Models of Networked Analysis at Regional Centres Harvey Newman MONARC Workshop CERN May 10,

PROJECT On LHC COMPUTING MODELS MONARC Models of Networked Analysis at Regional Centres

MONARC Primary Goals u Determine which classes of Computing Models are feasible for the LHC Experiments Ý Match network capacity, data handling resources likely to be available u Specify the main parameters characterizing this class of Models u Produce example “Baseline” models that fall into the “feasible” category ` COROLLARIES: Ý Help define Regional Center architecture and functionality Ý Help define the Analysis Process for LHC experiments Ý Provide guidelines to keep the final Computing Models in the feasible range

MONARC DELIVERABLES u Specifications for a set of feasible Models u Guidelines for the Collaborations to use in building their Computing Models u A set of Modeling Tools to enable the experiments to simulate and refine their CM

MONARC SCHEDULE (in PAP) u PHASE 1: to Summer 1999 Ý First-round set of Modeling tools u PHASE 2: to Submission of CTPR Ý Refined set of tools Ý Guidelines for constructing feasible CM, in time for the Computing TPR u PHASE 3: to 2001 (Future R&D) Ý Prototype designs and test implementations of the CM, for the second round Computing TDRs

MONARC PHASES (Ideal) u 0 COMPLETE Setup of WG 8/ /98 u 1A SETUP10/ /98 u 1B STARTUP 11/98 - 1/99 u 1C MODELING 2/99 - 5/99 u 2A REFINED MODELING 6/99 - 8/99 u 2B VERIFICATION and CONVERGENCE 9/ /99

MONARC Phase 1B STARTUP: 11/98 - 2/99 u 1. Validate chosen tool(s) with existing Model Ù Final choice of simulation tool by 1/99 u 2. Formulate first detailed model for simulation Integrate Configurations, Performances and Workloads From Phase 1A u 3. Code and test Model Study 1 (MS1) u 4. MS1 Simulation Runs: Monitor, Display and Refine... u 5. MS1 Analysis Results u 6. Validate MS1 Results: Spot-check using testbed u 7. MS1 Design Reviews: 1/15/99; 2/15/99 u 8. MS1 Conclusions, and Preparation for Phase 1C u 9. In parallel: Setup federated Objectivity/DB on testbed (with RD45, GIOD)

MONARC Phase 1C MODELING: 2/99 - 6/99 u 1. Refine coding for Sites (CERN, Regionals), networks, workloads, ODBMS, HPSS u 2. Choose Models to be studied: A range of working methods, and system input parameters “MS2 - MSn” Ù Setup Team to Run and Analyze Models in Parallel, with Cross-checks Ù Run and Analyze MS2 - MSn Ù Evaluate MS2 - MSn Ù Extract achievable workload (throughput), latency as a function of network bandwidth and site performance u 3. Identify key parameters, performance bottlenecks, sources of long latency; derive priorities to match the workload u 4. Identify preferred Analysis Process(es) u 5. Validate key results with testbed Ù Propose measurements to be done to narrow critical uncertainties in the Models u 6. Setup first Models with ODBMS and/or HPSS: MSOHn Ù Run and Analyze MSOHn Ù Evaluate Impact of Use of ODBMS, HPSS

MONARC PHASE 2A (1) REFINED MODELING 6/99 - 9/99 è MONARC Model Design/Evaluation Review in 5/99; LCB Progress Report/Review/Discussion by 6/99 u 1. Build Model-set with refinements for HPSS, ODBMS, network and dynamic workload behaviors u 2. Run, analyze and evaluate Model-set u 3. Focus on “Good” combinations of site configuration/Analysis Process/Data handling strategy (factorizable ?) u 4. Agree on a standard set of evaluation criteria: system performance and working efficiency u 5. Choose a promising set: the “Baseline Models” u 6. Run, analyze, evaluate in detail, using standard criteria u 7. Verify key features of the Baseline Model simulations using the testbed, including ODBMS and HPSS features

MONARC System Design and Development Task u Use the modeling and simulation constructs and tools Ù To design the overall system: this means defining and making choices for XThe Site Architecture(s) XThe Analysis Processes Ù Layout a Complete set of user, site, and inter-site “tasks” Ù Define Key Limits: Quotas, Max Transaction-Times; Priorities Ù Profile the behaviors of the site and network components, and the “Actors” (Sites and Users) Ù Set the scale of the analysis: this means defining X“how much data is accessed, processed and transmitted; by how many people; and how often” X the major intra-site and inter-site “events” X high water marks to trigger events X conditions that alter component response-time or performance Ù Match to the foreseen level of site and network resources: a feasible overall picture of “The Analysis”

MONARC System Design and Development (II) u OPTIMIZATION: Choosing a System Architecture and Deciding How Best to Use It (for a given level of resources) Ù Cost Versus Value Metrics X Recalculation versus data transport time and resource usage X Time to transaction completion XAffinity: concurrence of requests for data l Proximity in Space (File location) l Proximity in Time Ù Strategies to allow for dynamic non-local data access Ù Caching, mirroring, preemptive data movement Ù High water marks and “branch points” in system behavior Ù Turnaround time and/or workflow Targets Ù Isolation of Key Parameters: e.g. how much re-reconstruction; how much non-local data; how much access from tape

MONARC System Design and Development (III) u Complete Top-Down Design: Nothing Significant Left Out Tools for Building Architectures Deployed to the Team Ù Profiling Structures + Utilities Ù Profile Setup: X Site-Architecture Profile X Site-Task Profile X Site-Pair Interaction Profile (for all pairs; or a hierarchy) X User Workload Profiles: Batch Oriented and Rate Oriented X Response Time Profiles (Dynamic) X Network Profiles: Base capability; performance/load characteristics X Priority Profiles; Quotas; Marginal Utility Factors X Decision High Watermark Profile(s) Ù Adaptation Mechanisms; Algorithms Ù Recovery Mechanisms u Simulation Team Ù Tool developers (developers’ developers) Ù System Specification developers Ù Operations group: run, analyze, suggest modified and new tools

MONARC A SAMPLING OF ADDITIONAL ISSUES u Scaling tests of the system as a function of network bandwidth u ODBMS/HPSS Interactions; Custom software to enable the ODBMS to work in a WAN-distributed environment u Adaptability of the system to the architectures likely to exist in different countries u Impact on overall performance of the system, for different approaches to the analysis u Minimum level of flexibility required for the sake of physicists’ working efficiency, in spite of limited resources

MONARC PHASE 2A (2) REFINEMENTS 6/99 - 8/99 u 1. Refine coding for preferred site (CERN, Regionals), network, and workload configuration and priority schemes u 2. Refine the testbed in terms of its connections (to HPSS), network performance, and measurement capability u 3. Refine code for ODBMS, HPSS configuration and behaviors u 4. Refine code for network performance vs. load; QoS mechanisms; Congestion conditions u 5. Refine the Analysis Process Ù Production, Analysis by Groups, Individuals Ù Calibrations and Reruns of the Data Ù Role of the Desktop u 6. Implement d ata transport/recompute/delete strategies Ù Automatic and manual (?), speculative (?) replication Ù Caching and reclustering ?

MONARC PHASE 2B VERIFICATION and CONVERGENCE 8/ /99 u 1. In depth study of the Baseline Models chosen in Phase 2A u 2. Verification of Model Prototypes using the testbed u 3. Identify/extract the features that distinguish the Baseline Models as being “feasible” u 4. Investigate the consequences of variations in key parameters (e.g. WAN bandwidth, I/O bandwidth, desktop CPU) u 5. Evaluate Baseline Models for Ù Adaptability: of the Models system to architectures likely to exist in different countries Ù Responsiveness: turnaround time and ability to respond to peak loads for “urgent” analysis Ù Scalability: performance vs. time as the data volumes and component performances increase Ù Flexibility: to adapt and/or migrate to different Models over time. Ù Overall Performance Vs. Cost: Users’ throughput, working efficiency; operational infrastructure, manpower, maintainability u 6. Report: Guidelines and recommendations; Input for the CTPR by 12/99

MONARC WORKING GROUP TASKS (1) u SYSTEMS DESIGN Ý CERN Center Architectures: CPU, Storage, I/O, LAN Ý Regional Center Architectures: CPU, Storage, I/O, LAN Ý MAN (Regional) and WAN configurations Ý Site and Network Performance Parameters/Profiles Ý Queueing and Prioritization Mechanisms u ANALYSIS + NETWORK PROCESS DESIGN Ý Analysis Tasks (Flow Diagram) Ý Workloads: Frequency, Duration, Day/Week/Month Cycles Ý Other Network Loads Ý Formulate ODBMS and HPSS behavior for use in Simulations (with RD45 and GIOD) Ý Priorities Among Tasks

MONARC WORKING GROUPS (2) u STEERING GROUP Ù CLASSIFICATION of “Baseline Models”, Based on Figures of Merit Ù COORDINATE Comparative Model study cycles Ù OVERSEE Model Evolution Ù PERIODIC REVIEWS Ù GATHER and COORDINATE RESOURCES

MONARC Principles of Project Design: SPE u Formality: Systematic production of tools, documents and procedures; Some central support (software and ops.) u Completeness: Nothing important left out u Abstraction: Represent in a simplified general form (Extensible) u Decomposition: Break into tractable parts u Hiding: Encapsulation; Design for variable granularity u “Standards”: Tools, Interfaces, Methods u Design Techniques: Stepwise Refinement Localization (Modules; OO Approach)