Transcript Global Payroll Performance Optimisation - II David Kurtz

Global Payroll Performance
Optimisation - II
David Kurtz
Go-Faster Consultancy Ltd.
[email protected]
www.go-faster.co.uk
Who Am I?
• Oracle Database Specialist
– Independent consultant
• Performance tuning
– PeopleSoft ERP
– Oracle RDBMS
• Book
– www.psftdba.com
• UKOUG Director
• Server Tech & PeopleSoft
•
Oak Table
•
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
2
Agenda
•
•
•
•
•
•
‘Streaming’ –Parallel processing
Data Volume
Read Consistency
Partitioning
Reporting
Archiving
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
3
Warning
• This is an unashamedly technical session.
• I am going to talk about database internals.
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
4
Size matters!
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
5
Parallel processing
• All modern machines
have multiple
processors,
– most of the processors
have multiple cores.
– Even the CPU in my 4
year old laptop has a 2
core CPU.
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
6
Database Parallelism
• All objects in the PeopleSoft schema are
explicitly set NOPARALLEL
– Indexes are built parallel, but later reset.
– Can invoke parallel query with PARALLEL
hint
– Parallel insert in direct path model
– Parallel DML only works on partitioned objects
• 1 PQ slave per partition
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
7
PeopleSoft Batch Programs
• Only run on one CPU at any one time.
• Client Server processes
– Program (COBOL or Application Engine)
– Database (eg. Oracle)
• Either busy executing COBOL or waiting
for the database.
– If your payroll calculation is a single process
you are not getting value for money!
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
8
Payroll ‘Streaming’
• Several GP processes can be split up.
– Each piece processes a distinct set of
employees
• Range of EMPLID
– The pieces can be run concurrently.
– Maximum number of streams determined by
hardware.
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
9
Streamable Processes
• COBOL
– Payroll Calculation
• Application Engine
–
–
–
–
Banking Preparation
GL Preparation
EDI Preparation
Payslip Preparation
• Database Intensive
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
10
Payroll ‘Streaming’ Challenges
• Payroll isn’t over until the last stream
completes.
– Streams need to be evenly balanced.
– Employee churn?
• One global definition of streams
– Balance for largest payroll?
• Inter-stream contention
– Shared working storage tables in COBOL
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
11
Payroll Calculation Process Phases
• Identify
– Populate working storage and some result tables
• Database Intensive
• Calculation
– COBOL Intensive
• Cancellation
– Delete results
• Database Intensive
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
12
How Many Streams?
• In a well tuned systems, the payroll calculation
phase spends about
– 2/3 of its time in COBOL
– 1/3 on the database.
• Number of streams should not exceed
– 3 * CPU on database server
– 1.5 * CPU on Process Scheduler server
• Payroll identification process is database
intensive.
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
13
How Many Streams?
• First GP I ever worked on
– 20 CPUs on Application/Batch server
– 20 CPUs on Database server
• Maximum number of streams?
– 20 / 1/3 = 60 on database server
– 20 / 2/3 = 30 on Application server
• So we used 30 streams
– Application server fully utilised during payroll calc
– Database about 50% during calc,
– Probably overloaded during identification.
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
14
How Many Streams?
• Optimise number of streams for calculation phase.
• Restrict concurrency of database intensive process
on process scheduler.
– To limit CPU consumption, and possibly also I/O
contention.
• Consider use of Oracle Resource Manager
– Mainly for Payroll identification
– I’ve never had to do this myself.
– Cancellation will be restricted by I/O
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
15
Balancing Streams
• Balance employees across streams on basis
of
– 80% number of payroll segments per stream
– 20% number of JOB history rows
• Longer serving employees in earlier streams likely
to have more payroll segment and job history.
– Make allowance for employee churn.
• You will need to periodically rebalance the streams.
– Balance for the largest payroll.
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
16
Employee Churn
• EMPLID is allocated as an accession
number.
• Streams are a range of EMPLIDs
– New employees are hired into the last stream
– Employees are terminated across all streams
• Over time the streams will go out of balance
– Last stream will take longest
• Periodically rebalance the streams
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
17
Bulk Churn Effects
• Migration
– If migrated to GP in tranches then order of
migration could affect stream balance
• Company merger/divestment history can
affect balance of payroll.
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
18
Rebalancing the streams?
• Calculate new stream range values
– Allow space for estimated future growth
• Rebuild all range partitioned tables
– Half the I/O of partition merge/split
– About 42 tables in UK tables.
– Need working storage space to do this
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
19
Reversing the EMPLID
• Reverse the EMPLID
– Instead of EMPLID 0000012345
– Use EMPLID 543210000
• Streams stay balanced because new
employees hired across range
• Improved search performance across HCM
• BUT you must do this before you go live!
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
20
Reversing the EMPLID
•
•
•
•
•
•
•
•
0000012345
0000012346
0000012347
0000012348
0000012349
0000012350
0000012351
0000012352
Global Payroll Performance Optimisation
•
•
•
•
•
•
•
•
5432100000
6432100000
7432100000
8432100000
9432100000
0532100000
1532100000
2532100000
©2011 www.go-faster.co.uk
21
Inter-stream Contention
• Streams are just ranges of EMPLIDs.
• Oracle inserts data into the first available block (roughly
speaking)
• Multiple streams insert data simultaneously into the same
data blocks in result tables.
• Payroll cancel/recalculation deletes from result tables.
• Multiple transactions concurrently update different rows in
the same block.
– On Oracle/SQL Server >=2005: No locking, streams continue to
run, but read consistency processing is expensive
– Other database can experience page level locking
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
22
Working Storage Tables
• COBOL
– One shared instance of
each working storage
table
• Shared SQL
– Candidate for Global
– Temporary Table so
one instance per
session
Global Payroll Performance Optimisation
• Application Engine
– PeopleSoft Temporary
Record
– One instance of record
per process
• Different SQL
• Still consider GTT to
reduce redo
©2011 www.go-faster.co.uk
23
Read Consistency
• The data set that you query remains the same
throughout the life of your query.
– If somebody else updates data that you are reading (and
commits), after your query starts, then you see the
original value.
• Thus, readers do not block writers or vice versa.
• Oracle has always done this, like this since 1990.
• SQL Server 2005 has ‘read committed snapshot’
option
• Other databases either block or can permit ‘dirty
read’.
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
24
Read Consistency
• Oracle achieves this by storing ‘undo’
information for every change
– Recovers ‘read-consistent’ in-memory copy of
data block to point in time when query started.
– A good reason for buying Oracle
– Resource intensive process
– Performance problem if abused.
• Global Payroll is the perfect storm!
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
25
Read Consistency
• Query @ 10023
• Update @ 10024
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
26
Avoiding Inter-stream Contention
• Prevent different streams accessing the same data
blocks
– Range Partition result tables to match stream ranges
– Use Global Temporary Tables (Oracle) for working
storage tables
– Partition these also on other platforms.
• Now different streams access different partitions.
• No code change, a job for the DBA
– licensed option on most platforms
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
27
Partitioning
• Partitioned Table
– Different physical components
• Value of data determines physical location
– Logically still one table
– Transparent to application
– Rather like a multi-part encyclopaedia.
• Partition Elimination
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
28
What is Partitioning?
• Typically used in DSS
• But can also be effective in OLTP
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
29
Partitioning
• Keep similar things
together
– Employees for one
stream in on partition
Global Payroll Performance Optimisation
• Keep different things
apart
– Only one transaction in
each block of each
segment
– No need for read
consistency
©2011 www.go-faster.co.uk
30
Partitioning GP Recommendation
• Range Partitioning
– EMPLID – to match streams
• List Sub-partition
– CAL_RUN_ID – calendar group ID.
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
31
Secondary Benefits
• CAL_RUN_ID list sub-partition
• Easier to archive later
– Historical partitions
– Different Tablespaces
– Different Data Files
•
•
•
•
Old data on slower disk
Read Only
Less frequent back-up of read-only tables
Faster Backup
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
32
Global Temporary Tables
• Global because the
data is private
• Temporary because
the definition is
permanent
Global Payroll Performance Optimisation
• Global because
everyone can see the
definition
• Temporary because
physical existance of
the table is temporary
so it does not need to
be recovered.
©2011 www.go-faster.co.uk
33
Global Temporary Tables
• A temporary object
– No redo generation
• But there is undo, and there is redo on the undo!
– Each session gets its own physical copy.
• Again no read consistency problems
• No high water mark issues
• Lower high water marks – less I/O
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
34
Building the DDL
• Demonstrate GFCBUILD utility.
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
35
Group Lists
• Specify a list of individual EMPLIDs for
whom to run pay calc or another process.
• Some customers have experienced problems
when run groups shortly before or during
larger batch payroll calculations.
• Why?
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
36
Cost Based Optimizer
• SQL Execution Plan Caching
• Bind Variable Peeking during Parse
• Different Plan for Group List
– Because different bind variables
• But plan cached and gets used for main pay
calculation which then runs longer than
usual!
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
37
Plan Stability
• Remember the good plan used by large
payroll.
• Force it to be used for all payrolls including
group list.
– Data Volumes small so poor plan won’t really
matter.
• Oracle Stored Outline
– No code change, DBA can implement.
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
38
Plan Stability
• Collect and applied stored outline with
database trigger
– http://www.gofaster.co.uk/gpdoc.htm#gp.stored_outlines
• Use Active Session History to demonstrate
the problem and solution
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
39
Capture Stored Outline
CREATE OR REPLACE TRIGGER sysadm.gfc_create_stored_outlines
BEFORE UPDATE OF runstatus ON sysadm.psprcsrqst
FOR EACH ROW
WHEN (new.prcsname = 'GPPDPRUN' AND (new.runstatus = 7 OR
old.runstatus = 7))
DECLARE
l_sql VARCHAR2(100);
BEGIN
l_sql := 'ALTER SESSION SET create_stored_outlines = ';
IF :new.runstatus = 7 THEN
EXECUTE IMMEDIATE l_sql||:new.prcsname;
ELSIF :old.runstatus = 7 THEN
EXECUTE IMMEDIATE l_sql||'FALSE';
END IF;
--because I dont want to crash the process scheduler
EXCEPTION WHEN OTHERS THEN NULL;
END;
/
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
40
Apply Stored Outline
CREATE OR REPLACE TRIGGER sysadm.gfc_use_stored_outlines
BEFORE UPDATE OF runstatus ON sysadm.psprcsrqst
FOR EACH ROW
WHEN (new.prcsname = 'GPPDPRUN'
AND (new.runstatus = 7 OR old.runstatus = 7))
DECLARE
l_sql VARCHAR2(100);
BEGIN
l_sql := 'ALTER SESSION SET use_stored_outlines = ';
IF :new.runstatus = 7 THEN
EXECUTE IMMEDIATE l_sql||:new.prcsname;
ELSIF :old.runstatus = 7 THEN
EXECUTE IMMEDIATE l_sql||'FALSE';
END IF;
--because I dont want to crash the process scheduler
EXCEPTION WHEN OTHERS THEN NULL; END;
/
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
41
Three Scenarios Compared
Large / Small / Plan Stable Small
SQL_ID
SCENARIO 1
ASH_SECS SCENARIO 2
ASH_SECS SCENARIO 3
ASH_SECS
------------- ------------------- ---------- ------------ ---------- ------------ ---------4uzmzh74rdrnz
2514155560
280 3829487612
28750 **SAME**
5023
4n482cm7r9qyn
1595742310
680 869376931
140 **SAME**
889
2f66y2u54ru1v
1145975676
630
**SAME**
531
1n2dfvb3jrn2m
1293172177
150
**SAME**
150
652y9682bqqvp
3325291917
30
**SAME**
110
d8gxmqp2zydta
1716202706
10 678016679
10 **SAME**
32
2np47twhd5nga
3496258537
10
**SAME**
27
4ru0618dswz3y
2621940820
10
539127764
22
4ru0618dswz3y
539127764
100
**SAME**
22
4ru0618dswz3y
3325291917
10
539127764
22
4ru0618dswz3y
1403673054
110
539127764
22
gnnu2hfkjm2yd
1559321680
80
**SAME**
19
fxz4z38pybu3x
1478656524
30
4036143672
18
2xkjjwvmyf99c
1393004311
20
**SAME**
18
a05wrd51zy3kj
2641254321
10
**SAME**
15
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
42
Data Volume
• Payroll generates a lot of data.
• Every pay period it generates more data.
• Partitioning can offer ways of accessing the
data you want quickly
– Without having to trawl through data you don’t
want.
• Need to consider how long you need data
– Do you still need data from last tax year?
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
43
Archiving
• Put the data you do need to keep into a reporting
table
– Remove data from the live result tables
– Partitioning can help you move/delete this data
efficiently
– May need to rebuild tables where you have to use
DELETE
• Reduced data volumes should improve
performance of reports.
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
44
Reporting
• Payroll result tables delivered with single
index
– Not suitably indexed for all reporting
requirements
• Particularly single PIN queries
– Adding more indexes would degrade
calculation performance
– Consider generating reporting table
• Subset of data, and indexed as necessary.
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
45
GFC_GPRPTGEN
• Reporting Table for Single Pin Queries
– List Partitioned by Pin
– One Partition for each Pin
• Incremental Maintenance by Application Engine
– Uses Parallel DML to maintain reporting table.
– Sub-Paritioned GP Result Tables may still be faster for
single employee, single calendar group ID queries!
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
46
Further Reading
• Configuring and Operating Streamed Processing in
PeopleSoft Global Payroll
– www.gofaster.co.uk/gpdocs.htm#Configuring_Operating_Streamed_Payrol
l
• Managing Oracle Table Partitioning in PeopleSoft
Applications with GFC_PSPART Package
– www.gofaster.co.uk/gpdocs.htm#Managing_Oracle_Table_Partitioning
• Use of Oracle Plan Stability (Stored Outlines) in
PeopleSoft Global Payroll
– www.go-faster.co.uk/gpdocs.htm#gp.stored.outlines
Global Payroll Performance Optimisation
©2011 www.go-faster.co.uk
47
Questions?