DB14

1
(No Transcript)
2
DB-14 Tales of the Bunker - 2005

• Gus Björklund, Progress Software Corporation
• John Harlow, Bravepoint, Inc.
• Dan Foreman, Bravepoint, Inc.
• Rich Banville, Progress Software Corporation

3
Goals of the Bunker Test

• Find the optimal way to run Progress on Linux
• Test various ideas and theories
• Have fun
• Bunker 2005
• Pre-release 10.1A
• 64-bit AMD
• Performance of Utilities
• Investigate coma problem
• Network speed effects

4
Why Linux?
5
Bunker 2005 Team

• Gus Björklund, Wizard, Progress Software
• Progress User since 1989
• John Harlow, President of BravePoint
• Progress User since 1984
• Dan Foreman
• Progress User since 1984
• Rich Banville, Fellow, Progress Software
• Progress User since 1993

Newbie !
6
The ATM Benchmark Environment
7
The ATM Benchmark

• Simulates teller machine transactions
• deposit or withdrawal
• heavy database update workload
• Each transaction
• does 3 fetches, 3 updates, 1 create
• retrieve and update account, branch, and teller
  rows
• create a history row
• Run n transaction generators
• concurrently
• for fixed time period
• count total number of transactions performed

8
Test Database (logical)
9
Test Database (physical)
10
Test Database (other info)
11
Equipment
12
Server 1 hostname uniblab
13
Server 3 hostname hal
14
Server 3 hostname jumbo
15
AMD-64 Hyper-Transport Design
16
Bunker Network Components

• 3 server machines
• Netgear GS 105 Gigabit switch
• SMC Barricade WAP
• LinkSys WVC11b bunker cam
• Various laptops running
• Windoze
• Linux
• Mac OS X 10.3

17
Bunker Network (partial)
Internet
router
wap
bunker Cam
gigabit switch
laptops
hal
uniblab
jumbo
servers
18
Other Equipment
19
Past Results
20
Lessons from Past Bunkers

• Type II Data Areas are faster
• Dont use Reiser File System
• Use EXT3 or XFS File System
• Dont use the Anticipatory Scheduler
• Deadline or CFQ is better
• 2.6 Kernel is faster than 2.4 Kernel
• For RAID 10, the Largest Possible Stripe Size was
  always the fastest, both Software Hardware
  Striping
• Very good performance at low cost

21
This Years (October 2005)Bunker Results
22
Setup Results
Database is about 11 gigabytes
23
Baseline Server Configuration

• Data extents on striped array, BI log on own disk
• BI cluster size 16384
• BI blocksize 16
• Server options
• -n 200 -L 10240
• -B 64000
• -spin 50000
• -bibufs 32
• Page writers 4
• BI writer yes
• AI writer no

24
Linux I/O schedulers - 64-bit AMD SuSE 10
What do we learn from this?
25
Dump/Load

• -index 0 Option on Binary Dump
• Excellent Performance Improvement
• But order of records may not be what you want

26
Dump/Load with -RO

• OpenEdge 10 -RO Faster than V9
• V9 with -RO faster than without
• OpenEdge 10 with -RO performance same as without
• OpenEdge 10 -RO Clients now write entries in the
  .lg file

27
Dump/Load

• Logical Scatter Factor is very important
• Performance Difference of 400 to 1000

28
Remote Clients

• -Ma
• The Lower Value, the Better the Performance
• -Mm
• No Negative Impact on ATM Benchmarks

29
Network Speed
30
Coma Problem

• We have experienced this problem in every Bunker
  Test
• We still dont know whats wrong
• A customer on RH AS 4 Kernel 2.6.9-5.Elsmp
  reports problem solved - for him
• There are an infinite number of things and
  combinations of things that can be changed in the
  kernel
• We need to do some work with the aggressiveness
  of the APWs to helpbut also more testing

31
Coma problem -directio helps
32
We still think about this problem
33
64-Bit

• We sawNo difference in general
  performancebetween 32 and 64 bit Progress

34
Strange Problem

• WEIRD PROBLEM ON ONE MACHINE
• The bigger B, the lower the TPS rate
• True with both 32 and 64 bit Progress/Linux
• Could be caused by
• Enterprise versus Desktop version of Linux
• 10.1A Beta problem
• SUSE Linux 10 issue (unsupported OS)
• Something else
• HyperTransport Effects
• All the above
• DID NOT OCCUR ON OTHER MACHINES
• HAVE NOT SEEN AGAIN

35
TPS vs Response Time
Avg Response time 0.2 seconds
36
V10.0B versus V10.1A Beta

• No Difference in general performance

37
SATA versus SCSI

• SCSI was faster
• SATA is less expensive
• Beware desktop drives not rated for 24x7
  operation

38
Online backup time
workload
39
Online backup rate (approximate)
workload
40
Online backup performance impact
150 user atm workload
41
Adding extent online elapsed time

• Add 2 GB extent on same disk array

42
Adding extent online performance impact
150 user workload, add 2 GB data extent
online,Extent on same striped array as other
extents
43
Secret Bunker Web Pages

• March 2002 October 2002
• http//www.myfloridacottage.com/benchmark.html
• April 2004
• http//www.myfloridacottage.com/bunker3.html
• Oct 2005
• http//www.myfloridacottage.com/bunker4/

44
Join Us in the Bunker

• Film

45
?
Want Answers
Also see Guss RDBMS Tuning Guide on conference CD
46
(No Transcript)