Capacity Planning for the Newer Workloads: Capacity Planning for the Newer Workloads Linwood Merritt Capital One Services, Inc. linwood.merritt@capitalone.com


Disclaimer: Disclaimer These generic issues are addressed by this presentation: Vendor capacity ratings e-Commerce Continuous availability Data warehousing Growth rates This presentation contains no specific business-related information.


Introduction: Environment: Introduction: Environment Capital One 5th largest card issuer in the United States Capital One to S&P 500 in 1998 Fortune 500 company (#260) Managed loans at $48.6 billion as of Q1 2002 Accounts at 46.6 million as of Q1 2002 Fortune 100 “Best Places to Work in America” CIO 100 Award “Master of the Customer Connection” Information Week “Innovation 100” Award Winner ComputerWorld “Top 100 places to work in IT”


Outline of Approach: Outline of Approach Understand behavior and issues around workloads, hardware, and data Create projections and build recommendations. Report the findings.


Outline of Presentation: Outline of Presentation Discussion of workload types and capacity projection approaches Overall summary of issues and approaches Examples


What Workloads?: What Workloads? E-Commerce Relational database systems Mainframe-class UNIX Multiple platforms New characteristics


e-Commerce WorkloadsDirect to Client (business-to-business): e-Commerce Workloads Direct to Client (business-to-business) Access Internet Leased line Services Point of Care / Point of Sale Value-added analysis


e-Commerce WorkloadsDirect to Customer: e-Commerce Workloads Direct to Customer Access Internet Dial-in Services Marketing Account query


e-Commerce WorkloadsHow to Predict: e-Commerce Workloads How to Predict Take business projections of volumes or users (include fudge factor) Estimate transaction volumes and CPU/transaction Convert to normalized unit such as MIPS


Relational Databases: Relational Databases Sub-second (OLTP), decision support / data mining Distributed gateways Database machines Redundant data with extracts How to predict: estimate a factor over current database demand or take usage estimates


Mainframe-Class Unix: Mainframe-Class Unix Types: Mainframe USS or Linux, Future UNIX vendor offerings Candidate applications Web server Vendor-ported applications User-ported / new applications How to predict: Estimate by timeframe Add factor to growth rates


Multiple Platforms: Multiple Platforms Mainframe: plan like existing applications (#users, transactions * CPU/transaction, application look-alikes, sizing tools) Distributed: use vendor sizing, modeling tools, existing applications Network: use network simulation tools, rules-of-thumb, bandwidth calculations


New Characteristics: New Characteristics External users Continuous availability New user interfaces Cross-platform


External Users: External Users Drive need for continuous availability Different access patterns (e.g., doctor’s office vs. call center) Service level measurement - harder to put agent on external workstations


Continuous Availability: Continuous Availability Driven by external users 24x7 schedule Application redesign Data Sharing: CPU overhead Coupling Facility Expansion of “prime shift” 99.999% “up time” Redundancy, overhead Availability reporting


User Interfaces: User Interfaces TCP/IP - no “definite response” (end-to-end response time measurement) Multiple internal transactions per “mouse click” Response time measurement: Agent on workstations Scripting from “robots”


Cross Platform Applications: Cross Platform Applications Only unified view: simulation package Each platform (“silo”) can be analyzed separately. Different application development groups May be able to cross-validate user numbers


Types of Implementation (1): Types of Implementation (1) Standalone / “shrink-wrap” Layered onto legacy applications New mainframe application code GUI front-end Browser Middle-tier (Unix or NT) MQSeries - can add middle-tier and new mainframe applications


Types of Implementation (2): Types of Implementation (2) Legacy extracts Re-engineered legacy applications Convergence of business rules / applications Re-usable components Redundant access Salvage investment, fix Band-Aids Simplify logic, reduce platform complexity


What Are We Analyzing?(Mainframe): What Are We Analyzing? (Mainframe) MIPS - growth, latent demand, software cost Memory - track and watch 2 GB limit on central storage (goes away with 64-bit) I/O - channels, gigabytes of disk, tape Coupling Facility - Parallel Sysplex, Shared Data, continuous availability Vendor upgrade paths New partitions


What Are We Analyzing?(Distributed): What Are We Analyzing? (Distributed) Number and types of platforms CPU, memory, disk space Bandwidth Location of applications / processes Platform limitations (CPU, memory) Software pricing considerations Porting opportunities


Measurement of New Workloads : Measurement of New Workloads Summarize by platform: Workload rules (process or user names) Processes by descending CPU% Resources: CPU, memory, disk space, Coupling Facility, network traffic Growth: Resources/user/application Number of users + application changes


Distributed Approach: Distributed Approach Consider tiers of service (not currently at Capital One) Address service level measurement issue Implement reporting Add to Capacity Plan “Silo” vs. “Application”


Tiers of Service“Platinum”: Tiers of Service “Platinum” Most expensive Modeling product Install in one server for each major application, use collection product for other servers


Tiers of Service“Gold”: Tiers of Service “Gold” Collection product Capacity planning with Rules of Thumb


Tiers of Service“Brass”: Tiers of Service “Brass” Least expensive (man-hours only) “Native” Unix scripts NT PerfMon


Service Level Measurement: Service Level Measurement API call at workstation - “Applications Response Measurement” (ARM) or Windows 2000 trace API calls Agents: software tracing of Windows API calls - can be installed in a subset of end-user base (sampling) Scripting (“robots”) Stop watch sampling and logging


Distributed Reporting: Distributed Reporting


Add to Capacity Plan: Add to Capacity Plan


Scope of Analysis: Scope of Analysis Silos Look at each hardware/application environment independently. Applications Look at each application as a whole. Application instrumentation Inference: put platform silos together.


Analyzing the DataGrowth Rates: Analyzing the Data Growth Rates General list of business plans List of technical scenarios Timeline Estimate median and maximum likely MIPS/CPU/users/business units Derive scenario growth rates


Analyzing the DataAdditional Resources: Analyzing the Data Additional Resources Parallel Sysplex (Coupling Facility): important for continuous availability, level set functionality Disk / channels / tape: disk megabytes, channel maximum, tape connectivity Communications connectivity: new partitions for availability Memory: 2 GB constraint, 64-bit


Growth: Growth “Baseline” growth “Scenario” growth Independent events (merger/acquisition, potential major project)


Example 1: Mainframe Upgrade: Example 1: Mainframe Upgrade Task force, led by Capacity Planner Driven by expiring three-year lease (CPU replacement, three-year planning horizon) “Vendor parade” - presentations and dialogues Upgrade paths Technology / service differences References / site visits Capacity sizing: MIPS charts, LSPR / sizing tools


Mainframe Upgrade Deliverables: Mainframe Upgrade Deliverables Document Business drivers and technical scenarios Growth forecasts Vendor options and growth paths Coupling Facility / Parallel Sysplex Evaluation Difference thresholds: MIPS claims, price/MIPS, ICF Differentiators


Business and Technical: Business and Technical Business Drivers Cost management External business Improved data access Business expansion Technical Scenarios Consolidation of distributed servers Continuous availability Significant external business Data Warehousing Acquisition/merger


Projections: Projections Make educated guess by timeframe for each scenario Add to “baseline” growth Convert to growth rate Use both “baseline” and “scenario growth” Compare maximum scenario growth to maximum for platform family


Impact Analysis: Impact Analysis


Scenario Timeline: Scenario Timeline


Vendor Upgrade PathsDetail: Vendor Upgrade Paths Detail Use logarithms: Start*CAGR^x = Threshold x years = log(Threshold/Start)/log(CAGR) Model MIPS MSU +40%/Yr +25%/Yr GS2068E 952 160 Aug-00 Sep-00 GS2074E 1013 171 Oct-00 Dec-00 GS2084E 1141 193 Apr-01 Jul-01 GS2094E 1260 213 Sep-01 Dec-01 GS2104E 1378 234 Nov-01 May-02


Vendor Upgrade PathsSummary: Vendor Upgrade Paths Summary


Upgrade Document: Upgrade Document


Example 2: UNIX Modeling: Example 2: UNIX Modeling Modeling product installed on MQSeries server Application running with a known number of users Projected rollout schedule used to drive model Mainframe side: CICS application, IMS load


UNIX Platform Workloads: UNIX Platform Workloads Two primary workloads: MQSeries userids (mqm*) - memory intensive Messaging application processes (MDA*) - “CPU intensive”


Workload Modeling Methodology: Workload Modeling Methodology MQSeries - Calculate relative workload intensity, enter model ratio. Messaging application processes - Keep constant until application is removed from platform (“design loop” - always uses 1 CPU). Must adjust across CPU upgrade to continue using 1 CPU.


Track Across Upgrade: Track Across Upgrade


Model Spreadsheet: Model Spreadsheet


Model Presentation: Model Presentation Timeframe: April 2000 #Users: 180, 100 Ratios: 1.27, 1.00 Config: F50/02,2GB Comment: Add Event1 Users


Validation - Tracking Users(on mainframe): Validation - Tracking Users (on mainframe) //ECLUSRS EXEC SASV8,REGION=0M //ECLD1 DD DSN=XYZ.PRD.A.AAAPRD.I.VOLFIL,DISP=SHR //ECLDPDB DD DSN=CAPLAN.PRD.ECLDPDB,DISP=OLD //SYSIN DD *,DLM=@@ data ecld1; format date date.; format dt datetime.; INFILE ECLD1 MISSOVER; INPUT @1 RECNUM $CHAR5. @6 RECTYPE $CHAR8. @14 USERCT $CHAR5. @19 USERMAX $CHAR5.; if recnum =: '99999' and rectype =: 'TCSCONFG'; dt = datetime(); date = datepart(dt); hour = hour(dt); data ecldpdb.users; update ecldpdb.users ecld1; by date hour; proc print; title 'Ecloud1 Users';


Example 3: Server Replacement: Example 3: Server Replacement Project: replace “old” NT servers Application: Imaging servers Capacity sizing data: Rules-of-thumb analysis by vendor, using projected claims/minute and processor clock speeds Benchmark information


Server Replacement Process: Server Replacement Process Multiple servers: each server is a workload, must be sized separately. Enumerate and measure servers. Apply growth rates and determine processing power requirements for the replacements. Research available configurations and order appropriate server configurations. Track CPU utilization across the upgrades. Update relative capacity specs for next upgrade.


Server Sizing: Server Sizing Find (or derive) benchmark capacity ratings for starting and replacement configurations. Apply an estimate of current CPU utilization, a growth percentage, and a “peak/average” and performance buffer (+100% for this study). Output: estimated percentages of a standard configuration. The number of estimated CPUs needed (23) came very close to the vendor’s original number of 24.


Sizing Spreadsheet: Sizing Spreadsheet


Example 4: Hundreds of Servers: Example 4: Hundreds of Servers Data capture Reporting Business drivers


Data Capture: Data Capture Time-based scheduling product Script-based data “pull” Issue: data loss, time to find and rebuild Potential fixes: Product Data “push” from servers


Data Reporting, Analysis: Data Reporting, Analysis Color-based “health index” (Concord NetHealth metric). Statistical Analysis (over two standard deviations from mean) Thumbnail drilldown graphs Automatic generation of html “Treemap” graphs


Health Index *: Health Index * * Concord NetHealth metric


Statistical Process Control: Statistical Process Control cmg


Thumbnail Html: Thumbnail Html


Automatic Generation of Html: Automatic Generation of Html Driven by “matrix” Originally spreadsheet Converted to relational database Ultimate capacity planning solution: information by server, application, platform, business driver SAS code - builds web pages and hyperlinks


Treemap: Treemap Paper by Ben Shneiderman, University of Maryland, http://www.cs.umd.edu/hcil/treemaps ASSDSDFVVBNM XSDFFGFRRFHFHJKJKLLXXXXX XESDGFKOKJHHMM XESDGFKOKJ DERFFVBBNHGFF XESDG XES SDEFBJMGG XESDG


Business Drivers: Business Drivers Capacity Councils - business units responsible for capacity planning of “demand” side Capacity Planners - build projections based on business drivers and historical trending


Business Driver Based Forecasts: Business Driver Based Forecasts Server Application Application Application Business Driver Business Driver Projections Projections


Regression Analysis: Regression Analysis Widgets Gadgets Customers CPU By month (input = Widgets, Gadgets, Customers): projection =Widgets*f1 + Gadgets*f2 + Customers*f3; f1 f2 f3 Output = Coefficients Input = CPU and Business Drivers by month


Graphical Output: Graphical Output Widgets Gadgets Customers


Enterprise “Capacity at a Glance”: Enterprise “Capacity at a Glance”


SummaryIssues: Summary Issues Access patterns and schedules Platforms (more types and numbers) Resources (what to track) Levels of capacity management Reporting of utilization and service levels, for large numbers of platforms Higher availability (redundancy, reporting) Deriving and reporting projections


SummaryDeriving Projections: Summary Deriving Projections Basic capacity planning: Growth rates Upgrade thresholds Aggressive estimate of “scenario” demand Bracket growth: Lower end: “baseline” Upper end: “scenarios”


SummaryTypes of Projections: Summary Types of Projections Number of transactions Number of users Number of platforms Application sizing input Application complexity Fraction of an existing workload Growth rate


SummaryCapacity Planning: Summary Capacity Planning Projections based on application and platform Levels of capacity planning service Report on all enterprise resources Organize data with “matrix” database