Download presentation
Presentation is loading. Please wait.
1
VGrADS Execution System or “the Virtual Grid”
Andrew Chien, Henri Casanova, Yang-suk Kee, Richard Huang, Fran Berman with input from the entire VGrADS Team Department of Computer Science and Engineering San Diego Supercomputing Center Center for Networked Systems UC San Diego September 13-14, 2004 VGrADS Workshop University of Tennessee, Knoxville
2
Meta-Goals for VGrADS Execution System Architecture
Leverage GrADS Learning Provide Organizing Framework for VGrADS Execution System Efforts Enable Integration of All of the Efforts Provide Working Framework which Enables Myriad Research Efforts and Experiments Enables Leverage Across Different Teams Leverages Globus Infrastructure Support Application and PPS Research and Progress We’re Leading and Organizing… not doing all of it! VGrADS Workshop 9/13/2004
3
Focused Execution System Goals
Provide Simple Resource Abstraction for PPS Enable Simpler and Better Optimization and Scheduling Enable Application-driven Adaptation Notify on Significant Changes in Resource Behavior Scale Good Resource Selection and Binding to Large and Dynamic Resource Pools Dedicated Sharing, Competitive Sharing, Malicious Behavior Present Dynamic Resource Information Better Presentation/Organization More Up-to-Date Information at Large Scale VGrADS Workshop 9/13/2004
4
Virtual Grids as a Resource Abstraction
Application / PPS Virtual Grid Resource Classification/Characterization Resource Properties Resource Mgmt Application: Resource Requirements, Comp and Scheduling Opt, Adaptation Virtual Grid: Resource Abstraction, Information, Monitoring, Adaptation Underlying Resource: Characterization and Classification Properties and Management VGrADS Workshop 9/13/2004
5
Basic Virtual Grid Operations
Application Interfaces VGManage() Create a virtual grid, add resources, delete resources VGReplaceResources() Swap a specified set of resource for some currently in the system Application reconfiguration must be dealt with by application VG System Management VGCreateBroker() VGTerminateBroker() Terminate Agent VGrADS Workshop 9/13/2004
6
VG System Elements Client: Application
Broker: Forms Virtual Grids, Modifies them, Negotiates with Resource Managers Monitor: Collects Dynamic VG Information, Notifies Application Broker and Monitor are active parts of the Virtual Grid There may be more in the future… Brokers compete with each other on behalf of their application VGrADS Workshop 9/13/2004
7
VG Evolution I: Acquire and Use
VGrADS Workshop 9/13/2004
8
VG Evolution II: Acquire, Adapt, App Reconfigure
VGrADS Workshop 9/13/2004
9
VG Evolution III: Acquire, Transparent Adaptation, FT?
VGrADS Workshop 9/13/2004
10
Refining Virtual Grid Concepts
Good start, but what do they mean? How do they compose? How to describe them, precisely… VGrADS Workshop 9/13/2004
11
Virtual Grid Description Language
Vgrid ::= Identifier = Rdl-expression | Rdl-expression at time/event Rdl-expression ::= Rdl-subexpression | ( Rdl-subexpression ) | Rdl-subexpression op Rdl-subexpression Rdl-subexpression ::= Associator-expression | Node-expression Associator-expression ::= Bag-of-expression | Cluster-of-expression Bag-of-expression ::= LooseBagof "<" Identifier ">" "[" MinNode ":" MaxNode "]" [ "[" Number [ “su” | “sec” ] "]" ] ";" Node-expression | TightBagof "<"Identifier ">" "[" MinNode ":" MaxNode "]" [ "[" Number [ “su” | “sec” ] "]" ] ";" Node-expression Identifier ::= String Min ::= Integer Max ::= Integer Node-expression ::= Identifier "=" Node-constraint Node-constraint ::= "{" Attribute-constraint | Rdl-expression "}" | Rdl-expression Attribute-constraint ::= (see Redline) Cluster-of-expression ::= Clusterof "<" identifier ">" "[" MinNode ":" MaxNode [ “,” MaxTime “:” “MinTime”] "]" ";" Node-expression op := close | far | highBW | lowBW VGrADS Workshop 9/13/2004
12
Virtual Grid Description Concepts
Hierarchical Aggregators ClusterOf (homogeneous and close) TightBagOf (heterogeneous and close) LooseBagOf (heterogeneous and far) Composers Close, Far, LowBW, HighBW Resource Attributes (extensible) Performance, Security, Owner, Pricing Policy, etc. Advanced and Extent Reservation (start time, stop time) Quantity of Resources (service units) VGrADS Workshop 9/13/2004
13
Application Examples: GrADS
Fasta = LooseBagOf<Node3>[8:32] { Node3 = [Software=DB] } Scalapack = ClusterOf<Node2>[8:32][3600 SEC] { Node2 = [Processor=Pentium4] VGrADS Workshop 9/13/2004
14
Application Examples: VGrADS
EOL1 = { Node=[Software=DB,Memory=[1024:INF], Rank=ClockSpeed] } EOL2= {SNode=[Software=DB] } far LooseBagOf<CNode>[32:4096] { CNode = [Memory=[1024:INF]] … and more complex versions … VGrADS Workshop 9/13/2004
15
Application Examples: VGrADS (cont.)
EMAN = LooseBagOf<Node4>[1:64] { Node4 = ClusterOf<Node5>[8:32] { Node5 = [Processor=Opteron, Rank=ClockSpeed] } VGrADS Workshop 9/13/2004
16
Virtual Grid Interface = Annotated VG Resource Request
Rname: foo.ucsd.edu Load: 0.6 Pred Load: 2.0 … Request Response & Active Representation Virtual Grid is an Explicit, Active Entity Bound Resources and their relation to the Request Information: Static and Dynamic Expression Attributes (AST with Attributes) Resources, Static Info (proc type, speed, location, etc.) Dynamic Information (load, prediction, security state, failure state, NWS info, etc.) Characterization / Classification Information VGrADS Workshop 9/13/2004
17
Discussion Virtual Grid as Explicit and Active entity enables explicit query, use, adaptation Annotated VG Description provides convenient, structured presentation of attribute information New types of Resource Information easily integrated in a pluggable attribute framework Internal VG Expression Nodes are natural places for some aggregate properties VGrADS Workshop 9/13/2004
18
Challenges in Realizing the Virtual Grid
Selection and Binding in Complex Resource Environments Scaling, Quality of Results, Complex Application Requirements, Rich Resource Descriptions, Competitive Sharing Scalable and Convenient Presentation of Resource Information Interfaces and Ease of Use, Integrating Group/Interrelation Information, Scaling and Accuracy of Dynamic Information Monitoring and Reacting Customization to Resource Description, Practically Good Monitoring, Application-Aided and Automatic Adaptation Approaches and Offline activities Resource Classification Resource Characterization VGrADS Workshop 9/13/2004
19
An Abstract Virtual Grid Execution System Architecture
VG Creation, Management Brokers Applications VG Monitoring Trigger Adaptation Monitors Resource Managers Information Services Brokers Select, Bind, Implement Operations on VGrid Monitors Watch Resources, Trigger on Changes, Notify Apps Application Needs ==== Execution System Intelligence (Resource Descriptions) VGrADS Workshop 9/13/2004
20
An Initial Approach to Selection and Binding
Resource Broker Simplify/Reduce Resource Description Assemble a Surplus of Candidates Bind Candidates to Form Virtual Grid If Succeed Return Virtual Grid Discard excess Candidates If Fail, many Strategies Collect more Candidates Next Time (Aggression) Application Change Resource Spec (less picky) Return Partial Virtual Grid (application adapt) … VGrADS Workshop 9/13/2004
21
VGrADS Execution System Prototype: Local Broker Architecture
Capture Resource Information from MDS/Generator Local Resource Broker DB Queries DVC Launcher Globus/GRAM Backend Monitoring Not Implemented VGrADS Workshop 9/13/2004
22
Execution System Prototype: Local Broker
PPS/Application Resource Broker VGrADS APIs VG RB APIs Specification Parser/SQL Generator (javacc) SQL phrase DVC JNI Interface JDBC (MySQL Connector/J) Launch DVC/Globus DB (MySQL) Resource Information (Generator) VGrADS Workshop 9/13/2004
23
Resource Selection Procedure
Parse Description Parse Description Simplify Parse tree Optimize Parse tree Generate SQL phrase Generate SQL phrase Annotate Description Annotate Description VG ClusterOf node 1, 100 Processor Opteron VG ClusterOf cluster 1, 10 CluterOf node Processor Opteron SELECT ID FROM Cluster WHERE CPUs BETWEEN 1 AND 100; SELECT * FROM HOST WHERE Cluster = ID AND Processor = “Opteron”; VG= ClusterOf[1:10]<cluster> { Cluster[1:10]<node> { node = [Processor=Opteron] } VG: Cluster0: Node0: Opteron, 2.4Ghz, 1024 KB, 4096 MB, 200 GB, Node1: Opteron, 2.4Ghz, 1024 KB, 4096 MB, 200 GB, Node2: Opteron, 2.4Ghz, 1024 KB, 4096 MB, 200 GB, Node3: Opteron, 2.4Ghz, 1024 KB, 4096 MB, 200 GB, VGrADS Workshop 9/13/2004
24
Resource Description Simplification
Simplify Redundant Aggregators and Expensive Composers Aggregator Simplification ClusterOf<>[n1, n2] { ClusterOf<> [m1,m2] } -> ClusterOf<>[n1*m1,n2*m2] TightBagOf<>[n1,n2] { ClusterOf/TightBagOf<>[m1,m2] } -> TightBagOf<>[n1*m1,n2*m2] LooseBagOf<>[n1,n2] { ClusterOf/TightBagOf/ LooseBagOf<>[m1,m2]} -> n1 instance of ClusterOf/TightBagOf/LooseBagOf<>[m1,m2] Composer Simplification Aggregator<>[n1,n2] close Aggregator<>[m1,m2] ->TightBagOf<>[n1+m1,n2+m2] VGrADS Workshop 9/13/2004
25
Database Table Structure (Schema)
TightBag TightBag Cluster Cluster Cluster Cluster Cluster Host Host Host Host Host Host Host Host Host Host Host VGrADS Workshop 9/13/2004
26
Demo I: Description & Selection in a Synthetic Grid
VGrADS PPS VGrADS Broker EMAN, EOL, Iter 2. Resources 15,000 tight bags 50,000 clusters 1,000,000 hosts Synthetic Grid Resource Generator VGrADS Workshop 9/13/2004
27
Demo II: Select, Bind, and Use in a Real Grid Testbed
VGrADS PPS 4. Script output 2. Script launch 1. Description/ Script 3. Resources DVC VGrADS Broker Globus UCSD Cluster VGrADS Workshop 9/13/2004
28
Summary and Prototype Status
Designed Virtual Grid Description Language Developed Concepts and Integration Architecture Incorporates input from VGrADS team, supports breadth of research Built and Demonstrated 1st Working Prototype Resource Description language Accepts full language, not all semantics implemented Local Resource Broker Many parameters to be explored; Dynamic Data, Resource Availability Resource Binder Utilizes DVC (Distributed Virtual Computer) Utilizes Globus toolkit (GSI,GRAM) to bind resources Prototype Functions End-to-End: Describe resources, request, bind, execute application Change Modules, Evolve to Support Research VGrADS Workshop 9/13/2004
29
More Virtual Grid Issues
Quantitative Access and Control Coupling, Matching Detailed Control Complex Reasoning Resource Properties Confidence and Quality Interaction w/ Resource Management/sharing Selection vs. Selection&Binding Advance and Extent Reservation Quantity of Resource (speed*number) Batch Queues and Variable Latency Adaptation Explicit vs. Transparent Application Initiated VG Monitoring Initiated <Any other way> Initiated Attributes Security Properties Statistical Properties Fault Tolerance Classification Characterization Aggregate Properties VGrADS Workshop 9/13/2004
30
VGrADS Execution System Collaboration/Coordination
API’s/Services for presentation to PPS/Applications Selection and Binding Adaptation API’s/Services for integration to information services API’s/Services for integration/ties to Fault-Tolerance Work Characterization, Prediction – Connect under Attribute System VGrADS Workshop 9/13/2004
31
Core Virtual Grid Execution System Plans I
Resource Brokers, Algorithms and Implementations Local Approach: Challenges Scaling and Quality of Results Evaluation on Synthetic Environments, Real Grid Environments (VGrADS) Experiments with Kernels and Full-scale Applications Local Approach: Schedule First Prototype, UCSD Team (9/2004) VGrADS Team Usable Prototype (1Q2005) Distributed Approach: Challenges Distributed Information and Accuracy Distributed Decision Making and Quality of Results Distributed Approach: Schedule First Prototype, UCSD Team (??) VGrADS Team Usable Prototype (??) Experiments with Kernels and Full-scale Applications (??) VGrADS Workshop 9/13/2004
32
Core Virtual Grid System Plans II
Virtual Grid Monitoring: Challenges Per VGrid Description Generated Monitor Implicit Specification of Performance Requirements Experiments with Detection, Notification, and Robustness Virtual Grid Monitoring: Schedule First Prototype System (2Q2005) VGrADS Team Usable Prototype (??) VGrADS Workshop 9/13/2004
33
Additional Virtual Grid Activity
MicroGrid Modeling Tools (Andrew talk) Large-scale Experiments Increasingly Robust Operation Enable Application Experiments with Virtual Grid Scheduling (Henri C. and Richard H. talks) Volatility and Information Quality Scheduling for Virtual Grids Application Driven Evaluation EMAN, LEAD, EOL Kernels Fault Tolerance (Dongarra), Reasoning Systems (Reed), Characterization (Wolski) VGrADS Workshop 9/13/2004
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.