Understanding Buffer Pool Performance and Tuning in DB2 UDB v8.

Jeff Brokaw Product Manager Quest Central for DB2

November 2, 2005

You Can Expect More

Agenda Buffer pool overview Buffer pool management Collecting performance metrics Evaluating performance metrics
Physical vs. logical IO Asynchronous vs. synchronous IO Page cleaner tuning

Q&A

You Can Expect More

Introduction Q: With so many tuning areas to focus on why should we pay special attention to the buffer pool? A: DB2 uses the buffer pool to offset the performance disparity between CPU and disk, so they are vital for optimizing transaction throughput. From the 8.2 Admin Guide: Performance (emphasis mine):
Because most data manipulation takes place in buffer pools, configuring buffer pools is the single most important tuning area.

You Can Expect More

You Are Here DB2 Process Model

Client Application

Prefetchers

External Storage

Listener

Coordinator

Subagents

Buffer Pool

Page Cleaners

Logger

Log

You Can Expect More

Buffer Pool Overview

Basics and Terminology

Area of memory into which database pages are read, modified, and held during processing
A hit occurs when required page is found in buffer pool A miss occurs when required page not found in buffer pool

Pages in buffer pool exist in one of three states:

1. In use 2. Dirty 3. Clean

You Can Expect More

Buffer Pool Overview

Basics and Terminology (contd)

Pages read into buffer pool from disk by:
Agents, using synchronous I/O Prefetchers, using asynchronous I/O

Pages written to disk from buffer pool by:

Agents, using synchronous I/O Page cleaners, using asynchronous I/O

You Can Expect More

Buffer Pool Overview

Tuning Goals
Recovery vs. Performance
Recovery wants everything on disk Database manager wants everything in memory Which goal is more important largely depends on data Which combination works best for both goals

High hit ratio means logical reads usually successful, avoiding disk access Normally want to maximize asynchronous I/O to minimize waits

You Can Expect More

Buffer Pool Overview

Prefetchers and Page Cleaners

DB2 uses prefetch and page cleaner processes to reduce the time that agents have to wait for IO operations to complete New in 8.1.4: Proactive Page Cleaning, AKA Alternate Page Cleaning
DB2_USE_ALTERNATE_PAGE_CLEANING=ON

You Can Expect More

Buffer Pool Overview

Prefetch
Prefetchers try to ensure that agents doing tablespace scans never wait for disk I/O Agents send asynchronous read-ahead request to common prefetch queue Prefetchers service the request by bringing the requested pages into the buffer pool Should result in logical read, drive hit rate up Two types:
List prefetch, efficiently reads a set of non-consecutive data pages Sequential prefetch, reads consecutive pages

You Can Expect More

Buffer Pool Overview

Page Cleaners
Two main goals
Ensure that agents reading a page into the buffer pool never need to first flush a dirty page to disk to free up a slot Speed recovery time by reducing # log files needed

Configuration parameters
num_iocleaners number of page cleaners for a database chngpgs_thresh percentage of changed pages at which the page cleaners will be started, if not currently active

Alternate page cleaning (8.1.4 and later)

List of good victim pages (dirty pages just written out) kept, prevents searching in many cases LSN gaps anticipated, prevents I/O spikes chngpgs_thresh is ignored

You Can Expect More

You Are Here DB2 Process Model

Client Application

Prefetchers

External Storage

Listener

Coordinator

Subagents

Buffer Pool

Page Cleaners

Logger

Log

You Can Expect More

Buffer Pool Management

Buffer pools not used with all data types

LONG VARCHAR, BLOB and CLOB retrieved directly from disk using non-buffered (direct) IO Keeping large object pages out of buffer pools helps DB2 maintain effective caching for regular data types
Without use of direct IO, large data types would force DB2 to flush all resident pages to disk ruining the hit rate

1 GB BLOB

1 GB BP

Disk

You Can Expect More

Buffer Pool Management

Creating/Altering Buffer Pools

IBMDEFAULTBP automatically created with each database Additional buffer pools added with the CREATE BUFFERPOOL statement SYSCTRL or SYSADM privileges are required to work with buffer pools
CREATE BUFFERPOOL bufferpool_name SIZE number of pages [IMMEDIATE | DEFERRED] [PAGESIZE integer [K]] [[NOT] EXTENDED STORAGE | NUMBLOCKPAGES number of pages [BLOCKSIZE number of pages]] [ALL DBPARTITIONNUMS | DATABASE PARTITION GROUP db_partition_group_name [,]] [EXCEPT ON DBPARTITIONNUM[S] (db_partition_number1 [TO db_partition_number2] SIZE number of pages [,])]

You Can Expect More

Buffer Pool Management

Memory Usage
Version 8 makes DB2 easier to manage by using database global memory (DATABASE_MEMORY) instead of database heap (DBHEAP) Prevents resizing of DBHEAP with bufferpool resizing

CREATE BUFFERPOOL bufferpool_name SIZE number of pages [IMMEDIATE | DEFERRED] [PAGESIZE integer [K]] [[NOT] EXTENDED STORAGE | NUMBLOCKPAGES number of pages [BLOCKSIZE number of pages]] [ALL DBPARTITIONNUMS | DATABASE PARTITION GROUP db_partition_group_name [,]] [EXCEPT ON DBPARTITIONNUM[S] (db_partition_number1 [TO db_partition_number2] SIZE number of pages [,])]

You Can Expect More

Buffer Pool Management

Extended Storage
32-bit DB2 cant support creating buffer pools larger than 4G due to addressable storage limits Extended storage allows DB2 to flush dirty pages to memory rather than disk, so future access to the pages is much faster Requires NUM_ESTORE_SEGS and ESTORE_SEG_SIZE configuration parameters
CREATE BUFFERPOOL bufferpool_name SIZE number of pages [IMMEDIATE | DEFERRED] [PAGESIZE integer [K]] [[NOT] EXTENDED STORAGE | NUMBLOCKPAGES number of pages [BLOCKSIZE number of pages]] [ALL DBPARTITIONNUMS | DATABASE PARTITION GROUP db_partition_group_name [,]] [EXCEPT ON DBPARTITIONNUM[S] (db_partition_number1 [TO db_partition_number2] SIZE number of pages [,])]

You Can Expect More

Buffer Pool Management

Block-Based Buffer Pools

Contiguous blocks of pages to be moved into contiguous portions of memory Block-based buffer pools will be wasted if your application doesnt perform sequential prefetch

CREATE BUFFERPOOL bufferpool_name SIZE number of pages [IMMEDIATE | DEFERRED] [PAGESIZE integer [K]] [[NOT] EXTENDED STORAGE | NUMBLOCKPAGES number of pages [BLOCKSIZE number of pages]] [ALL DBPARTITIONNUMS | DATABASE PARTITION GROUP db_partition_group_name [,]] [EXCEPT ON DBPARTITIONNUM[S] (db_partition_number1 [TO db_partition_number2] SIZE number of pages [,])]

You Can Expect More

Buffer Pool Management

Hidden Buffer Pools

For improved availability only DB2 creates small (16 page) buffer pools of each page size at database activation (4, 8, 16, 32k) Hidden from the user, and used only when ordinary buffer pools cannot be allocated Performance will be noticeably impacted, and DB2 will write a warning message to the administration notification log

You Can Expect More

Buffer Pool Management

Related Objects
CREATE TABLESPACE specifies which buffer pool to use Page sizes of buffer pool and tablespace must match valid values are 4, 8, 16, and 32 kilobytes CREATE BUFFERPOOL statement can specify partition group, to control partitions in which a given buffer pool is created

You Can Expect More

Collecting Performance Metrics

Tuning An Iterative Process

Collect performance data
Snapshot monitoring Event monitoring

Review the results

Understand SQL workload Evaluate performance (calculations)

Take action
Resize buffer pools Define more page cleaners Break out tables with like workloads

Start the process again

You Can Expect More

Collecting Performance Metrics Snapshot monitor Event Monitor

Point in time representation of data Event based representation of data Less overhead (~5%) Need to reissue over time Monitor Switches need to be turned on to collect data Real time balance of application and database performance More overhead (~10-20%) Create it once and it runs until its told to stop Create takes care of everything

Historical main focus on Application statistics

You Can Expect More

Collecting Performance Metrics

DB2 Snapshot Switches

Seven Switches Buffer Pool (dft_mon_bufpool) Lock (dft_mon_lock) Sort (dft_mon_sort) Statement (dft_mon_stmt) Unit of Work (dft_mon_uow) Table (dft_mon_table) Timestamp (dft_mon_timestamp)

You Can Expect More

Collecting Performance Metrics

DB2 Snapshot Switches (contd)

Turning switches on at the instance take care here
UPDATE DATABASE MANAGER CONFIGURATION USING dft_mon_bufpool ON dft_mon_lock ON dft_mon_sort ON dft_mon_stmt ON dft_mon_table ON dft_mon_uow ON IMMEDIATE

Turning switches on for session

UPDATE MONITOR SWITCHES USING bufferpool ON

You Can Expect More

Collecting Performance Metrics Issuing a SNAPSHOT request Instance level (requires ATTACH)
For all active databases
GET SNAPSHOT FOR ALL BUFFERPOOLS

For a specific database

GET SNAPSHOT FOR BUFFERPOOLS ON database_name

For a specific database partition

GET SNAPSHOT FOR BUFFERPOOLS ON database_name AT DBPARTITIONNUM db_partition_number

BUFFERPOOL and DATABASE snapshot calls are essential for tuning buffer pools

You Can Expect More

Collecting Performance Metrics

BUFFERPOOL Snapshot
Bufferpool name Database name Database path C:\DB2\NODE0000\SQL00005\ Input database alias Buffer pool data logical reads Buffer pool data physical reads Buffer pool data writes Buffer pool index logical reads Buffer pool index physical reads . . . Node number Tablespaces using bufferpool Alter bufferpool information: Pages left to remove Current size Post-alter size = BASEBP1 = BASEBALL = = = = = = = BASEBALL 300 3490 3050 67 34

= 0 = 4 = 0 = 1000 = 1000

You Can Expect More

Collecting Performance Metrics

DATABASE Snapshot
Database name Database path . . . Buffer pool data logical reads Buffer pool data physical reads Asynchronous pool data page reads Buffer pool data writes Asynchronous pool data page writes Buffer pool index logical reads Buffer pool index physical reads Asynchronous pool index page reads Buffer pool index writes Asynchronous pool index page writes Total buffer pool read time (ms) Total buffer pool write time (ms) Total elapsed asynchronous read time Total elapsed asynchronous write time Asynchronous read requests LSN Gap cleaner triggers Dirty page steal cleaner triggers Dirty page threshold cleaner triggers Time waited for prefetch (ms) = AUTO_PRD = C:\DB2\NODE0000\SQL00009\

= = = = = = = = = = = = = = = = = = =

165079 36264 10546 3991 3873 4014 558 332 26 26 8329 121852 5353 121292 1681 24 63 33 626

You Can Expect More

Evaluating Performance Metrics

Know your workload

Is the SQL workload primarily:
Update/Insert/Delete (OLTP) Select (ad hoc, data warehouse)

Event Monitoring can help if unsure

You Can Expect More

Evaluating Performance Metrics Logical IO vs. physical IO minimize number of times DB2 must access physical disks Key I/O Metrics: Overall hit rate Data hit rate Index hit rate Physical IO read rate

You Can Expect More

Evaluating Performance Metrics

Overall Hit Rate

Includes all data and index reads into buffer pool
(bp.data.logical.reads + bp.index.logical.reads) *100 BuffePoolOHR = BufferpoolOHR bp.index.logical.reads + bp.index. physical.reads + bp.data.logical.reads + bp.data. physical.reads

Low values = more physical IO High values = more logical IO 85 90% is a good starting point

You Can Expect More

Evaluating Performance Metrics

Data Hit Rate

Includes all data reads into buffer pool
bp.data .logical .reads Bufferpool DHR = bp.data .logical .reads + bp.data . physical .reads *100

Low values = more physical IO High values = more logical IO 85 90% is a good starting point

You Can Expect More

Evaluating Performance Metrics

Index Hit Rate

Includes all index reads by buffer pool
bp .index .logical .reads Bufferpool IHR = bp .index .logical .reads + bp .index . physical .reads * 100

Low values = more physical IO High values = more logical IO 85 90% is a good starting point

You Can Expect More

Evaluating Performance Metrics

Physical IO Read Rate

Total number of physical reads by buffer pool over time
bufferpool.data.logical.reads + bufferpool.data. physical.reads PhysicalIOReads = total.collection.time

Use:
GET SNAPSHOT FOR TABLESPACES ON database_name

Could indicate the need for

Additional containers Buffer pool breakout

You Can Expect More

Evaluating Performance Metrics

Prefetch Asynchronous vs. Synchronous I/O

Waiting is always bad Make DB2 read ahead for you as much as possible, given application mix

Approach:
Evaluate prefetch ratio Know the workload Evaluate prefetcher configuration Use block I/O

You Can Expect More

Evaluating Performance Metrics

Prefetch Ratio
async.pool .data.read s + async.pool .index.rea ds *100 PrefetchRa tio = bp.logical .reads + bp.index.l ogical.rea ds

Lower values = more synchronous Higher values = more asynchronous Are there enough defined?

You Can Expect More

Evaluating Performance Metrics

Block-based Buffer Pools

If using sequential prefetch, enable buffer pools to utilize
block IO
ALTER BUFFERPOOL basebp2 NUMBLOCKPAGES 30; ALTER BUFFERPOOL basebp2 BLOCKSIZE 10;

NUMBLOCKPAGES should be a multiple of block size

You Can Expect More

Evaluating Performance Metrics

Evaluating Page Cleaner Performance:

LSN Gap Triggers Dirty Page Threshold Triggers Dirty Page Steal Triggers

Approach:
Evaluate workload and recovery goals Evaluate each metric against all others Modify parameters influencing page cleaner activity Enough page cleaners defined? More automated in V8.1.4 and above

You Can Expect More

Evaluating Performance Metrics

Log Cleans Ratio

lsn.gap.cleaner .triggers LCR = *100 dirt . page.steal .cleaner .triggers + lsn.gap.cleaner .triggers + dirty . page.threshold .cleaners .triggers

Has an affect on recovery times Too small or too large could be detrimental to buffer pool performance Affected by SOFTMAX DB CONFIG parameter 0 if ALTERNATE_PAGE_CLEANING=ON

You Can Expect More

Evaluating Performance Metrics

Threshold Cleans Ratio

dirty . page.threshold .cleaner .triggers TCR = *100 dirt . page.steal .cleaner .triggers + lsn.gap.cleaner .triggers + dirty . page.threshold .cleaner .triggers

CHNGPGS_THRESH DB CONFIG parameter Finding a happy medium is key 0 if ALTERNATE_PAGE_CLEANING=ON

You Can Expect More

Evaluating Performance Metrics

Victim Cleans Ratio

dirty . page.steal .cleaner .triggers VCR = *100 dirt . page.steal .cleaner .triggers + lsn.gap.cleaner .triggers + dirty . page.threshold .cleaner .triggers

Above 40%, typically means buffer pool needs to be larger Could also mean SOFTMAX parameter is too high (not enough page turnover) 0 if ALTERNATE_PAGE_CLEANING=ON

You Can Expect More

Conclusion

Focus on three main areas

1. BP hit ratios (logical vs. physical) 2. Prefetcher tuning (asynchronous vs. synchronous) 3. Page cleaner tuning (if not using newer alternate cleaning)

Collect performance data over time

Sampling methodology snapshots every 5-15 minutes over the course of the day, or during a batch run Decipher data and plug into equations Know the workload Chart performance over time Various solutions depending on findings

Review the results

Take action

You Can Expect More

Quest Software: You Can Expect More

http://www.quest.com/db2/