SlideShare a Scribd company logo
Total Cost of Ownership
Zsolt Fabian
Database Engineering
2	
  
1.  Who	
  is	
  Spil	
  Games?	
  
2.  What	
  is	
  TCO?	
  
3.  What	
  drives	
  the	
  costs?	
  
4.  Possible	
  improvements?	
  
Outline of the talk
Who are we?
Who	
  is	
  Spil	
  Games?	
  
	
  
4	
  
•  Company	
  founded	
  in	
  2001	
  
•  350+	
  employees	
  world	
  wide	
  
•  200M+	
  unique	
  visitors	
  per	
  month	
  
•  45	
  portals	
  in	
  19	
  languages	
  
•  Casual	
  games	
  
•  Social	
  games	
  
•  Real	
  Kme	
  mulKplayer	
  games	
  
•  Mobile	
  games	
  
•  35+	
  MySQL	
  clusters	
  
Facts
TCO
Total	
  Cost	
  of	
  Ownership	
  
	
  
6	
  
•  1980	
  –	
  “Life	
  Cycle	
  CosKng”	
  
•  1987	
  –	
  Gartner	
  Group	
  
•  Total	
  Cost	
  of	
  Ownership	
  
•  Analysis	
  of	
  enterprise	
  spending,	
  vendor	
  evaluaKon	
  
•  Many	
  inter-­‐related	
  direct	
  and	
  indirect	
  costs	
  
•  Problems	
  with	
  TCO	
  
What is TCO and why it is important?
CAPEX	
  –	
  Capital	
  Expenses	
   OPEX	
  –	
  OperaKng	
  Expenses	
  
Relevant Costs
7	
  
Infrastructure life-cycle
8	
  
•  CollecKng	
  data	
  
•  Different	
  base	
  systems	
  
•  4,3	
  TB	
  of	
  RAM	
  
•  402	
  drives	
  wriKng	
  3	
  ExaBytes	
  
•  3	
  DBA	
  
•  12,3	
  kW	
  power	
  consumpKon	
  
•  Asset	
  value	
  exceeds	
  €	
  500,000	
  
•  Cables	
  are	
  available	
  in	
  all	
  colors	
  of	
  the	
  spectrum	
  J	
  	
  
A quick inventory
9	
  
Aging servers
•  Component	
  failures	
  
•  Power	
  consumpKon	
  
•  Volume	
  of	
  deployment	
  
10	
  
•  The	
  most	
  obvious	
  
•  Base	
  server	
  +	
  CPU	
  +	
  Mem	
  +	
  Disk	
  +	
  Controller	
  
•  Replacement	
  parts	
  and	
  Warranty	
  
•  NPV	
  –	
  Net	
  Present	
  Value	
  
•  Mining	
  invoices	
  
CAPEX: Purchase costs
AFR	
  –	
  Annual	
  Failure	
  Rate	
   ALE	
  –	
  Annual	
  Loss	
  Expectancy	
  
11	
  
•  Business	
  need	
  
•  Cost	
  of	
  downKme	
  
•  Levels	
  of	
  HA	
  and	
  Fault	
  Tolerance	
  
•  Server	
  downKme	
  
•  CPU	
  failure	
  
•  MySQLd	
  crash	
  
•  Solar	
  acKvity	
  
•  Combined	
  availability	
  formula:	
  
CAPEX: Purchase costs: HA
A	
  =	
  1	
  –	
  (1	
  –	
  Ax)	
  ^	
  n	
  
12	
  
•  Giving	
  back	
  J	
  
•  MySQL	
  Standard	
  EdiKon:	
  $	
  2,000	
  
•  Enterprise:	
  $	
  5,000	
  
•  Cluster	
  Carrier	
  Grade:	
  $	
  10,000	
  
•  Microsoj	
  SQL	
  Server:	
  $	
  6k+	
  /	
  core	
  	
  
•  Oracle	
  10/11g:	
  $	
  40k+	
  
•  AddiKonal:	
  Technology	
  extensions,	
  Dashboards	
  
CAPEX: License costs
13	
  
•  Work	
  hours	
  spent…	
  
•  SLA,	
  Responsibility,	
  Stress,	
  Insider	
  Threat	
  
•  MySQL	
  annual	
  subscripKon:	
  $	
  2,000	
  
•  Specialist	
  support	
  from	
  $	
  1,500	
  /	
  server	
  /	
  year	
  
•  Unlimited	
  servers	
  for	
  $	
  30,000	
  /	
  year	
  
•  Depends	
  on	
  complexity	
  
•  €	
  2,700	
  /	
  server	
  /	
  year	
  
	
  
OPEX: MySQL Professional support
It's	
  cheaper	
  to	
  add	
  more	
  servers	
  than	
  hire	
  more	
  engineers.	
  
14	
  
•  Cage	
  –	
  Rack	
  –	
  Server	
  
•  Networking	
  
•  Deployment	
  (+	
  fuel	
  +	
  Kme)	
  
•  Power	
  density	
  constraints	
  
•  Let’s	
  say	
  €	
  60	
  /	
  server	
  /	
  month	
  
OPEX: Hosting: Colocation
15	
  
•  Wao	
  =	
  Ampere	
  *	
  Volt	
  
•  €	
  0.2	
  /	
  kWh	
  
•  12,31	
  kWh	
  (~	
  16%	
  of	
  total)	
  
•  €	
  21,300	
  annually	
  
•  Measurements	
  
•  IPMI	
  
•  Power	
  meter	
  J	
  
•  Contributors	
  
•  HDD	
  (25%)	
  
•  Roles	
  (5-­‐10W)	
  
OPEX: Hosting: Power
16	
  
OPEX: Hosting: Power
Putting it together
Pu@ng	
  it	
  together	
  	
   The	
  model	
  	
  
Cost	
  Item	
   CAPEX	
   OPEX	
   TCO	
  5y	
  
Base	
  System	
   ?	
  
Extra	
  memory	
   ?	
  
Extra	
  disks	
   ?	
  
Licence	
  fees	
   ?	
  
Replacement	
  parts	
  /	
  
Warranty	
  
?	
   ?	
  
Co-­‐locaKon	
   ?	
  
Power	
   ?	
  
Professional	
  support	
   ?	
  
TOTAL	
   ?	
   ?	
   ?	
  
Pu@ng	
  it	
  together	
  	
   Cluster	
  of	
  6	
  nodes	
  
Cost	
  Item	
  (6x)	
   CAPEX	
   OPEX	
   TCO	
  5y	
  
Purchase:	
  HP	
  DL380	
  G7	
   €	
  25,000	
  
Purchase:	
  Extra	
  memory	
  	
   €	
  6,000	
  
Professional	
  support	
   €	
  16,200	
  
ColocaKon	
  cost	
   €	
  4,320	
  
Power	
  costs	
  @200W	
   €	
  2,100	
  
Replacement	
  parts	
  
	
  
€	
  1,500	
  
TOTAL	
   €	
  32,500	
   €	
  22,620	
   €	
  145,600	
  
Pu@ng	
  it	
  together	
  	
   Cluster	
  of	
  6	
  nodes	
  
Purchase	
  
costs	
  
18%	
  
Professional	
  
support	
  
58%	
  
ColocaKon	
  
15%	
  
Power	
  
8%	
  
Replacement	
  
parts	
  
1%	
  
TCO 5 years
Improvements
“Insanity:	
  doing	
  the	
  same	
  
thing	
  over	
  and	
  over	
  again	
  and	
  	
  
	
  expecSng	
  different	
  results”	
  -­‐	
  
Albert	
  Einstein	
  
	
  
22	
  
Purchase prioritization
Extend	
  warranty	
  
MTBF	
  
Obsolete,	
  Not	
  Urgent	
  
Limited	
  reparability	
  
Repairable	
  
New	
  
23	
  
•  SSD	
  reduces	
  the	
  cost	
  of	
  IOPS	
  
•  SSD	
  is	
  fast	
  for	
  reads,	
  but	
  not	
  for	
  writes	
  
•  Has	
  limited	
  lifespan	
  
•  NAND	
  Flash	
  and	
  RAM	
  based	
  SSDs	
  
•  MLC	
  (MulK	
  Level	
  Cell)	
  vs.	
  SLC	
  (Single	
  Level	
  Cell)	
  
•  R/W	
  in	
  pages,	
  but	
  always	
  erased	
  in	
  blocks	
  
•  The	
  main	
  advantage	
  that	
  it	
  doesn’t	
  have	
  to	
  spin	
  
HDD vs. SSD: Inner workings
24	
  
•  Dropping	
  over	
  Kme	
  
•  Facts	
  
•  Tsunami	
  
•  Rare-­‐earth	
  minerals	
  
•  Cost	
  per	
  GB	
  is	
  cheaper	
  at	
  HDD	
  
•  Some	
  predicts	
  that	
  in	
  5	
  to	
  10	
  years,	
  SSD	
  will	
  replace	
  
HDD	
  completely	
  
HDD vs. SSD: Pricing
25	
  
HDD vs. SSD: Power consumption
Type	
   Model	
   Idle	
  Power	
   AcSve	
  Power	
   Annual	
  cost	
  
contribuSon	
  
SSD	
   Intel	
  520	
   600	
  mW	
   850	
  mW	
   €	
  1,49	
  
SSD	
   Intel	
  X25-­‐M	
   60	
  mW	
   150	
  mW	
   €	
  0,26	
  
SSD	
   Intel	
  X25-­‐E	
   60	
  mW	
   4,2	
  W	
   €	
  7,35	
  
HDD	
   ST9146852S	
   5	
  W	
   7	
  W	
   €	
  12,2	
  
26	
  
•  MTBF:	
  ~	
  200	
  years	
  
•  HDD	
  
•  AFR:	
  0,55%	
  	
  (-­‐	
  1,5%	
  -­‐	
  2%	
  -­‐	
  …)	
  
•  ALE!	
  
•  SSD	
  have	
  LIMITED	
  write	
  cycles	
  
•  SLC:	
  100,000	
  writes	
  /	
  block	
  
•  MLC:	
  10,000	
  writes	
  /	
  block	
  
•  Typical	
  MLC:	
  5,000	
  writes	
  /	
  block	
  
•  Future	
  developments	
  
HDD vs. SSD: Endurance
27	
  
•  Intel	
  520	
  SSD:	
  5	
  years	
  (with	
  normal	
  desktop	
  usage)	
  
•  35	
  TB	
  write	
  endurance	
  (TBW)	
  
•  Official	
  endurance	
  is	
  overprovisioned	
  (~	
  50	
  %)	
  
•  SaturaKon	
  
•  Filesystem	
  overhead	
  (0.25	
  btrfs-­‐>	
  0.015	
  (ext4))	
  
•  Readout	
  
•  S.M.A.R.T.:	
  Percentage	
  of	
  rated	
  lifeKme	
  used	
  
•  UKliKes	
  
•  RAID	
  controllers	
  
HDD vs. SSD: Endurance
28	
  
•  Determined	
  during	
  the	
  design	
  
•  The	
  equaKon	
  is	
  about:	
  Buffers	
  x	
  ConnecKons	
  
•  DIMM	
  populaKon	
  guide	
  
•  DRAM	
  market:	
  OVERSUPPLY	
  
Memory configuration
29	
  
•  Extending	
  Linux	
  block-­‐cache	
  with	
  SSD	
  
•  Block-­‐cache	
  works	
  by	
  caching	
  accessed	
  blocks	
  
•  Perfect	
  sizing	
  -­‐>	
  performance	
  similar	
  to	
  SSD	
  only	
  
•  Failures	
  
•  Writethrough	
  	
  	
  	
  	
  cache	
  ||	
  storage	
  	
  
•  Writearound	
  	
  	
  	
  	
  	
  	
  –(w)–>	
  storage	
  –(r)–>	
  cache	
  –>	
  
•  Writeback	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  –(w)–>	
  cache	
  –(w)–>	
  storage	
  –>	
  
Flashcache
30	
  
•  6	
  x	
  300	
  GB	
  RAID10	
  -­‐>	
  2	
  x	
  100	
  GB	
  SSD	
  +	
  2	
  x	
  1	
  TB	
  RAID1	
  
•  ConsideraKons	
  
•  I/O	
  rates	
  
•  Write	
  savings	
  
•  Overprovisioning	
  
•  ParKKoning	
  
•  How	
  long	
  to	
  cache	
  
•  What	
  to	
  cache	
  
•  Less	
  HDD	
  purchases,	
  More	
  SSD	
  purchases	
  (High	
  ALE)	
  
Flashcache
31	
  
•  4	
  nodes	
  of	
  6	
  x	
  146	
  GB	
  HDD	
  
•  CAPEX:	
  €	
  4,700	
  
•  OPEX:	
  €	
  293	
  
•  TCO	
  5y:	
  €	
  6,165	
  
	
  
•  4	
  nodes	
  of	
  2	
  x	
  600	
  HDD	
  +	
  2	
  x	
  Intel	
  520	
  SSD,	
  30%	
  opt.	
  writes	
  
•  Replacing	
  them	
  every	
  year	
  (TBWy:	
  100	
  TB)	
  
•  CAPEX:	
  €	
  11,680	
  
•  OPEX:	
  €	
  110	
  
•  TCO:	
  €	
  12,230	
  
Flashcache: Case study
32	
  
•  TCO	
  is	
  a	
  very	
  handy	
  (someKmes	
  misleading)	
  tool	
  
•  RCO	
  -­‐	
  Relevant	
  Cost	
  of	
  Ownership	
  
•  CAPEX:OPEX	
  raKo	
  
•  Giving	
  back	
  is	
  OPEX	
  
•  Keep	
  eye	
  on	
  your	
  power	
  consumpKon	
  
•  Buy	
  DRAM	
  
•  Consider	
  ALE	
  when	
  you	
  switch	
  to	
  SSD	
  
Summary
Questions?
34	
  
•  PresentaKon	
  can	
  be	
  found	
  at:	
  
hop://spil.com/tcomeetup2013	
  
	
  
	
  
•  You	
  can	
  contact	
  me	
  at:	
  	
  
zsolt.fabian@spilgames.com	
  
Thank you!

More Related Content

PPT
Spil Games @ FOSDEM: Galera Replicator IRL
PPTX
Percona Live London 2014: Serve out any page with an HA Sphinx environment
PDF
3 周彦偉-隨需而變 我所經歷的my sql架構變遷﹣周彥偉﹣acmug@2015.12台北
PDF
MySQL High-Availability and Scale-Out architectures
PDF
Run Cloud Native MySQL NDB Cluster in Kubernetes
PPTX
Overview of some popular distributed databases
PDF
MySQL NDB Cluster 101
PDF
MySQL NDB Cluster 8.0 SQL faster than NoSQL
Spil Games @ FOSDEM: Galera Replicator IRL
Percona Live London 2014: Serve out any page with an HA Sphinx environment
3 周彦偉-隨需而變 我所經歷的my sql架構變遷﹣周彥偉﹣acmug@2015.12台北
MySQL High-Availability and Scale-Out architectures
Run Cloud Native MySQL NDB Cluster in Kubernetes
Overview of some popular distributed databases
MySQL NDB Cluster 101
MySQL NDB Cluster 8.0 SQL faster than NoSQL

What's hot (20)

PDF
MySQL 5.6 Performance
PPTX
MariaDB Galera Cluster
PDF
Scylla Summit 2016: Outbrain Case Study - Lowering Latency While Doing 20X IO...
PDF
Performance Monitoring: Understanding Your Scylla Cluster
PDF
MySQL highav Availability
PDF
The InnoDB Storage Engine for MySQL
PDF
MySQL in the Hosted Cloud - Percona Live 2015
PDF
Modern MySQL Monitoring and Dashboards.
DOCX
Master master vs master-slave database
PDF
MySQL High Availability Solutions
PDF
Galera Cluster for MySQL vs MySQL (NDB) Cluster: A High Level Comparison
PDF
MHA: Getting started & moving past quirks percona live santa clara 2013
PDF
Choosing a MySQL High Availability solution - Percona Live UK 2011
PDF
Boosting Machine Learning with Redis Modules and Spark
PDF
Mesosphere and Contentteam: A New Way to Run Cassandra
PDF
Ramp-Tutorial for MYSQL Cluster - Scaling with Continuous Availability
PDF
Highly Available MySQL/PHP Applications with mysqlnd
PDF
Introduction to Cassandra Architecture
PDF
Advanced Operations
PDF
Best practices for MySQL/MariaDB Server/Percona Server High Availability
MySQL 5.6 Performance
MariaDB Galera Cluster
Scylla Summit 2016: Outbrain Case Study - Lowering Latency While Doing 20X IO...
Performance Monitoring: Understanding Your Scylla Cluster
MySQL highav Availability
The InnoDB Storage Engine for MySQL
MySQL in the Hosted Cloud - Percona Live 2015
Modern MySQL Monitoring and Dashboards.
Master master vs master-slave database
MySQL High Availability Solutions
Galera Cluster for MySQL vs MySQL (NDB) Cluster: A High Level Comparison
MHA: Getting started & moving past quirks percona live santa clara 2013
Choosing a MySQL High Availability solution - Percona Live UK 2011
Boosting Machine Learning with Redis Modules and Spark
Mesosphere and Contentteam: A New Way to Run Cassandra
Ramp-Tutorial for MYSQL Cluster - Scaling with Continuous Availability
Highly Available MySQL/PHP Applications with mysqlnd
Introduction to Cassandra Architecture
Advanced Operations
Best practices for MySQL/MariaDB Server/Percona Server High Availability
Ad

Similar to Database TCO (20)

PDF
Ceph Day Tokyo -- Ceph on All-Flash Storage
PDF
Presentation database on flash
PDF
Evoluzione dello storage
PDF
Nimble Storage Series A presentation 2007
PPTX
Deploying ssd in the data center 2014
PPTX
MemVerge: Memory Expansion Without Breaking the Budget
PDF
How Ceph performs on ARM Microserver Cluster
PDF
S3
PDF
Demystifying SSD, Mark Smith, S3
PDF
2013.11.14 Big Data Workshop Michael Browne
PDF
Stabilizing Ceph
PPTX
DevOps for ETL processing at scale with MongoDB, Solr, AWS and Chef
PPTX
stackArmor presentation for DevOpsDC ver 4
PDF
Big Data Hadoop Briefing Hosted by Cisco, WWT and MapR: Cisco UCS For Big Dat...
PDF
3.INTEL.Optane_on_ceph_v2.pdf
PDF
Ceph Day Beijing - Optimizing Ceph Performance by Leveraging Intel Optane and...
PDF
Ceph Day Beijing - Optimizing Ceph performance by leveraging Intel Optane and...
PPTX
QCT Ceph Solution - Design Consideration and Reference Architecture
PPTX
QCT Ceph Solution - Design Consideration and Reference Architecture
PDF
Pragma torino 20mar2014 flash&tools vr03
Ceph Day Tokyo -- Ceph on All-Flash Storage
Presentation database on flash
Evoluzione dello storage
Nimble Storage Series A presentation 2007
Deploying ssd in the data center 2014
MemVerge: Memory Expansion Without Breaking the Budget
How Ceph performs on ARM Microserver Cluster
S3
Demystifying SSD, Mark Smith, S3
2013.11.14 Big Data Workshop Michael Browne
Stabilizing Ceph
DevOps for ETL processing at scale with MongoDB, Solr, AWS and Chef
stackArmor presentation for DevOpsDC ver 4
Big Data Hadoop Briefing Hosted by Cisco, WWT and MapR: Cisco UCS For Big Dat...
3.INTEL.Optane_on_ceph_v2.pdf
Ceph Day Beijing - Optimizing Ceph Performance by Leveraging Intel Optane and...
Ceph Day Beijing - Optimizing Ceph performance by leveraging Intel Optane and...
QCT Ceph Solution - Design Consideration and Reference Architecture
QCT Ceph Solution - Design Consideration and Reference Architecture
Pragma torino 20mar2014 flash&tools vr03
Ad

Recently uploaded (20)

PPTX
20250228 LYD VKU AI Blended-Learning.pptx
PPTX
VMware vSphere Foundation How to Sell Presentation-Ver1.4-2-14-2024.pptx
PDF
Architecting across the Boundaries of two Complex Domains - Healthcare & Tech...
PPTX
Digital-Transformation-Roadmap-for-Companies.pptx
PDF
Reach Out and Touch Someone: Haptics and Empathic Computing
PDF
Dropbox Q2 2025 Financial Results & Investor Presentation
PDF
Empathic Computing: Creating Shared Understanding
PDF
Encapsulation theory and applications.pdf
PDF
Review of recent advances in non-invasive hemoglobin estimation
PPT
Teaching material agriculture food technology
PDF
Network Security Unit 5.pdf for BCA BBA.
PPTX
Detection-First SIEM: Rule Types, Dashboards, and Threat-Informed Strategy
PPTX
sap open course for s4hana steps from ECC to s4
PDF
Advanced methodologies resolving dimensionality complications for autism neur...
PDF
Profit Center Accounting in SAP S/4HANA, S4F28 Col11
DOCX
The AUB Centre for AI in Media Proposal.docx
PDF
Build a system with the filesystem maintained by OSTree @ COSCUP 2025
PPT
“AI and Expert System Decision Support & Business Intelligence Systems”
PDF
Diabetes mellitus diagnosis method based random forest with bat algorithm
PDF
Approach and Philosophy of On baking technology
20250228 LYD VKU AI Blended-Learning.pptx
VMware vSphere Foundation How to Sell Presentation-Ver1.4-2-14-2024.pptx
Architecting across the Boundaries of two Complex Domains - Healthcare & Tech...
Digital-Transformation-Roadmap-for-Companies.pptx
Reach Out and Touch Someone: Haptics and Empathic Computing
Dropbox Q2 2025 Financial Results & Investor Presentation
Empathic Computing: Creating Shared Understanding
Encapsulation theory and applications.pdf
Review of recent advances in non-invasive hemoglobin estimation
Teaching material agriculture food technology
Network Security Unit 5.pdf for BCA BBA.
Detection-First SIEM: Rule Types, Dashboards, and Threat-Informed Strategy
sap open course for s4hana steps from ECC to s4
Advanced methodologies resolving dimensionality complications for autism neur...
Profit Center Accounting in SAP S/4HANA, S4F28 Col11
The AUB Centre for AI in Media Proposal.docx
Build a system with the filesystem maintained by OSTree @ COSCUP 2025
“AI and Expert System Decision Support & Business Intelligence Systems”
Diabetes mellitus diagnosis method based random forest with bat algorithm
Approach and Philosophy of On baking technology

Database TCO

  • 1. Total Cost of Ownership Zsolt Fabian Database Engineering
  • 2. 2   1.  Who  is  Spil  Games?   2.  What  is  TCO?   3.  What  drives  the  costs?   4.  Possible  improvements?   Outline of the talk
  • 3. Who are we? Who  is  Spil  Games?    
  • 4. 4   •  Company  founded  in  2001   •  350+  employees  world  wide   •  200M+  unique  visitors  per  month   •  45  portals  in  19  languages   •  Casual  games   •  Social  games   •  Real  Kme  mulKplayer  games   •  Mobile  games   •  35+  MySQL  clusters   Facts
  • 5. TCO Total  Cost  of  Ownership    
  • 6. 6   •  1980  –  “Life  Cycle  CosKng”   •  1987  –  Gartner  Group   •  Total  Cost  of  Ownership   •  Analysis  of  enterprise  spending,  vendor  evaluaKon   •  Many  inter-­‐related  direct  and  indirect  costs   •  Problems  with  TCO   What is TCO and why it is important? CAPEX  –  Capital  Expenses   OPEX  –  OperaKng  Expenses   Relevant Costs
  • 8. 8   •  CollecKng  data   •  Different  base  systems   •  4,3  TB  of  RAM   •  402  drives  wriKng  3  ExaBytes   •  3  DBA   •  12,3  kW  power  consumpKon   •  Asset  value  exceeds  €  500,000   •  Cables  are  available  in  all  colors  of  the  spectrum  J     A quick inventory
  • 9. 9   Aging servers •  Component  failures   •  Power  consumpKon   •  Volume  of  deployment  
  • 10. 10   •  The  most  obvious   •  Base  server  +  CPU  +  Mem  +  Disk  +  Controller   •  Replacement  parts  and  Warranty   •  NPV  –  Net  Present  Value   •  Mining  invoices   CAPEX: Purchase costs AFR  –  Annual  Failure  Rate   ALE  –  Annual  Loss  Expectancy  
  • 11. 11   •  Business  need   •  Cost  of  downKme   •  Levels  of  HA  and  Fault  Tolerance   •  Server  downKme   •  CPU  failure   •  MySQLd  crash   •  Solar  acKvity   •  Combined  availability  formula:   CAPEX: Purchase costs: HA A  =  1  –  (1  –  Ax)  ^  n  
  • 12. 12   •  Giving  back  J   •  MySQL  Standard  EdiKon:  $  2,000   •  Enterprise:  $  5,000   •  Cluster  Carrier  Grade:  $  10,000   •  Microsoj  SQL  Server:  $  6k+  /  core     •  Oracle  10/11g:  $  40k+   •  AddiKonal:  Technology  extensions,  Dashboards   CAPEX: License costs
  • 13. 13   •  Work  hours  spent…   •  SLA,  Responsibility,  Stress,  Insider  Threat   •  MySQL  annual  subscripKon:  $  2,000   •  Specialist  support  from  $  1,500  /  server  /  year   •  Unlimited  servers  for  $  30,000  /  year   •  Depends  on  complexity   •  €  2,700  /  server  /  year     OPEX: MySQL Professional support It's  cheaper  to  add  more  servers  than  hire  more  engineers.  
  • 14. 14   •  Cage  –  Rack  –  Server   •  Networking   •  Deployment  (+  fuel  +  Kme)   •  Power  density  constraints   •  Let’s  say  €  60  /  server  /  month   OPEX: Hosting: Colocation
  • 15. 15   •  Wao  =  Ampere  *  Volt   •  €  0.2  /  kWh   •  12,31  kWh  (~  16%  of  total)   •  €  21,300  annually   •  Measurements   •  IPMI   •  Power  meter  J   •  Contributors   •  HDD  (25%)   •  Roles  (5-­‐10W)   OPEX: Hosting: Power
  • 18. Pu@ng  it  together     The  model     Cost  Item   CAPEX   OPEX   TCO  5y   Base  System   ?   Extra  memory   ?   Extra  disks   ?   Licence  fees   ?   Replacement  parts  /   Warranty   ?   ?   Co-­‐locaKon   ?   Power   ?   Professional  support   ?   TOTAL   ?   ?   ?  
  • 19. Pu@ng  it  together     Cluster  of  6  nodes   Cost  Item  (6x)   CAPEX   OPEX   TCO  5y   Purchase:  HP  DL380  G7   €  25,000   Purchase:  Extra  memory     €  6,000   Professional  support   €  16,200   ColocaKon  cost   €  4,320   Power  costs  @200W   €  2,100   Replacement  parts     €  1,500   TOTAL   €  32,500   €  22,620   €  145,600  
  • 20. Pu@ng  it  together     Cluster  of  6  nodes   Purchase   costs   18%   Professional   support   58%   ColocaKon   15%   Power   8%   Replacement   parts   1%   TCO 5 years
  • 21. Improvements “Insanity:  doing  the  same   thing  over  and  over  again  and      expecSng  different  results”  -­‐   Albert  Einstein    
  • 22. 22   Purchase prioritization Extend  warranty   MTBF   Obsolete,  Not  Urgent   Limited  reparability   Repairable   New  
  • 23. 23   •  SSD  reduces  the  cost  of  IOPS   •  SSD  is  fast  for  reads,  but  not  for  writes   •  Has  limited  lifespan   •  NAND  Flash  and  RAM  based  SSDs   •  MLC  (MulK  Level  Cell)  vs.  SLC  (Single  Level  Cell)   •  R/W  in  pages,  but  always  erased  in  blocks   •  The  main  advantage  that  it  doesn’t  have  to  spin   HDD vs. SSD: Inner workings
  • 24. 24   •  Dropping  over  Kme   •  Facts   •  Tsunami   •  Rare-­‐earth  minerals   •  Cost  per  GB  is  cheaper  at  HDD   •  Some  predicts  that  in  5  to  10  years,  SSD  will  replace   HDD  completely   HDD vs. SSD: Pricing
  • 25. 25   HDD vs. SSD: Power consumption Type   Model   Idle  Power   AcSve  Power   Annual  cost   contribuSon   SSD   Intel  520   600  mW   850  mW   €  1,49   SSD   Intel  X25-­‐M   60  mW   150  mW   €  0,26   SSD   Intel  X25-­‐E   60  mW   4,2  W   €  7,35   HDD   ST9146852S   5  W   7  W   €  12,2  
  • 26. 26   •  MTBF:  ~  200  years   •  HDD   •  AFR:  0,55%    (-­‐  1,5%  -­‐  2%  -­‐  …)   •  ALE!   •  SSD  have  LIMITED  write  cycles   •  SLC:  100,000  writes  /  block   •  MLC:  10,000  writes  /  block   •  Typical  MLC:  5,000  writes  /  block   •  Future  developments   HDD vs. SSD: Endurance
  • 27. 27   •  Intel  520  SSD:  5  years  (with  normal  desktop  usage)   •  35  TB  write  endurance  (TBW)   •  Official  endurance  is  overprovisioned  (~  50  %)   •  SaturaKon   •  Filesystem  overhead  (0.25  btrfs-­‐>  0.015  (ext4))   •  Readout   •  S.M.A.R.T.:  Percentage  of  rated  lifeKme  used   •  UKliKes   •  RAID  controllers   HDD vs. SSD: Endurance
  • 28. 28   •  Determined  during  the  design   •  The  equaKon  is  about:  Buffers  x  ConnecKons   •  DIMM  populaKon  guide   •  DRAM  market:  OVERSUPPLY   Memory configuration
  • 29. 29   •  Extending  Linux  block-­‐cache  with  SSD   •  Block-­‐cache  works  by  caching  accessed  blocks   •  Perfect  sizing  -­‐>  performance  similar  to  SSD  only   •  Failures   •  Writethrough          cache  ||  storage     •  Writearound              –(w)–>  storage  –(r)–>  cache  –>   •  Writeback                      –(w)–>  cache  –(w)–>  storage  –>   Flashcache
  • 30. 30   •  6  x  300  GB  RAID10  -­‐>  2  x  100  GB  SSD  +  2  x  1  TB  RAID1   •  ConsideraKons   •  I/O  rates   •  Write  savings   •  Overprovisioning   •  ParKKoning   •  How  long  to  cache   •  What  to  cache   •  Less  HDD  purchases,  More  SSD  purchases  (High  ALE)   Flashcache
  • 31. 31   •  4  nodes  of  6  x  146  GB  HDD   •  CAPEX:  €  4,700   •  OPEX:  €  293   •  TCO  5y:  €  6,165     •  4  nodes  of  2  x  600  HDD  +  2  x  Intel  520  SSD,  30%  opt.  writes   •  Replacing  them  every  year  (TBWy:  100  TB)   •  CAPEX:  €  11,680   •  OPEX:  €  110   •  TCO:  €  12,230   Flashcache: Case study
  • 32. 32   •  TCO  is  a  very  handy  (someKmes  misleading)  tool   •  RCO  -­‐  Relevant  Cost  of  Ownership   •  CAPEX:OPEX  raKo   •  Giving  back  is  OPEX   •  Keep  eye  on  your  power  consumpKon   •  Buy  DRAM   •  Consider  ALE  when  you  switch  to  SSD   Summary
  • 34. 34   •  PresentaKon  can  be  found  at:   hop://spil.com/tcomeetup2013       •  You  can  contact  me  at:     zsolt.fabian@spilgames.com   Thank you!