GreenDisc: A HW/SW Energy Optimization Framework in Globally Distributed Computation Marina  Zapater†,  José  L.  Ayala*,  José  M.  Moya‡   †CEI  Campus  Moncloa,  UCM-­‐UPM   *Universidad  Complutense  de  Madrid   ‡  Universidad  Politécnica  de  Madrid     Marina  Zapater  |  UCAMI  2012   1   

Outline •  Mo?va?on   •  Proposed  solu?on   –  The  GreenDisc  PlaIorm   •  Holis?c  Op?miza?on  Approach   –  Power  op?miza?on  of  all  implied  agents   •  Conclusions     •  Future  Work   Marina  Zapater  |  UCAMI  2012   2   

Motivation •  The  key  to  next-­‐genera?on  e-­‐Health  solu?ons  are   wearable  personal  health  systems   –  Biomedical  monitoring    (European  ini?a?ves)   •  World-­‐wide  sensor  deployment:   –  Very  large  amount  of  data   •  Electrocardiogram  sensor  (ECG),  Electromyogram  (EMG),  O2  and   CO2 ,  temperature,  lactate,  movement  sensors.   –  Raw  data  must  be  turned  into  useful  informa?on   Marina  Zapater  |  UCAMI  2012   3    FET      “Guardian  Angels”.  European  ini?a?ve:  FET  Flagships  2013  -­‐  hap://www.ga-­‐project.eu/home   

Motivation •  Need  for  an  accurate,  integrated  and  long-­‐term   assessment  and  feedback.   –  Acquire,  monitor  and  analize  data  24/7   •  Need  for  an  applica?on-­‐specific  architecture     –  But  also  provide  flexibility  and  enough  performance   •  Current  issues:  energy  and  heat   –  High  energy  consump?on  and  short  baaery  lifespan  of  sensor   nodes.     •  Heat  produced  by  sensor  nodes  for  biomedical  applica?ons   –  Tackle  the  computa?on  needs  to  obtain  useful  informa?on   from  all  data   •  Energy  consump?on  at  the  Data  Center  level  must  not  put  at  stake     e-­‐Health  deployments   Marina  Zapater  |  UCAMI  2012   4   Current needs and issues 

Contributions •  A  HW/SW  energy  op?miza?on  framework   –  Versa?le  plaIorm  that  integrates  processing,  analysis  and   wireless  communica?on  of  biomedical  data.   –  Both  at  the  WSN-­‐level  and  at  the  Data  Center  level   •  Supports  e-­‐Health  applica?ons  (and  future   evolu?ons)  with  lower  costs  and  shorter  ?me-­‐to-­‐ market   Marina  Zapater  |  UCAMI  2012   5   GreenDisc Platform 

GreenDisc Platform •  Based  on  Wireless  Body  Sensor  Networks  (WBSN)   –  All  sensors  transmit  to  a  PDA   Marina  Zapater  |  UCAMI  2012   6   Context and System Architecture 

GreenDisc Platform •  Based  on  Wireless  Body  Sensor  Networks  (WBSN)   –  All  sensors  transmit  to  a  PDA   •  Computa?on  takes  place  at  Data  Centers   –  Data  storage  and  processing   Marina  Zapater  |  UCAMI  2012   7   Context and System Architecture 

GreenDisc Platform •  Power  op?miza?on  in  the  processing  nodes  of  the   WBSN   –  Design  of  embedded  processors  for  signal  processing   –  Op?miza?on  at  the  radio  interface   –  Design  automa?on  of  applica?ons  for  the  processing  node     •  Power  op?miza?on  in  data  centers   Marina  Zapater  |  UCAMI  2012   8   Holistic Optimization Approach 

Holistic Optimization •  Design  of  embedded  processors  for  signal  processing   –  Architectural  modifica?ons  considering  the  applica?on   mapping,  the  execu?on  profile,  and  the  compiler   op?miza?ons   –  Reducing  the  energy  consump?on  of  the  main  energy   consump?on  sources   •  Selec?on  of    instruc?on  memory  architecture   •  Design  of  func?onal  units  with  tunable  architecture  (dynamic   reconfigura?on)   Marina  Zapater  |  UCAMI  2012   9   Processing nodes of the WBSN (I) 

Holistic Optimization •  Instruc?on  memory   produces  highest  energy   consump?on   •  Proper  selec?on  of   instruc?on  memory   architecture  impacts  energy   consump?on   –  SPM  -­‐  scratch  pad  memory     –  CELB  -­‐  central  loop     buffer   –  CLLB-­‐  clustered  loop  buffer   Marina  Zapater  |  UCAMI  2012   10   Processing nodes of the WBSN (I) 

Holistic Optimization •  Power  op?miza?on  in  the  radio  interface   –  Reducing  the  amount  of  informa?on  to  transmit:   •  Framework  for  signal  analysis  to  develop  compressed  sensing   techniques  for  several  bio-­‐signals   •  Case  studies  for  monitoring  bio-­‐signals  with  a  QoS  study   –  Reduce  the  overhead  of  the  transmision  protocol:   •  Study  of  the  impact  of  tuning  several  parameters  of  the  MAC  layer   (development  of  802.15.4  MAC  analy?cal  model)   Marina  Zapater  |  UCAMI  2012   11   Processing nodes of the WBSN (II)  Compressed  sensing:    signal  processing  technique  for  efficiently  acquiring  and  reconstruc?ng  a   signal.  Uses  the  signal  sparseness  or  compressibility  in  some  domain,  allowing  the  en?re  signal  to  be   determined  from  rela?vely  few  measurements.   Candes,  E.  J.  and  Wakin,  M.  B.  (2008)  An  IntroducEon  to  Compressive  Sampling.   IEEE  Signal  Processing  Magazine.  Vol  2  (pp.  21-­‐30)   

Holistic Optimization •  Design  automa?on  of  applica?ons  in  the  processing   node   –  Aims  to  reduce  the  amount  of  data  transmiaed  to  the   backbone   –  Providing  a  generic  high-­‐level  model  of  the  architecture   •  Analysis  of  the  impact  of  design  parameters  in  power   consump?on   •  Framework  for  automa?c  design  and  op?miza?on  of   applica?ons       Marina  Zapater  |  UCAMI  2012   12   Processing nodes of the WBSN (III) 

Holistic Optimization •  Implementa?on  of  several  resource  managing   techniques  at  different  abstrac?on  levels   •  Exploi?ng  the  heterogeneity  of  applica?ons  and   compu?ng  resources  for  energy  minimiza?on.   –  Proper  usage  of  heterogeneity  can  lead  to  significant   energy  savings   •  Analysis  of  cooling  mechanisms  and  development  of   control  techniques     Marina  Zapater  |  UCAMI  2012   13   Energy Optimization in Data Centers 

Holistic Optimization •  Poten?al  benefits  of  workload  characteriza?on  and   dynamic  assignment  policies   Marina  Zapater  |  UCAMI  2012   14   Energy Optimization in Data Centers 

Conclusion and Future Work •  This  paper  shows  how  the  GreenDisc  plaIorm  can   op?mize  the  energy  consump?on  of  next-­‐genera?on   e-­‐Health  applica?on  by  combining  the  usage  of:   –  Op?miza?on  policies  at  several  levels  of  the  WBSN   –  Agressive  energy  efficiency  policies  at  the  Data  Center   •  Future  work  will  integrate  all  steps  of  the  plaIorm   and  show  the  overall  savings  for  a  par?cular  e-­‐Health   workload.   Marina  Zapater  |  UCAMI  2012   15   

Questions? Marina  Zapater  |  UCAMI  2012   16   Thank  you  for    your  a9en:on   Marina  Zapater   Laboratorio  de  Sistemas  Integrados  (LSI)   Universidad  Politécnica  de  Madrid   [email protected]   hap://greenlsi.die.upm.es