1. LEAD Workflow Orchestration Lavanya Ramakrishnan Renaissance Computing Institute University of North Carolina – Chapel Hill Duke University North Carolina State University

2. Outline Background and LEAD Service Architecture —web services, BPEL LEAD Workflows —status, site visit example run —characteristics, workflow structures LEAD Virtual Grid Integration —resource management

3. Linked Environments for Atmospheric Discovery Rationale —each year, mesoscale weather – floods, tornadoes, hail, strong winds, lightning, hurricanes and winter storms – causes hundreds of deaths, routinely disrupts transportation and commerce, and results in annual economic losses in excess of $13B. From “offline” to “online” forecasting —data assimilation and adaptive evaluation

4. Static Forecasting Large-scale tightly coupled components —not adaptable to weather or resource behavior —modifications are not trivial NetRad Radar ingest Fetch data products Convert to format for assimilation Assimilate into 3D grid forecast model execution Plan ensemble run Store results to personal catalog Analyze final product Center for Analysis and Prediction of Storms (CAPS, OU): It took an MS student 6 months, working with a research scientist, to learn how to modify the operational forecast system for her needs and run it using real data

5. Dynamic Adaptive LEAD System Meteorology goal —to provide timely and accurate forecasts using dynamic adaptation Computer Science goal —map application requirements to resource capabilities –redundant runs, scheduling policies —adapt to weather as well as resource behavior Need real time monitoring to make adaptation decisions

6. Define Domain, Data Requirements and Query for Desired Data ADAS Analysis Processing -producing- ADAS Analysis (3D Gridded Fields) + Background Fields ESML & LDM Decoding ADAS Remapping, Gridding, Conversion, Quality control ADAS-to-WRF Converter Yielding 3D Gridded Fields in WRF Height Coordinate STOP Adjust Forecast Configuration and Schedule Resources START Allocate Computational Resources and Move/Stream Data to Appropriate Location (e.g., PACI Center) Single or Multiple Copies of WRF Forecast Model WRF Gridded Output Meta Data Creation and Cataloging myLEAD Storage WRF Ensemble Initial Conditions Visualization Data Mining with ADaM Define Forecast Configuration, Mining Configuration and Thresholds Monitor Education & Meteorology LEAD Control and Data Flow

7. Service Oriented LEAD Workflows NetRad Radar ingest Fetch data products Convert to format for assim Assimilate into 3D grid Forecast model execution Plan ensemble run Store results to personal catalog Analyze final product BPEL Workflow Source: Beth Plale

8. Business Process Execution Language (BPEL) BPEL (aka WSBPEL and BPEL4WS) — enable cross-enterprise automated business processes —multiple vendor participation –Oracle, Microsoft, IBM, SUN, SAP, BEA, EDS, Adobe, … —multiple implementations –Oracle BPEL, IBM BPWS4J and Microsoft BizTalk Process Managers, Active BPEL (www.activebpel.org) Open Source Process Manager Features —loop and control logic, synchronous/asynchronous communication —composition and concurrent execution

9. LEAD Site Visit Experimental Ensemble Data Assimilation WRF IDV … IU (chinkapin) UNC (dante) ARPSPLT Unidata (lead1) OU (aquaman)

10. Characteristics of LEAD Workflows Coupled analysis and assimilation tools, data repositories —change configuration rapidly and automatically in response to weather —Streaming data, steer remote observing technologies Multilevel monitoring and intelligent control —workflow, resource, application, service —performance and reliability guarantees of the resources Model Resources Data Need resources Have more resources Get data Have more data Have more data

11. Adapting Workflow Structure Reactive Adaptation —e.g. service failure, resource behavior Proactive prediction and decisions —adjust i, j, k (weather science meets the infrastructure science) to meet individual workflow guarantees —global optimization of number of workflows being serviced …… … Adaptation in Time (i) Adaptation in Space (j) Adaptation based on Initial Conditions (k)

12. Outline Background and LEAD Service Architecture web services, BPEL LEAD Workflows status, site visit example run characteristics, workflow structures LEAD Virtual Grid Integration —resource management

13. Distributed Resources Computation Specialized Applications Steerable Instruments Storage Data Bases Resource Access Services GRAM Grid FTP SSH Scheduler LDM OPenDAP Generic Ingest Service User Interface Desktop Applications IDV WRF Configuration GUI LEAD Portal Portlets Visualization Workflow Education Monitor Control Ontology Query Browse Helper Crosscutting Services Authorization Authentication Monitoring Notification Configuration and Execution Services Workflow Monitor MyLEAD Workflow Engine/Factories Resource Cat THREDDS Application Resource Broker (Scheduler) Host Environment GPIR Application Host Execution Description WRF, ADaM, IDV, ADAS Application Description Application & Configuration Services Client Interface Observations Streams Static Archived Data Services Workflow Services Catalog Services RLS OGSA- DAI Geo-Reference GUI Control Service Query Service Stream Service Ontology Service Decoder/ Resolver Service Transcoder Service/ ESML LEAD Architecture

14. Distributed Resources (compute, disk) Protocols (e.g. web services, GridFTP, Gatekeeper) Crosscutting Services Authorization Authentication Monitoring Notification MyLEAD Client Interface LEAD Architecture Resource and Data Management (VGrid, myLEAD, etc) Application Services (e.g. WRF, IDV, etc) BPEL Workflow Engine LEAD Portal

15. Virtual Grid provides LEAD … Dynamic, scalable resource abstraction framework —scheduler, resource broker Integrated monitoring and notification of resource behavior —NWS, NWS-HAPI DVCW vgFAB vgES APIs vgMON Information Services vgLAUNCH VG vgAgent Grid Resources Resource Managers

16. DVCW vgFAB Proposed LEAD – VGrADS Architecture LEAD VGrid Manager vgES APIs vgMON vgDL Information Services Resource Managers vgLAUNCH VG vgAgent Grid Resources LEAD Event System BPEL Workflow Engine Expectations Violations Decision Process Other LEAD System Events (e.g. application, service)

17. Some Thoughts … LEAD VGrid Manager vgDL VG LEAD Event System BPEL Workflow Engine Expectations Violations Decision Process Web Service Progress of workflow Resource behavior Weather, Application behavior Performance, Fault tolerance, Availability Steer Workflow Find and bind resources Workflow Contract {time, accuracy (performance,reliability)} Workflow description Meet Guarantees

18. Streaming data —Unidata LDM Service monitoring —web service load Application —behavior on resources Resource —performance —reliability BPEL Workflow Engine LDM Service GridFTP Service WRF Service IDV Viz Service Data arrives 1. “Data has arrived” 3. Next step in workflow – Launch WRF 0. Possible advanced reservation 5. Move data to compute (virtual) nodes 4. Get resources Resource Broker Data Mining 6. Run WRF 8. WRF complete 7. More WRF runs needed? If yes repeat 5 to 6 2. Get required resources Start End Resources - compute, network storage LEAD Dynamic Workflows Multilevel monitoring Multiple decision points

19. vgDL Specification for LEAD LEADSpec = LDMNode = {isLDMNode} // Note the loose coupling far WRFNode = {memory >= 500MB, cpu > 2000, diskspace > 4GB} far VizNode = {memory >=4GB, cpu > 4000} vgidLEAD = vgCreateVG(vgESsrv.renci.org, LEADspec, 1000, null); vgRoot = vgGetRoot(vgidLEAD); LDMNode WRFNode VizNode vgRoot WRFspec = LooseBagof [1:32]; C = Clusterof [4:256]; node = {node.memory > 500MB, node.cpu > 2000}; C = {C.hasSharedFileSystem = true} vgidLEAD = vgGetMyVG(); wrfNode = vgGetMyNode(); status = vgAddToVG(vgidLEAD, WRFspec, wrfNode,1000, init_WRF); … 1 0 32 31 1 0 255 2 … 1 0 2 1010 … Provisioning

20. LEAD Virtual Grid Research Implications Streaming data in resource scheduling —fixed point makes scheduling complicated Persistent and transient services —service directory, monitoring, scheduling Data management across the virtual grid —run time knowledge of data location Integrated decision process —performance and reliability guarantees —application, weather

21. Comments?

22. LEAD Monitoring Architecture Real-time Resource Monitoring (NWS-HAPI) Steer the workflow, etc Workflow (triggered by real-time data arrival) LEAD Monitoring and Orchestration Target Resource Set Resource Behavior Prediction Application Code Analysis (SvPablo) Historical data Workflow Engine Decision Process Scheduler Real-time Resource Monitoring (NWS-HAPI) Steer the workflow, etc Workflow (triggered by real- time data arrival) LEAD Monitoring and Orchestration Target Resource Set Application and Resource Behavior Prediction Application Code Analysis (SvPablo) Historical data Workflow Engine Decision Process Scheduler Weather, application change

23. LEAD Control and Data Flow LEGEND Control Flow Data Resources Input Data/Model grids in LEAD usable form ADAS Output in standard form ADAS Output in WRF input form WRF Output ADaM Output LEAD Metadata (Color Determines Resource) Experiment Metadata