Presentation is loading. Please wait.

Presentation is loading. Please wait.

Case Study 2: Scientific Workflow for Computational Economics Tiberiu Stef-Praun Gabriel Madeira Ian Foster Robert Townsend.

Similar presentations


Presentation on theme: "Case Study 2: Scientific Workflow for Computational Economics Tiberiu Stef-Praun Gabriel Madeira Ian Foster Robert Townsend."— Presentation transcript:

1 Case Study 2: Scientific Workflow for Computational Economics Tiberiu Stef-Praun Gabriel Madeira Ian Foster Robert Townsend

2 eSS 2007Service-Oriented Science: Globus Software in Action2 The Challenge l Expand capability of economists to develop and validate models of social interactions at large scales u Harness large computation systems u Simplify programming model (eye toward easy integration of science code) u Improve automation l Requires an end-to-end approach, but through integration, not the silo model

3 eSS 2007Service-Oriented Science: Globus Software in Action3 Moral Hazard Problem l An entity in control of some resources (the entrepreneur) contracts with other entities that use these resources to produce outputs (the workers) l Two organizational forms are available u The workers cooperate on their efforts and divide up their income (thus sharing risks) u The workers are independent of each other, and are rewarded based on relative performance l Both are stylized versions of what is observed in tenancy data in villages such as in Maharastra, India (Townsend and Mueller 1998)

4 eSS 2007Service-Oriented Science: Globus Software in Action4 Moral Hazard Solver l Five stages, each solved by linear programming u Balance between promises for future and consumption to optimally reward agents l In each stage: Given a set of parameters: consumption, effort, technology, output, wealth u Do a linear optimization to find out the best behavior u Parameter sweep (grid of parameter values) u Linear solver is run independently on each point of the parameter grid u Results are merged at end of the stage l Across stages: Different organization (parameters) for similar stage structure u Most stages depend on results of other stages

5 eSS 2007Service-Oriented Science: Globus Software in Action5 Stage One 26 x StageOne.${i}.out *.mat input data files Stage Five MergedStageOne.out MergedStageTwo.out MergedStageThree.out MergedStageFour.out MergedStageFive.out Stage Two 52 x StageTwo.${i}.out Stage Four 40 x StageOne.${i}.out Stage Three 40 x StageThree.${i}.out Remote Execution Local Execution Legend 50 Min 30 Min 3 Min 40 Min 2 Min

6 eSS 2007Service-Oriented Science: Globus Software in Action6 Issues - Technical l Language u Science code written in MATLAB/Octave u End to end system must be language-independent l Code prerequisites u Each solver task requires MATLAB/Octave pre- installed on the execution node, and solver code staged in prior to execution u Each solver task requires files from previous stages l Automation u ~200 tasks must be executed u This is a lot of babysitting if performed manually

7 eSS 2007Service-Oriented Science: Globus Software in Action7 Issues - Social l Licensing u MATLAB licensing has a per-node cost u Expensive if youre using O(10)+ nodes l Provenance u Task execution, data integrity u Not a huge concern at this scale, but for larger scales (10,000 tasks) it is important to record how the work is performed l Provisioning, resource sharing u This problem used a shared campus cluster (at U Chicago) u We know of problems with 2-3 orders of magnitude more tasks, which require (inter)national-scale resources to accomplish in a timely fashion

8 eSS 2007Service-Oriented Science: Globus Software in Action8 Swift System l Clean separation of logical/physical concerns u XDTM specification of logical data structures + Concise specification of parallel programs u SwiftScript, with iteration, etc. + Efficient execution on distributed resources u Karajan threading, Falkon provisioning, Globus interfaces, pipelining, load balancing + Rigorous provenance tracking and query u Virtual data schema & automated recording Improved usability and productivity u Demonstrated in numerous applications

9 eSS 2007Service-Oriented Science: Globus Software in Action9 Workflow Language - SwiftScript l Goal: Natural feel to expressing distributed applications u Variables (basic, data structures) u Conditional operators (if, foreach, ) u Functions (atomic / compound) l Used to connect outputs to inputs l It does not specify invocation order, only dependencies l It can be seen as a metadata for expressing experiments

10 eSS 2007Service-Oriented Science: Globus Software in Action10 Execution Engine l Karajan engine (event-based execution) l Has a scheduler to map tasks to resources u Score-based planning u Recovers from failures (retries) l Falkon resource manager creates a virtual private cluster u Uses Globus GRAM4 (PBS/Condor/Fork) to acquire resources from Grid systems

11 Virtual Node(s) SwiftScript Abstract computation Virtual Data Catalog SwiftScript Compiler SpecificationExecution Virtual Node(s) Provenance data Provenance data Provenance collector launcher file1 file2 file3 App F1 App F2 Scheduling Execution Engine (Karajan w/ Swift Runtime) Swift runtime callouts C CCC Status reporting Provisioning Falkon Resource Provisioner Amazon EC2 Dynamic Provisioning: Swift Architecture Yong Zhao, Mihael Hatigan, Ioan Raicu, Mike Wilde, Ben Clifford

12 eSS 2007Service-Oriented Science: Globus Software in Action12 The Solution l Code changes u Solver code was broken into modules (atomic blocks) to allow parallel execution u Code ported from MATLAB to Octave to avoid per- node licensing fees u Workflow was described in SwiftScript l Software installation u Swift engine, Karajan, Falkon deployed locally l Shared resource (already available) u Existing compute cluster with GRAM4, GridFTP, etc.

13 eSS 2007Service-Oriented Science: Globus Software in Action13 Moral Hazard SwiftScript Code Excerpts // A second atomic procedure: merge (file mergeSolutions[]) econMerge (file merging[]) { } } // We define the stage one procedure–a compound procedure (file solutions[]) stageOne (file inputData[], file prevResults[]) { file script ; int batch_size = 26; int batch_range = [0:25]; string inputName = "IRRELEVANT"; string outputName = "stageOneSolverOutput"; // The foreach statement specifies that the calls can be performed concurrently foreach i in batch_range { int position = i*batch_size; solutions[i] = moralhazard_solver(script,batch_size,position, inputName, outputName, inputData, prevResults); } } // These get used in the main program as follows stageOneSolutions = StageOne(stageOneInputFiles,stageOnePrevFiles); stageOneOutputs = econMerge(stageOneSolutions);

14 eSS 2007Service-Oriented Science: Globus Software in Action14 Execution on 40 Processors

15 eSS 2007Service-Oriented Science: Globus Software in Action15 Results - Moral Hazard Solver l Performance u Original run time: ~2 hrs u Swift run time: ~28 min u Depending on the stage structure, speedup up to 10x, or slowdown (because of overhead) u Only used one grid site (UC), on multiple sites could get better performance l Execution has been automated u Human labor greatly reduced u Separation of human concerns (science code, system operation, task management) u Easy to repeat, modify & rerun, etc.

16 eSS 2007Service-Oriented Science: Globus Software in Action16 Other Applications Application#Jobs/computationLevels ATLAS* HEP Event Simulation 500K1 fMRI DBIC* AIRSN Image Processing 100s12 FOAM Ocean/Atmosphere Model 2000 (core app runs CPU jobs) 3 GADU* Genomics: (14 million seq. analyzed) 40K4 HNL fMRI Aphasia Study 5004 NVO/NASA* Photorealistic Montage/Morphology 1000s16 QuarkNet/I2U2* Physics Science Education 10s3-6 RadCAD* Radiology Classifier Training 1000s5 SIDGrid EEG Wavelet Proc, Gaze Analysis, … 100s20 SDSS* Coadd, Cluster Search 40K, 500K2, 8


Download ppt "Case Study 2: Scientific Workflow for Computational Economics Tiberiu Stef-Praun Gabriel Madeira Ian Foster Robert Townsend."

Similar presentations


Ads by Google