350 likes | 485 Views
Industrial "Power Grid" Computing: The Next High Performance Challenge. Dr. Kenneth Neves Senior Technical Fellow Director, Computer Science Boeing, Seattle WA. Slides Available at: http://homepages.go.com/~drneves/index.html. Outline. “Grid Computing”: The Concept Background
E N D
Industrial "Power Grid" Computing: The Next High Performance Challenge Dr. Kenneth Neves Senior Technical Fellow Director, Computer Science Boeing, Seattle WA Slides Available at: http://homepages.go.com/~drneves/index.html
Outline • “Grid Computing”: The Concept • Background • Parallelism - winning battles! • Application Frameworks • Grid Frameworks: Virtual Net-Machine • Enabling tools • Challenges JSF
Grid Computing: The Concept • “The Grid” - Computational nodes and connections of an Internet or Intranet • Grid Computing - Untilization of the grid assests (memory, CPU power, connectivity) to acheive high performance computing and information access • The Analogy (A Vision) - Just as we “plug” into the electrical power network when we want electricity, we should be able to “plug” into the “Internet/Intranet” and “compute” from grid Why do this? Is it possible? What applications require this?
Background • Ultimatelywe need to • integrate • all three PDM Simu- late CAD International Space Station • Fortune 500 companies have enterprise-wide computing challenges • Challenging scientific computingsimulations are still required to meet future competitive product design needs, particularly in multi-discipline approaches • CAD systems must be integrated, distributed, and support simulation of physical “end products” • Business systems (people management, MRP, PDM) are approaching tens of terabytes of storage, and geographic distribution and synchronization
Focus - Scientific Computing • Today, I will focus on scientific computing, but consider this an example area • The scenarios proposed for scientific computing can be developed for other areas, e.g.: • Data rich applications, particularly of large data sets • Knowledge discovery frameworks where a series of techniques can be linked to large data sets “in situ” • Process specific collaboration • Data fitting and reduction through multidimensional techniques • Multimedia access and dissemination
Scientific Computing - Background • Developing parallel, distributed or “grid based” applications requires investment • Investment requires stability • Industry invests in software for decades, not 18 months • Computational infrastructure has been changing too rapidly • Nevertheless, in recent years, many application codes have been (modestly) paralleled on distributed machines, a start to grid computing Let’s look at some examples from the Boeing High Performance Computing Benchmark Suite
8 X slower, 1000 X cheaper Sun E4000 Compaq Pent. 200mhz HP Pentium II 300MHz IBM SP2 TLNS3D Thin Layer Navier Stokes Cost, Always Good Incentive Cray Triton Single CPU Performance 3500 DEC Alpha 3000 SGI Origin 2500 HP V2200 $800 2000 Dell Pentium II 400mhz CPU Time 1500 1000 500 $1M 0 Computer Model
Using Myrinet SGI Origin Fast Multipole MethodPARADYM (radar cross section) WARNING! Network Latency cannot be ignored! 1995 200MHz PC CPU Time No. Processors
OVERFLOW Wing Body (3.5M pts, 6 zones)(Overflow HSCT CFD) Excellent algorithm scalability on even larger clusters CPU Time No. Processors
Multiple CPU Comparison (OVERFLOW HSCT CFD) CPU Time No. Processors
Grid Clusters Show Possibilities, but Connectivity Key • High speed networks enable “payoffs from” cluster computing, but private protocol networks add cost [note: similar statements can be made for data access applications of large distributed data driven by non-partitionable data bases] • Web usage and media content are driving bandwidth up, and costs down • Consequently, clustering of resources promises to be common and cheap: • NGI, Internet II will exceed today’s Myrinet-type speeds even over long distances • Access to data (science, weather, CAD, etc.) will be fast and cheap, even if quite remote
Scientific Application Challenges • Many industrial applications are one or two decades old -- why? • They are continually enhanced and validated by testing and use • Newcodes arenot trusted (nor should they be) • What pays the bills is the process being supported, not the application’s isolated results • More resolution, higher model fidelity, while important, don’t necessarily improve the process results • Rather than refine the analysis, we desire to optimize against often conflicting constraints, and multiple goals • Complexity is enormous, tradeoffs are not understood
Stack & Batch Approach CFD Nastran DCAC Visualization Catia Optimizer (executive) App 2 CAD to finite element gridder App 1 Input & Setup - CAD def. Current Industrial Approach to MDO CPU Time & Human Effort We require a more orderly process! . . . A Framework! Cost- flow time
Application Frameworks - A definition • Goals • improved processes and quality of the final design • easy collaborationamong disciplines • gain insight, not simply produce results • help for the human in the loop with statistical and cognitive aids • lower cost and shorten process cycle time • take advantage of distributed resources, data, and expertise • flexible and extensible usage • Characteristics • Systematic use of existing analysis codes • Provides tools for integrating multiple disciplines • Provides tools for data manipulation and viewing • Algorithm choices if appropriate • Reuse of middleware, libraries, common data
Stack & Batch Approach Visualization Optimizer (executive) App 2 CAD to finite element gridder Input & Setup Stat. Design Visualization App App 1 middleware Input & Setup - CAD def. Optimizer old Grid gen. new Design Explorer (DE) AN EXAMPLE OF AN APPLICATION FRAMEWORK Design Explorer: focus of a multi-year collaboration between researchers at Boeing and Rice University Ref.: Andrew Booker, Paul Frank, John Dennis, Doug Moore, and David Serafini, "Managing Surrogate Objectives to Optimize a Helicopter Rotor Design" , AAIA MDO 98-4717
Design Explorer: Framework Features • Can be configured to the problem type • Exploits decision tools • Statistical design techniques • Global domain behavior • Parameter sensitivity analysis • Decouples the actual application from the executive process • can “wrap” the function evaluation into the system • can couple multiple applications • can provide insight • Utilizes new approaches to optimization • Surrogate model (to save computational overhead and gain insight) • Meta-algorithm optimization (to achieve accurate “true” solution) • Flexible and applicable to a myriad of problems The first Boeing Plane
Optimization Techniques • Small-scale, calculus-based, local opt: • NPSOL - SQP Method • HDNLPR - SQP Method • Large-scale, calculus-based, local opt: • HDSNLP - Schur-complement method • Interior Point Method - prototype code • Small-scale, bounds constrained, global opt: • Globopt - Stochastic, multi-start local opt • Direct - Subdivision method DE’s Framework Features • Configurable to the problem type • Exploits decision tools • Statistical design techniques • Global optimization issues • Parameter sensitivity analysis • Decouples the actual application from the systems • can “wrap” the function evaluation into the system • can couple multiple applications • can connect to other frameworks • Utilizes new approaches to optimization • Surrogate model • Meta-algorithm optimization • Flexible and applicable to a myriad of problems
3-D Fighter Aerodynamics Shot peen forming of wing skins Engine Nozzle Performance Rotor Design Multidisciplinary wing platform design & 777 Engine Duct Seals Machining, riveting, and drilling (simulation) Widely Dispersed Applications--but One Framework
An Approach to a Design Framework • Expensive to evaluate • Many variables • Sensitivity to parameters unknown • One function evaluation is a supercomputing problem Outer Dielectric • Multiple Objectives • find absolute max • minimize the max • tradeoffs among competing • objectives
Surrogate Model Validate 0.2 Surrogate Model 0.0 x2 -0.2 Statistical analysis of global modeling evaluation pts. Y X -0.4 X X X 0.4 0 X X 0.2 X 0.4 0 X x2 0.2 -0.2 0 X x1 -0.4 -0.2 0.0 0.2 0.4 -0.4 -0.2 x1 Build surrogate multidimensional model -0.4
The DE Framework Initialize (Build and/orread model in) Global Surrogate Model Global Statistical Methods Algorithmic Framework (Executive) Calibrate Surrogate Model Local Optimization Save the State of the Opt Process & Sensitivities "Optimize" the Model Expensive Valid Code(s) Execution
More loosely coupled process can be distributed more heterogeneously Supercomputer analysis, maps to large MPPs where tight parallelism must be managed Parameter evaluation Independent MPP class jobs can be distributed to remote MPPs Computational Opportunities in Frameworks Not only does a framework increase the degree of parallelism, but mapping to distributed resources should be easier Initialize (Build and/orread model in) Global Surrogate Model Global Statistical Methods Algorithmic Framework (Executive) Calibrate Surrogate Model Save the State of the Opt Process & Sensitivities "Optimize" the Model Local Optimization Expensive Valid Code
Other Boeing Frameworks • EASY5 continuous simulation system (control oriented) • 25 year history • current version is interactive, distributed, library components, and user defined and wrapped functions • commercially available • Coupling CAD to Simulation • Easy5 as a simulation tool • Genesis (hydraulics from CAD) • Factory Assembly Modeling • Workflow Planning and Collaboration • Interactive and Haptic Visualization • L3: Lines, Loads, Laws • Simulation & Knowledge Based Design
Non-scientific Frameworks • Metadata: The ”backbone” of information about the data in the IDS environment. • It is used by developers and administrators to manage and deploy data. • It provides the business user context and legibility of the data they are accessing.
Collaboration: People, Frameworks, Systems Team walk through of International Space Station mission, with simulated operation Structural analysis of damaged airplane at remote location Interactive design review of detailed system assembly with suppliers Recreation of flight into bad weather based on NCAR stored storm data and authentic CAD data Search for cause of repeated air conditioning failure from multi-airline operational data
Outline • Situation - Opportunity • Parallelism - winning battles! Wars? • Application Frameworks • Grid Frameworks - A Virtual Net-Machine • Enabling tools • Challenges Next: A required infrastructure JSF
NIS Campus Server Room FDDI Ring Grid Concept - A Virtual Net-Machine Application Frameworks GRID INFRASTRUCTURE Virtual Services, Network, File System, Security, CPU Services, Transaction Processing Virtual Local Local Security Local Security DCOM CORBA SUNOS NT AIX Printers & Workstations
Super Reference (Well Edited): The Grid: Blueprint for a New Computing Infrastructure Edited by Ian Foster and Carl Kesselman July 1998 - ISBN 1-55860-475-8 Grid Frameworks • Grid frameworks vary in tools, philosophy, & adaptability • Application specific tools (e.g. SCIRun, Dongarra et al) • Object component based (e.g. Legion, Gannon, Grimshaw, et al) • Custom use of commodities (ORBs, Jini, Java, ActiveX . . .) • “Bag of Services”, (e.g., Globus Toolkit, Kesselman & Foster) • Scheduling, and network languages (e.g. IDL, Predictive Schedulers, Francine Berman) • Impact on application designers/users • Design and execution • Transition to grid paradigm is a key issue • User responsibilities vary: Do very little just supply the function box? And/or provide schedule? And/or develop framework? And/or schedule assets and download executables? . . . • A Grid Infrastructure may be useless, unless users provide application frameworks! Applications will never have widespread use & impact without grid infrastructures! (the former is a fact, the latter is my conjecture)
Common Grid Framework Concerns • Executive control • Throughput (of the job stream) vs. performance (of the individual application) a traditional rivalry NEW ISSUE - Framework throughput NEW ISSUE - Transaction throughput for an enterprise data server • Schedule and synchronization model (Dynamic P&S) • Control given by the application and user schedule or by system agents and reactive resource allocation agents Deterministic/repeatable VS serendipitous/variable • Management of “executables” and data • Application control vs. middleware control • Persistence or not • Resource management and asset control (including accounting) • Information (data) access and data synchronization (integrity) • System health, security, recovery, and QoS
Some Grid Related Boeing Activities • KAoS Agents Architecture (W Florida, Lawrence Berkeley, NASA, Darpa) • Structured frame work, extensible • Standard discourse • Agent based security • Example: NOMAD (next slide) • Security, Intrusion Detection and Health Maintenance • Global-mobile (active and hybrid) network, pervasive computing • Services tools • Example: SWAN Heralds (next slide) • Component based systems (Unger, Klawitter, Tyler) • Parallel computing and performance/scalability modeling • Data modeling and warehouse architecture • CAD independent visualization, display, haptics, immersion, & simulation of product data • Collaboration tools, work flow • Statistical methods applicable to resource measurements, DOE, Frameworks like DE • Natural language interfaces
Example: NOMAD • Collaboration between Boeing and Univ. of W. Florida (Suri, Bradsahw, Breedy,Ditzel, Hill, Pouliot, and Smith. Darpa Supported). • Agent based infrastructure • Persistent with “strong” mobility • Context mobility (captures state independent of machine) • Supports security AND policy • Capacity permissions • Agent initiated check pointing to other VMs for reliability • Moves philosophically from “orchestrated control” to “serendipitous control” • For example, consider a NOMAD based approach to resource scheduling
SWAN Heralds • Goal:provide a mechanism based on standard protocols to support scalable synchronized collaboration • Approach • Automatic and dynamic topology with a goal of quadruple paths • Minimal path depth (using a heuristic algorithm) • Maintain synchronization, in near real time • Advantages • Scales to 1000s • Weakest link doesn’t degrade others performance (e.g. NetMeeting • No central control (i.e. Distributed shared history & registry) that is failure resistant • Failed links cause no problems, and can be restored by remaining heralds (including collective history) • Commercially available licenses
Initialize (Build and/orread model in) "Optimize" the Model Data Center FDDI Ring Algorithmic Framework (Executive) Calibrate Surrogate Model NIS Campus Server Room FDDI Ring Save the State of the Opt Process & Sensitivities Local Optimization Expensive Valid Code Printers & Workstations Mapping App Frameworks to Grid Frameworks Shared Responsibility Between App. Users and Grid Developers • Executive control • Executable management • Schedule & synch model • Resource management • Communication services • Information access • Security • Health and status
Summary and Recommendations • Application frameworks are necessary for Boeing use of grid frameworks [we need to get going at Boeing] • Grid frameworks must provide stable models of computation, synchronization, with ease use [we need to engage the grid community: dialogue, partner, and assess!] • Raytheon, Aerospace, GM are already active • TRANSITION to grid computing by industry, requires an enduring model for grid frameworks. • THIS IS A RESEARCH FRONTIER • We could help set the standards (e.g. Agent language) • Industrial companies must take more central control of computing assets and provide strong strategic planning for (often reluctant) user communities
Thank you Q & A