Db2ZAI Machine Learning in Db2 for z/os

Transcription

1 Db2ZAI Machine Learning in Terry Purcell, IBM Dec 10th, 2018 IBM z Analytics

2 What is Machine Learning? A twist on traditional data processing Put another way, it s about computers that learn without being explicitly programmed data program Computer output Traditional data processing data Computer output program Identify patterns not readily discerned by humans Build models of behavior from those patterns Machine learning Score or predict behavior with deployed models 2

3 Organizations can exploit ML technology Machine learning Constantly learns and adapts Avoids making the same mistakes Provides faster, deeper, improved insights Resulting in ü ü ü More-favorable business outcomes Reduced risks and costs New opportunities Churn analysis helps identify cause of churn and facilitates implementation of effective retention strategies Detect and understand life-threatening medical conditions and design ever more effective treatment programs Learn, predict weather patterns and energy production from renewable sources and integrate into grid more effectively Product recommendation, next purchase prediction, targeted offers individually tailored shopping experience Identify suspicious behavior, predict and prevent threats / fraud continually reduce business risks and costs 3

4 How about Db2 optimizer exploitation of ML? Machine learning Constantly learns and adapts Avoids making the same mistakes Provides faster, deeper, improved insights Resulting in ü ü ü More-favorable business outcomes Reduced risks and costs New opportunities How many optimizer mistakes can you tolerate while we learn? NOTE: This is a rhetorical question we know optimizer mistakes aren t tolerated Stakes are higher versus using machine learning to cut down on irrelevant online ads You wouldn t want a self-driving car to make a mistake Potential benefits: Enhanced reliability (less risk of performance regression) Improved performance 4

5 Machine Learning in - The query optimization challenge 5

6 IBM z Analytics What is the Db2 Optimizer? An analogy. The optimizer is responsible for Choosing the most efficient method of accessing the data for a given SQL statement Think of your transportation choices Start/end location, time of day, construction, traffic, available options/routes All can impact the quickest route???? 6

7 An Optimization Challenge Determining the degree of filtering for each predicate (i.e. the filter factor) is a key input to the optimizer s access path cost formulas Question: how many of you were born before parameter marker? BIRTHDATE <? It depends What if the substitution value is ? What if the substitution value is ? When we have a range predicate with a parameter marker or host variable we use a default filter factor (these are documented) 7

8 Tuning just ONE query: the user challenge Access path analysis can be a very complex, time-consuming task Depends on several factors (user skill level, available tooling, etc.) Solutions that we like to see: Determine RUNSTATS options for targeted database objects Source: SYSIBM.SYSSTATFEEDBACK (system-level) or DSN_STAT_FEEDBACK (query-level) Indexing add or modify Add OPTIMIZE FOR n ROWS to SQL ( paging-style subset of result set fetched) Solutions that are less preferred: BIND/REBIND with APREUSE or OPTHINTS Query rewrite Fudging catalog statistics Query tricks (0=1 for example) Organizations often revert to these Skill often missing 8

9 Large-scale REBIND can present another challenge for users Suppose you REBIND 10,000 static packages using APCOMPARE(WARN) After migration to new Db2 version, or after applying maintenance If 90% of packages show no access path changes, and 10% (1000 packages) do - how do you validate that changed access paths are OK? Review changes in REMARKS column of PLAN_TABLE if REBIND executed with EXPLAIN(YES) Do access paths look better (e.g., increase in MATCHCOLS)? Any red flags? If access path change looks OK, go with new instance of package What if access path change does NOT look OK? REBIND with SWITCH to revert to previous instance of package? Use OPTHINT to force use of an alternate access path? Try to tune affected queries? Approaches most commonly taken 9

10 Users want performance, but also stability We ve been at this a long time IBM invented cost-based SQL optimization The optimizer has been continually enhanced over 35+ years, based on user feedback and internal analysis by the Db2 development team Stability and reliability have long been priorities for query optimization A contributing factor: current and accurate catalog statistics, provided by RUNSTATS utility, provide inputs to the optimizer cost model Access path stability: a core characteristics of static SQL, also provided for dynamic SQL via prepared statement cache and (with Db2 12) dynamic plan stability Plan management (for static SQL) enables quick and easy switch to earlier instance of a package in the event of an access path regression 10

11 Optimizer development is guided by a twofold objective Deliver new access path choices that enhance performance and stability without compromising the quality and stability of existing access path choices 11

12 The Solution - IBM Db2 AI for z/os (Db2ZAI) IBM Db2 AI for z/os (Db2ZAI) empowers the optimizer in your engine to determine the best-performing query access paths based on your workload characteristics using machine learning Learns the patterns from the collected data from workloads in customer s unique operating environment and determine the optimal paths for SQL statements entering Built on top of the Machine Learning for z/os stack leveraging all the services without requiring data scientist support Db2 generates the training data, deploys and monitors/retrains models with MLz What is the Db2 Optimizer? It s job is to determine the best way to return data results from a query in Db2 12

13 IBM z Analytics Db2ZAI Business Value Up to 25% CPU Savings* Db2z infused with ML MLz IBM Z Value Proposition Enable optimizer to leverage Machine Learning for z/os ( MLz ) services and IBM Z Reduce CPU consumption and IT cost through optimization for best query access paths Improve Db2 application performance Rapid model learning specific to the data/application behavior per subsystem without requiring data science skills *CPU savings metric based on IBM internal benchmarks; actual savings will vary according to customer workloads and environment13

14 Db2ZAI and Query Optimization V1.1 (Sept 2018) Solution Targeting key inputs to the Optimizer s access path choice That Db2 must estimate today Actual value of Host variable/parameter markers Number of rows required by the application (OPTIMIZE FOR n ROWS) Results in more accurate estimation of SQL costs More reliable access path choices More accurate IDAA offload decision Better accuracy for predictive governor Learn from the actual customer workload By collecting query execution history and feeding information into ML models Available for the optimizer at next bind/rebind/prepare 14

15 More Detail on the Db2ZAI 1.1 Solution Using filter factors that are customized for an organization s particular query workload, better-performing access paths can be generated by the optimizer Internal IBM benchmarks have shown up to 25% reduction in CPU time when machine learning technology was employed via Db2ZAI for a set of Db2 queries compared to access paths selected by the Db2 optimizer without the aid of machine learning For static SQL, using ML-enhanced optimization requires package rebind The Db2ZAI performance dashboard provides a list of packages for which model-augmented access path selection is available Note: it is understood that improving the BIND/REBIND experience is important for enabling the fullest possible exploitation of machine-learning-enhanced optimization for static queries this is a parallel effort within development 15

16 Db2ZAI and Query Optimization V (Dec 2018) V Adds Sort - Improved memory allocation for (SQL) sort Learned from prior SQL execution SQL that will benefit most SQL with DISTINCT ORDER BY with larger FETCH FIRST n ROWS values Longer sort rows (> 4K row length) Parallelism - Smarter parallelism exploitation Relies upon (prior) actual costs rather than estimates Allows shorter running SQL to benefit from (modest) parallelism DEGREE Limited to static packages bound with DEGREE(ANY) 16

17 Db2ZAI - Architecture SQL Applications Db2 12 for z/os Db2 Optimizer Learning Task (ziip Offloaded*) Model Life Cycle Automation Subtasks Performance History ML Scoring Service* IBM Open Data Analytics* (IZODA) Machine Learning for z/os V1.2 Db2ZAI User Interface MLz Model Services - Db2ZAI - Machine Learning for z/os - z/os - x86 or z/linux * ziip Offload 17

18 Key Db2ZAI Processes (1 of 2) Execution History in Db2 Executes on general CPs Overhead typically 1-2% per SQL SQL execution information written out to new Db2 pseudo-catalog tables per execution No traces to be enabled - Information is NOT from trace records. Internal Db2 data (similar to statement cache stats). Db2 tables are created as part of User Interface setup of Db2ZAI Db2ZAI manages data summarization/purge (see learning task next slide) User managed for REORG, RUNSTATS 18

19 Key Db2ZAI Processes (2 of 2) Learning task(s) in Db2 ziip Eligible Individual tasks that sleep/wakeup to process Model creation/training/deployment Execution history summarization and cleanup User Interface in MLz Create connection Start/Stop ML per Db2 subsystem or data sharing group Display package and SQL performance metrics Display model statistics 19

20 Key Db2ZAI Externals Externals all surfaced from the UI Creating Db2 connections Start/Stop ML 20

21 Key Db2ZAI Externals Once enabled via UI Db2ZAI will Collect execution history for each SQL* execution *Limited to those SQLs from selected packages Once SQL execution hits threshold (currently 100) ML models will be created/trained/deployed for that SQL Next BIND/REBIND/Prepare/Explain will use that model Goal is to integrate into customer s existing BIND/REBIND processes REBIND without APREUSE to choose a new access path NO BIND commands or zparm control UI reports packages using ML and/or having a model ready 21

22 Key Db2ZAI Externals Externals all surfaced from the UI Identifying packages that are using and/or candidates for ML SQL Performance Metrics 22

23 Key Db2ZAI Externals Performance Metrics 23

24 Key Db2ZAI Externals Model Dashboard Since Db2ZAI manages (deploys, trains) models, model statistics are an FYI to show number of SQL statements using each model type 24

25 Db2ZAI V (Dec 2018) Usability Enhancements Scope/Performance Package list Allows User to provide package list for Db2ZAI Reduces overhead (limited to only those packages) Reduces risk by allowing customers to ease into Machine Learning Allows customer to choose packages with THEIR list of packages with best ROI Db2ZAI provides a recommended list for customer to choose from Static packages that are expected to benefit Model training concurrency control Single threading of model training on MLz Reduces overhead of concurrent MLz processing Most Db2 bind/rebind processes occur in batches, thus model training does not need to be performed now Pause Db2 history collection after initial model training To reduce Db2 (non-ziip) overhead history collection is paused as soon as model training is complete May re-initiate collection when re-training is required Score caching To improve performance of ML scoring calls 25

26 Key Db2ZAI Externals V Update Package Selection Search/Select and unselect packages for Db2ZAI 26

27 Db2ZAI Software Pre-requisites IBM Db2 AI for z/os, V requires: z/os 2.1 or later, with the Product Registration Services component. IBM Db2 12 for z/os (either 5650-DB2 or 5770-AF3), plus APAR PH05323 (for V features). Note: FL500 must be activated. IBM Machine Learning for z/os, V1.2 (5698-ML1) plus APARs PI99400 and PI99401 and fix pack Including IBM Open Data Analytics for z/os 27

28 How do I know if my workload(s) will benefit? If you re asking Can you tell me what the Db2 Optimizer doesn t know before Machine Learning has learned what the Db2 Optimizer doesn t know? Db2 development has SQL script that can identify the number of SQL statements in your system with risk of a sub-optimal access path due to the combination of Host variables/parameter markers WHERE clause predicates with data skew or range predicates AND Index design which exposes the optimizer to index choices that have above risk OR multi-table queries where risk of choosing the wrong join sequence or method exists Regularly updated to include latest Db2ZAI features Dec 2018 includes Sort optimizations and improved parallelism exploitation Contact your IBM rep for the Db2ZAI Analyzer 28

29 Long Term Goal of Db2ZAI Make Db2 Self-Tuning / Self-Managing Query optimization System health Targeted parallelism Query level and subsystem memory allocations (BP, wf, ridpool, sort, etc.) PCTFREE/FREEPAGE allocations based upon INSERT patterns IDAA Offload decision Db2 system memory configuration Utilities - REORG minimization, RUNSTATS/Optimizer integration Applying maintenance Opportunities are (currently) unlimited And in Continuous Delivery Db2ZAI continues to add use cases frequently 29

30 IBM z Analytics Db2ZAI Complements the Accelerator 100% of SQL statements at 100% of customers Including Accelerator offload OLTP Transactions High concurrency OLAP Adhoc queries/reports IBM Db2 Analytics Accelerator Standard reports Processing Hybrid Transactional & Analytical Processing The optimizer is responsible for data access decisions The Accelerator is merely another choice for the optimizer Like adding an airplane to your transportation options Db2ZAI improves the performance of SQL statements executing on Db2 30

31 Db2ZAI Complements all Performance Offerings Db2ZAI Improves the optimizer s intelligence by learning patterns from the collected data from workloads in customer s unique operating environment and determine the optimal paths for SQL statements entering Query Workload Tuner Reduces skill requirements for tuning Db2 queries and improving application performance. Advisors give actionable tuning recommendations. Query Monitor Low overhead SQL monitor. Can proactively identify SQL statements consuming excessive resources and inadequate object structures. Can determine which tables and indexes are actually being used by queries for tuning. IBM Db2 Analytics Accelerator Optimizes the performance of longer running (data warehouse or analytics) queries 31

32 Where to find more information on Db2ZAI The announcement letter: 01.ibm.com/common/ssi/ShowDoc.wss?docURL=/common/ssi/rep_ca/1/897/ENUS /index.html&request_locale=en The IBM Knowledge Center: Overview video: 32

33 Thank You! Terry Purcell 33