Mastering

 Performance  with  Oracle  Extended  SQL  Trace

Cary  Millsap
Method  R  Corporation,  Southlake,  Texas,  USA
http://twitter.com/CaryMillsap
[email protected]

Revised  2011/01/13
The  most  recent  update  of  this  paper  is  available  free  of  obligation  at  http://method-­‐r.com/downloads.  

Oracle's   extended  SQL   trace  data   stream   contains  a   linear  sequential  record   of   every  
database   call   and   every   operating   system   call   that   the   Oracle   kernel   executes   in  
response  to  the  code  that   you   write.  The  trace  Iile  contains  timings  that   enable  you   to  
precisely  measure  how  long   your  code  takes  to  run   and  why.   The  detailed  information  
about  individual  database  calls  and  operating  system  calls   is  vital  to  building   scalable  
applications   (as  well  as  troubleshooting  them   in  production).  The  information   in  there  
is  a  gold  mine,  and  some  of  it  is  available  nowhere  else.

TABLE  OF  CONTENTS 1 PERFORMANCE  AND  TRACING

1. Performance  and  Tracing  .................................................1 You  click  a  button  on  a  web  form.  It’s  a  button  that  you  
2. Why  You  Should  Trace  .......................................................2 click  all  the  time.  Normally,  the  system  you’re  using  
3. What  Happens  When  You  Trace  ....................................2 responds  to  that  click  almost  instantaneously,  but  
4. Controlling  the  Trace  ..........................................................2 now…  Now,  it’s  taking  more  than  5  seconds  every  time  
5. Session  Tracing  .....................................................................3 you  do  it,  and  sometimes  it  takes  more  than  10.  The  
6. End-­‐to-­‐End  Tracing  .............................................................3 obvious  and  very  simple  question  is:  why?
7. End-­‐to-­‐End  Prerequisites  .................................................4 The  vast  majority  of  people  who  experience  a  problem  
8. Finding  Your  Trace  File  .....................................................5 like  this  never  aind  out  the  answer.
9. Performance  Intrusion  ......................................................6
10. Tracing  in  Multiplexed  Architectures  ..........................6 Think  for  a  minute  about  what  the  answer  ought  to  
11. Tracing  While  You  Develop  .............................................7 look  like.  If  you’re  a  IT  professional,  then  you’ve  
12. Trace  File  Guided  Tour  ......................................................7 probably  seen  so  many  reports  that  the  simple  answer  
13. Trace  File  Alphabet  .............................................................9 doesn’t  even  occur  to  you.  Here’s  what  I  want:  I  want  a  
14. Trace  File  Alphabet:  DB  Calls  ..........................................9 report  telling  me  where  my  time  went.
15. Trace  File  Alphabet:  OS  Calls  ...........................................9 I  don’t  want  to  see  how  busy  the  whole  system  was  
16. Trace  File  Flow  ...................................................................10 when  my  click  took  5  seconds;  I  don’t  want  to  see  all  
17. Trace  File  Arithmetic:  Part  1  ........................................10 the  “bad”  SQL  that  executed  at  or  near  the  time  when  I  
18. Trace  File  Arithmetic:  Part  2  ........................................11 had  my  problem;  and  I  don’t  want  to  see  all  the  
19. Trace  File  Arithmetic:  Part  3  ........................................11 different  kinds  of  waiting  that  the  system  was  doing  
20. Tracing  Traps  ......................................................................11 when  I  clicked.  I  just  want  to  see  where  my  time  went,  
21. Commercial  Trace  File  Tools  ........................................12 untainted  by  information  about  anything  not  directly  
22. Acknowledgments  ............................................................12 relevant  to  my  response  time.
23. About  the  Author  ...............................................................12
24. Appendix:  “Hello  World”  Trace  ...................................13 With  Oracle,  you  can  have  it.  The  feature  that  gives  it  
to  you  is  included  with  every  release  and  every  edition  
of  the  Oracle  Database,  from  Express  Edition  (XE)  to  
Enterprise  Edition  (EE),  since  1992.  You  don’t  have  to  

©  2011  Method  R  Corporation.  All  rights  reserved.   1

buy  an  Option  or  a  Pack  to  use  it.  It’s  called  extended   3 WHAT  HAPPENS  WHEN  YOU  TRACE
SQL  trace.
Tracing  is  an  attribute  of  an  Oracle  session. 2  For  each  
session  with  trace  enabled,  any  Oracle  kernel  process3  
2 WHY  YOU  SHOULD  TRACE
that  executes  instructions  for  that  session  writes  trace  
Oracle  extended  SQL  trace  helps  you  diagnose,  it  helps   data  for  that  session  to  the  process’s  currently  open  
you  learn,  and  it  helps  you  debug: 1 trace  aile.  Some  facts:

• System  administrators,  diagnosticians,  and  trouble-­‐ • Tracing  is  a  session  attribute.

shooters  will  aind  it  invaluable  for  solving  problems  
• But  Oracle  kernel  processes  are  what  write  your  
with  programs  for  which  the  source  is  not  readily   trace  data.
available  (or  not  easily  navigable)  since  they  do  not  
need  to  be  recompiled  in  order  to  trace  them. • Since  more  than  one  process  can  execute  
instructions  for  a  given  Oracle  session  (for  example,  
• Students,  hackers,  and  the  overly-­‐curious  will  aind   by  using  Oracle  Parallel  Execution  features,  or  by  
that  a  great  deal  can  be  learned  about  a  system  and   enabling  Oracle  Shared  Server),  your  trace  data  can  
its  system  calls  by  tracing  even  ordinary  programs. appear  in  more  than  one  aile.
• And  programmers  will  aind  trace  data  very  useful   • The  more  exotic  your  architecture  (for  example,  the  
for  bug  isolation,  sanity  checking,  and  performance   more  tiers  you  use,  the  more  pooling  features  you  
optimization.  Some  problems  cannot  be  solved   use),  the  more  difaicult  it  may  be  to  map  your  trace  
without  it. data  to  a  speciaic  user’s  request.
What  makes  tracing  so  good  compared  to  the  other   Exhibit  1  shows  a  simple  case  in  which  a  user  traces  an  
performance  information  you  can  get  from  Oracle  is   Oracle  session  connected  through  sqlplus,  causing  an  
that  with  tracing,  you  get  a  linear  sequential  record  of   Oracle  kernel  process  to  write  trace  data  to  a  .trc  aile.
everything  your  Oracle  kernel  process  does  while  it’s  
trying  to  compute  the  answer  you’ve  asked  for.  You  
don’t  have  to  worry  about  gaps  in  your  data  (like  with  
Oracle’s  Active  Session  History)  or  aiguring  out  how  to  
isolate  your  process’s  work  in  the  huge  picture  of  
everything  the  system  is  doing  right  now  (like  with  
some  of  Oracle’s  v$  aixed  views  and  all  of  the  GUI  tools  
that  use  them).
With  trace  data,  you  see  every  database  call  your  
application  makes  to  the  Oracle  kernel,  and  you  see  
every  system  call  your  kernel  makes  to  its  host  
operating  system.  And  it’s  the  perfect  tool  for  
developers,  because  you  don’t  have  to  have  loads  of   Exhibit  1.  The  Oracle  kernel  process,  to  which  a  
complicated  permissions  to  use  it.  I  believe  you  can   user  connects  through  some  type  of  application,  
learn  more  about  Oracle  performance  with  trace  data   is  the  process  that  writes  the  trace  Iile.
in  a  day  than  you  can  learn  studying  v$  data  for  a  
month,  because  trace  data  so  directly  maps  to  the  
4 CONTROLLING  THE  TRACE
reality  of  how  the  user,  the  application  code,  and  the  
database  interact. You  can  trace  any  Oracle  session  on  your  system  if  you  
have  permission  to  execute  the  standard  Oracle  PL/
Trace  ailes  are  the  foundation  of  almost  everything  you  
SQL  package  called  dbms_monitor.  This  package  gives  
need  to  know  about  Oracle  performance.

1  Oracle  extended  SQL  trace  is  useful  for  the  same  kinds  of  reasons  that  the  Unix  strace  tool  is  useful.  My  description  here  is  

paraphrased  from  the  strace  manual  page  at  http://linuxcommand.org/man_pages/strace1.html.

2  An  Oracle  session  is  an  entity  with  a  unique  sid-­‐serial#  combination  within  an  Oracle  instance,  which  you  can  see  in  v$session.

3  An  Oracle  kernel  process  is  an  operating  system  process  (which  you  can  inspect  with,  for  example,  a  Linux  command  like  ps,  lsof,  

or  strace)  that  executes  the  instructions  required  by  an  Oracle  session.

©  2011  Method  R  Corporation.  All  rights  reserved.   2

you  a  lot  of  control  over  which  sessions  you  trace.   Triggers
Here  are  some  of  its  procedures: For  fast-­‐moving  sessions  that  don’t  give  you  time  
to  identify  their  session  ids  before  they  begin  
session_trace_enable
doing  work  (like  that  Pro*C  program  that  runs  for  
Enable  tracing  for  either  the  database  session  
20  seconds,  but  that  should  run  in  2),  triggers  are  
from  which  the  procedure  is  called,  or  a  speciaic  
a  convenient  tool  for  inserting  code  to  enable  
session  identiaied  by  its  session  identiaier  and  
tracing.  You  can,  for  example,  create  an  after  logon  
serial  number.
trigger  whose  body  uses:
database_trace_enable session_trace_enable(
Enable  tracing  for  a  speciaic  instance  or  a  whole   session_id => null,
database. serial_num => null,
…
Oracle  provides  a  disable  procedure  to  match  each   )

enable  procedure. This  code  will  enable  tracing  for  the  session  that’s  
logging  on.  There’s  no  need  for  a  corresponding  
When  you  use  dbms_monitor,  I  recommend  that  you  
before  logoff  trace  disablement,  by  the  way,  
use  the  default  value  of  waits=>true.  If  you  use  the  
because  logging  off  automatically  disables  the  
value  waits=>false,  you’ll  be  using  SQL  trace,  but  not  
trace.
extended  SQL  trace.  You’ll  get  a  trace  of  the  Oracle  
database  calls  your  application  makes,  but  not  the  OS   sys_context
calls  that  Oracle  makes.  Without  a  record  of  those  OS   The  sys_context  function  gives  you  dozens  of  hooks  
calls  in  your  trace  aile,  you’ll  not  be  able  to  account  for   you  can  use  inside  of  a  trigger  body  to  identify  the  
potentially  signiaicant  amounts  of  response  time. session  that’s  logging  in.  For  example,
sys_context('USERENV','IP_ADDRESS')
I  also  recommend  that  you  use  the  non-­‐default  value  
binds=>true.  Setting  binds=>true  causes  the  Oracle   …returns  the  IP  address  of  the  machine  from  
Database  to  emit  the  values  that  the  application  bound   which  the  client  is  connected.  With  sys_context,  
to  its  placeholder  variables  into  the  trace  data  stream.   you  can  write  PL/SQL  code  in  your  after  logon  
For  example,  if  someone  has  executed  this  SQL   trigger  body  that  can  decide  to  trace  only  logins  
statement: originating  from  a  speciaic  IP  address,  only  logins  
requested  by  certain  client  OS  users,  or  logins  
update salary set amount = :v1 where id = :v2
matching  any  of  the  other  criteria  that  sys_context  
…then  your  trace  data  will  show  what  values  were   makes  available  to  you.
assigned  to  v1  and  v2.  This  is  invaluable  information   Running  suspicious  SQL  in  your  own  session
for  creating  reproducible  test  cases  for  solving   There  are  lots  of  ways  you  can  identify  suspicious  
problems.  However,  be  aware  of  the  data  security   SQL  statements:  you’ll  aind  them  in  OEM,  in  your  
implication  of  emitting  bind  data  into  a  trace  aile;   AWR  or  ADDM  reports,  or  by  querying  v$sqlstat.  
when  you  do  it,  you  are  copying  potentially  sensitive   Copy  them  and  run  them  in  SQL*Plus  or  Oracle  
data  from  your  database  to  your  aile  system. SQL  Developer.  You  can  plug  in  whatever  bind  
variable  values  you  need  and  trace  to  your  heart’s  
5 SESSION  TRACING content.
One  of  the  nice  things  about  Oracle  tracing  is  that   dbms_monitor.database_trace_enable
you’re  not  required  to  change  an  application  in  order   If  all  else  fails,  there’s  always  the  nuclear  bomb  
to  use  it.  Even  if  the  application  you’re  using  is   option  of  tracing  everything  that  happens  over  
completely  un-­‐modiaiable  and  was  in  no  way  designed   some  time  interval  you  choose.  Beware,  however:  
to  cooperate  with  tracing,  you  still  have  the  following   enabling  trace  for  a  whole  instance  or  database  
very  powerful  tools  at  your  disposal: will  only  activate  tracing  for  new  sessions  that  log  
in  after  you  enable  the  trace.  It  won’t  enable  
dbms_monitor.session_trace_enable
tracing  for  sessions  that  are  already  logged  in.
With  session_trace_enable,  you  can  trace  any  
session  on  your  system  if  you  know  its  session  id  
and  serial  number.  You  just  have  to  be  meticulous   6 END-­‐TO-­‐END  TRACING
to  enable  the  trace  immediately  before  the  user   There  are  three  major  milestones  in  the  history  of  
issues  the  slow  request,  and  then  disable  the  trace   tracing  in  the  Oracle  Database:
immediately  after  the  slow  request  is  fulailled.

©  2011  Method  R  Corporation.  All  rights  reserved.   3

• In  1988,  Oracle  version  6  introduced  SQL  trace,   the  business  task  that  the  user  is  experiencing  right  
which  let  us  discover  the  start  and  end  time  of  every   there  at  her  keyboard.  When  you  trace  a  session  based  
database  call  an  application  executed. on  its  session  id,  you  almost  always  capture  either  
more  trace  data  than  you  really  want,  or  not  enough.  
• In  1993,  Oracle  version  7  introduced  extended  SQL  
trace,  which  let  us  discover  the  duration  of  every   That’s  because  it  takes  you  a  little  time  to  identify  the  
system  call  the  database  executed. session’s  id  before  you  can  execute  the  
session_trace_enable  call.
• In  2003,  Oracle  version  10  introduced  end-­‐to-­‐end  
extended  SQL  trace,  which  let  us  activate  tracing,   However,  when  you  activate  tracing  for  a  speciaic  
not  just  for  sessions  whose  session  ids  we  knew,  but   client  id  or  service-­‐module-­‐action  name  combination,  
for  sessions  with  speciaic  attribute  values. you  create  a  standing  order  to  Oracle,  instructing  the  
kernel  to  begin  tracing  any  session  that  changes  its  
The  third  milestone—end-­‐to-­‐end  tracing—makes  it   client  id  to  your  speciaic  value,  or  whenever  it  changes  
possible  for  you  to  specify  the  intent  to  trace  a   its  server-­‐module-­‐action  name  combination  to  your  
business  task  in  advance  of  that  task  ever  running.   speciaic  values.  And  then  Oracle  will  stop  tracing  
Then,  the  Oracle  Database  will  automatically  enable   immediately  when  that  session  changes  its  identifying  
tracing  for  any  session  that  executes  the  task  you  have   information  to  something  different.
speciaied.
Pure  magic.
The  way  you  specify  the  standing  order  to  trace  
certain  tasks  is  with  the  following  dbms_monitor   7 END-­‐TO-­‐END  PREREQUISITES
procedures:
But  there’s  a  catch  with  end-­‐to-­‐end  tracing.  How  can  
client_id_trace_enable
dbms_monitor  know  a  session’s  client  id  or  service-­‐
Enable  tracing  for  any  database  session  whose  
module-­‐action  names?  The  answer:  your  application  
client  identiaier  matches  the  one  you  specify.
has  to  provide  those  attribute  values.  If  your  
serv_mod_act_trace_enable application  doesn’t  name  its  tasks  or  set  the  client  id  
Enable  tracing  for  any  database  session  whose   attribute  for  its  sessions  in  a  useful  way,  then  you  can’t  
service,  module,  and  action  names  match  the  ones   use  the  dbms_monitor  end-­‐to-­‐end  tracing  features.
you  specify.
If  you’re  using  an  off-­‐the-­‐shelf  application,  your  only  
For  example,  if  you  wanted  to  trace  any  user  whose   recourse  may  be  to  live  without  the  end-­‐to-­‐end  tracing  
client  id  is  192.168.1.104,  you’d  use   features  (though  you  can  still  use  the  techniques  
dbms_monitor.client_id_trace_enable( described  in  section  5)  and  lobby  your  vendor  to  
client_id => '192.168.1.104', instrument  a  future  version  of  your  application  in  the  
waits => true,
binds => true,
manner  I’m  about  to  describe.  If  you  can  change  your  
plan_stat => 'ALL_EXECUTIONS' application  source  code  yourself,  instrumenting  for  
) end-­‐to-­‐end  tracing  is  not  hard  to  do.  The  only  thing  
…And  that’s  that.  No  matter  when  the  next  person   you  really  have  to  do  is  have  each  session  set  what  
logged  in  from  192.168.1.104,  his  work  would  be   Oracle  calls  its  user  session  handle  attributes.  There  are  
traced.  If  you  wanted  to  trace  any  business  task  whose   several  ways  to  do  it.
service  is  called  FINAPPS,  whose  module  is  called   One  way  you  can  set  those  session  handle  attributes  
GLPPOS,  and  whose  action  is  called  Concurrent for  your  code  is  to  use  these  standard  Oracle  PL/SQL  
Request,  you’d  execute: packages:
dbms_monitor.serv_mod_act_trace_enable(
service_name => 'FINAPPS',
dbms_session
module_name => 'GLPPOS', The  set_identiIier  procedure  lets  you  set  the  calling  
action_name => 'Concurrent Request' session’s  client  identiaier.
waits => true,
binds => true, dbms_application_info
plan_stat => 'ALL_EXECUTIONS' The  set_action  procedure  lets  you  set  the  calling  
)
session’s  action  name,  and  the  set_module  
…Then  you’d  be  set  to  trace  your  next  several  GL   procedure  lets  you  set  the  calling  session’s  module  
Posting  runs. and  action  names  together.
Before  end-­‐to-­‐end  tracing,  the  hardest  thing  about   If  you  prefer  a  simpler  approach  to  setting  your  
tracing  was  getting  a  trace  aile  that  represents  exactly   session  handle  attributes,  you  can  use  the  

©  2011  Method  R  Corporation.  All  rights  reserved.   4

Instrumentation  Library  for  Oracle  (ILO),  a  free  open-­‐ procedures  instead,  until  you  can  convince  the  people  
source  project  that  Method  R  Corporation  maintains  at   who  built  your  application  to  instrument  the  
SourceForge.  ILO  gives  you  PL/SQL  packages  that   application  so  that  you  can  measure  and  manage  its  
make  it  dead  simple  for  an  application  developer  to   performance  while  you  operate  it.  Until  then,  not  only  
deaine  where  in  an  application’s  code  path  each   will  you  not  be  able  to  trace  as  alexibly,  you  also  won’t  
business  task  begins  and  ends.  It  sets  all  the  “hooks”   be  able  to  group  application  performance  statistics  
that  your  systems  DBAs  will  need  later  to  trace   using  other  applications  like  Oracle  Enterprise  
speciaic  business  tasks  by  name.  ILO  gives  you  several   Manager.
extra  features  as  well,  which  are  detailed  in  the  
project’s  documentation. 8 FINDING  YOUR  TRACE  FILE
If  it  bothers  you  to  insert  additional  database  calls  into   Once  you’ve  traced  your  code,  you’ll  need  to  aind  your  
your  application  (and  if  you’re  building  an  application   trace  aile.  The  process  running  your  Oracle  kernel  code  
that  will  serve  thousands  of  database  requests  per   writes  your  trace  aile  to  the  operating  system  directory  
second,  it  should),  then  you  should  choose  another   named  by  an  Oracle  instance  parameter.  In  Oracle  
option.  Some  Oracle  APIs  allow  you  to  pass  service-­‐ version  11,  the  Oracle  kernel  will  write  your  ailes  into  
module-­‐action  names  and  a  client-­‐id  from  the   a  directory  that  you  can  identify  by  using  the  following  
application  to  the  database  as  an  attribute  of  each  of   SQL  statement:
the  database  calls  that  you  were  already  determined   select * from v$diag_info where name='Diag Trace'
to  make.
In  older  versions  of  Oracle,  you  can  identify  your  trace  
The  Oracle  user  session  handle  attributes  you’ll  need   aile  directories  by  using  this  statement:
to  set  are  these:
select * from v$parameter where name in
OCI_ATTR_CLIENT_IDENTIFIER ('user_dump_dest','background_dumnp_dest')
OCI_ATTR_SERVICE
OCI_ATTR_MODULE In  Oracle  version  10  and  prior,  the  user_dump_dest  
OCI_ATTR_ACTION
directory  is  probably  where  most  of  your  trace  ailes  
You  can  set  them,  for  example,  with  the  Oracle  Call   will  be.  If  you  use  Oracle  parallel  execution  features,  
Interface  (OCI)  OCIAttrSet  function,  or  with  the  JDBC   then  you’ll  aind  some  of  your  trace  ailes  in  the  
setEndToEndMetrics  method.  These  are  client-­‐side   background_dump_dest  directory.
function  calls  that  don’t  require  a  round-­‐trip  to  the   Different  ports  of  Oracle  use  different  naming  
database.  Once  you  have  made  a  call  to  set  your  user   conventions  for  trace  ailes.  Your  trace  aile  names  will  
session  handle  attributes,  each  database  call  that  your   probably  look  something  like  one  of  the  following:
application  code  makes  after  that  is  marked  with  the  
xe_ora_10840.trc
identifying  information  that  dbms_monitor  needs. prod7_23389_ora.trc
ora_1492_delta1.trc
Setting  your  Oracle  user  session  handle  attributes   ORA01215.trc
yourself  enables  you  to  instrument  your  application   fin1_ora_11297_POSTING.trc
with  no  extra  database  calls  at  all,  which  means  no   MERKUR_S7_FG_ORACLE_013.trc
extra  network  round-­‐trips  between  the  application   Trace  ailes  may  look  different  on  different  platforms  
and  the  database.  It’s  the  most  scalable  way  to   and  on  different  versions  of  Oracle,  but  you  can  count  
instrument  your  code. on  your  trace  aile  names  containing  at  least  some  of  
To  summarize,  you  can  instrument  your  application   the  following  elements:
code  for  end-­‐to-­‐end  tracing  by  calling  standard  Oracle  
PL/SQL  procedures  (or  calling  them  through  someone   • The  string  ora.
else’s  package,  like  ILO),  or  you  can  do  it  with  no  extra   • Your  Oracle  instance  name.
database  calls  by  setting  session  handle  attributes   • Your  Oracle  kernel  process  id  (on  Microsoft  
directly  with  an  Oracle  API  like  OCI  or  JDBC. Windows,  it  will  be  your  process’s  thread  id).
If  your  application  is  not  designed  to  set  these   • If  you  set  a  traceIile_identiIier  within  your  Oracle  
attributes  for  your  sessions,  then  you  won’t  be  able  to   session  (with  an  alter  session  command),  then  the  
use  the  client_id_trace_enable  or   string  value  of  that  parameter.
serv_mod_act_trace_enable  procedures  in  
• The  sufaix  .trc.
dbms_monitor.  You’ll  have  to  use  only  the  
session_trace_enable  and  database_trace_enable  

©  2011  Method  R  Corporation.  All  rights  reserved.   5

If  you’re  writing  code  that  will  connect  to  an  Oracle   they’ll  get  into  trouble  if  they  do  anything  that  might  
instance  that  someone  else  manages,  you’ll  need  to   make  performance  worse  than  it  already  is.  However,  
coordinate  with  that  person  to  get  permissions  to  read   your  goal  in  tracing  is  to  learn  how  much  faster  a  
your  trace  ailes.  Without  access  to  your  trace  ailes,   program  ought  to  be  running.  Isn’t  that  information  
optimizing  the  code  you  write  is  going  to  be  a  lot  more   worth  at  least  as  much  as  a  one-­‐time  response  time  
expensive  for  your  company. penalty  induced  by  a  trace?  I’ve  seen  people  work  on  
problems  for  months  on  end  without  trace  data,  when  
9 PERFORMANCE  INTRUSION they  could  have  cleared  everything  up  with  just  a  
single  one-­‐time  trace.  Even  if  tracing  imposed  aive  
One  thing  that  everyone  always  wants  to  know  is,   times  the  performance  penalty  that  it  does,  I  wouldn’t  
“What  is  the  performance  penalty  of  tracing?”  People   use  it  any  less.
get  a  little  nervous  about  all  those  extra  timer  calls  
that  their  Oracle  kernel  processes  must  execute  when   10 TRACING  IN  MULTIPLEXED  ARCHITECTURES
you  turn  trace  on.
I  mentioned  in  section  3  that  tracing  is  an  attribute  of  
What  they  don’t  realize  is  that  their  Oracle  kernel  
the  Oracle  session.  However,  many  systems  these  days  
process  is  making  all  those  timer  calls  anyway,  
have  many  layers  of  software  between  the  user  and  
whether  tracing  is  activated  or  not.  How  else  could  it  
the  Oracle  session,  and  many  of  those  layers  employ  
publish  the  timing  information  you  see  in  aixed  views  
pooling  technologies  that  conceal  the  user’s  identity  
like  v$sesstat,  v$session_event,  and  v$session_wait?  The  
from  her  Oracle  session.
actual  performance  intrusion  of  tracing  is  the  cost  of  
the  additional  I/O  calls  required  to  persist  to  your   The  solution,  in  general,  is  this:  where  there’s  a  will,  
trace  aile  the  information  Oracle  is  already  collecting. there’s  a  way.  If  you  are  committed  to  tracing  a  
speciaic  user’s  speciaic  business  task,  it  can  be  done.  
Most  people  should  trace  only  a  few  sessions  at  a  time.  
Here  are  some  of  the  tools  you  can  use  to  do  it:
If  you  don’t  have  specialized  tools,  you’ll  never  in  your  
life  be  able  to  sift  through  all  the  trace  ailes  that  an   User  session  handle  attributes
Oracle  Database  system  can  generate  in  a  good,  busy   The  service,  module,  action,  and  client  id  
hour  of  work.  If  you  want  to  trace  a  lot  of  sessions  at   attributes  are  the  right  and  proper  cure  for  tracing  
once  (and  there  are  legitimate  reasons  for  wanting  to),   in  pooled  architectures.  When  the  application  
you  can  trace  as  many  sessions  as  your  hardware   identiaies  its  users  to  the  Oracle  database,  the  
capacity  will  allow.  If  you  trace  infrequently,  then  you   dbms_monitor  package  lets  you  enable  tracing  for  
don’t  have  to  be  too  careful  about  your  trace  aile   exactly  the  tasks  you  want  traced,  and  then  tools  
destination.  If  you  trace  a  lot  of  sessions   like  Oracle’s  trcsess  help  you  do  the  work  of  
simultaneously,  then  just  make  sure  you  write  your   concatenating  the  various  fragments  of  trace  data  
trace  ailes  to  a  drive  array  that  has  all  the  storage   that  will  describe  how  your  task  spent  its  time.
capacity  and  throughput  capacity  you  need,  and  then   Session  isolation
trace  all  you  want. If  your  application  doesn’t  set  its  session  handle  
Unless  you’re  using  a  version  of  Oracle  that  has  a   attributes  that  allow  you  to  distinguish  the  
trace-­‐related  performance  bug  (Oracle  bug  3009359,   business  task  you’re  interested  in  from  the  
introduced  in  version  9.2.0.2,  was  a  famous  one),  your   thousands  of  other  business  tasks  running  at  the  
program  will  probably  run  with  ±5%  of  the  duration   same  time,  then  isolate  your  user  into  a  
that  it  takes  to  run  it  without  tracing.4  If  a  program’s   conaiguration  where  you  can  make  the  distinction.  
response  time  gets  more  than  10%  worse  when  you   For  example,  instead  of  your  user  connecting  to  
trace  it,  then  you  should  check  whether  your  I/O   her  usual  http://www.app.com  site,  where  her  task  
subsystem  to  which  you’re  writing  trace  ailes  is  ok,   gets  stripped  of  her  identity  as  it  enters  some  
check  whether  your  session’s  setting  of  statistics_level   session  pool  or  connection  pool  with  a  thousand  
is  what  you  want  it  to  be,  and  check  your  Oracle   other  tasks,  have  her  connect  to  http://
Support  database  for  bugs. www.app.com:88,  where  her  task  is  the  only  task  
in  a  session  pool  or  connection  pool  with  a  pool  
People  who  ask  about  the  performance  intrusion  of   size  of  exactly  one.  You  can  trace  every  session  
tracing  generally  ask  because  they’re  fearful  that  

4  That’s  right,  plus  or  minus.  Sometimes  a  program  will  run  faster  when  you  trace  it.  It  sounds  nice,  but  it’s  usually  not  what  you  

want.

©  2011  Method  R  Corporation.  All  rights  reserved.   6

 connecting  to  the  database  through  port  88,   4. Developers  and  database  administrators  
which,  at  the  time  you’re  doing  your  analysis,  you   participate  together  during  integration  testing  to  
know  will  be  nobody  else  but  the  one  user  you’re   trace  the  application  code  (which  is  easy  with  the  
analyzing. user  session  handle  attributes  all  set)  and  study  
the  trace  ailes  to  ensure  that  the  code  is  efaicient  
11 TRACING  WHILE  YOU  DEVELOP under  load  and  that  there  are  no  unwanted  
performance  interdependencies  among  competing  
The  most  important  time  for  a  developer  to  trace  is   sessions.
early  in  the  software  development  lifecycle,  where  
defects  are  cheaper  and  easier  to  aix  than  any  other   5. Production  system  administrators  (database  
time  in  your  project.5  Fortunately,  tracing  is  easiest  in   administrators,  for  example)  trace  the  application  
the  development  environment. code  periodically  in  production  when  the  
application  is  behaving  well,  to  record  a  baseline  
I  can  remember  since  my  earliest  days  of  working  with   of  how  each  application  task  is  supposed  to  
Oracle  having  some  kind  of  sqlplus  script  that  will  turn   perform  (which  is  easy  with  the  user  session  
trace  on  or  off  when  you  call  it.  It  could  be  as  simple  as   handle  attributes  all  set).
two  scripts  traceon.sql  and  traceoff.sql  that  look  like  
6. During  performance  problem  episodes,  
this:
production  operators  trace  problem  business  
$ cat traceon.sql tasks  (session  handle  attributes  make  this  easy)  
exec dbms_monitor.session_trace_enable(binds=>true);
$ and  work  together  with  application  developers—
$ cat traceoff.sql who  are  already  familiar  with  what  the  trace  aile  
exec dbms_monitor.session_trace_disable(); for  the  given  task  normally  looks  like—to  solve  
When  you  write  your  code,  it  works  like  this: problems  quickly.
@traceon With  habitual  tracing  throughout  a  software  lifecycle,  
-- Your code goes here applications  tend  to  have  fewer  performance  
@traceoff
problems,  and  they  tend  to  get  aixed  more  quickly  on  
Nowadays,  I  use  the  MR  Trace  extension  for  Oracle   the  rare  occasion  when  they  do  have  a  problem.  They  
SQL  Developer.  When  MR  Trace  is  enabled,  every  Run   tend  to  evolve  toward  being  fast,  optimized  systems  
Script  button  click  creates  a  trace  aile,  which  MR  Trace   that  can  last  a  long  time.  …Much  like  the  Oracle  
automatically  fetches  for  you  from  your  database   Database,  which  is  itself  heavily  performance  
server  to  your  workstation.  It  also  lets  you  tag  your   instrumented.
local  trace  ailes  to  make  them  easy  to  aind  later.
The  best  thing  a  developer  can  do  to  promote  high   12 TRACE  FILE  GUIDED  TOUR
performance  over  an  application’s  long  lifespan  is  to   Once  you  have  your  trace  aile  in  hand,  the  real  fun  
make  it  easy  to  trace  the  application  throughout  its   begins.  Now  you  have  a  call-­‐by-­‐call  description  of  what  
production  lifecycle.  Of  course,  you  do  this  by  coding   the  Oracle  Database  did  throughout  your  entire  
the  application  to  set  its  own  user  session  handle   response  time.
attributes,  as  I’ve  described  in  section  7.
If  your  only  interface  to  the  Oracle  Database  has  been  
Applications  work  better  when  the  world  works  like   SQL  or  PL/SQL  executed  through  client  software  like  
this: sqlplus  or  Oracle  SQL  Developer,  you  may  not  know  
1. Application  developers  trace  their  SQL  and  PL/ about  the  distinct  steps  that  Oracle  uses  to  process  a  
SQL  code  as  they  write  it.  They  study  their  trace   single  statement.  Exhibit  2  shows  the  steps  from  the  
ailes  to  ensure  that  the  code  they’re  writing  is   perspective  of  the  Oracle  Call  Interface.
efaicient.
2. Developers  become  aluent  in  understanding  what  
the  trace  ailes  mean,  which  results  in  more  
efaicient  application  code.
3. Application  developers  code  applications  to  set  
their  user  session  handle  attributes.

5  Barry  Boehm,  1981,  Software  Engineering  Economics,  Englewood  Cliffs  NJ:  P  T  R  Prentice  Hall,  p40.

©  2011  Method  R  Corporation.  All  rights  reserved.   7

 line  38,  the  Oracle  kernel  process  issued  an  OS  write  
call  to  send  the  result  of  the  prepare-­‐execute  call  pair  
back  to  the  client  (which  Oracle  calls  SQL*Net  message  
to  client).
On  line  39  is  a  fetch  call,  which  consumed  8  µs  and  
returned  1  row.  Line  40  shows  the  execution  plan  that  
Oracle  used  to  process  my  query.  Line  41  shows  the  
Oracle  kernel  blocked  on  an  OS  read  call  for  375  µs,  
awaiting  another  database  call  from  the  client  (which  
Oracle  calls  SQL*Net  message  from  client).
On  line  42,  the  kernel  has  processed  a  second  fetch  
call,  consuming  1  µs  and  fetching  no  rows.  Line  43  
shows  that  the  kernel  issued  another  OS  write  call  to  
send  the  zero-­‐row  result  set  back  to  the  client  (which  
of  course  informs  the  client  that  there  are  no  more  
rows  to  be  had  from  the  query).  Line  44  shows  the  
kernel  blocked  for  1,627  µs  on  another  OS  read  call,  
awaiting  instructions  for  what  to  do  next.  Line  45  
Exhibit  2.  Steps  in  processing  a  SQL  statement,   shows  that  the  next  action  was  a  commit.
from  the  OCI  Programmer’s  Guide.
The  remainder  of  the  trace  aile  (shown  in  its  entirety  
You  can  see  this  sequence  of  activity  whenever  you   in  section  24)  is  interesting,  because  it  shows  
trace  even  a  simple  SQL  statement  like  this: something  you  might  not  have  expected:  the  session  
select 'Hello world' from dual
inserted  a  row  into  sys.aud$.  From  this,  I  can  tell  that  
my  instance  is  set  up  for  session-­‐level  auditing.  I  didn’t  
An  Oracle  11.2.0.2.0  trace  aile  for  an  execution  of  this   have  to  work  my  way  down  a  checklist  to  aigure  that  
SQL  statement  is  shown  in  its  entirety  in  section  24.   out;  plus,  my  trace  data  will  show  me  exactly  the  cost  
Here  are  some  highlights. of  that  overhead  upon  my  response  time.
On  line  34,  you  can  see  the  “Hello  world”  query  itself.   On  line  48,  you  can  see  the  SQL  statement.  Line  50  
Lines  36–45  show  how  Oracle  processed  the  query: shows  that  the  prepare  call  for  that  statement  
32. ===================== consumed  93  µs.  On  lines  51–156,  you  can  see  the  
33. PARSING IN CURSOR #7285244 len=30 dep=0 uid=5 oct=3 values  that  were  bound  into  the  statement’s  
lid=5 tim=1294688002747560 hv=1604528588 ad='3d526cd8'
sqlid='guthms1gu6afc'
placeholder  variables  (:1,  :2,  …,  :22)  within  that  
34. select 'Hello world' from dual statement.  Line  157  shows  that  the  execute  call  for  the  
35. END OF STMT insert  took  1,733  µs.  Line  158  shows  the  execution  
36. PARSE …,e=70,…
37. EXEC …,e=112,…
plan  for  the  insert,  and  line  159  shows  that  the  close  
38. WAIT …'SQL*Net message to client' ela= 2 call  for  the  insert  took  5  µs.  Line  160  shows  that  the  OS  
39. FETCH …,e=8,…,r=1,… call  Oracle  uses  to  implement  its  log  Iile  sync  event  
40. STAT …op='FAST DUAL (…)'
41. WAIT …'SQL*Net message from client' ela= 375
consumed  153  µs,  and  the  ainal  two  lines  of  the  trace  
42. FETCH …,e=1,…,r=0,… aile  show  two  more  close  calls,  which  consumed  9  µs  
43. WAIT …'SQL*Net message to client' ela= 1 and  30  µs.
44. WAIT …'SQL*Net message from client' ela= 1627
45. XCTEND rlbk=0,… I  hope  you  can  see  from  just  this  little  tour  how  it  is  
Here,  you  can  see  the  query  being  prepared  (the   possible  to  use  trace  data  to  see  exactly  where  your  
PARSE  line),  which  consumed  70  µs.6  Next,  it  was   time  has  gone.7  You  just  need  to  know  how  to  read  
executed  (regarding  Exhibit  2,  there  was  no  bind   them.  Let’s  begin  with  the  alphabet.
necessary,  since  my  query  didn’t  contain  any  
placeholders),  which  consumed  112  µs.  Then,  on  

6  Call  durations  in  trace  ailes  are  expressed  in  microseconds,  abbreviated  µs.  1  µs  =  .000001  seconds,  or  1/1,000,000th  of  a  second.

7  You  can  see  another  trace  aile  guided  tour  in  “For  Developers:  Making  Friends  with  the  Oracle  Database  for  Fast,  Scalable  

Applications.”

©  2011  Method  R  Corporation.  All  rights  reserved.   8

13 TRACE  FILE  ALPHABET p=1
This  call  obtained  1  Oracle  block  via  OS  read  calls.
Trace  ailes  look  plenty  ugly  when  you  airst  meet  them,   cr=7
but  it’s  not  too  difaicult  to  get  a  hang  of  what  they   This  call  executed  7  “consistent  mode”  accesses  
mean.  There  are  really  only  two  types  of  lines  in  the   upon  the  Oracle  database  buffer  cache.
trace  aile  that  tell  you  where  time  has  gone:
cu=0
Database  call  lines This  call  executed  0  “current  mode”  accesses  upon  
Database  call  lines  begin  with  the  keyword  PARSE,   the  Oracle  database  buffer  cache.
EXEC,  FETCH,  CLOSE,  UNMAP,  or  SORT  UNMAP.   mis=0
Such  a  line  indicates  that  an  application  client  has   This  call  encountered  0  misses  upon  the  Oracle  
made  a  database  call,  and  that  the  Oracle  Database   library  cache.
has  responded  to  it. r=1
System  call  lines This  call  returned  1  row.
System  call  lines  begin  with  the  keyword  WAIT.   dep=2
Such  a  line  indicates  that  the  Oracle  kernel   There  were  2  levels  of  database  calls  below  this  
process  has  made  an  OS  call,  and  that  the  OS  has   call  in  the  database  call  stack.  That  is,  this  call  is  a  
responded  to  it. child  of  another  db  call,  which  in  turn  is  a  child  of  
That’s  it;  just  two. another  db  call.
og=4
14 TRACE  FILE  ALPHABET:  DB  CALLS This  call’s  Oracle  optimizer  goal  is  choose.  The  
four  choices,  in  order,  are  all,  Iirst,  rule,  and  
With  exception  of  the  CLOSE  call,  which  doesn’t   choose.
contain  all  the  aields  that  the  other  database  calls  do,   plh=3992920156
all  contain  the  same  aields.  They  look  like  this: This  call’s  execution  plan  hash  value  is  
FETCH #12390720:c=8000,e=1734,p=1,cr=7,cu=0,mis=0,r=1, 3992920156.
dep=2,og=4,plh=3992920156,tim=1294438818783887
tim=1294438818783887
Here’s  what  the  aields  mean: This  call  completed  at  time  1,294,438,818,783,887  
FETCH ±1  µs.  The  tim  value  is  given  in  µs,  but  the  answer  
This  was  a  fetch  call  (not  a  parse  or  exec  or  some   to  “µs  since  what?”  varies  by  platform.  All  that  
other  kind). really  matters  is  for  the  tim  values  to  be  consistent  
within  a  given  trace  aile,  which  they  usually  are.9
#12390720
The  call’s  cursor  handle  id  is  12390720.  You  can  
search  backwards  (toward  the  beginning)  in  your   15 TRACE  FILE  ALPHABET:  OS  CALLS
trace  aile  to  aind  the  SQL  or  PL/SQL  statement  for   All  OS  call  lines  resemble  this:
which  this  call  was  made. WAIT #12397272: nam='db file sequential read' ela= 221
c=8000 file#=1 block#=2735 blocks=1 obj#=423
This  call  consumed  8,000  ±10,000  µs  of  CPU  time.   tim=1294438818791494

Yes,  Oracle  measured  the  call  as  having  consumed   Here’s  what  the  aields  mean:
[far]  more  CPU  time  than  the  call’s  actual  elapsed  
WAIT
duration.8
This  call  is  an  OS  call,  not  a  database  call.10
e=1734
#12397272
This  call  consumed  1,734  ±1  µs  of  elapsed  time.
The  call’s  cursor  handle  id  is  12397272.

8  For  a  detailed  description  of  why  the  CPU  consumption  statistic  is  accurate  to  only  ±10,000  µs  (except  on  Solaris  machines  on  

which  microstate  accounting  is  enabled),  see  Optimizing  Oracle  Performance,  by  Cary  Millsap  and  Jeff  Holt  (O’Reilly  2003),  pp161–
165.

9  They’re  most  likely  not  to  be  consistent  within  a  given  aile  on  Microsoft  Windows  systems,  where  tim  values  wrap  (that  is,  get  as  

large  as  they  can  go  and  then  start  over  at  0)  more  frequently  than  on  other  systems.

10  I’ve  had  good  fun  explaining  why  the  word  wait  is  the  wrong  word  to  use  in  this  context.  See  for  example  my  essay,  “Dang  it,  

people,  they’re  syscalls,  not  ‘waits’...”  

©  2011  Method  R  Corporation.  All  rights  reserved.   9

nam='db file sequential read' each  dep=1  call  as  the  child  of  the  dep=0  call  that  
This  call  was  a  read  of  database  aile  that  retrieved   followed  it  in  the  trace  data  stream.  With  tim  values  on  
one  or  more  blocks  and  stored  them  contiguously   all  our  call  lines,  it  is  no  longer  necessary  to  do  that.  
somewhere  in  memory.  Oracle  Corporation  gives   You  can  determine  all  the  parent-­‐child  information  
each  OS  call  its  kernel  makes  a  distinct  name,  and   you  need  now  just  by  noticing  at  what  time  each  call  
the  company  documents  the  meanings  of  those   begins  and  ends.
names  so  you  can  more  easily  understand  what’s  
going  on. 17 TRACE  FILE  ARITHMETIC:  PART  1
ela= 221
This  call  consumed  221  ±1  µs  of  elapsed  time. Now  you  know  the  trace  aile  alphabet  and  how  trace  
aile  lines  alow  forth  at  the  completion  of  the  database  
file#=1 block#=2735 blocks=1
calls  your  client  application  makes  and  the  OS  calls  
This  read  call  (probably  a  pread  call—you  can  tell  
your  Oracle  kernel  makes.  The  next  step  is  to  
by  using  strace)  read  1  block  from  Oracle  aile  id  1,  
understand  a  few  details  about  how  time  works  in  
beginning  at  block  id  2735.  Each  call  has  three  
Oracle  trace  ailes.
parameters  like  these,  which  are  named  
differently  depending  on  the  call  name.  These   First,  a  database  call’s  total  duration  (e)  is  
three  parameters  give  you  contextual  information   approximately  the  sum  of  the  time  it  spends  using  CPU  
about  what  the  call  was  meant  to  accomplish. (c)  plus  the  duration  of  the  OS  calls  made  by  the  call.  
obj#=423 The  equation  is  only  approximate,  primarily  because  
This  call  operated  upon  the  Oracle  object  with   of  the  ±10,000  µs  on  the  CPU  time  measurement.  
id  423.  When  a  call  operated  upon  no  object  in   Expressed  as  a  precise  equality,  we  have,  for  a  given  
particular,  the  Oracle  kernel  will  emit  obj#=-1. database  call  i:
tim=1294438818791494   ei = ci + ∑ j ∈Children(i ) ela j + Δ i   (1)
This  call  completed  at  time  1,294,438,818,791,494  
±1  µs.
The  term  ∆i  there  at  the  end  is  the  time  within  the  
known  duration  of  the  call  that  is  unaccounted  for  
16 TRACE  FILE  FLOW (unexplained,  if  you  prefer)  by  Oracle.  It  can  be  
The  Oracle  kernel  emits  trace  lines  as  calls  complete.   positive  or  negative  or  zero.  As  much  as  you’d  like  for  
You  have  to  know  this  to  determine  the  parent-­‐child   this  unaccounted-­‐for  duration  to  be  zero,  it  just  won’t  
relationships  among  calls.  For  example,  what  can  you   be.  However,  in  the  aggregate,  unaccounted-­‐for  time  
know  about  the  following  (simpliaied)  trace  data   won’t  usually  be  a  material  contributor  to  total  
stream? response  time.  When  it  is,  it’s  an  indication  that:
PARSE #2:…,e=1,…,tim=5 • Either  you  have  forgotten  to  trace  the  OS  calls  (as  
EXEC #2:…,e=2,…,tim=8
with  waits=>false);
FETCH #2:…,e=4,…,tim=15
PARSE #1:…,e=15,…,tim=17
• Or  you  have  an  un-­‐patched  Oracle  bug;
From  just  the  data  shown  here,  you  can  determine  the   • Or  you  have  a  process  preemption  problem  
start  and  end  time  of  each  call,  and  from  that,  you  can   (perhaps  too  much  paging  or  swapping).
work  out  the  parent-­‐child  relationships  among  these  
Here’s  how  the  arithmetic  works  out  in  a  two-­‐line  
calls:
trace  aile  excerpt:
tim call
--- -------------- 1. WAIT #124367516: nam='Disk file operations I/O'
2 PARSE #1 { ela= 124 FileOperation=2 fileno=4 filetype=2 obj#=-1
4 PARSE #2 { tim=1294702369858445
5 } 2. FETCH #124367516:c=0,e=386,p=0,cr=6,cu=0,mis=0,r=14,
6 EXEC #2 { dep=0,og=1,plh=3956160932,tim=1294702369858666
8 }
11 FETCH #2 { Here,  you  have  an  OS  call  on  line  1  that  is  a  child  of  the  
15 } database  call  on  line  2.  You  can  prove  that  by  
17 } comparing  tim  and  duration  values:
Before  Oracle  10.2,  when  there  were  no  tim  values  on  
the  OS  call  lines,  we  inferred  call  parent-­‐child  
hierarchy  from  the  dep  values  on  database  call  lines  
and  from  the  lines’  relative  locations.  We  attributed  

©  2011  Method  R  Corporation.  All  rights  reserved.   10

 tim
----------------
call
---------------------
19 TRACE  FILE  ARITHMETIC:  PART  3
1294702369858280 FETCH #124367516 {
1294702369858321 WAIT #124367516 {
The  ainal  category  of  response  time  to  describe  is  the  
1294702369858445 } time  that  is  unaccounted-­‐for  between  the  top-­‐level  
1294702369858666 } database  and  OS  calls.  This  one  is  simple:
For  the  database  call  (a  fetch),  we  have  e  =  386,  and     Δ between = R − ∑ i ∈ TopLevelCalls ei   (3)
c  =  0.  The  sum  of  the  OS  call  durations  is  simply  124.  
Using  the  equation  above,  this  leaves  an  unaccounted-­‐
That  is,  the  total  between-­‐call  unaccounted-­‐for  
for  duration  of  ∆  =  262.  Here’s  what  the  call’s  response  
duration  is  the  total  response  time  that  we  know  has  
time  looks  like  in  the  form  of  a  proaile  table:
passed  minus  the  time  accounted  for  by  top-­‐level  calls  
Duration  (µs) Call (both  database  and  OS)  in  the  trace  aile.  So  Equation  1  
262 67.9% unaccounted-­‐for  within  calls
gives  us  most  of  the  detail  that  you  need  to  make  up  
your  proaile.  Equation  2  gives  you  the  total  response  
124 32.1% Disk  aile  operations  I/O time  that  you  need  to  account  for.  And  Equation  3  
0 0.0% CPU  service,  fetch  calls gives  you  the  label  to  which  to  assign  the  remainder  of  
386 100.0% Response  time the  unexplained  duration.  These  are  the  three  
elements  you  need  to  create  a  complete  response  time  
Exhibit  3.  Response  time  proIile  for  the  fetch   proaile.
database  call,  using  Equation  1.
20 TRACING  TRAPS
This  is  only  a  simple  example.  You  can  extend  the  
same  logic  to  database  calls  that  are  parents  of  many   When  performance  is  the  question,  tracing  is  usually  
more  OS  calls,  or  even  calls  that  are  parents  of  one  or   my  answer.  However,  it’s  not  always  easy.  Here  are  
more  levels  of  other  database  calls. some  of  the  pitfalls  to  avoid.
Data  collection.  By  far,  the  most  common  problem  we  
18 TRACE  FILE  ARITHMETIC:  PART  2 encounter  in  the  aield  is  the  trace  aile  whose  response  
The  second  relationship  you  need  to  understand  is   time  doesn’t  match  the  response  time  of  the  end  user’s  
that  the  total  response  time  described  by  a  trace  aile  is   experience  that  you’re  trying  to  diagnose.  We  get  a  lot  
simply  the  time  elapsed  from  the  beginning  of  the  airst   of  trace  ailes  whose  overall  response  times  are  
call  to  the  end  of  the  ainal  call.  Remember  that  the   dominated  by  either  a  small  number  of  long-­‐duration  
Oracle  kernel  doesn’t  emit  call  begin  times—only  call   SQL*Net  message  from  client  calls  or  a  small  number  of  
end  times.  But  you  can  compute  the  begin  time  of  a  call   unaccounted-­‐for  gaps  in  the  action.  The  most  
by  subtracting  its  duration  from  its  end  time: important  thing  to  remember  about  collecting  trace  
data  is  you  want  the  response  time  of  your  trace  aile  to  
  (
R = timend − timbegin − ebegin   ) (2) exactly  match  the  response  time  of  your  end  user’s  
experience.
Note  that  although  Equation  2  refers  to  the  duration  of  
Oracle  microseconds.  A  microsecond  is  
the  airst  call  as  e,  if  the  airst  call  in  the  aile  is  an  OS  call,  
1/1,000,000th  of  a  second,  or  1,000  nanoseconds.  It  
you’ll  need  to  use  its  ela  aield.
may  surprise  you  to  learn  that,  on  some  platforms,  an  
Be  aware  that  the  values  timbegin  and  ebegin  refer  to   Oracle  microsecond  is  1,024  nanoseconds,  or  
aields  associated  with  the  airst  call  to  begin  in  the   1/976,563th  of  a  second.  Being  off  by  2.4%  may  not  
process  being  traced.  Because  the  Oracle  kernel  emits   sound  like  too  big  of  a  problem,  but  if  you  want  to  
trace  lines  when  calls  end  (not  when  they  begin),  the   regard  an  Oracle  tim  value  as  a  number  of  
airst  call  to  begin  is  not  necessarily  the  airst  call  listed   microseconds  that  have  elapsed  since  the  Unix  epoch  
in  the  trace  aile. (1970-­‐01-­‐01T00:00:00.000000Z  in  ISO  8601  form),  
you  need  to  multiply  your  tim  value  by  1.024  before  
calculating  the  date  and  time  it  represents.  Otherwise,  
for  tim  values  generated  in  2011,  you  will  be  off  by  
almost  a  year.

©  2011  Method  R  Corporation.  All  rights  reserved.   11

Oracle  bugs.  The  Oracle  11.1  era  has  not  been  kind  to   mrskew
trace  aile  users.  Bug  7522002  causes  database  call  time   The  Method  R  skew  analyzer  is  a  data-­‐mining  tool  
values  to  run  off  a  1,024-­‐nanosecond-­‐per-­‐microsecond   for  trace  ailes.  It  can  do  nearly  everything  the  
clock  and  OS  call  time  values  to  run  off  a  1,000-­‐ Proailer  can  do,  and  more.  With  mrskew,  you  can  
nanosecond-­‐per-­‐microsecond  clock.  This  causes  the   group  by  pretty  much  anything  in  a  trace  aile  to  
sequence  of  calls  in  a  trace  aile  to  zigzag  back  and  forth   drill  with  practically  any  analysis  perspective  you  
in  time.  This  problem  is  correctible,  but  then  bug   want.  For  example,  you  can  group  a  whole  
8342329  isn’t.11  It  causes  the  Oracle  kernel  to  emit  the   directory  full  of  trace  ailes  by  SQL  statement  hash  
same  tim  value  repeatedly  for  multiple  distinct   values,  or  database  aile  ids,  or  hash  latch  address  
database  calls,  which  is  debilitating.  Happily,  these   values,  or  library  cache  miss  counts,  or  block  
bugs  in  11.1  have  patches  and  are  aixed  in  version   counts  per  read  call,  or  service-­‐module-­‐action  
11.2.0.2.0. names,  or  prepare  call  durations,  …anything  you  
like.
21 COMMERCIAL  TRACE  FILE  TOOLS
22 ACKNOWLEDGMENTS
There’s  generally  too  much  detail  in  a  trace  aile  to  look  
at  all  at  once.  To  make  sense  out  of  what’s  there   Thank  you  to  Jeff  Holt,  Ken  Ferlita,  Harold  Palacio,  and  
requires  software  assistance.  The  most  famous  trace   Ron  Crisco  at  Method  R  Corporation  for  their  help  in  
aile  software  tool  of  all  is  Oracle  Corporation’s  tkprof   preparing  this  work,  and  to  Andrew  Zitelli  of  Thales  
utility,  which  Oracle  ships  with  the  Oracle  Database   Raytheon  Company  for  sharing  his  research  with  us.
(all  releases,  all  editions).  Tkprof  aggregates  by  SQL  
statement,  and  it  allows  you  to  sort  those  SQL   23 ABOUT  THE  AUTHOR
statements  in  its  output  report  by  any  of  several  
criteria.  This  can  work  out  ok,  as  long  as: Cary  Millsap  is  the  founder  and  president  of  Method  R  
Corporation,  a  small  business  that  builds  and  
• You  care  about  inefaicient  SQL  more  than  you  care   optimizes  software  all  over  the  world  (http://method-­‐
about  response  time  (which  is  often  a  mistake). r.com).  Cary  designs  and  writes  software  and  
• You  sort  by  the  right  thing   educational  material.  He  is  the  author  of  Optimizing  
(sort=prsela,exeela,fchela  is  almost  always  the   Oracle  Performance  (O’Reilly  2003),  for  which  he  and  
right  thing). Method  R  colleague  Jeff  Holt  were  named  Oracle  
Magazine’s  Authors  of  the  Year.  He  has  presented  at  
• Your  problem  doesn’t  involve  parent-­‐child   hundreds  of  conferences  and  courses  worldwide,  and  
relationships  among  database  calls  (so-­‐called  
he  is  also  published  in  Communications  of  the  ACM.
recursive  SQL).
Cary  spent  the  1990s  as  a  consultant  and  leader  of  
I  used  tkprof  extensively  for  the  airst  ten  years  of  my  
consulting  teams  within  Oracle  Corporation.  In  1999,  
career.  However,  I  spent  a  lot  of  time  looking  at  raw  
Cary  resigned  as  vice  president  of  Oracle’s  System  
trace  data,  too,  because  there’s  a  lot  of  information  
Performance  Group  to  start  a  career  as  a  business  
inside  a  trace  aile  that  tkprof  ignores,  and  the  format  of  
owner.  Cary  is  an  Oracle  ACE  Director  and  a  founding  
the  tkprof  output  is  not  particularly  friendly  to  my  
partner  of  the  Oak  Table  Network,  an  informal  
response  time  focus.  So,  I’ve  helped  to  design  and  
association  of  “Oracle  scientists”  that  are  well  known  
write  a  few  software  tools  in  the  past  ten  years,  
throughout  the  Oracle  community.  Cary  blogs  at  
including:
http://carymillsap.blogspot.com  and  tweets  at  http://
Method  R  Proailer twitter.com/CaryMillsap.
The  Method  R  Proailer  is  software  that  Jeff  Holt  
and  I  wrote  about  in  our  book  Optimizing  Oracle  
Performance  (O’Reilly  2003).  It  produces  an  HTML  
report  that  unambiguously  explains  the  response  
time  accounted  in  a  trace  aile,  no  matter  what  the  
cause  (whether  SQL  or  not),  using  color  to  direct  
your  attention.

11  See  also  patches  9415425  for  Oracle  11.1.0.7.1  on  Solaris  and  9941786  for  Oracle  11.1.0.7.4  on  Solaris.

©  2011  Method  R  Corporation.  All  rights  reserved.   12

24 APPENDIX:  “HELLO  WORLD”  TRACE
1. Trace file /opt/oracle/diag/rdbms/v11202/V11202/trace/V11202_ora_21089.trc
2. Oracle Database 11g Enterprise Edition Release 11.2.0.2.0 - Production
3. With the Partitioning, OLAP, Data Mining and Real Application Testing options
4. ORACLE_HOME = /opt/oracle/product/11.2.0.2
5. System name: Linux
6. Node name: rhel01.dev.method-r.com
7. Release: 2.6.18-92.1.13.el5xen
8. Version: #1 SMP Thu Sep 4 04:20:55 EDT 2008
9. Machine: i686
10. VM name: Xen Version: 3.1 (PVM)
11. Instance name: V11202
12. Redo thread mounted by this instance: 1
13. Oracle process number: 30
14. Unix process pid: 21089, image: [email protected] (TNS V1-V3)
15.
16.
17. *** 2011-01-10 13:33:22.746
18. *** SESSION ID:(19.1247) 2011-01-10 13:33:22.746
19. *** CLIENT ID:() 2011-01-10 13:33:22.746
20. *** SERVICE NAME:(SYS$USERS) 2011-01-10 13:33:22.746
21. *** MODULE NAME:(SQL*Plus) 2011-01-10 13:33:22.746
22. *** ACTION NAME:() 2011-01-10 13:33:22.746
23.
24. =====================
25. PARSING IN CURSOR #7285244 len=60 dep=0 uid=5 oct=47 lid=5 tim=1294688002745892 hv=1548041990 ad='3d51df50'
sqlid='31wf1v9f4ags6'
26. BEGIN dbms_monitor.session_trace_enable(binds=>true); END;
27. END OF STMT
28. EXEC #7285244:c=0,e=845,p=0,cr=0,cu=0,mis=0,r=1,dep=0,og=1,plh=0,tim=1294688002745889
29. WAIT #7285244: nam='SQL*Net message to client' ela= 2 driver id=1650815232 #bytes=1 p3=0 obj#=-1 tim=1294688002746391
30. WAIT #7285244: nam='SQL*Net message from client' ela= 931 driver id=1650815232 #bytes=1 p3=0 obj#=-1
tim=1294688002747354
31. CLOSE #7285244:c=0,e=55,dep=0,type=0,tim=1294688002747449
32. =====================
33. PARSING IN CURSOR #7285244 len=30 dep=0 uid=5 oct=3 lid=5 tim=1294688002747560 hv=1604528588 ad='3d526cd8'
sqlid='guthms1gu6afc'
34. select 'Hello world' from dual
35. END OF STMT
36. PARSE #7285244:c=0,e=70,p=0,cr=0,cu=0,mis=0,r=0,dep=0,og=1,plh=1388734953,tim=1294688002747559
37. EXEC #7285244:c=0,e=112,p=0,cr=0,cu=0,mis=0,r=0,dep=0,og=1,plh=1388734953,tim=1294688002747793
38. WAIT #7285244: nam='SQL*Net message to client' ela= 2 driver id=1650815232 #bytes=1 p3=0 obj#=-1 tim=1294688002747865
39. FETCH #7285244:c=0,e=8,p=0,cr=0,cu=0,mis=0,r=1,dep=0,og=1,plh=1388734953,tim=1294688002747910
40. STAT #7285244 id=1 cnt=1 pid=0 pos=1 obj=0 op='FAST DUAL (cr=0 pr=0 pw=0 time=4 us cost=2 size=0 card=1)'
41. WAIT #7285244: nam='SQL*Net message from client' ela= 375 driver id=1650815232 #bytes=1 p3=0 obj#=-1
tim=1294688002748453
42. FETCH #7285244:c=0,e=1,p=0,cr=0,cu=0,mis=0,r=0,dep=0,og=0,plh=1388734953,tim=1294688002748493
43. WAIT #7285244: nam='SQL*Net message to client' ela= 1 driver id=1650815232 #bytes=1 p3=0 obj#=-1 tim=1294688002748520
44. WAIT #7285244: nam='SQL*Net message from client' ela= 1627 driver id=1650815232 #bytes=1 p3=0 obj#=-1
tim=1294688002750168
45. XCTEND rlbk=0, rd_only=1, tim=1294688002750230
46. =====================
47. PARSING IN CURSOR #7386176 len=447 dep=1 uid=0 oct=2 lid=0 tim=1294688002750383 hv=1097020010 ad='3d7e77d4'
sqlid='f711myt0q6cma'
48. insert into sys.aud$( sessionid,entryid,statement,ntimestamp#, userid,userhost,terminal,action#,returncode, logoff
$lread,logoff$pread,logoff$lwrite,logoff$dead, logoff$time,comment$text,spare1,clientid,sessioncpu,proxy$sid,user$guid,
instance#,process#,auditid,dbid) values(:1,:2,:3,SYS_EXTRACT_UTC(SYSTIMESTAMP), :4,:5,:6,:7,:8, :9,:10,:11,:12,
cast(SYS_EXTRACT_UTC(systimestamp) as date),:13,:14,:15,:16,:17,:18, :19,:20,:21,:22)
49. END OF STMT
50. PARSE #7386176:c=0,e=93,p=0,cr=0,cu=0,mis=0,r=0,dep=1,og=4,plh=0,tim=1294688002750382
51. BINDS #7386176:
52. Bind#0
53. oacdty=02 mxl=22(22) mxlc=00 mal=00 scl=00 pre=00
54. oacflg=08 fl2=0001 frm=00 csi=00 siz=24 off=0
55. kxsbbbfp=0070ac04 bln=22 avl=04 flg=05
56. value=63843
57. Bind#1
58. oacdty=02 mxl=22(22) mxlc=00 mal=00 scl=00 pre=00
59. oacflg=08 fl2=0001 frm=00 csi=00 siz=24 off=0
60. kxsbbbfp=0070abe0 bln=24 avl=02 flg=05

©  2011  Method  R  Corporation.  All  rights  reserved.   13

61. value=2
62. Bind#2
63. oacdty=02 mxl=22(22) mxlc=00 mal=00 scl=00 pre=00
64. oacflg=08 fl2=0001 frm=00 csi=00 siz=24 off=0
65. kxsbbbfp=0070abbc bln=24 avl=02 flg=05
66. value=1
67. Bind#3
68. oacdty=01 mxl=32(06) mxlc=00 mal=00 scl=00 pre=00
69. oacflg=18 fl2=0001 frm=01 csi=178 siz=32 off=0
70. kxsbbbfp=006df248 bln=32 avl=06 flg=09
71. value="SYSTEM"
72. Bind#4
73. oacdty=01 mxl=32(23) mxlc=00 mal=00 scl=00 pre=00
74. oacflg=18 fl2=0001 frm=01 csi=178 siz=32 off=0
75. kxsbbbfp=006df1c2 bln=32 avl=23 flg=09
76. value="rhel01.dev.method-r.com"
77. Bind#5
78. oacdty=01 mxl=32(05) mxlc=00 mal=00 scl=00 pre=00
79. oacflg=18 fl2=0001 frm=01 csi=178 siz=32 off=0
80. kxsbbbfp=635ffd04 bln=32 avl=05 flg=09
81. value="pts/0"
82. Bind#6
83. oacdty=02 mxl=22(22) mxlc=00 mal=00 scl=00 pre=00
84. oacflg=08 fl2=0001 frm=00 csi=00 siz=24 off=0
85. kxsbbbfp=0070ab98 bln=24 avl=03 flg=05
86. value=101
87. Bind#7
88. oacdty=02 mxl=22(22) mxlc=00 mal=00 scl=00 pre=00
89. oacflg=08 fl2=0001 frm=00 csi=00 siz=24 off=0
90. kxsbbbfp=0070ab74 bln=24 avl=01 flg=05
91. value=0
92. Bind#8
93. oacdty=02 mxl=22(22) mxlc=00 mal=00 scl=00 pre=00
94. oacflg=08 fl2=0001 frm=00 csi=00 siz=24 off=0
95. kxsbbbfp=0070ab40 bln=24 avl=02 flg=05
96. value=44
97. Bind#9
98. oacdty=02 mxl=22(22) mxlc=00 mal=00 scl=00 pre=00
99. oacflg=08 fl2=0001 frm=00 csi=00 siz=24 off=0
100. kxsbbbfp=0070ab1c bln=24 avl=01 flg=05
101. value=0
102. Bind#10
103. oacdty=02 mxl=22(22) mxlc=00 mal=00 scl=00 pre=00
104. oacflg=08 fl2=0001 frm=00 csi=00 siz=24 off=0
105. kxsbbbfp=0070aaf8 bln=24 avl=02 flg=05
106. value=4
107. Bind#11
108. oacdty=02 mxl=22(22) mxlc=00 mal=00 scl=00 pre=00
109. oacflg=08 fl2=0001 frm=00 csi=00 siz=24 off=0
110. kxsbbbfp=0070aad4 bln=24 avl=01 flg=05
111. value=0
112. Bind#12
113. oacdty=01 mxl=32(00) mxlc=00 mal=00 scl=00 pre=00
114. oacflg=18 fl2=0001 frm=01 csi=178 siz=32 off=0
115. kxsbbbfp=00000000 bln=32 avl=00 flg=09
116. Bind#13
117. oacdty=01 mxl=32(06) mxlc=00 mal=00 scl=00 pre=00
118. oacflg=18 fl2=0001 frm=01 csi=178 siz=32 off=0
119. kxsbbbfp=635ffc98 bln=32 avl=06 flg=09
120. value="oracle"
121. Bind#14
122. oacdty=01 mxl=32(00) mxlc=00 mal=00 scl=00 pre=00
123. oacflg=18 fl2=0001 frm=01 csi=178 siz=32 off=0
124. kxsbbbfp=00000000 bln=32 avl=00 flg=09
125. Bind#15
126. oacdty=02 mxl=22(22) mxlc=00 mal=00 scl=00 pre=00
127. oacflg=08 fl2=0001 frm=00 csi=00 siz=24 off=0
128. kxsbbbfp=0070aab0 bln=24 avl=02 flg=05
129. value=1
130. Bind#16
131. oacdty=02 mxl=22(00) mxlc=00 mal=00 scl=00 pre=00
132. oacflg=08 fl2=0001 frm=00 csi=00 siz=24 off=0

©  2011  Method  R  Corporation.  All  rights  reserved.   14

133. kxsbbbfp=0070aa8c bln=24 avl=00 flg=05
134. Bind#17
135. oacdty=01 mxl=32(00) mxlc=00 mal=00 scl=00 pre=00
136. oacflg=18 fl2=0001 frm=01 csi=178 siz=32 off=0
137. kxsbbbfp=00000000 bln=32 avl=00 flg=09
138. Bind#18
139. oacdty=02 mxl=22(22) mxlc=00 mal=00 scl=00 pre=00
140. oacflg=08 fl2=0001 frm=00 csi=00 siz=24 off=0
141. kxsbbbfp=0070aa58 bln=24 avl=01 flg=05
142. value=0
143. Bind#19
144. oacdty=01 mxl=32(05) mxlc=00 mal=00 scl=00 pre=00
145. oacflg=18 fl2=0001 frm=01 csi=178 siz=32 off=0
146. kxsbbbfp=bfc1b840 bln=32 avl=05 flg=09
147. value="21089"
148. Bind#20
149. oacdty=01 mxl=32(00) mxlc=00 mal=00 scl=00 pre=00
150. oacflg=18 fl2=0001 frm=01 csi=178 siz=32 off=0
151. kxsbbbfp=00000000 bln=32 avl=00 flg=09
152. Bind#21
153. oacdty=02 mxl=22(22) mxlc=00 mal=00 scl=00 pre=00
154. oacflg=08 fl2=0001 frm=00 csi=00 siz=24 off=0
155. kxsbbbfp=0070a7f8 bln=24 avl=06 flg=05
156. value=518865609
157. EXEC #7386176:c=0,e=1733,p=0,cr=1,cu=2,mis=0,r=1,dep=1,og=4,plh=0,tim=1294688002752264
158. STAT #7386176 id=1 cnt=0 pid=0 pos=1 obj=0 op='LOAD TABLE CONVENTIONAL (cr=1 pr=0 pw=0 time=155 us)'
159. CLOSE #7386176:c=0,e=5,dep=1,type=0,tim=1294688002752341
160. WAIT #0: nam='log file sync' ela= 153 buffer#=5673 sync scn=1046841 p3=0 obj#=-1 tim=1294688002752632
161. CLOSE #7285244:c=0,e=9,dep=0,type=0,tim=1294688002752690
162. CLOSE #7248564:c=0,e=30,dep=0,type=0,tim=1294688002752740

©  2011  Method  R  Corporation.  All  rights  reserved.   15