HPC	
  +	
  Virtualiza.on?
Joseph	
  Antony	
   (NCI)	
  
joseph.antony@anu.edu.au	
   	
  
	
   	
  
	
  

Disclaimer:	
  Views	
  expressed	
  are	
  en.rely	
  mine	
  	
  
WHAT	
  IS	
  NCI?	
  
NCI	
  –	
  an	
  overview	
  
Mission:	
  

	
  
	
  
Research	
  Outcomes	
  

• 	
  to	
  foster	
  ambi8ous	
  and	
  aspira8onal	
  
research	
  objec8ves	
  and	
  to	
  enable	
  their	
  
realisa8on,	
  in	
  the	
  Australian	
  context,	
  through	
  
world-­‐class,	
  high-­‐end	
  compu8ng	
  services	
  
• 	
  being	
  driven	
  by	
  research	
  objec8ves	
  	
  
• 	
  a	
  comprehensive,	
  ver8cally-­‐integrated	
  
research	
  service	
  
• 	
  providing	
  na8onal	
  access	
  on	
  priority	
  and	
  
merit,	
  and	
  
• 	
  being	
  built	
  on,	
  and	
  sustained	
  by,	
  a	
  
collabora8on	
  of	
  na8onal	
  organisa8ons	
  and	
  
research-­‐intensive	
  universi8es	
  

Research Objectives

NCI	
  is:	
  

	
  
Communi.es	
  and	
  	
  
Ins.tu.ons/	
  
Access	
  and	
  Services	
  
	
  
Exper.se	
  Support	
  	
  
and	
  	
  
Development	
  
	
  
Digital	
  Laboratories	
  
	
  
Data	
  Centric	
  Services	
  	
  
	
  
Compute	
  (HPC/Cloud)	
  
and	
  	
  
Data	
  Infrastructure	
  
Climate Science has a solution
Integrated
Intimately connected
Robust
Accessible

Raijin

NCI
cloud

NCI + CoE
technical
Impact !
ACCESS
• A collaborative tool
• Under svn
• Co-support
• CoE/BoM/CSIRO PhDs
• Shared research(ers)

CMIP-5
• A collaborative data set
• Co-supported
• Shared analyses
• CoE/BoM/CSIRO PhDs
• Shared research(ers)

Raijin

NCI
cloud

NCI + CoE
technical
In	
  case	
  you’re	
  wondering	
  where	
  are	
  we	
  located?	
  

•  In	
  the	
  na.on’s	
  capital,	
  at	
  its	
  na.onal	
  
university	
  …	
  
HPC	
  Virtualiza.on?	
  

•  HPC	
  procurements	
  typically	
  involve	
  major	
  CAPEX	
  
spend	
  on	
  Big	
  Iron	
  for	
  aWacking	
  grand	
  challenge	
  
problems	
  
•  Typically	
  most	
  large	
  HPC	
  centers	
  have	
  
–  Capability	
  machines:	
  lands	
  in	
  the	
  TOP500	
  around	
  
#10	
  to	
  20.	
  These	
  have	
  special	
  purpose	
  
architectures	
  and	
  accelerators	
  
–  	
  Work-­‐horse	
  machines:	
  usually	
  x86	
  +	
  IB	
  
HPC	
  Virtualiza.on?	
  (1)	
  

•  HPC	
  procurements	
  typically	
  involve	
  major	
  CAPEX	
  
spend	
  on	
  Big	
  Iron	
  for	
  aWacking	
  grand	
  challenge	
  
problems	
  
•  Typically	
  most	
  large	
  HPC	
  centers	
  have	
  
–  Capability	
  machines:	
  lands	
  in	
  the	
  TOP500	
  around	
  
#10	
  to	
  20.	
  These	
  have	
  special	
  purpose	
  
architectures	
  and	
  accelerators	
  
–  	
  Work-­‐horse	
  machines:	
  usually	
  x86	
  +	
  IB	
  
HPC	
  Virtualiza.on?	
  (2)	
  

From ‘CERN Data Center Evolution’
Looming	
  Iceberg	
  ….	
  

Dribble.com – “Mr. Iceberg meets the Titanic”
Changing	
  Environment	
  for	
  HPC	
  Centers	
  

HMAS HPC
File system storage

Use/Re-use: Metadata
Apriori Analytics
Deep Storage Retrieval
Long Lived Artifacts
Multiple communities
Publishing
Data Replication
From http://www.exascale.org/
Is	
  there	
  life	
  beyond	
  a	
  batch-­‐oriented	
  system?	
  

•  HPC	
  Centers	
  will	
  be	
  forced	
  to	
  evolve	
  beyond	
  
batch-­‐oriented	
  systems	
  due	
  to	
  an	
  immovable	
  
iceberg	
  –	
  ‘Big	
  Data’	
  
•  The	
  NCI	
  moved	
  to	
  virtualiza.on	
  in	
  1999	
  to	
  handle	
  
non-­‐tradi.onal	
  workloads	
  due	
  to	
  complex	
  data	
  
lifecycles	
  
–  Satellite	
  Image	
  Processing	
  
–  CMIP5	
  Climate	
  Data	
  Processing	
  
–  Genomics	
  Assembly	
  Pipelines	
  
–  N-­‐to-­‐N	
  Cancer	
  Genomics	
  Comparisons	
  
–  Interac.ve	
  Volume	
  Rendering	
  
–  Trawling	
  YouTube	
  and	
  Analyzing	
  Birthday	
  Videos	
  
Engaging	
  with	
  Priority	
  Research:	
  Environment.	
  
Goal:	
  
•  To	
  provide	
  a	
  single	
  high-­‐performance	
  
compu.ng	
  for	
  environment	
  research	
  
	
  
Partners:	
  	
  
•  CSIRO,	
  GA,	
  Bureau,	
  Research	
  Community	
  
•  Lockheed-­‐Mar.n,	
  GA,	
  NCI,	
  VPAC	
  
	
  
Requirements:	
  
•  Provide	
  na.onal	
  processing	
  environment	
  
for	
  key	
  satellite	
  data	
  (eg.	
  SEADAS)	
  
•  Provide	
  collabora.ve	
  environment	
  for	
  
tools	
  that	
  produce	
  reference	
  Digital	
  
Eleva.on	
  Maps	
  
•  Provide	
  	
  data	
  environment	
  for	
  fast,	
  easy	
  
na.onal-­‐nested	
  grid.	
  
	
  
	
  
	
  
	
  	
  
Engaging	
  with	
  Priority	
  Research:	
  Environment.	
  
Data	
  Intensive	
  Ac8vity	
  
•  Data	
  Processing	
  Intensive	
  Pipelines	
  (SEADAS)	
  over	
  large	
  data	
  raw	
  imagery	
  
Key	
  ini8al	
  datasets	
  
•  LANDSAT	
  archive	
  
•  MODIS	
  
•  DEMs	
  (9s,	
  3s,	
  1s)	
  
•  LIDAR	
  
•  Deriva.ve	
  products	
  
Data	
  Intensive	
  query	
  and	
  analysis	
  environment	
  Eg.	
  Hadoop	
  over	
  nested	
  grids.	
  
Engaging	
  with	
  Priority	
  Research:	
  Environment.	
  
• 
• 
• 

Collabora.on	
  to	
  provide	
  beWer	
  and	
  common	
  processing	
  environments	
  
Next	
  genera.on	
  of	
  tools	
  (able	
  to	
  operate	
  at	
  na.onal	
  scale)	
  
New	
  aggrega.on	
  of	
  tools	
  and	
  techniques	
  under	
  TERN	
  e-­‐MAST	
  project.	
  	
  
Engaging	
  with	
  Priority	
  Research:	
  Climate	
  cont.	
  
Research	
  Highlights:	
  Life	
  Sciences	
  
Research	
  Highlights:	
  Materials/Nanotechnology	
  
Research	
  Highlights:	
  Physical	
  and	
  Chemical	
  Sciences	
  
Research	
  Highlights:	
  Physical	
  and	
  Chemical	
  Sciences	
  
Research	
  Highlights:	
  Physical	
  and	
  Chemical	
  Sciences	
  
Research	
  Highlights:	
  Earth	
  Sciences	
  
From http://www.exascale.org
NCI’s	
  Cloud	
  Node	
  Architecture	
  from	
  Dell	
  
NCI’s	
  Science	
  Cloud	
  Building	
  Blocks	
  from	
  Dell	
  

•  From	
  Dell	
  using	
  C8000	
  chassis	
  building	
  blocks	
  for	
  
OpenStack	
  compute,	
  Swik	
  (S3)	
  and	
  Ceph	
  (EBS)	
  
•  Hyperscale	
  meets	
  Exascale	
  …	
  (?)	
  
Summary	
  

•  HPC	
  in	
  the	
  Cloud	
  –	
  Clusters-­‐in-­‐the-­‐cloud	
  
–  Offload,	
  Burs.ng	
  

•  Big	
  Data	
  needs	
  
–  Complex,	
  long-­‐lived	
  data	
  processing	
  
–  Community	
  ecosystems	
  

•  Service	
  provider	
  abstrac.on	
  
END	
  

Virtualization for HPC at NCI

  • 1.
    HPC  +  Virtualiza.on? Joseph  Antony   (NCI)   joseph.antony@anu.edu.au           Disclaimer:  Views  expressed  are  en.rely  mine    
  • 2.
  • 3.
    NCI  –  an  overview   Mission:       Research  Outcomes   •   to  foster  ambi8ous  and  aspira8onal   research  objec8ves  and  to  enable  their   realisa8on,  in  the  Australian  context,  through   world-­‐class,  high-­‐end  compu8ng  services   •   being  driven  by  research  objec8ves     •   a  comprehensive,  ver8cally-­‐integrated   research  service   •   providing  na8onal  access  on  priority  and   merit,  and   •   being  built  on,  and  sustained  by,  a   collabora8on  of  na8onal  organisa8ons  and   research-­‐intensive  universi8es   Research Objectives NCI  is:     Communi.es  and     Ins.tu.ons/   Access  and  Services     Exper.se  Support     and     Development     Digital  Laboratories     Data  Centric  Services       Compute  (HPC/Cloud)   and     Data  Infrastructure  
  • 4.
    Climate Science hasa solution Integrated Intimately connected Robust Accessible Raijin NCI cloud NCI + CoE technical
  • 5.
    Impact ! ACCESS • A collaborativetool • Under svn • Co-support • CoE/BoM/CSIRO PhDs • Shared research(ers) CMIP-5 • A collaborative data set • Co-supported • Shared analyses • CoE/BoM/CSIRO PhDs • Shared research(ers) Raijin NCI cloud NCI + CoE technical
  • 6.
    In  case  you’re  wondering  where  are  we  located?   •  In  the  na.on’s  capital,  at  its  na.onal   university  …  
  • 7.
    HPC  Virtualiza.on?   • HPC  procurements  typically  involve  major  CAPEX   spend  on  Big  Iron  for  aWacking  grand  challenge   problems   •  Typically  most  large  HPC  centers  have   –  Capability  machines:  lands  in  the  TOP500  around   #10  to  20.  These  have  special  purpose   architectures  and  accelerators   –   Work-­‐horse  machines:  usually  x86  +  IB  
  • 8.
    HPC  Virtualiza.on?  (1)   •  HPC  procurements  typically  involve  major  CAPEX   spend  on  Big  Iron  for  aWacking  grand  challenge   problems   •  Typically  most  large  HPC  centers  have   –  Capability  machines:  lands  in  the  TOP500  around   #10  to  20.  These  have  special  purpose   architectures  and  accelerators   –   Work-­‐horse  machines:  usually  x86  +  IB  
  • 9.
    HPC  Virtualiza.on?  (2)   From ‘CERN Data Center Evolution’
  • 10.
    Looming  Iceberg  ….   Dribble.com – “Mr. Iceberg meets the Titanic”
  • 11.
    Changing  Environment  for  HPC  Centers   HMAS HPC File system storage Use/Re-use: Metadata Apriori Analytics Deep Storage Retrieval Long Lived Artifacts Multiple communities Publishing Data Replication
  • 12.
  • 13.
    Is  there  life  beyond  a  batch-­‐oriented  system?   •  HPC  Centers  will  be  forced  to  evolve  beyond   batch-­‐oriented  systems  due  to  an  immovable   iceberg  –  ‘Big  Data’   •  The  NCI  moved  to  virtualiza.on  in  1999  to  handle   non-­‐tradi.onal  workloads  due  to  complex  data   lifecycles   –  Satellite  Image  Processing   –  CMIP5  Climate  Data  Processing   –  Genomics  Assembly  Pipelines   –  N-­‐to-­‐N  Cancer  Genomics  Comparisons   –  Interac.ve  Volume  Rendering   –  Trawling  YouTube  and  Analyzing  Birthday  Videos  
  • 15.
    Engaging  with  Priority  Research:  Environment.   Goal:   •  To  provide  a  single  high-­‐performance   compu.ng  for  environment  research     Partners:     •  CSIRO,  GA,  Bureau,  Research  Community   •  Lockheed-­‐Mar.n,  GA,  NCI,  VPAC     Requirements:   •  Provide  na.onal  processing  environment   for  key  satellite  data  (eg.  SEADAS)   •  Provide  collabora.ve  environment  for   tools  that  produce  reference  Digital   Eleva.on  Maps   •  Provide    data  environment  for  fast,  easy   na.onal-­‐nested  grid.            
  • 16.
    Engaging  with  Priority  Research:  Environment.   Data  Intensive  Ac8vity   •  Data  Processing  Intensive  Pipelines  (SEADAS)  over  large  data  raw  imagery   Key  ini8al  datasets   •  LANDSAT  archive   •  MODIS   •  DEMs  (9s,  3s,  1s)   •  LIDAR   •  Deriva.ve  products   Data  Intensive  query  and  analysis  environment  Eg.  Hadoop  over  nested  grids.  
  • 17.
    Engaging  with  Priority  Research:  Environment.   •  •  •  Collabora.on  to  provide  beWer  and  common  processing  environments   Next  genera.on  of  tools  (able  to  operate  at  na.onal  scale)   New  aggrega.on  of  tools  and  techniques  under  TERN  e-­‐MAST  project.    
  • 18.
    Engaging  with  Priority  Research:  Climate  cont.  
  • 19.
  • 20.
  • 21.
    Research  Highlights:  Physical  and  Chemical  Sciences  
  • 22.
    Research  Highlights:  Physical  and  Chemical  Sciences  
  • 23.
    Research  Highlights:  Physical  and  Chemical  Sciences  
  • 24.
  • 25.
  • 26.
    NCI’s  Cloud  Node  Architecture  from  Dell  
  • 27.
    NCI’s  Science  Cloud  Building  Blocks  from  Dell   •  From  Dell  using  C8000  chassis  building  blocks  for   OpenStack  compute,  Swik  (S3)  and  Ceph  (EBS)   •  Hyperscale  meets  Exascale  …  (?)  
  • 28.
    Summary   •  HPC  in  the  Cloud  –  Clusters-­‐in-­‐the-­‐cloud   –  Offload,  Burs.ng   •  Big  Data  needs   –  Complex,  long-­‐lived  data  processing   –  Community  ecosystems   •  Service  provider  abstrac.on  
  • 29.