic program. –Writes random numbers to files. –Can vary the number of files generated and run time. 4 sets of jobs with run time and number of files based on the scientific programs from the size experiment. –Run time per job: 5 or 20 minutes. –Files per job: 10 or 300.

28. 28 Overhead of PAC Two Clusters: PAC and Condor without Quill or FileTrace. For each set of jobs, submit as many jobs as will take ~12 hours to complete. No significant difference in average time per job or total run time.

29. 29 Outline of Talk Motivation and Introduction PAC - Provenance Aware Condor Is PAC Practical? (Storage requirements and overhead) DBLife: Benefits and Limitations of PAC Conclusions

30. 30 DBLife Community Information Management (CIM) Creates and maintains an ER graph of the database research community. Gathers data by crawling web. Interesting example for PAC: –Large, complex workflow. –Accesses many files.

31. 31 Overview of DBLife List of Sources Domain Knowledge Provided by Community Expert Data Page Layer Mention Layer Conference University Paper Company Joe ER Layer Joe Jane DBLife Home CIM Application

32. 32 DBLife Front Page

33. 33 DBLife Person Homepage

34. 34 Questions PAC Can Answer In general - questions developers and system administrators would ask. Examples: –Is this the version of the program that was used to create the current data set? –Was this machine used to create any part of the DBLife data set? –Does the DBLife ER model reflect the recent changes on my web page?

35. 35 Questions PAC Cannot Answer In general - questions that require knowledge of the semantics of DBLife. Examples: –Why does the DBLife portal show that person X is affiliated with institution Y? –Why doesn’t my page on the DBLife portal include that I presented a paper at this workshop?

36. 36 Problem: Large FileTrace Table Ran DBLife with 10 crawl start points: –FileTrace table: 100 MB Full scale run uses 1000 start points –FileTrace table est.: 10 GB Run DBLife daily for 1 year: –FileTrace table est.: 3.7 TB Possibilities for reducing table size: –Could be smarter about how data is stored. –Table compression.

37. 37 Outline of Talk Motivation and Introduction PAC - Provenance Aware Condor Is PAC Practical? (Storage requirements and overhead) DBLife: Benefits and Limitations of PAC Conclusions

38. 38 Limitations of PAC File level is coarse granularity. Files are not under PAC’s management. Workflows are not explicitly recorded. Doesn’t know about the semantics of an application.

39. 39 Contributions of PAC Many different applications can use PAC for provenance. Has little impact on Condor. Could be combined with other systems to yield a stronger provenance system. –Workflow management system. –File management system.

40. 40 Future Work Integrate PAC with an Xlog version of DBLife. –Xlog is similar to Datalog –Record information when Xlog rule is triggered. –Provides fine grained provenance. –Can answer questions that depend on the semantics of the application.

41. 41 Acknowledgements Supported in part by National Science Foundation Award SCI-0515491. Thank you CondorDB, DBLife, and Condor groups!