1
Pegasus is a system for mapping and execu4ng abstract applica4on workflows over a range of execu4on environments. The same abstract workflow can, at different 4mes, be mapped different execu4on environments such as XSEDE, OSG, commercial and academic clouds, campus grids, and clusters. Pegasus can easily scale both the size of the workflow, and the resources that the workflow is distributed over. Pegasus runs workflows ranging from just a few computa4onal tasks up to 1 million. Workflows oJen consume tens of thousands of hours of computa4on and involve transfer of many terabytes of data. Workflows have a DAG model A node in the DAG is started only when all the parent nodes have successfully finished. Event-Based Triggering and Management of Scientific Workflow Ensembles Suraj Pandey 1, Karan Vahi 2 , Rafael Ferreira da Silva 2 , Ewa Deelman 2 , Ming Jiang 3 , Albert Chu 3 , Cyrus Harrison 3 , Henri Casanova 1 1 University of Hawaii, 2 USC Information Sciences Institute, 3 Lawrence Livermore National Laboratory h=ps://github.com/pegasus-isi/pegasus-llnl Project is supported by the Department of Energy under contract DE-AC52-07NA27344 and by LLNL under project 16-ERD-036. Pegasus is currently funded by the NaRonal Science FoundaRon(NSF) under the OAC SI2-SSI grant 1664162 . Pegasus WMS System Architecture Experimental Setup at LLNL Catalyst Cluster APIs Users Interfaces Pegasus Dashboard OpenStack, Eucalyptus, Nimbus Pegasus WMS Mapper Engine Scheduler Monitoring & Provenance Workflow DB Logs Clouds Notifications j 1 j 2 j n Job Queue Cloudware Amazon EC2, Google Cloud, RackSpace, Chameleon Compute Amazon S3, Google Cloud Storage, OpenStack Storage Distributed Resources Campus Clusters Local Clusters Open Science Grid XSEDE HTCondor GRAM PBS LSF SGE Middleware C O M P U T E GridFTP Storage Other workflow composition tools: Submit Host HTTP FTP SRM IRODS SCP Problems mapping certain computaRons to DAG workflows With push towards extreme scale compu4ng it is possible run tradi4onal HPC simula4on codes simultaneously on tens of thousands of cores. Generated data oJen needs to be periodically analyzed using Big Data Analy4cs frameworks. Integra4ng Big Data analy4cs with HPC simula4ons is a major challenge for the current genera4on of scien4fic workflow management systems. Need an ability to automa4cally spawn and manage the analysis workflows, as the long running simula4on workflow executes. Ensemble Manager SoluRon Pegasus has an Ensemble Manager Service that allows user to submit a collec4on of workflows called ensembles. We extended the Ensemble Manager to support event triggers that can trigger addi4on of new workflows to an exis4ng ensemble. We support the following types of triggers a) File based event triggers – a file gets modified b) Directory based event triggers – files appear in a directory Ini4ally, ensemble has a single workflow consis4ng of the long running HPC simula4on workflow. The HPC simula4on workflow periodically generates output data that in a directory that is tracked by the ensemble manager. A new analysis workflow is launched automa4cally as the output data is detected. Reliably and repeatedly test that implemented solu4on works. Tested the implementa4on on LLNL Catalyst Cluster ( 150 teraFLOP/s system with 324 nodes, each with 128 GB of DRAM and 800 GB of non vola4le memory ) . Experimental Setup 1. On catalyst, a Magpie SLURM job is submiaed that does - Determines which nodes will be “master” nodes, “slave”nodes, or other types of nodes. - Sets up, configures, and starts appropriate Big Data daemons to run on the allocated nodes. In our setup , we used the Magpie SPARK template to setup a dynamic Spark cluster - Reasonably op4mizes configura4on for the given cluster hardware that it is being run on. Magpie then executes a user specified script to give control back to the user. 2. The user script sets up Pegasus WMS and starts the Ensemble Manager. 3. Ensemble Manager submits the HPC Simula4on Workflow consis4ng of LULESH applica4on - Every 10 simula4on cycles LULESH writes out outputs to a directory on the shared filesystem. - This directory is tracked by the ensemble manager as part of the event trigger specified. - Ensemble Manager invokes a script that generates the Big Data Analy4cs Workflow on the newly generated datasets. HPC Simula+on Workflow Ensemble Manager Big Data Analy+cs Workflow Magpie Setup Submit Slurm Job Run Magpie Daemons Run Magpie User Script Pegasus WMS Setup Launch Pegasus & Condor Launch Ensemble Manager Read Event Configura@on Trigger Event? Generate and submit analy@cs workflow LULESH MPI Job Generate Trigger File Copy Data into HDFS Run Spark Analy@cs Workflow Ensemble Execu:on Timeline

Pegasus WMS - 情報処理学会sighpc.ipsj.or.jp/HPCAsia2018/poster/post102s2-file2.pdf · • Pegasus can easily scale both the size of the workflow, and the resources that the

  • Upload
    others

  • View
    3

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Pegasus WMS - 情報処理学会sighpc.ipsj.or.jp/HPCAsia2018/poster/post102s2-file2.pdf · • Pegasus can easily scale both the size of the workflow, and the resources that the

•  Pegasusisasystemformappingandexecu4ngabstractapplica4onworkflowsoverarangeofexecu4onenvironments.

•  Thesameabstractworkflowcan,atdifferent4mes,bemappeddifferentexecu4onenvironmentssuchasXSEDE,OSG,commercialandacademicclouds,campusgrids,andclusters.

•  Pegasuscaneasilyscaleboththesizeoftheworkflow,andtheresourcesthattheworkflowisdistributedover.Pegasusrunsworkflowsrangingfromjustafewcomputa4onaltasksupto1million.

•  WorkflowsoJenconsumetensofthousandsofhoursofcomputa4onandinvolvetransferofmanyterabytesofdata.

•  WorkflowshaveaDAGmodel

•  AnodeintheDAGisstartedonlywhenalltheparentnodeshavesuccessfullyfinished.

Event-Based Triggering and Management of Scientific Workflow Ensembles Suraj Pandey1, Karan Vahi2, Rafael Ferreira da Silva2, Ewa Deelman2, Ming Jiang3, Albert Chu3, Cyrus Harrison3, Henri Casanova1

1University of Hawaii, 2USC Information Sciences Institute,

3Lawrence Livermore National Laboratory

h=ps://github.com/pegasus-isi/pegasus-llnl ProjectissupportedbytheDepartmentofEnergyundercontractDE-AC52-07NA27344andbyLLNLunderproject16-ERD-036.

PegasusiscurrentlyfundedbytheNaRonalScienceFoundaRon(NSF)undertheOACSI2-SSIgrant1664162.

PegasusWMSSystemArchitecture

ExperimentalSetupatLLNLCatalystCluster

APIs

Users

Interfaces

Pegasus Dashboard

OpenStack, Eucalyptus, Nimbus

Pegasus WMS

Mapper

Engine

Scheduler

Monitoring & Provenance

Workflow DB

Logs

Clouds

Not

ifica

tions

j1

j2

jn Job Queue

Cloudware

Amazon EC2, Google Cloud, RackSpace, Chameleon Compute

Amazon S3, Google Cloud Storage, OpenStack Storage

Distributed Resources

Campus Clusters

Local Clusters

Open Science Grid

XSEDE

HTCondor GRAM

PBS LSF SGE

Middleware C O M P U T

E

GridFTP

Storage

Other workflow composition tools:

Submit Host

HTTP

FTP SRM

IRODS SCP

ProblemsmappingcertaincomputaRonstoDAGworkflows•  Withpushtowardsextremescalecompu4ngitispossibleruntradi4onalHPCsimula4oncodessimultaneouslyontensofthousandsofcores.•  GenerateddataoJenneedstobeperiodicallyanalyzedusingBigDataAnaly4csframeworks.•  Integra4ngBigDataanaly4cswithHPCsimula4onsisamajorchallengeforthecurrentgenera4onofscien4ficworkflowmanagementsystems.•  Needanabilitytoautoma4callyspawnandmanagetheanalysisworkflows,asthelongrunningsimula4onworkflowexecutes.

EnsembleManager

SoluRon•  PegasushasanEnsembleManagerServicethatallowsusertosubmit

acollec4onofworkflowscalledensembles.•  WeextendedtheEnsembleManagertosupporteventtriggersthat

cantriggeraddi4onofnewworkflowstoanexis4ngensemble.•  Wesupportthefollowingtypesoftriggers

a)  Filebasedeventtriggers–afilegetsmodifiedb)  Directorybasedeventtriggers–filesappearinadirectory

•  Ini4ally,ensemblehasasingleworkflowconsis4ngofthelongrunningHPCsimula4onworkflow.•  TheHPCsimula4onworkflowperiodicallygeneratesoutputdatathat

inadirectorythatistrackedbytheensemblemanager.•  Anewanalysisworkflowislaunchedautoma4callyastheoutputdata

isdetected.

•  Reliablyandrepeatedlytestthatimplementedsolu4onworks.•  Testedtheimplementa4ononLLNLCatalystCluster(150teraFLOP/ssystemwith324nodes,

eachwith128GBofDRAMand800GBofnonvola4lememory).ExperimentalSetup1.  Oncatalyst,aMagpieSLURMjobissubmiaedthatdoes-Determineswhichnodeswillbe“master”nodes,“slave”nodes,orothertypesofnodes.-Setsup,configures,andstartsappropriateBigDatadaemonstorunontheallocatednodes.Inoursetup,weusedtheMagpieSPARKtemplatetosetupadynamicSparkcluster-Reasonablyop4mizesconfigura4onforthegivenclusterhardwarethatitisbeingrunon.Magpiethenexecutesauserspecifiedscripttogivecontrolbacktotheuser.2.  TheuserscriptsetsupPegasusWMSandstartstheEnsembleManager.3.  EnsembleManagersubmitstheHPCSimula4onWorkflowconsis4ngofLULESHapplica4on-Every10simula4oncyclesLULESHwritesoutoutputstoadirectoryonthesharedfilesystem.-Thisdirectoryistrackedbytheensemblemanageraspartoftheeventtriggerspecified.-EnsembleManagerinvokesascriptthatgeneratestheBigDataAnaly4csWorkflowonthenewlygenerateddatasets.

HPCSimula+onWorkflow

EnsembleManager

BigDataAnaly+csWorkflow

MagpieSetup

SubmitSlurmJob

RunMagpieDaemons

RunMagpieUserScript

PegasusWMSSetup

LaunchPegasus&Condor

LaunchEnsembleManager

ReadEventConfigura@on

TriggerEvent?

Generateandsubmitanaly@csworkflow

LULESHMPIJob

GenerateTriggerFile

CopyDataintoHDFS

RunSparkAnaly@cs

WorkflowEnsembleExecu:onTimeline