TechTarget Event - Storage Architectures for the Modern Data Centre – Chris Evans

35
STORAGE ARCHITECTURES FOR THE MODERN D ATA CENTRE Chris M Evans Langton Blue Ltd Copyright © 2016 Langton Blue Ltd 1

Transcript of TechTarget Event - Storage Architectures for the Modern Data Centre – Chris Evans

STORAGEARCHITECTURES FOR THE MODERNDATACENTRE

Chris  M  EvansLangton  Blue  Ltd

Copyright  ©  2016  Langton  Blue  Ltd 1

A  WORDABOUTME

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 2

§ 28  years  experience  in  the  IT  industry  across  all  platforms,  including  IBM  mainframe,  Windows  &  Open  Systems.    

§ Co-­‐founder  and  independent  consultant  at  Langton  Blue  Ltd,  a  specialist  consulting  company  in  the  UK.

§ Blogger  and  part-­‐time  analyst  at  Architecting  IT§ Twitter:  @chrismevans,  @architectingIT,  

@langtonblue§ Web:  www.architecting.it,  

www.langtonblue.com

AGENDA

§ Evolution  of  The  Data  Centre§ Understanding  Customer  Needs§ Storage  Architectures§ Product  Marketplace§ Q&A

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 3

EVOLUTION  OF  THE  DATA  CENTREA  quick  primer  on  storage  heritage

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 4

20TH CENTURYDESIGN

§ Monolithic§ Bespoke  hardware§ Expensive  to  acquire,  maintain  &  

manage§ Static  designs  and  configuration

§ (Think  time  taken  to  run  binfilechange)

§ Limited  or  no  data  services§ No  RAID,  replication,  snapshots§ All  features  that  developed  over  time

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 5

20TH CENTURYARCHITECTURE

§ Custom  hardware  designs  &  components  still   reign§ Hosts  own  LUNs  mapped  from  physical  RAID  groups§ Performance  defined  by  RAID  group  size  and  disk  

speed§ Wide  striping  for  performance

§ Throughput  &  I/O  latency  influenced   (and  restricted)  by  controller  capabilities

§ Lots  of  manual  effort  to  rebalance  systems,  especially  at  scale§ LUN  migration,  rebuilds,  reliance  on  host-­‐level  LVMs  for  a  

lot  of  the  work

§ Lots  of  risk  fragmentation  and  orphan  storage§ Wasted  space  in  RAID  groups

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 6

CONTROLLER  A CONTROLLER  B

BED BED BED BED BED BED BED BED

21ST CENTURYTECHNOLOGY

§ Vendors  have  moved  away  from  bespoke  hardware  designs§ x86  Standardisation§ Commodity,  reliable  components§ High  performance  processors  &  memory  – Moore’s  Law

§ Interoperability§ Components  generally  work  together  without  problems§ Storage  protocols  are  robust  and  mature  (still  also  developing  – NVMe)

§ New  Technology§ Flash§ NVDIMM§ 3D  Xpoint

§ Migration  of  “smarts”  to  software    -­‐ Software  Defined  Storage§ Features  implemented   in  code,  not  microcode

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 7

FLASH &  NEW MEDIA

§ Designs§ SLC  (Single  Level  Cell),  MLC  (Multi-­‐level  Cell),  TLC  (Triple  Level  Cell)§ 3D-­‐NAND  – 3-­‐D  rather  than  planar  (2D)  technology  for  higher  density

§ 3D-­‐Xpoint  – faster  and  more  scalable  persistent  storage  than  Flash,  slower  than  DRAM  and  in  between  on  price

§ Intel/Micron  joint  development§ NVRAM/NVDIMM  – DIMM  form  factor  products  delivering  

persistent  flash  memory  on  the  motherboard§ NVMe – new  connectivity  standard  for  writing  to  persistent  storage

§ PCIe and  SDD  type  form  factors  (not  yet  widely  adopted)

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 8

PROCESSOR &  MEMORYPERFORMANCE

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 9

§ Processor,  memory  and  bus  speeds  are  increasing

§ Processor  performance  and  bus  speeds  follow  Moore’s  Law

§ Bus  speeds  double  with  each  release§ PCIe  v1.x  =  250MB/s  (per  lane)  -­‐ 2003§ PCIe  v2.x  =  500MB/s  (per  lane)  -­‐ 2007§ PCIe  v3.0  =  985MB/s  (per  lane)  -­‐ 2010§ PCIe  v4.0  =  1969MB/s  (per  lane)  – 2014

§ Features  much  easier   to  deliver   from  software  on  x86§ New  instruction  sets   like  SSSE3  and  AVX2  

allow  storage  computations  like  erasure  coding/Reed  Solomon  to  be  done  in  software

§ Fewer  reasons  to  develop  bespoke  ASICs  and  FGPAs

SOFTWARE DEFINED STORAGE(KEY FEATURES)

§ Abstraction – I/O  services  should  be  delivered  independent  of  the  underlying  hardware,  through  logical  constructs  like  LUNs,  volumes,  file  shares  and  repositories.

§ Automation – resources  should  be  consumed  using  CLIs  and  APIs  rather  than  manually  allocated  through  a  GUI.

§ Policy/Service  Driven  – the  service  received  (IOPS,  latency)  should  be  established  by  policies  that  implement  Quality  of  Service,  availability  and  resiliency.

§ Scalable – solutions  should  enable  performance  &  capacity  scaling  independent  of  I/O  delivery.

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 10

MAJOR STORAGETHEMES

§ More  choice  in  products   (both  hardware  and  software  than  ever  before)

§ Divergence  – storage  platforms   for  specific  purposes  (no  single  platform  for  all  requirements)§ Object  storage,  cloud,  backup,  primary

§ Convergence  – collapsing   the  storage/server  hardware  into  a  single  unit  integrating  new  technology  (converged  &  hyper-­‐converged)

§ Integration  – VVOLs,  APIs,  platform  drivers  (e.g.  Cinder),  hypervisor  extensions   (VASA,  VAAI)

§ New  Media§ Flash  and  its  successors

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 11

UNDERSTANDING  CUSTOMER  NEEDSThe  changing  face  of  application  deployment

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 12

WHAT ARE CUSTOMERS’  PAIN POINTS

§ Cost  § Performance  § Operational  Complexity  § Reliability§ Floor  Space§ Power  Consumption

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 13

Source:  Tintri State  of  Storage  Report,  March  2015https://www.tintri.com/sites/default/files/fie ld/pdf/wh itepapers/state_of_storage_ infographic_150513t10216.pdf

WHAT DO CUSTOMERS CARE ABOUT

§ Efficiency§ Cost  ($/GB,  effective  or  raw)§ Physical  media

§ Performance§ Latency,  IOPS,  throughput,  Quality  of  Service

§ Management§ Multi-­‐tenancy,  APIs  (Cinder),  Automation,  Abstraction  (VVOLs)

§ Reliability§ Resiliency,  Availability

§ Features§ Data  Services,  Data  Protection

§ Breaking  the  buying   cycle§ No  forklift  upgrades  or  3-­‐4  year  refresh  projects

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 14

PREDICTABLYUNPREDICTABLE

§ Workloads  becoming  more  random  in  nature§ I/O  blender

§ Peaks  and  troughs  in  demand§ VDI  (Bootstorms,  logoff  storms)

§ Increase  in  I/O  Density§ Driven  by  virtualisation  and  soon  containers

§ Agility§ Desire  to  move  data  around,  spin  up  copies,  create  test  dev environments,  destroy  

and  recreate§ Persistence

§ Applications  expecting  to  be  transient  in  nature,  but  more  need  for  storage  to  be  100%  available  across  the  entire  infrastructure

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 15

VIRTUAL FIRST STRATEGY

§ Provided   the  ability  to  consolidate   over-­‐provisioned   server  resources§ Systems  typically  25%  allocated

§ Reduced  delivery  time  for  (virtual)  servers  from  days/months   to  hours§ Assuming  sufficient  capacity   available

§ Enabled  high  availability  through  infrastructure§ No  need  to  deploy  clustered  systems§ Centralised  data  protection  (e.g.   VADP)

§ Centralised  storage  and  virtualisation   don’t   play  well  together§ Storage  doesn’t   understand  VMs

§ Storage  sees  LUNs  &  volumes§ Virtualisation  sees  VMs§ LUNs  encapsulate  many  volumes  – no  visibility  of  VMs§ Problems  being  worked  around  with  VVOLs  and  other  solutions  that   integrate  directly  

into  hyper-­‐converged  platforms.

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 16

TECHNICAL REQUIREMENTS– QUALITY OF SERVICE

§ Eliminate  the  “noisy  neighbour”  problem§ Provide  storage  resources  based  on  policy§ Ensure  storage  isn’t  under-­‐delivering

§ Guarantee  IOPS,  performance  minima§ Ensure  storage  isn’t  over-­‐delivering

§ 1st on  the  array  gets  best  performance  (until  there  are  more)

§ Set  limits  and  prioritisation§ Apply  to  VM/VMDK  where  possible

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 17

TECHNICAL REQUIREMENTS -­‐ API  DRIVEN

§ Admin-­‐based  functionality  driven  by  API§ Ability  to  create,  manage,  destroy  LUNs  

from  code§ Integrate  storage  functions  into  

workflow§ Cloud,  Automation,  Self  Service

§ Reduce  dependence  on  manual  process§ Agility

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 18

TECHNICAL REQUIREMENT– SUPPORTVIRTUAL

§ Reduce  waste  from  over-­‐provisioning§ Support  – object  (e.g.  VM-­‐based)  

management§ Apply  policies  at  the  VM  level

§ Data  location,  performance,  latency,  backup§ VM  friendly  for  data  protection  

(snapshots/replication)§ Offload  heavy  lifting  tasks  (e.g.  VAAI)

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 19

STORAGE  ARCHITECTURESA  few  words  about  design

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 20

DIVERSITY OF STORAGEPLATFORMS

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 21

HYBRID

COMMODITY OPEN-­‐SOURCE

APP/DATA  AWARE

SMART  STORAGE

OBJECT  STORES

SCALE-­‐OUT  SAN

SCALE-­‐OUT  NAS

ALL-­‐FLASH

SOFTWARE  DEFINED

SCALE-­‐OUT  BACKUP

ULTRA-­‐FAST  FLASH

HIGH-­‐CAPACITY  FLASH

HPC

INFRASTRUCTURE CONSUMPTION CONTINUUM

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 22

As-­‐a-­‐Service Hyper-­‐converged Converged Proprietary  

ApplianceSoftware  on  Commodity

SMB/Start-­‐ups Large  Enterprise Hyperscale &  SP

Ease  of  Implementation

Implementation  Flexibility

Degree  of  Vendor  Lock-­‐in

Cost  Efficiency

Harder

More  Flexible

Less  Lock-­‐in

At  Large  Scale

Easier

Less  Flexible

More  Lock-­‐in

At  Small  Scale

SCALE UP

§ Typically  Dual  controller  implementation§ Active/Active  &  Active/Passive

§ Throughput   limited  by  controller  performance§ More  throughput  means  better  controller  

capability   (CPU/memory)§ I/O  performance  limited  by  media§ Flash  based  implementations   failed  to  use  

flash  to  fullest   capability§ Older  designs  don’t  cater  for  write  endurance  

issues§ Automated  tiering capabilities   to  place  data  on  

most  appropriate  technology

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 23

CONTROLLER  A CONTROLLER  B

SCALE OUT (CLOSELYCOUPLED)

§ Scale  out  through  adding  nodes  and/or  capacity

§ Tight  node  integration  either  paired  or  with  bespoke  backplane  (e.g.  PCIe/Infiniband)

§ Symmetric  design  – scale  with  nodes  of  similar  size/capacity

§ LUNs/Volumes  delivered  from  specific  nodes  – not  fully  load  balanced

§ Scale  out  limited  by  backplane  traffic§ Add/remove  nodes  dynamically

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 24

CTRL   A CTRL   B CTRL   C CTRL   D CTRL   E CTRL   F

SCALE OUT (LOOSELY COUPLED)

§ Fully  distributed  “Shared  Nothing”  architecture§ Loose  node  coupling   (usually  IP-­‐based)§ All  nodes  contribute  storage  (no  disk  shelves)

§ At  least  one  node  of  “spare”  capacity  required  across  the  cluster§ More  nodes  means  more  efficiency  in  redundancy

§ Data  resiliency  based  on  replicating  data  between  nodes§ Standard  hardware  (typically  1U  server)  design

§ Node  power  efficiency  &  size  become  important§ High  scalability§ Smaller  increments  in  adding  capacity/performance§ More  complex  design  

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 25

NODE  A

NODE  B

NODE  C

NODE  D

NODE  E

NODE  F

HYPER-­‐CONVERGED

§ Single  hardware  device  for  storage  and  server§ Node-­‐based  scale-­‐out§ Integrated  hypervisor  &  storage  services

§ Storage  implemented  in  VM  or  hypervisor  kernel§ Distributed  file  system  (scale-­‐out  shared  nothing)

§ Efficient  use  of  resources§ Simplified  management§ At  scale  probably  more  expensive  than  separate  

components§ Reduced  VM  density§ Issues  managing  performance  vs  capacity  scaling

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 26

NODE  A

NODE  B

NODE  C

NODE  D

NODE  E

NODE  F

HYPER-­‐CONVERGENCE -­‐ TCO

§ Simplified§ Less  “plan”  time   -­‐ no  requirement   to  certify  and  test  components§ Less  “build”  time  – deploy  a  new  node,  power  up  and  add  to  the  

configuration§ Less  “operate”   time  – single  management  console,  typically  

integrated  with  the  hypervisor,  tight  hypervisor  awareness§ Efficient

§ Scale  per  node;  less  wastage,  no  up-­‐front  hardware  acquisition§ “Decommission”   time  almost  eliminated;   simply  add  new  nodes  

and  evacuate/remove  old  nodes  over  time

§ Provides  overall  TCO  savings  (not   storage  specific)§ Not  typically  suited  to  high-­‐scale  environments

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 27

STORAGEVIRTUALISATION -­‐ EVOLUTION

§ Monolithic   – single  central  controller§ Inline  in  the  data  path§ Controller  maps  logical   to  “physical”  storage§ data  management  features  built  into  the  controller§ Not  highly  scalable

§ CentralisedMetadata  – Distributed  Data§ Central  metadata  functions§ Data  distributed/replicated  across  many  devices§ System  not  in  the  data  path§ Separation  of  control  and  data  planes

§ Totally  Distributed§ No  central  metadata  or  data§ Data  distributed  across  many  devices§ Fully  scalable  architecture

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 28

ADVANCED SERVICES

§ Space  Efficiency§ Compression§ De-­‐duplication§ Thin  Provisioning

§ Quality  of  Service§ Object  abstraction  (VVOLS,  per  VM  services)§ Application  consistent  snapshots/backup

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 29

SPACE OPTIMISATION/EFFICIENCY

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 30

Feature Compression De-­‐duplication Thin  Provisioning

Physical  space  reduction ✔ ✔ ❌

Architecture/metadata  dependent ❌ ✔ ✔

CPU Friendly ❌ ✔ ✔

Traditional  OLTP  Database  friendly ✔ ❌ ❌

VDI/VSI Friendly ❌ ✔ ✔

HDD  Friendly ✔ ❌ ✔

Flash  Friendly ✔ ✔ ✔

QUALITY OF SERVICE (AGAIN)

§ Assign  service-­‐based  metrics  to  application  workloads§ Reduce/eliminate  the  “noisy  neighbour”  effect§ Set  application  owner  “expectations”  on  performance

§ Avoid  first/last  application  syndrome  from  legacy  arrays§ Metrics

§ Latency§ IOPS§ Throughput

§ Establish  maximum/minimum/burst  values  &  priority

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 31

OBJECTABSTRACTION

§ Implement  per-­‐VM  services§ Abstract  away  from  LUNs/Volumes

§ Implement  object-­‐specific  services§ Snapshots§ Replication§ Cloning

§ Application-­‐consistent  data  management§ Examples

§ VVOLs§ Object  Stores

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 32

VENDOR  LANDSCAPEA  look  at  the  product  marketplace

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 33

STORAGENASCAR  SLIDE!

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 34

PUTTING IT ALL TOGETHER

§ Start  with  Requirements§ Don’t  buy  what  you  don’t  need§ Focus  on  your  sensitivities  (cost,  performance,  scalability)

§ See  the  “bigger  picture”§ Don’t  just  storage  in  isolation

§ Don’t  buy  a  technology§ Look  past  the  glamour  of  shiny  boxes!

§ Evaluate  your  options§ Use  tools  to  create  an  equitable  test  environment

§ Use  online  resources  and  other  views§ Wisdom  of  the  crowd  applies

Copyright  ©  2009-­‐2016 Langton  Blue  Ltd 35