Database Death Match

Oracle
vs. 
 
 
 
 
 

SQL
Server


®
 Jeremiah
Wilton









[email protected]

•  17
years
working
with
Oracle


•  Amazon.com’s
first
DBA
~1997


•  Technical
Lead,
Blue
Gecko,
Inc.

–  Remote
administraNon,
support
and
experNse

–  Oracle,
E‐Business
Suite,
MySQL,
PostgreSQL,
SQL
Server


•  Univ.
of
Washington
Oracle
instructor


•  President,
Western
Washington
Oracle
Users
Group


•  Co‐author:
Expert
Oracle
PracNces
(Oak
Table)


•  Frequent
speaker
at
Oracle
conferences


•  Current
focus:

–  Oracle
performance
on
the
Cloud
(Amazon
EC2)

–  Time‐based
tuning
for
SQL
Server

®
 RaNonale

•  Fun

•  InteresNng

•  Not
boring

•  Learn
about
SQL
Server

•  Break
things

•  Feel
superior

•  Engage
in
pretend
science
(à
la
Mythbusters)


®
 Caveats

•  Oracle
and
SQL
Server
are
not
the
same

•  I
am
not
a
Windows
or
SQL
Server
expert

•  My
tests
may
have
a
pro‐Oracle
bias

•  I
used
a
minimum
configuraNon
and

engineering
standard

–  No
standby,
replicaNon,
log
shipping,
etc.

–  Basic
out‐of‐box
funcNonality


®
 Method

Windows
Server

2008
R2
64‐bit

Windows
Server


2008
R2
32‐bit

Oracle
Driver
 Oracle

11gR2
64‐bit

Enterprise

Dell
DVD
Store
2.0

SQL
Server
2008

64‐bit

SQL
Server
Driver
 Enterprise


Amazon
Elas=c
Compute
Cloud


®
 Tools
I
Use

•  GNU
Core
UNliNes

–  tail

–  dd

–  sed

•  Microsoe
SysInternals
Process
Explorer

•  sqlcmd

•  sqlplus


®
 The
Four
Ordeals

1.  Loss
of
volume
containing
transacNon
log


2.  Hang/crash
of
data
file
writer
thread


3.  Corrupt
a
single
8k
block


4.  Side‐by‐side
compeNNon
for
resources


®
 SQL
Server
TransacNon
Log

•  Similar
to
Oracle’s
redo
log
(circular)

•  Also
provides
undo
funcNon

•  One
per
database
(mulNple
databases
per
server)

•  Usually
one
giant
file,
n
‘virtual’
logs
within
the
one
physical
file

•  Regular
T‐Log
backups
provide
“archive”
funcNonality

•  If
no
backup
before
T‐log
is
full,
file
grows
(Full
recovery
model)

•  No
na=ve
log
redundancy
mechanism


®
 SQL
Server:
Loss
of
log
volume

•  Each
contender
has
a
dedicated
volume
for
logs


SQL
Server

TransacNon
Log

Database

Data
File

Server

Storage
(D:\)

Oracle

Redo
Logs


1.  Apply
load
with
DS2

ds2sqlserverdriver
‐‐target=dm
‐‐n_threads=5
‐‐pct_newcustomers=0

2.  Tail
the
error
log

tail
‐f
"c:\Program
Files\Microsoft
SQL
Server\MSSQL10.MSSQLSERVER\MSSQL\Log\ERRORLOG"







|
sed
s/\x00//g

3.  Detach
log
storage
volume

4.  Observe

®
 Oracle
Redo
Logs

•  “Oracle
recommends
that
you
mulNplex
the

online
redo
log.”
‐11gR2
Docs
(standard

pracNce)


®
 Oracle:
Loss
of
log
volume

•  Each
contender
has
a
dedicated
volume
for
logs

Database
 SQL
Server

Server
 TransacNon
Log

Oracle
Log

Group
B

Oracle
Log

Oracle

Group
A


1.  Apply
load
with
DS2



c:\ds2\ds2oracledriver
‐‐target=dm
‐‐n_threads=5
‐‐pct_newcustomers=0
2.  Tail
the
alert
log

tail
‐f
c:\app\Administrator\diag\rdbms\dm\dm\trace\alert_dm.log

3.  Detach
log
group
A
storage
volume

4.  Observe

5.  Drop
missing
logfile
group
members

®
 SQL
Server
Lazy
Writer

•  Not
the
same
as
the
Oracle
Database
Writer

•  Writes
some
dirty
buffers
between
checkpoints

•  Worker
threads
also
write
dirty
buffers

•  Checkpoint
also
writes
dirty
buffers
(at
checkpoint)

•  How
to
find
it?
 session_id
command









os_thread_id

‐‐‐‐‐‐‐‐‐‐
‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐
‐‐‐‐‐‐‐‐‐‐‐‐


1









RESOURCE
MONITOR








2776


2









XE
TIMER
















2908

1>
select
session_id,
command,
os_thread_id

 3









XE
DISPATCHER











2416

2>
from
sys.dm_exec_requests
as
r

 4









LAZY
WRITER













1316

5









LOG
WRITER














2404

3>
join
sys.dm_os_workers
as
w


6









LOCK
MONITOR












1592

4>


on
r.task_address
=
w.task_address

 7









SIGNAL
HANDLER










1516

5>
join
sys.dm_os_threads
as
t

 9









TRACE
QUEUE
TASK








1544

6>


on
t.thread_address
=
w.thread_address

 10








BRKR
TASK















1540

7>
where
session_id
<=
50

 11








TASK
MANAGER












1524

8>
order
by
session_id

 13








CHECKPOINT














1520

14








BRKR
EVENT
HNDLR








1556


9>
go

15








BRKR
TASK















1552

16








BRKR
TASK















1536


®
SysInternals
Process
Explorer


®
 SQL
Server:
Hang/Crash
Lazy
Writer

1.  Apply
load
with
DS2

2.  Tail
the
error
log

3.  Suspend
the
Lazy
Writer
thread
(procexp)

4.  Observe

5.  Resume
the
Lazy
Writer
thread
(procexp)

6.  Kill
the
Lazy
Writer
thread
(procexp)

7.  Observe


®
 Oracle
Database
Writer(s)

•  Responsible
for
all
 SPID
NAME
DESCRIPTION

‐‐‐‐
‐‐‐‐
‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐


writes
of
dirty
blocks
 2116
PMON
process
cleanup

2120
VKTM
Virtual
Keeper
of
TiMe
process

2124
GEN0
generic0

2144
DIAG
diagnosibility
process

2148
DBRM
DataBase
Resource
Manager


•  How
to
find
it
in
 2160
PSP0
process
spawner
0

2164
DIA0
diagnosibility
process
0


Windows?
 2168
MMAN
Memory
Manager

2172
DBW0
db
writer
process
0

2176
LGWR
Redo
etc.

2180
CKPT
checkpoint

2184
SMON
System
Monitor
Process

2188
RECO
distributed
recovery

2192
MMON
Manageability
Monitor
Process

SQL>
select
spid,
b.name,

 2200
MMNL
Manageability
Monitor
Process
2

2372
RVWR
Recovery
Writer



2








description
 3972
QMNC
AQ
Coordinator



3

from
v$process
p,
 2488
ARC1
Archival
Process
1

2440
ARC0
Archival
Process
0



4






v$bgprocess
b
 2500
ARC2
Archival
Process
2



5

where
p.addr
=
b.paddr
 2572
ARC3
Archival
Process
3

3960
CJQ0
Job
Queue
Coordinator



6

and
p.serial#
=
b.pserial#;
3896
SMCO
Space
Manager
Process


®
 Oracle:
Hang/Crash
Database
Writer

1.  Apply
load
with
DS2

2.  Tail
the
alert
log

3.  Suspend
the
DBW0
thread

4.  Observe

5.  Resume
the
DBW0
thread

6.  Kill
the
DBW0
thread

7.  Observe


®
 Single‐Block
CorrupNon

•  SQL
Server
calls
them
pages

•  Oracle
calls
them
blocks

•  Both
are
8k
by
default


1.  Find
the
first
block
of
the
CUSTOMERS
table

2.  SQL
Server
and
Oracle
“trade”
the
block

3.  Try
to
read
the
data

4.  Try
to
recover
the
block


®
 Physical
Address
of
a
SQL
Server
Row

•  %%PHYSLOC%%
is
the
equivalent
of
ROWID

–  File:Block:Slot
(sound
familiar?)

1>
select
top
10

 
physloc


customerid

firstname

2>


sys.fn_physlocformatter(%%PHYSLOC%%)


‐‐‐‐‐‐‐‐‐
‐‐‐‐‐‐‐‐‐‐‐
‐‐‐‐‐‐‐‐

3>


customerid,
firstname
 
(4:896:0)










1
VKUUXF

4>
from
ds2.dbo.customers
 
(4:896:1)










2
HQNMZH

5>
go
 
(4:896:2)










3
JTNRNB


(4:896:3)










4
XMFYXD

1>
select
physical_name
 
(4:896:4)










5
PGDTDU

2>
from
ds2.sys.database_files
 
(4:896:5)










6
FXDZBW

3>
where
file_id
=
4
 
(4:896:6)










7
WVZTXZ

4>
go
 
(4:896:7)










8
LIWLAI


(4:896:8)










9
NCGWRC

physical_name
 
(4:896:9)









10
FUOHXX

‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐

C:\sql\dbfiles\ds2_2.ndf

®
 Physical
Address
of
an
Oracle
Row

•  We
use
ROWID
/
dbms_rowid

SQL>
select
 
FILE
BLOCK
CUSTOMERID
FIRSTNAME


2


dbms_rowid.rowid_relative_fno(rowid),
 
‐‐‐‐
‐‐‐‐‐
‐‐‐‐‐‐‐‐‐‐
‐‐‐‐‐‐‐‐‐


3


dbms_rowid.rowid_block_number(rowid),




6


131









1
VKUUXF


4


customerid,
firstname
 



6


131









2
HQNMZH


5


from
ds2.customers
 



6


131









3
JTNRNB


6


where
rownum
<=
10;
 



6


131









4
XMFYXD





6


131









5
PGDTDU

SQL>
select
name
 



6


131









6
FXDZBW


2


from
v$datafile
 



6


131









7
WVZTXZ


3


where
file#
=
6;
 



6


131









8
LIWLAI





6


131









9
NCGWRC

NAME
 



6


131








10
FUOHXX

‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐‐

D:\APP\ADMINISTRATOR\ORADATA\DM\CUST.DBF


®
 Swap
the
first
block
of
CUSTOMERS

1>
alter
database
ds2
set
offline

2>
go


C:\>dd
bs=8k
skip=896
count=1



if=d:\sql\dbfiles\cust1.ndf
of=c:\sb1

C:\>dd
bs=8k
skip=131
count=1



if=d:\app\administrator\oradata\dm\cust.dbf
of=c:\ob1


C:\>dd
bs=8k
conv=notrunc
seek=896



if=c:\ob1
of=d:\sql\dbfiles\cust1.ndf

C:\>dd
bs=8k
conv=notrunc
seek=131



if=c:\sb1
of=d:\app\administrator\oradata\dm\cust.dbf


1>
alter
database
ds2
set
online

2>
go


®
 SQL
Server
&
Oracle:

Read
Corrupted
Page

1>
select
top
10
firstname

2>
from
ds2.dbo.customers

3>
go

Msg
824,
Level
24,
State
2,
Server
IP‐0AF647E0,
Line
1


SQL
Server
detected
a
logical
consistency‐based
I/O
error:
incorrect

pageid
(expected
4:896;
actual
3:0).
It
occurred
during
a
read
of
page

(4:896)
in
database
ID
5
at
offset
0x00000000700000
in
file
'C:\sql
\dbfiles\cust1.ndf'.

Additional
messages
in
the
SQL
Server
error
log
or
system
event
log
may

provide
more
detail.


SQL>
alter
system
flush
buffer_cache;

SQL>
select
firstname



2

from
ds2.customers



3

where
rownum
<=
10;

*

ERROR
at
line
1:

ORA‐01578:
ORACLE
data
block
corrupted
(file
#
6,
block
#
131)

ORA‐01110:
data
file
6:
'D:\APP\ADMINISTRATOR\ORADATA\DM\CUST.DBF'


®
SQL
Server:
Recovery
of
Corrupt
Block

1>
backup
log
ds2


2>
to
disk='c:\bak\log.bak’

3>
go


1>
restore
database
ds2
page='4:896’


2>
from
disk='c:\bak\db.bak'
with
norecovery

3>
go


1>
restore
log
ds2


2>
from
disk='c:\bak\db.bak'
with
norecovery

3>
go


1>
restore
log
ds2


2>
from
disk='c:\bak\log.bak'
with
recovery

3>
go


1>
alter
database
ds2
set
online

2>
go

®
 Oracle:
Recovery
of
Corrupt
Block

•  Oracle
has
had
blockrecover
since
9i

•  One
line
does
it
all!

RMAN>
blockrecover
datafile
6
block
131;


®
 THE
FINAL
ORDEAL

•  Side‐by‐side
compeNNon
for
resources


1.  Apply
load
to
both
Oracle
and
SQL
Server

2.  Verify
100%
CPU
uNlizaNon

3.  Observe
TPM
numbers


®
Ques=ons?