Workflowsim escience12
3 likes1,614 views
WorkflowSim is a toolkit for simulating scientific workflows in distributed environments. It models workflow overhead, failures, and the hierarchical nature of workflows with tasks and jobs. WorkflowSim extends CloudSim to be workflow-aware and supports modeling diverse overhead distributions, failure models, and fault tolerant techniques like reclustering and job retry. It helps researchers evaluate workflow optimization techniques more accurately. Validation experiments show WorkflowSim can accurately simulate overhead and failures and their impact on workflow scheduling heuristics and fault tolerant clustering approaches.
Workflowsim escience12
- 1. WorkflowSim: A Toolkit for Simulating Scientific Workflows in Distributed Environments Weiwei Chen, Ewa Deelman
- 2. Outline § Introduction – Scientific workflows? – Distributed environments? § Challenge – Large scale, system overhead § Solution – Workflow Overhead and Failure Model § Validation and Application – Overhead Robustness – Fault Tolerant Clustering 2
- 3. Scientific Applications § Scientists often need to – Integrate diverse components and data – Automate data processing steps – Reproduce/analyze/share previous results – Track the provenance of data products – Execute processing steps efficiently – Reliably execute applications Scientific Workflows provide solutions to these problems 3
- 4. Scientific Workflows § DAG model (Directed Acyclic Graph) – Node: computational activities – Directed edge: data dependencies – Task: a process that users would like to execute – Job: a single unit for execution with one of more tasks – Task clustering: the process of grouping tasks to jobs 4
- 5. Workflow Management System § Common Features of WMS – Maps abstract workflows to executable workflows – Handles data dependencies – Replica selection, transfers, registration, cleanup – Task clustering, workflow partitioning, scheduling – Reliability and fault tolerance – Monitoring and troubleshooting § Existing WMS – Pegasus, Askalon, Taverna, Kepler, Triana 5
- 6. Workflow Simulation § Benefits – Save efforts in system setup, execution – Repeat experimental results – Control system environments (failures) § Trace based Workflow Simulation – Import trace from a completed execution – Vary workflow structures and system environments § Challenges (CloudSim, GridSim, etc.) – System overhead and failures – Multiple levels of computational activities(task/job) – A hierarchy of management components 6
- 7. Comparison CloudSim WorkflowSim Task, Bag of Execution Model Task, Job, DAG Tasks Failure and Monitoring No Yes Data Transfer Delay Data Transfer Overhead Model Workflow Engine Delay Delay Clustering Delay … Site selection Single Multiple Scheduling, Job retry Optimization Scheduling Clustering, Techniques Partitioning … WorkflowSim is an extension of CloudSim, but it is workflow aware 7
- 8. System Architecture § Submit Host – Workflow Mapper – Clustering Engine – Workflow Engine – Local Scheduler § Execution Site – Remote Scheduler – Worker Nodes – Failure Generator – Failure Monitor 8
- 9. Workflow Overhead – Workflow Engine – Postscript Delay Delay – Clustering Delay – Queue Delay – Data Transfer Delay 9
- 10. Example: Clustering Delay n is the number of tasks per level. k is number of jobs per level. !"#$%&'()*!!"#$%|!!! ! ! ! = = !"#$%&'()*!!"#$%|!!! ! ! ! mProjectPP, mDiffFit, and mBackground are the major jobs of Montage Overhead is not a constant variable. It has diverse distribution and patterns 10
- 11. Validation !"#$%&'#$!!"#$%&&!!"#$%&' Ideal Case: Accuracy=1.0 !!!"#$!% = !"#$!!"#$%&&!!"#$%&' k: Maximum jobs per level Overheads have the biggest impact 11
- 12. Application: Overhead Robustness § Overhead robustness: the influence of overheads on the workflow runtime for DAG scheduling heuristics. § Inaccurate estimation (under- or over-estimated) of workflow overheads influences the overall runtime of workflows. § Research Merits: – Sensitivity of heuristics – Overhead friendly heuristics 12
- 13. Application: Overhead Robustness § Increase or Reduce Overhead by a Factor (Weight) – Under estimation and over estimation § Heuristics Evaluated – FCFS: First Come First Serve – MCT: Minimum Completion Time – MinMin: The job with the minimum completion time is selected and assigned to the fastest resource. – MaxMin: The job with the maximum completion time and assigns it to its best available resource. 13
- 14. Application: Overhead Robustness MCT<MinMin MCT<MinMin MCT>MinMin MCT<MinMin Accurate estimation of Clustering Delay is more important 14
- 15. Workflow Failure § Failures have significant influence on the performance § Classifying Transient Failures – Task Failure: A task fails, other tasks within the same job may not fail – Job Failure: A job fails, all of its tasks fail 15
- 16. System Architecture § Submit Host – Job Retry – Reclustering § Execution Site – Failure Generator – Failure Monitor 16
- 17. Application: Fault Tolerant Clustering § Task clustering can reduce execution overhead § A job composed of multiple tasks may have a greater risk of suffering from failures § Reclustering and Job Retry are proposed – No Optimization (NOOP) retries the failed jobs. – Dynamic Clustering (DC) decreases the clusters.size if the measured job failure rate is high. 17
- 18. Application: Fault Tolerant Clustering – Selective Reclustering (SR) retries the failed tasks in a job – Dynamic Reclustering (DR) retires the failed tasks in a job and also decreases the clusters.size if the measured job failure rate is high. 18
- 19. Performance § Reclustering (DR/DC/SR) reduces the influence of failures significantly compared to NOOP § DR outperforms other techniques. The lower the better 19
- 20. Conclusion § WorkflowSim assists researchers to evaluate their workflow optimization techniques with better accuracy and wider support. – Modeling Overhead and Failures – Distributed and Hierarchical components – Workflow Techniques § It is necessary to consider both data dependencies, workflow failures and system overhead. 20
- 21. Acknowledgements § Pegasus Team: http://pegasus.isi.edu § FutureGrid & XSEDE § Funded by NSF grants IIS-0905032. § More info: wchen@isi.edu § Available on request § Q&A 21
- 22. Overhead Distribution Montage 22
- 23. Overhead Distribution CyberShake 23
- 24. Broadband 24
- 25. Task Failure Model and Job Failure Model n 4d k* = −d + d 2 − r ln(1− α ) k* = , if n >> r 2t * (kt + d) ttotal = * n(k *t + d) 1− β t = total * rk(1− α )k 25 25
- 26. Related Papers § Integration of Workflow Partitioning and Resource Provisioning, CCGrid 2012. § Improving Scientific Workflow Performance using Policy Based Data Placement, IEEE Policy 2012. § Fault Tolerant Clustering in Scientific Workflows, SWF, IEEE Services 2012 § Workflow Overhead Analysis and Optimizations, WORKS 2011. § Partitioning and Scheduling Workflows across Multiple Sites with Storage Constraints, PPAM 2011 § Pegasus: a Framework for Mapping Complex Scientific Workflows onto Distributed Systems, Scientific Programming Journal 2005 26