Leveraging AI to Optimize Database Architectures in a Multicloud World [LRN3299].pdf

1、1|Copyright 2024 Deloitte Development LLC.All rights reserved.Leveraging AI/ML to Optimize Architectures in a Multi-Cloud WorldLRN3299:September 11,2024|4:45 PM PDTO R A C L E C L O U D W O R L D 2 0 2 4what day is it?2|Copyright 2024 Deloitte Development LLC.All rights reserved.Introductions Not al

2、l workloads are“commodity cloud”capable due to on-premises infrastructure capability Establish corporate cloud shapes or patterns to increase application and database agility.Optimizing multi-cloud solutions to reduce cloud costs and improve application agility Machine Learning has many use cases ac

3、ross the industry,as well as projecting services models for all industries Establishing data sources to build on-premises and cloud based data metrics for predicting workload requirementsThings to considerThings to considerObjectives for Workload AssessmentObjectives for the SessionHenry(Hank)Tullis

4、Henry(Hank)TullisMult-Cloud Infrastructure and Data ArchitectDeloitte ConsultingDeloitte ConsultingMark SaltzmanMark SaltzmanConsumer Enterprise ArchitectDeloitte ConsultingDeloitte ConsultingMichelle Michelle MalcherMalcherDirector,Database Product ManagementOracle CorporationOracle CorporationColl

5、ect workload metrics with respect to database environmentsAnalyze workload metrics for performance and classificationsReview and establish standardized patterns reducing the architecture requirementsProject standardized patterns on available cloud shape and service catalogsReview the approach for st

6、rategic workload assessment and benefits of the establishing workload characteristicsDiscuss workload classification and workload projection for cloud deployment strategiesHow is Machine Learning changing the approach for strategic workload assessmentsHow the current strategic workload assessment pr

7、ocess is integrated into Oracles Cloud Infrastructure AI/ML service catalog3|Copyright 2024 Deloitte Development LLC.All rights reserved.Group Discussion .What are YOUR objectivesWe have three engineers/architects with a combined experience of 100 years What would YOU like to hear?4|Copyright 2024 D

8、eloitte Development LLC.All rights reserved.WelcomeAfter years of establishing Exadata workloads from consolidated commodity platforms,we are reversing the approach.How do we transform Exadata database environments to commodity solutions in the cloud?Most on-premise workloads consist of individual(h

9、undreds)workloads on commodity,converged and engineered systemsResults:Recommended Cloud Architecture Resource Pattern Definition Cloud Shape/Resource Pattern xRef Fit-Gap Analysis Cloud Transformation Roadmap StandardizeIsolateAnalyzeDistribute5|Copyright 2024 Deloitte Development LLC.All rights re

10、served.Data persistence includes MySQL and code developed on Macbook M3 DisclaimerDisclaimerComments and next steps will consider the approach for establishing an enterprise solution6|Copyright 2024 Deloitte Development LLC.All rights reserved.Multiple OptionsTransformation capable for Oracle,Postgr

11、eSQL,MySQL or other non-structured solutions.Oracle Specialized Workloads require performance and/or advanced ACID features that Oracle RDBMS providesWorkload Specialized(Exadata)Workloads require specialized hardware and software solutions to provide the required performance and throughputAre all o

12、n-premises workloads compatible with Cloud Solutions?Organizations are transitioning on-premises workloads to cloud service providers.These organizations cloud service provider engineers to assess and recommend a target solution.However,is using a single cloud provider any more efficient than mainta

13、ining a data center?Leveraging the benefits of all cloud providers elevates organization opportunities.Sometimes,refactoring applications is not the solution for high-throughput or high-compute applications.These situations require specific cloud services and specialized platforms.Workload Distribut

14、ionAnyone here currently multi-cloud/hybrid cloud?7|Copyright 2024 Deloitte Development LLC.All rights reserved.Escaping the constraints of the physical data centerEstablish a Multi-Cloud SolutionLeveraging AI/ML to establish a solutionEstablishing a virtual data center and distributed data architec

15、ture8|Copyright 2024 Deloitte Development LLC.All rights reserved.Data CollectionData LoadingData Preparation and PruningUtilizing Python programs to gather,condition and load data into an active data storeEvaluate data and time series for complete data.Evaluate data components for correct reporting

16、Data Analysis and ReviewAnalyze data segments and data components for performance and outliers based on reporting requirementsExtract workload and demographic details from active databases on-premisesPattern DefinitionReview data and application resource utilization characteristics.Define standard p

17、atterns defining consolidated resource utilizations for cloud shape associationCloud Mapping and Transition PlanA manual process for assessing Univariate Time Series and Multivariate Time Series.Univariate Time Series establishes the foundation for resource workload predictions for each of the workl

18、oad components.Multivariate Time Series combines values from the predictions to characterize the workloads into standardized patterns.An 8+Week Cloud Journey9|Copyright 2024 Deloitte Development LLC.All rights reserved.Establishing the on-premises workload patterns and simulating a cloud deployment

19、reduces risk during and after the migration ensuring a successful cloud migrationLife Cycle of the workload assessment and cloud strategy engagementCollection:Capturing the current resource and demographic requirements of applications and databases within the legacy environmentAnalysis:Consolidate w

20、orkload characteristics by category to establish trends and associations within the environmentEstablish Pattern Architecture(Categorize)The required resources are supported 100%Results:Recommended Cloud ArchitectureResource Pattern DefinitionCloud Shape/Resource Pattern xRefFit-Gap AnalysisCloud Tr

21、ansformation Roadmap Standardize/Modernize:Converts CPU resources to standard processor.Highly-Efficient,Tool-Supported Modernization StandardizeIsolateAnalyzeConsolidateUnivariate Time SeriesMost workload metrics are defined as single value time series,such as CPU,Throughput(MBPS),Memory11|Copyrigh

22、t 2024 Deloitte Development LLC.All rights reserved.Workload ComponentsSession CPUThe number of active sessions on CPU at the point in time.Establishes the number of CPU required and projected for workloads.Storage ThroughputDisk Space UtilizationMemory AllocationThe amount of data requested from th

23、e storage system by the database system for the point in time.Projects the required storage throughput for a database and application.The physical amount of storage required by the database for the point in time,includes both static and dynamic data sizes.The physical amount of memory allocated by t

24、he database for the point in time.The memory includes shared(anonymous)and private data points for the database,application and operating system.Network UtilizationThe amount of data requested from the network system by the database system for the point in time.Projects the required network throughp

25、ut for a database and application.What?Why arent you including IOPS in the workload assessment?12|Copyright 2024 Deloitte Development LLC.All rights reserved.High Throughput workload Most commodity platforms and cloud solutions struggle beyond 12GB/secFeb 152024Feb 16Feb 17Feb 18Feb 19Feb 20Feb 21Fe

26、b 22010k20k30k40k50kInstance 1Instance 2Throughput(MBPS)for Database:fceprd90Collection DateThroughput(MBPS)Storage Utilization over TimeMultiple periods of high throughput with persistent throughput for most of February 2113|Copyright 2024 Deloitte Development LLC.All rights reserved.Network worklo

27、ad outputHigh network throughput implies challenges with application and connectionFeb 152024Feb 16Feb 17Feb 18Feb 19Feb 20Feb 21Feb 22020M40M60M80M100MFrom Client To ClientNetwork Traffic to Instance 1 of rmnprd00Collection DateThroughput(MBPS)Network is pegged at 100MB/sec for both Oracle instance

28、s deployed on a quarter rack ExadataNetwork Utilization over Time14|Copyright 2024 Deloitte Development LLC.All rights reserved.CALYPSO(CALLIE)TULLISStill here?Technical Architecture and Solutioning15|Copyright 2024 Deloitte Development LLC.All rights reserved.High Consistent CPU ActivityHigh active

29、 and consistent CPU utilization Active Sessions on CPUActive Sessions over Time16|Copyright 2024 Deloitte Development LLC.All rights reserved.Nice Graphs but”So What?”These graphs represent a subset of queries supporting ALL database collectionsMaximumMeanMinimumProjectionworkload 164.535workload 28

30、526workload 385.534workload n74.524Each Workload Component Manually compile this information for more than 8 individual workload data points Establish queries for identifying the each column for the maximum,minimum and the mean Manually perform a projection based on the statistical view of the data

31、components17|Copyright 2024 Deloitte Development LLC.All rights reserved.Establishing Workload PatternsConsolidating workload predictions to establish a characterization of dataMultivariate Time SeriesTo establish Patterns,we look at a combination of metrics to define(classify)groupings of metrics.1

32、8|Copyright 2024 Deloitte Development LLC.All rights reserved.Sessions and ThroughputEstablish potential trends with respect to the number of active sessions and maximum throughput19|Copyright 2024 Deloitte Development LLC.All rights reserved.Database Size and Active SessionsEstablish potential tren

33、ds with respect to the number of active sessions and database size 20|Copyright 2024 Deloitte Development LLC.All rights reserved.Cohort 1Cohort 2Cohort 3OutliersDatabase Size and ThroughputEstablish potential trends with respect to the database size and the maximum throughput21|Copyright 2024 Deloi

34、tte Development LLC.All rights reserved.Establishing the workload distribution Focus on resource requirements associated with each of the workloads to drive the cloud migration path for data workloadsRe-Platform with OracleSituational CloudTransformation PotentialTransform to alternative database(Or

35、acle,MySQL,PostgreSQL)environments in AWS established by workload.The transformation focused on Oracle license reduction,leverages alternative database shapes for application support.Focus on specialized hardware and/or cloud services based on database and application resource requirements.Consider

36、Oracle Exadata on-premises(ExaCC)or in-cloud(OCI ExaCS,oracleAzure,oracleGoogle,oracleAWS)Transform to Oracle in the cloud.AWS manages database containers;Leverage workload analysis for sizing and transformation considerations.Migrate database and ETL services to Oracle RDS,supporting AWS based appl

37、ications.13Workload Distribution261Applications13Applications8Applications22|Copyright 2024 Deloitte Development LLC.All rights reserved.There has to be questions!Theodore(Teddy)SaltzmanContent Direction and Analysis23|Copyright 2024 Deloitte Development LLC.All rights reserved.Very Low Resource(VLR

38、)Throughput 100 MBPSCPU 10DB Size 100 GBSmall Resource(SR)Throughput 100 MBPSCPU 10DB Size 100&3,500 MBPSCPU 1,000 GB&100&10DB Size 10,000 GBHigh CPU and High ThroughputThroughput 12,500 MBPSCPU 10DB Size 10,000 GBPattern DistributionHigh CPU Resource(HCR)Throughput 100&10DB Size 1,000 GB&100&3,500

39、MBPSCPU 10DB Size 3,500&10DB Size 10,000 GB24|Copyright 2024 Deloitte Development LLC.All rights reserved.Why are Patterns Defined?Simplify the process for associating standard workloads to cloud shapes and services Establish cloud shapes that provide the resource requirements of the corporate patte

40、rn identify the appropriate data service supporting the requirements of the corporate patternMinimum number of standardized PATTERNSunlimitedshapes25|Copyright 2024 Deloitte Development LLC.All rights reserved.Generally this takes between 8 10 weeksWhat happens if we add Machine Learning 26|Copyrigh

41、t 2024 Deloitte Development LLC.All rights reserved.Lets add some Oracle Milo MalcherChief Happiness Officer27|Copyright 2024 Deloitte Development LLC.All rights reserved.Execute and evaluate ML ModelData CollectionData LoadingData Preparation and PruningEstablish”prediction”baselineThe same Journey

42、 outline that was presented earlier is now adjusted with considerations for Machine Learning integration.With Univariate Data points Establish a model focused on regression and data point prediction.With Multivariate Data points Define a model focused on categorization based on multiple workload dat

43、a points.Establish patterns addressing the resource utilization Considering the“Journey”Adjust prediction and time series focus to improve the model accuracy.Utilizing Python programs to gather,condition and load data into an active data storeEvaluate data and time series for complete data.Evaluate

44、data components for correct reportingAnalyze data segments and data components for performance and outliers based on reporting requirementsExtract workload and demographic details from active databases on-premisesExecute and evaluate ML Model for Classification analysisLeverage a multi-class classif

45、ier to leverage multiple workload predictions to define the cloud PatternsDetermine AccurracyEmploy Keras metrics for establishing model accuracy.Reduce“error”values based on the prediction function.Refine and persist prediction28|Copyright 2024 Deloitte Development LLC.All rights reserved.Establish

46、ing a Cloud Journey Assess and adjust data set for model accuracy Create a Sequential model based on data set and DNN settings(Neuron Density)Compile and fit model,focusing on reducing the loss value for tuning.Using machine learning to predict and plan a cloud solutionPrediction approach Establish

47、linear regression to identify outlier values Integrate prediction models to establish baseline data prediction Integrate statistical model predictions through Keras Mean Squared Error Mean Absolute ErrorUnivariant Values Consolidate data values and data labels Establish model for considering multipl

48、e data points for classification Compile and fit model,focusing on reducing the loss value for tuning.Multivariant ModelOn average 2 weeksSavings in more than 4 weeks for the assessment alters the use case for the assessment processThe assessment provides immediate feedback for cloud solution to inc

49、rease cloud agility29|Copyright 2024 Deloitte Development LLC.All rights reserved.29|Copyright 2024 Deloitte Development LLC.All rights reserved.Next Steps for refining AI driven multi-cloud design toolIntegrating enterprise class persistence and machine learning models for establishing a larger dat

50、a set to Generate Intelligent Cloud ModelsRefine persistence and Refine persistence and collection approachcollection approachIntegrate OCI based Integrate OCI based GenAIGenAI for establishing multifor establishing multi-cloud cloud models based on univariate models based on univariate and multivar

51、iate ML modelsand multivariate ML modelsPort database and Port database and application to Oracle 23aiapplication to Oracle 23aiEstablish focused data Establish focused data management and ML model management and ML model governancegovernancePersistentVolumeOCI ObjectStorageBlock Storage Currently,w

52、e employ a CSV and fixed format file structure A more portable file architecture for collections,such as JSON or parquet,would provide an effective collection and consumable media.Existing application is utilizing an open source copy of MySQL and Python/TensorFlow Deploy to a more enterprise archite

53、cture with integrated AI/ML capability Port TensorFlow application to existing Django framework All models,data and application code are single user and owner.An enterprise solution would expose the environment to multiple users and developers Establish DevOps pipeline to support code,model and para

54、meter managementOracle AI/GenAI Cloud provisioning and pattern to cloud shape association is still manual leverage GenAI to establish“complete”cloud solutions Investigate options for real time model support to make cloud deployments routine30|Copyright 2024 Deloitte Development LLC.All rights reserv

55、ed.About DeloitteDeloitte refers to one or more of Deloitte Touche Tohmatsu Limited,a UK private company limited by guarantee,and its network of member firms,each of which is a legally separate and independent entity.Please see for a detailed description of the legal structure of Deloitte Touche Toh

56、matsu Limited and its member firms.Please see for a detailed description of the legal structure of Deloitte LLP and its subsidiaries.Certain services may not be available to attest clients under the rules and regulations of public accounting.This publication contains general information only and Del

57、oitte is not,by means of this publication,rendering accounting,business,financial,investment,legal,tax,or other professional advice or services.This publication is not a substitute for such professional advice or services,nor should it be used as a basis for any decision or action that may affect yo

58、ur business.Before making any decision or taking any action that may affect your business,you should consult a qualified professional advisor.Deloitte shall not be responsible for any loss sustained by any person who relies on this publication.Copyright 2024 Deloitte Development LLC.All rights reserved.Thank you!