cc-4009, "optimizing hadoop deployments with seamicro sm15000" by satheesh nanniyur and...

20
OPTIMIZED HADOOP DEPLOYMENTS WITH SEAMICRO SM15000 PRESENTED AT AMD DEVELOPER SUMMIT, NOV 2013 SATHEESH NANNIYUR

Upload: amd-developer-central

Post on 26-Jan-2015

109 views

Category:

Technology


1 download

DESCRIPTION

Presentation, "Optimizing Hadoop Deployments with SeaMicro SM15000" at the AMD Developer Summit (APU13) Nov. 11-13, 2013.

TRANSCRIPT

Page 1: CC-4009, "Optimizing Hadoop Deployments with SeaMicro SM15000" by Satheesh Nanniyur and Anil Rao

OPTIMIZED  HADOOP  DEPLOYMENTS  WITH  SEAMICRO  SM15000  

PRESENTED  AT  AMD  DEVELOPER  SUMMIT,  NOV  2013  SATHEESH  NANNIYUR  

Page 2: CC-4009, "Optimizing Hadoop Deployments with SeaMicro SM15000" by Satheesh Nanniyur and Anil Rao

|      RIGHT  SIZING  HADOOP  DEPLOYMENTS    2  

Devices   Apps   Cloud  

BIG  DATA  IS  A  STRATEGIC  DECISION:  CAPEX  AND  OPEX  

Page 3: CC-4009, "Optimizing Hadoop Deployments with SeaMicro SM15000" by Satheesh Nanniyur and Anil Rao

|      RIGHT  SIZING  HADOOP  DEPLOYMENTS    3  

HIGHER  PERFORMANCE,  LESS  POWER  AND  SPACE  

Data  Storage  

Data  Processing  

Data  Access  

Management  

Data  Warehouse   Data  AnalyRcs  

Hadoop  Technology  Stack  

Page 4: CC-4009, "Optimizing Hadoop Deployments with SeaMicro SM15000" by Satheesh Nanniyur and Anil Rao

|      RIGHT  SIZING  HADOOP  DEPLOYMENTS    4  

SEAMICRO  SM15000™  ACCELERATES  APACHE™  HADOOP™  DEPLOYMENTS  

!  Superior  high  availability  ‒ AcRve/standby  NameNode  ‒ AcRve/standby  JobTracker  ‒ Highly  resilient  fabric  for  inter-­‐node  east-­‐west  traffic  

!  Reduced  down  Rme  ‒ Remap  or  rezone  disks  to  recover  data  

‒ Hot-­‐swappable  upgrades  or  component  replacements  

!  Hardware  redundancy  ‒ Power  supplies  ‒ Network  I/O  

Page 5: CC-4009, "Optimizing Hadoop Deployments with SeaMicro SM15000" by Satheesh Nanniyur and Anil Rao

|      RIGHT  SIZING  HADOOP  DEPLOYMENTS    5  

SM15000  OVERVIEW  

64  Industry  standard  x86  servers  •  AMD  Opteron™,  Intel  Xeon®,  Atom™  •  Energy  efficient  processor  •  20  Gbps  per  socket,  16X  tradiRonal  servers  

160  Gbps  Network  I/O  •  Share  network  I/O  across  all  servers  •  Eliminate  TOR  switch  •  Minimize  cabling  •  In  service  upgrades  as  needed  

64  HDDs/SDDs  •  Share  drives  across  all  servers  •  Assign  one  server  to  one  or  more  drives  as  needed  •  In  service  upgrades  as  needed  

960  terabytes  Fabric  Storage  •  Extends  supercompute  fabric  to  external  storage  •  Up  to  3.84  PB  storage  capacity;  up  to  960  3.5”  SAS/

SATA  drives  •  Map  to  any  CPU—same  as  internal  drives  

Page 6: CC-4009, "Optimizing Hadoop Deployments with SeaMicro SM15000" by Satheesh Nanniyur and Anil Rao

|      RIGHT  SIZING  HADOOP  DEPLOYMENTS    6  

SEAMICRO  FREEDOM™  FABRIC  ASIC  PROVIDES  MASSIVE  PERFORMANCE,  REDUCES  POWER  AND  SPACE    

Freedom™  

Provides massive bandwidth while eliminating power hungry

top of rack switches

Reduces the power used by any CPU by consolidating and

shutting off unused functionality

Eliminates 90% of the components on a motherboard shrinking power used, cost and space

Freedom Supercompute Fabric

SeaMicro IOVT

TIO

B E N E F I T S  

Page 7: CC-4009, "Optimizing Hadoop Deployments with SeaMicro SM15000" by Satheesh Nanniyur and Anil Rao

|      RIGHT  SIZING  HADOOP  DEPLOYMENTS    7  

FS  4060-­‐L  FABRIC  STORAGE  ENCLOSURE  WITH  ZONING  CAPABILITY  

!  High  density,  power  opRmized  4U  enclosure  with  60  3.5”  drives  

!  Up  to  16  enclosures  per  SM15000,  960  drives,  and  3.84  PB  storage  capacity  

!  Redundant  controllers,  ports,  fans,  and  PSUs  !  Support  cost  opRmized  24x7  operaRons  SATA  HDD  for  high  

density  Big  Data  and  Object  Storage  deployments  

!  OpRonal  configuraRon  to  logically  parRRon  an  enclosure  into  two  30  3.5”  drive  enclosures  

!  Balanced  disk  to  core  raRo  (1:1)  for  opRmizing  Hadoop  performance  

!  Field  configurable  to  provide  utmost    flexibility  to  balance  density  and  performance  

Page 8: CC-4009, "Optimizing Hadoop Deployments with SeaMicro SM15000" by Satheesh Nanniyur and Anil Rao

|      RIGHT  SIZING  HADOOP  DEPLOYMENTS    8  

FREEDOM  FABRIC  DISAGGREGATES  SERVER  RESOURCES  

!  SM15000  provides  independent  scaling  of  Compute,  Storage,  and  Network  

!  Centrally  managed  provisioning  of  storage  and  network  resources  to  compute  nodes  enabled  by  CLI  and  API  interfaces  

PROVIDES  FLEXIBILITY  FOR  EXPANSION  AND  INFRASTRUCTURE  OPTIMIZATION  

SeaMicro  SM15000  Server  

Fabric  Interconnect  

CPU   CPU  

CPU   CPU  

CPU   CPU  

Compute  and  Memory  Pool  

Shared  Storage  Pool  

Network  Pool  

Deploy  

Run  and  Analyze  

Tune  and  OpWmize  

Cost  and  Performance  

Hadoop  OpRmizaRon  

Page 9: CC-4009, "Optimizing Hadoop Deployments with SeaMicro SM15000" by Satheesh Nanniyur and Anil Rao

|      RIGHT  SIZING  HADOOP  DEPLOYMENTS    9  

SM15000  FLEXIBLE  STORAGE  ALLOWS  ITERATIVELY  OPTIMIZING  APACHE®  HADOOP™  DEPLOYMENT  

!  Flexible  shared  storage  with  commodity  hardware  enabled  by  SeaMicro  fabric  technology  

!  Decoupled  from  Compute  and  Network  to  grow  storage  independently  

!  IteraRvely  opRmize  Hadoop  disk  to  core  raRo  as  applicaRon  needs  evolve  

Traditional Rackmount

Freedom  Fabric  

Captive DAS with Rigid Storage to Compute Ratio

Flexible scale-out Fabric Storage up to 5PB

Intel  /AMD  X86  servers  

Page 10: CC-4009, "Optimizing Hadoop Deployments with SeaMicro SM15000" by Satheesh Nanniyur and Anil Rao

|      RIGHT  SIZING  HADOOP  DEPLOYMENTS    10  

APACHE®  HADOOP™  DEPLOYMENT  ON  THE  SEAMICRO  SM15000™  

HDFS  NameNode  

HDFS  NameNode  

MapReduce  JobTracker  

MapReduce  JobTracker  

DataNode/  TaskTracker  

DataNode/  TaskTracker  

DataNode/  TaskTracker  

DataNode/  TaskTracker  

DataNode/  TaskTracker  

DataNode/  TaskTracker  

DataNode/  TaskTracker  

DataNode/  TaskTracker  

ZooK

eepe

r  ZooK

eepe

r  

60  DataNode/  TaskTracker  Nodes      

Redundant  NameNode  and  JobTracker  

10Gb/s  network  bandwidth/node  

Up  to  160  Gb/s  bandwidth  for  data  

ingesWon  •  8GB/s  storage  bandwidth  •  Flexible  storage  capacity  

SM15000  

Page 11: CC-4009, "Optimizing Hadoop Deployments with SeaMicro SM15000" by Satheesh Nanniyur and Anil Rao

|      RIGHT  SIZING  HADOOP  DEPLOYMENTS    11  

SEAMICRO  REFERENCE  ARCHITECTURE  FOR  APACHE®  HADOOP™  

!  Hadoop  HA  with  NameNode  and  JobTracker  AcRve/Standby  ConfiguraRon  

!  Up  to  60  DataNode/TaskTracker  nodes,  512  x86  cores,  960  TB  raw  capacity,  10  Gb/s  Internode  bandwidth  and  160  Gb/s  uplink  bandwidth  in  28  RU  and  5.8    kW  

Highly  Available  AcRve/Standby  NameNode  

Highly  Available  AcRve/Standby  JobTracker  

DataNode/TaskTracker  

SM15000  Internal  Drives  for  NameNode  and  JobTracker  

SoluWon  Components  •  SeaMicro  SM15000  AMD  Opteron  or  

Intel  Xeon  (Ivy  Bridge)  CPU  chassis  •  32  GB  memory  per  node  •  2  10GbE  Network  cards  •  8  Storage  Controller  cards  •  4  FS  4060-­‐L  enclosures  with  SAS  

zoning  enabled  •  60  4TB  3.5”  SAS  or  SATA  drives  per  

enclosure    •  Any  Hadoop  distribuRon  (CDH,  HDP,  

MapR,  Apache  Hadoop  etc.)  •  ZooKeeper  for  NameNode  and  

JobTracker  HA  

Page 12: CC-4009, "Optimizing Hadoop Deployments with SeaMicro SM15000" by Satheesh Nanniyur and Anil Rao

|      RIGHT  SIZING  HADOOP  DEPLOYMENTS    12  

SM15000  HADOOP  DEPLOYMENT  

SoluWon  Components   SM15000  Intel  Xeon  

SM15000  AMD  Opteron  

Use  case   Performance  OpRmized  

Cost  OpRmized  

Racks   <1   <1  

Servers   64   64  

Cores   256   512  

DRAM   2TB   4  TB  

Hard  Drives   240/960  TB   240/960  TB  

Cable  Management   0  RU   0  RU  

ToR  SwRches   None   None  

Downlink  Network  Cables   None   None  

Data  Nodes   60   60  

Bandwidth  per  node   10  Gb/s   10  Gb/s  

Uplink  bandwidth   Up  to  160  Gb/s   Up  to  160  Gb/s  

Storage  bandwidth   8  GigaBytes/s   8  GigaBytes/s  

Page 13: CC-4009, "Optimizing Hadoop Deployments with SeaMicro SM15000" by Satheesh Nanniyur and Anil Rao

|      RIGHT  SIZING  HADOOP  DEPLOYMENTS    13  

SM15000  –  INDUSTRY’S  ONLY  FLEXIBLE  SYSTEM  FOR  OPTIMIZING  HADOOP  CLUSTERS  

Map  

Map  

Map  

Reduce  

Reduce  

Storage  Intensive  

Compute  Intensive  

Network  Intensive  

Compute  Intensive  

Storage  Intensive  

HDFS  Input  

Shuffle   HDFS  Output  

Map  and  Intermediate  Data  Write  

Reduce  

Flexible  scale-­‐out  storage  with  over  1400  spindles  and  5  Petabytes  of  capacity  

Up  to  512  x86  cores  with  4TB  DRAM  per  Fabric  Server  in  10RU  

10  Gpbs  Inter-­‐Node  Bandwidth  per  server    

160  Gbps    shared    uplink  for  Inter-­‐Rack  traffic  

Page 14: CC-4009, "Optimizing Hadoop Deployments with SeaMicro SM15000" by Satheesh Nanniyur and Anil Rao

|      RIGHT  SIZING  HADOOP  DEPLOYMENTS    14  

SM15000  HADOOP  PERFORMANCE  BETTER  THAN  COMPETITIVE  OFFERINGS  

!  77%  less  power  per  node  !  30%  less  power  per  core  !  63%  more  data  sorted  per  second  per  Wat  than  Large  Vendor  

SeaMicro  SM15000  with  64  Nodes  based  on    Intel  Xeon®  (Ivy  Bridge)  1265  L-­‐v2  CPU,  32  GB  memory,  and  2  3.5”  3TB  SAS  drives  per  node  

SM15000   Large  Vendor  

Terasort  CompleRon  

7  min  13  seconds  

8  min  33  seconds  

Nodes   62  incl.  HA   18  

CPU  Cores   248   216  

Power   5800  W   7200  W  

MB/s  per  Wat   0.4   0.24  

MB/s  per  CPU  core   9.3   9.0  

Wats/Node   94  W   400  W  

Wats/Core   23  W   33  W  

CompeRRve  soluRon  consists  of  dual  socket  2U  rack-­‐mount  servers  with  Intel  Xeon  E5-­‐2667  2.9  GHz  octal  core  CPUs  with  64  GB  memory,  16  disks,  and  4  GbE  network  links  per  node  

0  0.05  0.1  

0.15  0.2  0.25  0.3  0.35  0.4  

SM15000   Large  Vendor  

MB/s  p

er  W

a`  

Terasort  -­‐  Sort  rate  per  Wa`  

Page 15: CC-4009, "Optimizing Hadoop Deployments with SeaMicro SM15000" by Satheesh Nanniyur and Anil Rao

|      RIGHT  SIZING  HADOOP  DEPLOYMENTS    15  

WAYFAIR.COM:  PERSONALIZED  SHOPPING  EXPERIENCE  FOR  “A  ZILLION  THINGS”  

Applica'ons:  Apache®  Hadoop™,  SQL  server,  PHP  

!  Challenge  ‒  Space  and  power  constraints  hindered  availability  of  “shared  

nothing”  servers  for  development  ‒  Too  costly  in  space  and  power  to  use  tradiRonal  servers  for  the  

number  of  servers  required  and  accurately  test  applicaRon  performance  

!  SoluRon  ‒  SeaMicro  SM  high  density  server    ‒  256  Intel®  Xeon®  cores  in  10  RU  system  ‒  64  servers,  1.28  Tbps  SeaMicro  Freedom™  Supercompute  Fabric  

!  Results  ‒  Reduced  development  cycles  and  shortened  Rme  to  market  for  

new  products  ‒  Increased  producRvity  of  development  engineers  by  providing  

abundant  access  to  “shared  nothing”  servers  versus  developing  on  virtualized  server  farms  

‒  Eliminated  unnecessary  equipment  such  as  top  of  rack  switches  and  terminal  servers;  simplified  network  and  power  cabling  

“The  SeaMicro  SM  server  is  helping  us  operate  at  a  large  scale  and  fast  pace.    The  key  benefits  are  reduced  operaRng  costs  and  increased  efficiency  for  our  big  data  development  infrastructure.    It  provides  the  highest  density  and  flexibility  while  slashing  energy  consumpRon:  256  Intel  Xeon  cores,  64  hosts.    It  consumes  50  percent  less  power  and  doubled  our  compuRng  capacity...”    Ben  Clark,  Director  of  So@ware  Engineering  

Page 16: CC-4009, "Optimizing Hadoop Deployments with SeaMicro SM15000" by Satheesh Nanniyur and Anil Rao

|      RIGHT  SIZING  HADOOP  DEPLOYMENTS    16  

EHARMONY:  INCREASE  COMPUTING  WHILE    REDUCING  TOTAL  COST  OF  OWNERSHIP  

Applica'ons:  Apache®  Hadoop™  

!  Challenge  ‒ Provide  cost  effecRve  compuRng    plaxorm  for  Apache  Hadoop  

‒ Reduce  costs  incurred  from  external    cloud  compuRng  

!  SoluRon  ‒  SeaMicro  SM10000-­‐64  high  density  server    ‒ 512  Intel®  Atom™  cores  in  10  RU  system  

!  Results  ‒ Reduce  TCO  by  more  than  74  percent  ‒  Save  thousands  per  month  spent  on    cloud  compuRng  service  

‒ URlize  compuRng  resources  7  x  24  

“We  purchased  SeaMicro  servers  and  immediately  reduced  our  operaRng  expenses…The  system  has  been  in  place  for  over  two  years,  and  we  have  had  zero  down  Rme.”    Cormac  Twomey,  Data  Center  Opera'ons  

Page 17: CC-4009, "Optimizing Hadoop Deployments with SeaMicro SM15000" by Satheesh Nanniyur and Anil Rao

|      RIGHT  SIZING  HADOOP  DEPLOYMENTS    17  

AMD  SEAMICRO  PARTNERS  WITH  INDUSTRY  LEADERS  

Partner

Hadoop   OS/Hypervisor   OpenStack  

Page 18: CC-4009, "Optimizing Hadoop Deployments with SeaMicro SM15000" by Satheesh Nanniyur and Anil Rao

Discussion  

Page 19: CC-4009, "Optimizing Hadoop Deployments with SeaMicro SM15000" by Satheesh Nanniyur and Anil Rao

|      RIGHT  SIZING  HADOOP  DEPLOYMENTS    19  

Zone  1  

Zone  2  

FS  4060-­‐L  SAS  ZONING  

!  SAS  zoning  allows  the  logical  parRRoning  of  FS  4060-­‐L  enclosure  into  two  30  disk  enclosures  in  4U  ‒ Provides  fully  independent    end  to  end  path  to  all  30  drives  in  each  zone  ‒ Up  to  2  S-­‐cards  connected  to  a  storage  enclosure  

FS  4060-­‐L  

SM15000  S-­‐card   SM15000  S-­‐card  

FS  4060-­‐L  with  SAS  zoning  

Page 20: CC-4009, "Optimizing Hadoop Deployments with SeaMicro SM15000" by Satheesh Nanniyur and Anil Rao

|      RIGHT  SIZING  HADOOP  DEPLOYMENTS    20  

DISCLAIMER  &  ATTRIBUTION  

The  informaRon  presented  in  this  document  is  for  informaRonal  purposes  only  and  may  contain  technical  inaccuracies,  omissions  and  typographical  errors.    

The  informaRon  contained  herein  is  subject  to  change  and  may  be  rendered  inaccurate  for  many  reasons,  including  but  not  limited  to  product  and  roadmap  changes,  component  and  motherboard  version  changes,  new  model  and/or  product  releases,  product  differences  between  differing  manufacturers,  so{ware  changes,  BIOS  flashes,  firmware  upgrades,  or  the  like.  AMD  assumes  no  obligaRon  to  update  or  otherwise  correct  or  revise  this  informaRon.  However,  AMD  reserves  the  right  to  revise  this  informaRon  and  to  make  changes  from  Rme  to  Rme  to  the  content  hereof  without  obligaRon  of  AMD  to  noRfy  any  person  of  such  revisions  or  changes.    

AMD  MAKES  NO  REPRESENTATIONS  OR  WARRANTIES  WITH  RESPECT  TO  THE  CONTENTS  HEREOF  AND  ASSUMES  NO  RESPONSIBILITY  FOR  ANY  INACCURACIES,  ERRORS  OR  OMISSIONS  THAT  MAY  APPEAR  IN  THIS  INFORMATION.    

AMD  SPECIFICALLY  DISCLAIMS  ANY  IMPLIED  WARRANTIES  OF  MERCHANTABILITY  OR  FITNESS  FOR  ANY  PARTICULAR  PURPOSE.  IN  NO  EVENT  WILL  AMD  BE  LIABLE  TO  ANY  PERSON  FOR  ANY  DIRECT,  INDIRECT,  SPECIAL  OR  OTHER  CONSEQUENTIAL  DAMAGES  ARISING  FROM  THE  USE  OF  ANY  INFORMATION  CONTAINED  HEREIN,  EVEN  IF  AMD  IS  EXPRESSLY  ADVISED  OF  THE  POSSIBILITY  OF  SUCH  DAMAGES.  

 

ATTRIBUTION  

©  2013  Advanced  Micro  Devices,  Inc.  All  rights  reserved.  AMD,  the  AMD  Arrow  logo,  AMD  Opteron,  Freedom  and  combinaRons  thereof  are  trademarks  of  Advanced  Micro  Devices,  Inc.  in  the  United  States  and/or  other  jurisdicRons..  Other  names  are  for  informaRonal  purposes  only  and  may  be  trademarks  of  their  respecRve  owners.