• No results found

First of all, I would like to talk about the experiences we have made with several proof- of- concepts when comparing different Oracle platform

N/A
N/A
Protected

Academic year: 2021

Share "First of all, I would like to talk about the experiences we have made with several proof- of- concepts when comparing different Oracle platform"

Copied!
21
0
0

Loading.... (view fulltext now)

Full text

(1)
(2)

2

First  of  all,  I  would  like  to  talk  about  the  experiences  we  have  made  with  several  proof-­‐

of-­‐concepts   when   comparing   different   Oracle   platform   architectures   like   the   Exadata   versus  conventional  platforms.    

(3)

There   are   several   reasons   why  WK͛Ɛ   may   become   very   time   consuming,   complicated   and  expensive:  

‡ Owing   to   security   requirements   databases   with   sensitive   data   have   to   be   masked   (Oracle  provides  an  EM  pack  for  data  masking)  

‡ It   is   often   difficult   to   simulate   OLTP   Systems   with   a   large   user   community   (Oracle   provides  also  a  tool  for  this  requirement:  real  application  testing)  

‡ The  embedding  in  a  productive  environment  like  SOA  or  the  interface  to  messaging   Systems  for  data  transfer  is  complex  

‡ Several   customers   use   engineered   system,   like   the   Exadata,   for   database   consolidation.  How  can  we  simulate  this  without  migrating  many  databases  possibly   from   different   platforms   (different   ͙   operating   Systems,   byte   order,   database   releases)  to  the  target  system?    

‡ The  experiences  of  WK͛Ɛ  show,  that  you  have  to  invest  at  least  50+  labor  days.  And   the  whole  POC  will  probably  last  some  months.  

‡ In   order   to   execute   the   POC   one   have   to   travel   to   a   benchmark   center,   or   to   the   benchmark   center   of   selected   Oracle   partners.   In   both   cases   one   will   get   only   a   limited  time  window  for  tests.  

If   a   POC   is   run   with   a   single   application,   this   application   must   be   highly   scalable   to   challenge  the  performance  capabilities  of  a  new  platform  architecture.  

What   we   also   have   seen   is,   that   application   change   their   behavior   when   moving   to   a   system   with   complete  new   architecture   like   the   Exadata.   If  an   application  was   former   I/O-­‐bound,  it  may  now  becoming  more  CPU-­‐bound,  because  the  I/O  subsystem  delivers   much   better   I/O   performance.   Or   the   system   can   not   be   utilized,   because   the   application  is  not  designed  to  do  so.  

(4)

And   the   results   are   ambiguous,   because   an   application   changes   over   time   and   therefore   the   results  would  be  different  with  the  next  release  of  the  application  software.  The  POC  may  still  not   deliver  any  performance  data  suitable  for  capacity  planning  and  database  consolidation.  

And,  it  is  important  to  keep  in  mind:  a  POC  is  NOT  reality,  but  some  level  of  modeling  reality.  Like   each  benchmark.    

(5)

>Ğƚ͛Ɛ  take  a  look  at  other  industries  and  how  they  work  with  performance  numbers.  For   example  the  automobile  industry.    

Have  you  ever  seen  a  car  driver  running  a  proof-­‐of-­‐concept  to  validate  the  performance   data  of  his  car?  

(6)

The   automobile   industry   publishes   key   performance   metrics   for   cars,   like   speed,   acceleration,  horsepower  and  torque.    

If   I   ask   you,   how   fast   is   your   car,   you   will   probably   be   able   to   answer   my   question   immediately.   But   if   I   ask   you   about   the   key   performance   metrics   of   your   database   platform,  it  will  probably  be  more  difficult.    

Why?    

There  are  two  reasons:    

‡ First,  there  are  no  useful  standards  for  database  platform  key  performance  metrics  

‡ Second,  most  people  do  not  measure  the  performance  of  their  database  platforms   unless  users  complain  about  bad  performance  or  management  complains  about  high   cost  

(7)

6

Why   is   it   so   difficult   to   obtain   key   performance   metrics   of   our   database   platforms?   It  

(8)

Due  to  the  huge  amount  of  combinations  

‡ Components  

‡ Configurations  

‡ Architectures  (e.g.  for  HA  and  DR)  

there  are  probably  no  two  identical  Oracle  platforms  -­‐  even  in  the  same  company.  The   performance  of  such  an  complex  Oracle  platform  is  not  predictable.  Benchmarking  is   the  only  way  to  predict  performance  of  an  Oracle  database  platform.  

   

(9)

Lets  summarize  so  far:    

1. The  performance  of  a  database  platform  is  not  predictable  due  to  its  complexity.   2. Benchmarking  is  the  only  way  to  determine  the  key  performance  metrics.  

(10)

9

The   Transaction   Processing   Council   has   published   standards   for   database   platform  

benchmarks.   Unfortunately   these   benchmarks   are  only   used   by   vendors.   I   have   never   seen   a   customer   running   a   TPC   benchmark   to   certify   the   performance   or   to   evaluate   components  of  his  database  platform.  

(11)

Industry  benchmarks  like  TPC  are  only  used  by  IT  vendors,  but  not  by  IT  customers  because   they  are  unpractical.  

Certain   rules   have   to   be   followed   to   set   up   the   database   in   terms   of   sizing   and   usable   features.   But   a   customer   would   like   to   use   ALL   database   features   which   help   to   increase   overall  efficiency  and  to  get  the  most  performance  for  each  license  dollar.  A  benchmark  has   to  follow  customer  needs,  not  the  other  way  around.    

The   benchmark   must   be   operational   on   each   customer   system   with   any   database   size,   because  the  customer  wants  to  know  the  key  performance  metrics  of  his  platform.  

If  you  take  a  detailed  look  at  the  full  disclosure  report  of  some  of  these  benchmarks,  you   will   find   out,   that   for   a   10   TByte   database   about   3͛000   disks   are   used.   Completely   unrealistic.  Even  the  largest  banks  in  Switzerland  use  by  far  less  disks  for  a  single  database.   Take  a  look  at  the  TPC-­‐H  performance  metric.  Its  just  one  number.  Assume  the  value  is  17.   What  do  you  know  about  the  power  of  the  Wh͛Ɛ  from  this  value?  How  much  data  can  be   loaded  in  a  given  time  period?  Or  how  much  data  can  be  scanned  or  backed  up  in  a  given   time  window?  Does  this  value  of  17  help  anybody  in  the  project  team  to  size  the  system?     We   have  a   very   similar   situation  with  TPC-­‐C  benchmarks.   TPC-­‐C   measures   the   transaction   rate   at   highest   system   load   (>   95%   CPU   utilization).   But   this   is   a   situation   every   system   administrator   wants   to   avoid   in   OLTP   Systems.   This   high   load   is   an   exception   in   OLTP   Systems,  not  the  rule.    

Other  topics  concerning  TPC  benchmarks:  

‡ No  values  about  best-­‐case  behavior  (cache)  and  worst-­‐case  behavior  (disk)  

‡ No   values   about   system   behavior   with   increasing   load   (from   1   process   to   n   processes/system  saturation)  

 

(12)
(13)

Because  there  is  no  useful  benchmark  method  around,  during  the  past  few  years  we   developed  our  own  benchmark  method  and  benchmark  tool,  the  benchware  Loader.   We  are  measuring  Oracle  platform  performance  right  above  the  database  layer.  At  this   interface   we   identify   the   performance   of   the   complete   Oracle   platform.   The   key   performance   metrics   of   an   Oracle   platform   can   be   used   as   a   foundation   for   service   level  agreements  between  IT  operations  and  the  business  applications.  

     

(14)

The   Benchware   Loader   evaluates   the   key   performance   metrics   of   CPU   and   server   system,  storage  system  and  database  system.  The  benchmark  tests  are  representative   for   OLTP   and   DWH   systems.   They   analyze   system   behavior   for   all   major   database   operations   in   different   load   situations   from   a   single   process   up   to   complete   system   saturation.  

The   Benchware   Loader   examines   both   best   and   worst   case   performance.   All   components   of   the   platform   (server,   storage   and   database   system)   are   checked   for   their   performance   characteristics   and   limitations.   Bottlenecks   will   be   detected   immediately.    

The   Benchware   Loader   delivers   simple   and   understandable   key   performance   metrics,   this   is   an   important   difference   to   TPC   benchmarks.   The   key   performance   metrics   are   the  foundation  for  validating  the  chosen  system  architecture  and  for  capacity  planning   when  upgrading  systems  or  consolidating  servers  and  databases.  Performance  metrics   are   also   utilized,   taking   into   consideration   Oracle   license   costs,   to   conduct   price-­‐ performance  platform  comparisons.  

The   Benchware   Loader,   written   in   SQL,   PL/SQL   and   Java,   runs   on   all   Oracle   platforms   and  scales  from  small  server  to  very  large  high-­‐end  systems.  It  uses  synthetic  data  and   allows   benchmarks   to   be   repeated   and   compared   at   any   time   to   verify   the   impact   of   system  changes  on  overall  application  performance.    

 

(15)

The   Benchware   Loader   evaluates   the   key   performance   metrics   of   CPU   and   server   system,  storage  system  and  database  system.  The  benchmark  tests  are  representative   for   OLTP   and   DWH   systems.   They   analyze   system   behavior   for   all   major   database   operations   in   different   load   situations   from   a   single   process   up   to   complete   system   saturation.  

The   Benchware   Loader   examines   both   best   and   worst   case   performance.   All   components   of   the   platform   (server,   storage   and   database   system)   are   checked   for   their   performance   characteristics   and   limitations.   Bottlenecks   will   be   detected   immediately.    

The   Benchware   Loader   delivers   simple   and   understandable   key   performance   metrics,   this   is   an   important   difference   to   TPC   benchmarks.   The   key   performance   metrics   are   the  foundation  for  validating  the  chosen  system  architecture  and  for  capacity  planning   when  upgrading  systems  or  consolidating  servers  and  databases.  Performance  metrics   are   also   utilized,   taking   into   consideration   Oracle   license   costs,   to   conduct   price-­‐ performance  platform  comparisons.  

The   Benchware   Loader,   written   in   SQL,   PL/SQL   and   Java,   runs   on   all   Oracle   platforms   and  scales  from  small  server  to  very  large  high-­‐end  systems.  It  uses  synthetic  data  and   allows   benchmarks   to   be   repeated   and   compared   at   any   time   to   verify   the   impact   of   system  changes  on  overall  application  performance.    

 

(16)

Customers  are  using  Benchware  Loader  in  different  situations,  here  some  examples:  

‡ Testing   existing   architectures   (health   check)   to   identify   current   performance   bottlenecks  or  to  certify  platform  performance  capabilities  (usable  for  service  level   agreements  between  IT  and  Business)  

‡ Optimizing   storage   performance   by   comparing   different   product,   different   configurations,  different  architectures,  etc.    

‡ Evaluating  different  CPU  types  to  optimize  Oracle  license  cost  

‡ Comparing   the   impact   of   different   disaster   recovery   technologies   to   application   performance  

 

(17)
(18)
(19)

Based   on   benchmark   results,   Benchware   provides   a   sound   foundation   for   decisions:   components,  configurations,  architectures  and  complete  systems.  

 

In   an   evaluation  of   a   new  system   you   have  always  the  old   system   (blue)  with  relative   cost  of  1.0  and  relative  performance  of  1.0.  The  new  system  (red)  needs  an  investment,   in  this  case  factor  0.6  of  the  old  one  (blue),  and  provides  a  performance  improvement   of  factor  2.5,  based  on  key  performance  metrics.  

(20)

The  Benchware  Loader  has  several  value  propositions,  it  is  

‡ Comprehensive  ʹ  benchmarks  all  basic  database  operations  

‡ Simple  to  install  and  to  use    

‡ Fast  to  run,  within  a  few  hours  

‡ Scalable  -­‐  from  laptop  to  high-­‐end  system  

‡ Reproducible  ʹ  uses  its  own  data    

‡ Provides  understandable  performance  metrics  

‡ Uses   and   supports   important   Oracle   features   like   RAC,   Data   Guard,   Flash   Cache,   Exadata,  ͙    

(21)

Benchware   Ltd.   is   vendor   independent.   Benchware  does   not   sale   any  products  

from   other   vendors   nor   get   commission   from   any   vendor.   Benchware   is  

completely  committed  to  customers  interests.    

Benchware  offers  services  and  products  for  

‡

Performance  Analysis  &  Optimization  

‡

Benchmarking    

‡

System  Architecture  &  Component  Evaluation  of  Oracle  platforms  

Benchware  consultants  have  a  long  track  of  experience  in  designing,  tuning  and  

operating  large  OLTP  and  DWH  systems  with  high  requirements  to  performance,  

scalability   and   availability,   working   with   Oracle   database   software   since   1984  

(Oracle  Version  3.1)  and  running  benchmarks  und  tuning  applications  since  1993  

(Oracle  Version  7.1).    

 

Benchware  Ltd.  

Seestrasse  18  

CH-­‐8800  Thalwil  (Switzerland)  

Phone    +41  44  722  16  16  

Mail      [email protected]  

Web      www.benchware.ch  

20

References

Related documents

Deneux-Tharaux et al 7 concluded that when considering the mode of delivery, obstetricians and women should be cognisant of the increased risk of postpartum

This was calculated by the cost differential between the length of the transfer and the start of care in the subsequent designated place of care (admission area, labour ward,

Keywords: Augmentative and alternative communication, natural language processing, compansion (compression-expansion), telegraphic language, Catalan,

Insurance Business Objects Individual Product Lead Application Policy Commission Contract Claim Payment Invoice Sub-ledger General Accounts Treaty Illustration Prospect

He has over 30 years of management and leadership experience in the oil and gas industry, with strong customer orientation skills and ability to harness value from

The quantitative study involved an examination of MHL using novel online survey methodology and was focused on four disorders that often first occur or worsen during

The Clinical Practice Research Datalink (CPRD) is an ongoing primary care database of ano- nymised medical records from general practitioners, with coverage of over 11.3