How to find and fix your Oracle-based application performance problem

How to ﬁnd and ﬁx your
Java APEX ADF OBIEE .NET SQL PL/SQL
application performance
problem
Cary Millsap
Method R Corporation and Accenture Enkitec Group
@CaryMillsap
DOUG January Meeting · Richardson, Texas
5:00p–7:15p Thursday 28 January 2016
© 2006, 2016 Cary Millsap
1
TM
MeTHOD R
TM

@CaryMillsap
Q What is the most common
Oracle performance
problem you see?”
3
“

@CaryMillsap
What is the most common
Oracle performance
problem you see?”
4
“
Assuming that other
people’s common problems
must be your problem.
...
Q
A

@CaryMillsap 5
Java APEX ADF OBIEE .NET SQL PL/SQL

@CaryMillsap
What is a performance problem?
6

@CaryMillsap
Performance is not
an attribute of a system.
8

@CaryMillsap 9
ID USERNAME OPERATION R SLR
-- -------- --------- ------ ---
1 FCHANG OE BOOK 12.019 2.0
2 RSMITH OE SHIP 3.528 5.0
3 DJOHNSON OE PICK 1.211 5.0
4 FFORBES OE BOOK 0.716 2.0
5 FCHANG OE BOOK 1.917 2.0
6 LBUMONT PA MTCH 1.305 2.0
#define FASTid ((Rid ≤ SLRid)?”FAST”:”SLOW”)

@CaryMillsap 10
ID USERNAME OPERATION R SLR SPEED
-- -------- --------- ------ --- -----
1 FCHANG OE BOOK 12.019 2.0 SLOW
2 RSMITH OE SHIP 3.528 5.0 FAST
3 DJOHNSON OE PICK 1.211 5.0 FAST
4 FFORBES OE BOOK 0.716 2.0 FAST
5 FCHANG OE BOOK 1.917 2.0 FAST
6 LBUMONT PA MTCH 1.305 2.0 FAST
#define FASTid ((Rid ≤ SLRid)?”FAST”:”SLOW”)

@CaryMillsap
Performance is an attribute of
each individual experience
with a system.
11

@CaryMillsap 12
TASK
• id
• name
• ...
EXPERIENCE
• id
• task-id
• user-id
• ip-address
• start-time
• end-time
• ERROR-code
• WORK-done
SQL
• ID
• Task-id
• ...
N
1
N
1

@CaryMillsap 13
<experience
id = "b3196c98-906d-4394-bc55-0339518a63b2"
task-id = "7"
uid = "238"
ip = "142.128.130.186"
t0 = "2014-04-10T08:32:14.137886"
t1 = "2014-04-10T08:32:17.891173"
err = ""
work = "3"
/>

@CaryMillsap
has to ﬁnish quickly.”
click
button
link
row
query
report
job
}{“My
14
This is what performance is.

@CaryMillsap
has to ﬁnish quickly.”
click
button
link
row
query
report
job
}{“My
15
A performance problem is when it doesn’t.

@CaryMillsap 16
“How long does it take?”
Response time (R)
Duration from service request to
service fulﬁllment.
Sanjay Nancy Ken Jorge
R
t0
t1
R = t1 – t0
Two big questions...
1. How long did it take?
2. Why?

@CaryMillsap 17
Two big questions...
1. How long did it take?
2. Why?
“How long does it take?”
Response time (R)
Duration from service request to
service fulﬁllment.
Sanjay Nancy Ken Jorge
R
t0
t1
R = t1 – t0

@CaryMillsap
1. Select the experience you need to improve.
2. Measure its response time (R) in detail.
3. Execute the best net-payoﬀ remedy.
4. Repeat until economically optimal.
19
Method R

@CaryMillsap
20
Method R

@CaryMillsap
21
Method R

@CaryMillsap 22
Method R
OP
TIMIZE ANY
THING
M
e
TH O D R

@CaryMillsap
23
Method R

@CaryMillsap
24
Method R
How do you do this,
when the it is your code?

@CaryMillsap
EXADATAD A T A B A S E
ENTERPRISE EDITION
D A T A B A S E
STANDARD EDITION
D A T A B A S E
EXPRESS EDITION
Oracle extended SQL tracing
is a feature of every Oracle Database.
26
Oracle7 1992 Oracle8 1997 Oracle8i 2000 Oracle9i 2001 Oracle 10g 2004 Oracle 11g 2007 Oracle 12c 2013

@CaryMillsap
Measuring Oracle response times
27

@CaryMillsap 28
❶
Activate
tracing
❷
Get the
trace ﬁle
❸
Understand
its story

@CaryMillsap
❶
Activate
tracing
❷
Get the
trace ﬁle
❸
Understand
its story
29

@CaryMillsap 30
This is the hardest part.
...But only the ﬁrst time.
After that, you just lather, rinse, repeat.

@CaryMillsap
https://app.com/apex/f?p=150:1:5547991082303::NO:::&P_TRACE=YES
31
Well, it’s easy in Oracle APEX.
To decide at run time whether to trace your code...

@CaryMillsap
Other technologies require a little more work.
First, the basics.
32

@CaryMillsap 33
To decide at compile time to trace all your code...
dbms_session.session_trace_enable(
waits => true,
binds => true,
plan_stat => 'ALL_EXECUTIONS'
);
-- Your OE BOOK code
dbms_session.session_trace_disable();

@CaryMillsap
if (should_trace('OE BOOK', dbms_random.value(0,1)) {
dbms_session.session_trace_enable(
waits => true,
binds => true,
);
}
dbms_session.session_trace_disable();
34
To decide at run time whether to trace your code...

@CaryMillsap
sub should_trace(t, r) {
select trace_proportion from trace_control where task_name = :t;
return (r <= trace_proportion);
}
35
...where should_trace looks like this.
task_name trace_proportion
OE BOOK 0.05
OE PICK 0.02
OE SHIP 1.00
OE INVOICE 0.01
should_trace("OE BOOK", 0.00) → true
...
should_trace("OE BOOK", 0.06) → false
...
5%
95%
trace_control

@CaryMillsap
Oracle Database helps you implement
run time tracing decisions...
...without having to make your developers do the if block stuﬀ.
36

@CaryMillsap
dbms_monitor.serv_mod_act_trace_enable(
service_name => 'SYS$USERS',
module_name => 'OE BOOK',
action_name => dbms_monitor.all_actions,
waits => true,
binds => true,
);
37
The DBA does this, at run time.
But this works only if your code
sets its module name to “OE BOOK”.

@CaryMillsap
How you set your module name varies by technology.
SQL PL/SQL Java ADF .NET OBIEE
38

@CaryMillsap
-- PL/SQL example
dbms_application_info.set_module(
module_name => 'OE BOOK',
action_name => sys_guid()
);
dbms_application_info.set_module(
module_name => null,
action_name => null
);
39
SQL PL/SQL
To set your code’s module and action names...

@CaryMillsap
// JDBC 11g example
String metrics[] = new String[OraCxn.END_TO_END_STATE_INDEX_MAX];
metrics[END_TO_END_MODULE_INDEX] = "OE BOOK";
metrics[END_TO_END_ACTION_INDEX] = UUID.randomUUID().toString();
conn.setEndToEndMetrics(metrics, (short) 0);
// Your OE BOOK code
metrics[END_TO_END_MODULE_INDEX] = "";
metrics[END_TO_END_ACTION_INDEX] = "";
conn.setEndToEndMetrics(metrics, (short) 0);
40
Java ADF

@CaryMillsap
// JDBC 12c example
Properties p = new Properties();
p.put("OCSID.MODULE", "OE BOOK");
p.put("OCSID.ACTION", UUID.randomUUID().toString());
connection.setClientInfo(p);
p.put("OCSID.MODULE", "");
p.put("OCSID.ACTION", "");
connection.setClientInfo(p);
41
Java ADF

@CaryMillsap
// C# example
conn.ModuleName = "OE BOOK";
conn.ActionName = Guid.NewGuid().toString();
conn.ModuleName = "";
conn.ActionName = "";
42
ODP.NET

@CaryMillsap 43
OBIEE

@CaryMillsap
Here’s the goal.
44

@CaryMillsap
User’s R
experience
Oracle
trace ﬁle
45
AppUser Oracle DB
time
You want this to be small
You want this to be small

@CaryMillsap
Another
experience
An
experience
Not the
trace ﬁle
you want
46
AppUser Oracle DB
time

@CaryMillsap
Another
experience
An
experience
You want one
trace ﬁle per
experience
47
AppUser Oracle DB
time

@CaryMillsap
The goal:
Trace exactly each user experience you care about.
...So that you can see how your code consumes time
when it behaves properly,
and when it misbehaves.
48

@CaryMillsap
This is what
you’re
looking at
when you use
systemwide
aggregations.
50
AppUser Oracle DB
time

@CaryMillsap 51
❶
Activate
tracing
❷
Get the
trace ﬁle
❸
Understand
its story

@CaryMillsap 52
This is the boring part.
...But it’s an inexpensive problem to solve.

@CaryMillsap
Some things to know...
Your trace ﬁle is on the Oracle Database server(s), called:
select tracefile from v$process where addr=(
select paddr from v$session where sid=(
select sid from v$mystat where rownum=1
)
)
Sessions with DOP = k can create more than 2k + 1 trace ﬁles.
53

@CaryMillsap 54
Please, will you help me ﬁnd my trace ﬁle?

@CaryMillsap
There are lots of ways to fetch the trace data.
FTP
Samba
NFS mount
portable disk
USB thumb drive
Oracle Database directory objects
Method R Trace extension for Oracle SQL Developer
55

@CaryMillsap
Fetching trace ﬁles can be easy.
You can build tools, or you can buy them.
56
Fn’m [ mifp_^ jlif_g.

@CaryMillsap 57
❶
Activate
tracing
❷
Get the
trace ﬁle
❸
Understand
its story

@CaryMillsap 58
This is the FUN part.

@CaryMillsap 59
What’s in there?!

@CaryMillsap 60
An Oracle trace ﬁle is a log that shows
what your code did inside the Oracle Database.

@CaryMillsap
Some things to know...
Oracle writes a trace line when a call (db|os) ﬁnishes.
There are two primary line formats: one for db calls, one for os calls.
Each call is associated with a SQL or PL/SQL statement through a cursor id.
Each line contains a time stamp (tim) and a duration (e|ela).
R ≠ ∑(e|ela) because parent call durations include child call durations.
61

@CaryMillsap 62
method-r.com
For more details...

@CaryMillsap
Let’s look at some
trace lines...
63

@CaryMillsap 64
begin prepare
CPU
latch-related syscall
CPU
end prepare
begin exec
CPU
write(SQLNET_OUT, result_to_client);
end exec
read(SQLNET_IN, next_request_from_client);
begin fetch
CPU
CPU
end fetch
begin fetch
CPU
write(SQLNET_OUT, more_results);
end fetch
begin fetch
CPU
end fetch
begin fetch
CPU
end fetch
Oracle kernel code path
This is the
kind of stuﬀ
your code
causes the
Oracle kernel
to do.

@CaryMillsap 65
WAIT #42: nam='latch: library cache'…
PARSE #42:c=10000,…
WAIT #42: nam='SQL*Net message to client'…
EXEC #42:c=10000,…
WAIT #42: nam='SQL*Net message from client'…
WAIT #42: nam='latch: cache buffers chains'…
FETCH #42:c=20000,…
WAIT #42: nam='SQL*Net more data to client'…
FETCH #42:c=20000,…
FETCH #42:c=20000,…
FETCH #42:c=20000,…
Oracle extended SQL trace data
begin prepare
CPU
CPU
end prepare
begin exec
CPU
end exec
begin fetch
CPU
CPU
end fetch
begin fetch
CPU
end fetch
begin fetch
CPU
end fetch
begin fetch
CPU
end fetch
Oracle kernel code path
This is the
kind of
trace data
the kernel
produces.

@CaryMillsap 66
PARSE #42:c=10000,…
EXEC #42:c=10000,…
FETCH #42:c=20000,…
FETCH #42:c=20000,…
FETCH #42:c=20000,…
FETCH #42:c=20000,…
Of course,
you don’t
directly get
to see the
kernel code
path.

@CaryMillsap 67
PARSE #42:c=10000,…
EXEC #42:c=10000,…
FETCH #42:c=20000,…
FETCH #42:c=20000,…
FETCH #42:c=20000,…
FETCH #42:c=20000,…
...Or that
helpful grid
that I drew
for you.

@CaryMillsap 68
PARSE #42:c=10000,…
EXEC #42:c=10000,…
FETCH #42:c=20000,…
FETCH #42:c=20000,…
FETCH #42:c=20000,…
FETCH #42:c=20000,…
All you
get to see
is this.

@CaryMillsap
PARSE #42:c=10000,…
EXEC #42:c=10000,…
FETCH #42:c=20000,…
FETCH #42:c=20000,…
FETCH #42:c=20000,…
FETCH #42:c=20000,…
69
Oracle extended SQL trace dataOracle kernel code path
begin prepare
CPU
CPU
end prepare
begin exec
CPU
end exec
begin fetch
CPU
CPU
end fetch
begin fetch
CPU
end fetch
begin fetch
CPU
end fetch
begin fetch
CPU
end fetch
You can learn
to envision the
kernel’s code
path that
motivated
your trace ﬁle.

@CaryMillsap
There are lots of ways to summarize a trace ﬁle.
tkprof
SQL Developer [Trace] Viewer
Trace Analyzer
tvdxstat
xtrace
OraSRP
Method R Proﬁler
70

@CaryMillsap
Proﬁling trace ﬁles can be easy.
You can build tools, or you can buy them.
71
Fn’m [ mifp_^ jlif_g.

@CaryMillsap
What you can do with trace ﬁles
72

@CaryMillsap 74
mrskew "r1-fixed.trc"
CALL-NAME DURATION % CALLS MEAN MIN MAX
--------------------------- ------------ ------ ----- -------- -------- --------
SQL*Net message from client 1,403.927942 99.7% 2,161 0.649666 0.000000 0.927028
FETCH 3.013549 0.2% 2,161 0.001395 0.000000 0.005000
direct path read temp 1.259022 0.1% 83 0.015169 0.003287 0.046968
SQL*Net more data to client 0.141213 0.0% 2,460 0.000057 0.000005 0.001269
SQL*Net message to client 0.007964 0.0% 2,161 0.000004 0.000001 0.000376
--------------------------- ------------ ------ ----- -------- -------- --------
TOTAL (5) 1,408.349690 100.0% 9,026 0.156033 0.000000 0.927028
99.7% of the time is 2,161 network round-trips.
What SQL statements cause the round-trips?

@CaryMillsap 75
mrskew --group='($sqlid=~/^#/?"":"[".$sqlid."]")' --gl=SQLID --name='message from client' "r1-fixed.trc"
SQLID DURATION % CALLS MEAN MIN MAX
--------------- ------------ ------ ----- -------- -------- --------
[7d0bv6ds85q1f] 1,403.927942 100.0% 2,161 0.649666 0.000000 0.927028
--------------- ------------ ------ ----- -------- -------- --------
TOTAL (1) 1,403.927942 100.0% 2,161 0.649666 0.000000 0.927028
Just one. All 2,161 round-trips are executed
on behalf of just one SQL statement.

@CaryMillsap 76
mrskew --rc=p10.rc --name='SQL*Net message from client' "r1-fixed.trc"
RANGE {min ≤ e < max} DURATION % CALLS MEAN MIN MAX
----------------------------- ------------ ------ ----- -------- -------- --------
1. 0.000000 0.000001 0.000000 0.0% 1 0.000000 0.000000 0.000000
2. 0.000001 0.000010
3. 0.000010 0.000100
4. 0.000100 0.001000
5. 0.001000 0.010000
6. 0.010000 0.100000
7. 0.100000 1.000000 1,403.927942 100.0% 2,160 0.649967 0.547110 0.927028
8. 1.000000 10.000000
9. 10.000000 100.000000
10. 100.000000 1,000.000000
11. 1,000.000000 +∞
----------------------------- ------------ ------ ----- -------- -------- --------
TOTAL (11) 1,403.927942 100.0% 2,161 0.649666 0.000000 0.927028
Each round-trip consumes an average of .649967 ≈ .650 s.
Why?

@CaryMillsap
App Oracle DB
time
77
~.001 s
~.001 s
~.650 s~.648 s
Each SQL*Net message from client call (~.650 s) looks like this.
If round-trip network latency is ~.002 s, then this experience is
spending ~.648 s in the Java code executed between database calls.

@CaryMillsap 78
mrskew --name=':dbcall' --select='$row' --slabel=ROWS --precision=0 "r1-fixed.trc"
CALL-NAME ROWS % CALLS MEAN MIN MAX
--------- ------- ------ ----- ---- --- ---
FETCH 216,017 100.0% 2,161 100 17 100
--------- ------- ------ ----- ---- --- ---
TOTAL (1) 216,017 100.0% 2,161 100 17 100
One ﬁnal check...
The trace ﬁle shows that the application, at least, is fetching
an average of 100 rows per fetch call (per round-trip).
This helps explain the Java-side latency, but still, .648 s to
process just 100 rows needs some explaining.

@CaryMillsap 79
mrskew "r1-fixed.trc"
--------------------------- ------------ ------ ----- -------- -------- --------
SQL*Net message from client 1,403.927942 99.7% 2,161 0.649666 0.000000 0.927028
FETCH 3.013549 0.2% 2,161 0.001395 0.000000 0.005000
direct path read temp 1.259022 0.1% 83 0.015169 0.003287 0.046968
SQL*Net more data to client 0.141213 0.0% 2,460 0.000057 0.000005 0.001269
SQL*Net message to client 0.007964 0.0% 2,161 0.000004 0.000001 0.000376
--------------------------- ------------ ------ ----- -------- -------- --------
TOTAL (5) 1,408.349690 100.0% 9,026 0.156033 0.000000 0.927028
No matter how long you try to “ﬁx the database” here, you’re
going to see at most only a .3% diﬀerence in response time.
The problem here is in the Java.

@CaryMillsap 81
mrskew "prd1_ora_9031.trc"
----------------------------- ---------- ------ ----- -------- -------- --------
PARSE 735.426197 78.9% 698 1.053619 0.000000 4.498316
SQL*Net message from client 104.762229 11.2% 1,378 0.076025 0.000391 3.554818
FETCH 91.800028 9.8% 680 0.135000 0.000000 0.506923
db file sequential read 0.104670 0.0% 14 0.007476 0.001067 0.016408
EXEC 0.083988 0.0% 349 0.000241 0.000000 0.002000
gc cr block 2-way 0.073233 0.0% 96 0.000763 0.000280 0.001968
gc current block 2-way 0.031298 0.0% 47 0.000666 0.000361 0.001640
gc current grant busy 0.028037 0.0% 47 0.000597 0.000156 0.001508
SQL*Net more data from client 0.025819 0.0% 837 0.000031 0.000000 0.002564
CLOSE 0.018999 0.0% 698 0.000027 0.000000 0.001000
12 others 0.061576 0.0% 1,633 0.000038 0.000000 0.001687
----------------------------- ---------- ------ ----- -------- -------- --------
TOTAL (22) 932.416074 100.0% 6,477 0.143958 0.000000 4.498316
PARSE calls account for 78.9% of the
experience duration.
That is never appropriate.

@CaryMillsap 82
mrskew --rc=p10.rc --name=parse "prd1_ora_9031.trc"
RANGE {min ≤ e < max} DURATION % CALLS MEAN MIN MAX
----------------------------- ---------- ------ ----- -------- -------- --------
1. 0.000000 0.000001 0.000000 0.0% 307 0.000000 0.000000 0.000000
2. 0.000001 0.000010
3. 0.000010 0.000100
4. 0.000100 0.001000 0.007992 0.0% 8 0.000999 0.000999 0.000999
5. 0.001000 0.010000 0.033000 0.0% 33 0.001000 0.001000 0.001000
6. 0.010000 0.100000
7. 0.100000 1.000000
8. 1.000000 10.000000 735.385205 100.0% 350 2.101101 1.333797 4.498316
9. 10.000000 100.000000
10. 100.000000 1,000.000000
11. 1,000.000000 +∞
----------------------------- ---------- ------ ----- -------- -------- --------
TOTAL (11) 735.426197 100.0% 698 1.053619 0.000000 4.498316
That’s a lot of time spent parsing, and
these PARSE calls are really expensive.

@CaryMillsap 83
mrskew --name=parse --group='$sqlid' --gl=SQLID --sort=4nd "prd1_ora_9031.trc"
SQLID DURATION % CALLS MEAN MIN MAX
------------- ---------- ------ ----- -------- -------- --------
gkbss8w49204k 4.176363 0.6% 349 0.011967 0.000000 4.135371
66kf30526wrgy 3.153521 0.4% 1 3.153521 3.153521 3.153521
3r3dhkb0z824v 2.911558 0.4% 1 2.911558 2.911558 2.911558
3tzra8a2a7pny 1.605757 0.2% 1 1.605757 1.605757 1.605757
2hycpfzdzsu98 3.155520 0.4% 1 3.155520 3.155520 3.155520
6ppu3s1jszy3a 2.208665 0.3% 1 2.208665 2.208665 2.208665
66vkb784j9rcu 1.901711 0.3% 1 1.901711 1.901711 1.901711
5wamvs45j6nh4 1.492773 0.2% 1 1.492773 1.492773 1.492773
dj1buvhxg7h19 1.499772 0.2% 1 1.499772 1.499772 1.499772
41yrts4g94ghn 1.628753 0.2% 1 1.628753 1.628753 1.628753
340 others 711.691804 96.8% 340 2.093211 1.333797 4.498316
------------- ---------- ------ ----- -------- -------- --------
TOTAL (350) 735.426197 100.0% 698 1.053619 0.000000 4.498316
One statement was parsed 349 times; at least 348 of those are
unnecessary.*
There are 350 distinct SQL statements executed by this
report. ...Which is funny, because you know this report, and
you don’t remember there being that many.
*Actually all 349 are unnecessary, because the trace ﬁle shows that there’s never an EXEC call
associated with any of these PARSE calls, but that’s a story for another day.

@CaryMillsap 84
mrskew --rc=distinct-texts.rc "prd1_ora_9031.trc"
SSQLID DISTINCT-TEXTS % CALLS MEAN MIN MAX
---------- -------------- ------ ----- ---- --- ---
4151812497 70 20.0% 70 1 1 1
3642320257 70 20.0% 70 1 1 1
2047770123 70 20.0% 70 1 1 1
1928547239 70 20.0% 70 1 1 1
1138917066 69 19.7% 69 1 1 1
3957414185 1 0.3% 349 0 0 1
---------- -------------- ------ ----- ---- --- ---
TOTAL (6) 350 100.0% 698 1 0 1
For the ﬁrst 5 “shared SQL id” values shown here, there are
~70 distinct statements that could have been sharable.
You should be able to reduce the parse call count from 698
to 6, by writing sharable SQL statements, and pulling
PARSE calls out of loops.

@CaryMillsap 85
mrskew "prd1_ora_9031.trc"
----------------------------- ---------- ------ ----- -------- -------- --------
PARSE 735.426197 78.9% 698 1.053619 0.000000 4.498316
FETCH 91.800028 9.8% 680 0.135000 0.000000 0.506923
EXEC 0.083988 0.0% 349 0.000241 0.000000 0.002000
gc cr block 2-way 0.073233 0.0% 96 0.000763 0.000280 0.001968
CLOSE 0.018999 0.0% 698 0.000027 0.000000 0.001000
12 others 0.061576 0.0% 1,633 0.000038 0.000000 0.001687
----------------------------- ---------- ------ ----- -------- -------- --------
TOTAL (22) 932.416074 100.0% 6,477 0.143958 0.000000 4.498316
Before your boss will let you “ﬁx” this code, you have to predict the beneﬁt.
Reducing the parse count from 698 to 6 should reduce parsing duration
from ~735 to ~7, a savings of about 730 s. Response time should improve
from ~932 s to ~200 s, just from eliminating the PARSE calls only.

@CaryMillsap
OPTIM
IZE ANYT
HING
M
eTH O D
R
You might have known that you should
“use bind variables,” but you couldn’t
have quantiﬁed the R impact on the
experience without this trace ﬁle.
86

@CaryMillsap
BASELINE:
for each invoice number {
cursor = parse(“select ...where invoice_number = ” + number);
exec(cursor);
loop over the result set to fetch all the rows;
}
87
BAD
This is horriﬁc:
• Uses too much CPU for PARSE calls
• Serialization on library cache and shared pool latches
• Consumes too much memory in the library cache
• May execute too many network round-trips

@CaryMillsap
BASELINE: BAD
cursor = parse(“select ...where invoice_number = ” + number);
exec(cursor);
}
FIX 1 “Hey, let’s use bind variables”:
cursor = parse(“select ...where invoice_number = :a1”);
exec(cursor, number);
}
88
STILL BAD
A little better, but still really awful:
• Uses too much CPU for PARSE calls
• Serialization on library cache latches
• Maybe, too many network round-trips

@CaryMillsap
FIX 1 “Hey, let’s use bind variables”: STILL BAD
}
FIX 2:
}
89
BETTER
Better (only 1 PARSE call now!), but
still lots of network round-trips.

@CaryMillsap
FIX 2: BETTER
}
FIX 3:
cursor = parse(“
select ...where invoice_number
in (select invoice number from wherever your for each was getting them)
”);
exec(cursor);
90
Now, only 1 PARSE call, only 1 EXEC call, and
the minimum possible number of network
round-trips.*
*Unless there’s a way to return fewer rows.
BETTER YET

@CaryMillsap
Bad SQL
Bad PL/SQL
Slow network
Missing indexes
Parsing in a loop
Hot block problems
Not enough memory
Disk latency problems
Row locking problems
Row-at-a-time processing
Bad data structure choice
Hardware misconfigurations
Too much load on the system
OS parameters set inadequately
Oracle parameters set inadequately
SQL returns more rows than it should
Database buffer cache hot/cold problems
Oracle query optimizer choosing bad plans
Reports run with poorly limiting parameter values
Inefficient code between database calls in the application 92
A trace file shows you where your time has gone.
Performance problems cannot hide from that.

@CaryMillsap
There are only two possible root causes
for any response time problem:
❶ Call count is too big.
❷ Latency is too big.*
*Probably because someone else’s call counts are too big.
93
#ProTip

@CaryMillsap 94
mrskew --top=10 "prd1_ora_9031.trc"
----------------------------- ---------- ------ ----- -------- -------- --------
PARSE 735.426197 78.9% 698 1.053619 0.000000 4.498316
FETCH 91.800028 9.8% 680 0.135000 0.000000 0.506923
EXEC 0.083988 0.0% 349 0.000241 0.000000 0.002000
gc cr block 2-way 0.073233 0.0% 96 0.000763 0.000280 0.001968
CLOSE 0.018999 0.0% 698 0.000027 0.000000 0.001000
12 others 0.061576 0.0% 1,633 0.000038 0.000000 0.001687
----------------------------- ---------- ------ ----- -------- -------- --------
TOTAL (22) 932.416074 100.0% 6,477 0.143958 0.000000 4.498316
See how there are only two ways to reduce a DURATION?
You have the CALLS column, and the MEAN column.
Proﬁles like this make it easy to see how anything you do to
make something go faster must translate to a manipulation
of either CALLS or MEAN.

@CaryMillsap
With a good trace ﬁle, you can predict the
response time impact of a proposed change.*
*This is nearly impossible to do with systemwide aggregated statistics.
95
#ProTip

@CaryMillsap 96
It just takes practice.

@CaryMillsap
Your code does stuff.
Including some stuff inside Oracle.
The time this stuff takes is your user’s response time.
You can see exactly what it is.
It’s not that hard.
98

@CaryMillsap
Robyn Sands, et al. 2010.
Expert Oracle Practices.
Apress
Detailed information about
instrumenting your Oracle
application code.
Cary Millsap. 2011.
Mastering Oracle Trace Data.
Method R Corporation
Textbook for 1-day course that
teaches you how to master
Oracle trace data.
Ron Crisco, et al. 2011.
Expert PL/SQL Practices.
Apress
instrumenting your Oracle
application code.
Cary Millsap, Jeﬀ Holt. 2003.
Optimizing Oracle Performance.
O’Reilly
Oracle trace data and what to do
with it.
100

@CaryMillsap 101
method-r.com www.enkitec.com
method-r.com/facebook facebook.com/enkitec
@MethodR @Enkitec
cary.millsap@method-r.com cary.millsap@accenture.com
Q&A

How to find and fix your Oracle-based application performance problem

More Related Content

Similar to How to find and fix your Oracle-based application performance problem (20)

More from Cary Millsap (8)

Recently uploaded (20)

How to find and fix your Oracle-based application performance problem