Presentation is loading. Please wait.

Presentation is loading. Please wait.

Lessons from LEAD/VGrADS Demo Yang-suk Kee, Carl Kesselman ISI/USC.

Similar presentations


Presentation on theme: "Lessons from LEAD/VGrADS Demo Yang-suk Kee, Carl Kesselman ISI/USC."— Presentation transcript:

1 Lessons from LEAD/VGrADS Demo Yang-suk Kee, Carl Kesselman ISI/USC

2 Outline SC’06 Demo Summary New Features of VGES Year-5 Development and Research Plans VGES Support for SC’07 Demo

3 LEAD/VGrADS Demo at SC’06 The first integration of LEAD/VGrADS software stacks  Identified functionalities and requirements of core components  Demonstrated the resource slot concept and the QBETS (BQP) potential  Showed slot-based scheduling using performance model

4 LEAD BPEL Workflow Engine LEAD BPEL Workflow Engine Workflow Configuration Service Workflow Configuration Service Event Broker Event Broker Workflow Application Service (per task) Application Service (per task) Workflow and File Status DAG + Constraint myLEAD (subscribes to messages from the broker and knows what magic to do with input/output files and talks to RLS/DRS Run workflow one step at a time Run job Job Notification Adaptation Create Services App. Factory Launch Services Virtual Grid Execution System Virtual Grid Execution System Scheduler Mapper Scheduler Mapper Performance Model Performance Model LEAD Resource Broker LEAD Resource Broker Portal Schedule toward a workflow deadline Annotated DAG Batch Queue Prediction LEAD Linked Environments for Atmospheric Discovery

5 Virtual Grid Execution System Virtual Grid Execution System Scheduler Mapper Scheduler Mapper Performance Model Performance Model Resource Broker Resource Broker Schedule toward a workflow deadline Batch Queue Prediction DAG + Constraint Annotated DAG Here is the workflow and constraints + pointer to performance model. Give me a mapping Query the performance model for task’s resource requirements Find me two slots (vgFind) Constantly collecting data over time Use performance model and map the tasks to the slots. If deadline can’t be met, return. Query Batch Queue prediction about probabilities of getting slots Return slots above threshold Bind Resources (vgBind) GT4 GRAM (Reserved) If reserved submit PBS-glidin at slot start time else submit when BQP suggests Return mapping If not reserved resource, ask - Is it time to submit? PBS Slot PBS Slot Globus Gateway Run job Job Notification vgLaunchRun JobQuery status (vgStatus ) Send job notifications

6 New Features of Current VGES Language  Support of resource equivalence (limited implementation)  WS-GRAM schema wrapper for execution on the personalized resources Execution system  Probabilistic guarantee of resource binding  Resource orchestration and personalization

7 Resource Equivalence Specifying exchangeable constraints  Provides flexibility in resource discovery  Specifies constraints with precedence in order of appearance  PE = “Opteron” <> 4 * “Itanium”; vgdl = ClusterOf (node) [4] { node = [Processor == PE] }

8 WS-GRAM Schema Wrapper Providing abstract job description  Hides WS-GRAM schemas that are irrelevant for specifying applications Application-related WS-GRAM schema  argument, count, directory, environment, executable, job, jobType, library, path, stderr, stdin, stdout  cf) host, factoryEndpoint

9 Guarantee of Resource Binding Deterministic guarantee  Batch with advance reservation Probabilistic guarantee  Predicts resource availability for batch-scheduled resources  Models resource allocation of individual resource providers as a random variable with a binomial distribution

10 Resource Actualization WS-GRAM Resource actualization engine 2 4 notification 1 bind vgES Cluster 6 5 launch Application launcher 8 notification 3 7 acquire check submit update PBSLSF Condor PBS GRAM

11 described discovered boundinactive active unavailable 9:00 A.M 9:10 A.M 9:55 A.M 10:00 A.M 11:00 A.M select bind (actualize) (activate) (cleanup) vgdl=CluserOf (nd) [4] { nd=[Processor=“P4”] } vgdl=CluserOf (nd) [4] { nd=[Processor=“P4”] } sdsc (p=0.90) ncsa (p=0.85) iu (p=0.70) ada (p=0.65) sdsc (p=0.90) ncsa (p=0.85) iu (p=0.70) ada (p=0.65) submit P1P2 P4P3 sdscncsaiu sdsc (p=0.90) ncsa (p=0.85) iu (p=0.70) sdsc (p=0.90) ncsa (p=0.85) iu (p=0.70) P1P2 P4P3 Time

12 Year-5 Plans Extended implementation of slot allocation  Support of various resource managers (e.g., PBS, LSF, Load- leveler, Condor)  Personalization over multiple clusters Consistent resource slot provisioning  Provides efficient resource scheduling techniques  Tradeoffs between quality, availability, and cost Slot optimization  Optimization of inter/intra slot allocation Deploying to as many TeraGrid sites as possible

13 Consistent Resource Provisioning Motivation  Can we get a slot for a specified time period in practice? Limitation in both number of processors and wall time Goals  Exploring offline/online algorithms  Presents system sub-slot schedules to users

14 Resource Slot Provisioning Problem S-slot U-slot Slot duration Slot size MaxWallTime MaxCPU LooseBag for 2 days This slot will be never satisfied!

15 Resource Slot Provisioning Problem S-slot U-slot LooseBag for 2 days

16 VGES Support for the SC’07 Demo Resource equivalence  Enables flexible resource discovery  Provides more reliable resource discovery service New semantic of binding  Separates slot binding from actual resource allocations  Enables the LEAD workflow manager to exploit parallelism Probabilistic guarantee of binding  Provides high slot availability virtually  Minimizes resource allocation failures due to late resource arrivals

17 VGES Support for the SC’07 Demo Support of various resource managers  Plugs in Loadleveler (Bigred) and LSF (Tungsten)  Covers most resource managers in TeraGrid Callback mechanisms for resource arrivals  Provides asynchronous event notification  Lessens the burdens on both the client and the server Consistent resource provision for LooseBag slots  Provisions resources proactively  Realizes slot in practice


Download ppt "Lessons from LEAD/VGrADS Demo Yang-suk Kee, Carl Kesselman ISI/USC."

Similar presentations


Ads by Google