You are on page 1of 16

Mo vin g P erf or mance

Accounta bi li ty to the
De vel oper

Gudmundur Josepsson
Index Software Solutions and Consulting
Top ic
• Application performance –
whose responsibility?
• Case study:
The un-tunable application
• The right way
• Q&A

2
A ppl ica tion perf or mance
• Traditionally considered the DBA’s
responsibility
• “I don’t need to know this. This is DBA
stuff.”
• The code-tennis
• The end-users and the business suffer

3
DB A vs de vel oper
• Bickering about “who killed who”
• What it really comes down to is DBAs and
developers versus the end-users
• I’ve never seen a problem fixed by finger-
pointing
• The Metropolis analogy
(free movie)

4
Metr opol is / Or ac le
• City • Application
• Good life • Good response
• Big machine • Oracle
• Workers • IT pros

5
Case st udy:
T he un-tunable a pplic ation
• An interactive banking application
• Displayed summary information from
many other applications
• Ran thousands of times during business
hours
• Slow: waiting end-users, waiting
customers
• “As fast as it’s going to be”
• “Make it faster, anyway”

6
Case st udy:
T he un-tunable a pplic ation
• Previous attempts to optimize had failed
• The usual stuff did not work
– Statistics
– Re-org
– Session parameters
– Database parameters
• “Do you think it will run faster on SQL
Server?”

7
T he da ta ba se is fin e
• Nothing irregular in system-wide database
monitoring tools
• OS performance stats were good
– CPU load is fine
– Memory usage is fine
– Disk response times are fine
– Network is fine

8
“We can ’t chang e the code”
• Application developed in-house
• All supporting applications developed
in-house
• A decision was made that the application
had to use a layer of views to get data
from the other applications
• If more or different data was needed the
views were made bigger

9
Ar chitectur e
Un-tunable
application

Sub-app 1 Sub-app 2 Sub-app 3


...

Oracle

10
W ha t can we do?
• The database seems to be fine, the DBAs
swear it’s not their problem
• We can’t touch the code
• The manager is yelling
• Help...
• ...please

11
Too ls and methods
• Figure out which part of the application is
the slowest, where is the time being spent
• Instrument the code
• Extended SQL trace
• A profiler (well, THE profiler)

12
“T he wr ong castl e”
• Playing around with database parameters
was not the solution
• The right answer was to refuse to use the
generic views and specify exactly what
was needed
• The result was more views but each view
was smaller, faster and more manageable

13
Resul ts
• Lookup for biggest customers down from
30+ seconds to less than 8 seconds
• Average customer from ~8 seconds to less
than 1 second
• Happy developers
• Happy DBAs
• HAPPY END-USERS!

14
T he ri ght w ay
• Work together
• Know what your application is doing
• Get facts
• Don’t fight the symptoms –
solve the problem!

15
Thank you!

16

You might also like