Provisioning and Resource Allocaqon for Green Clouds

14 downloads 2226 Views 931KB Size Report
cloud to address oscillatory peaks of workload. -‐ In this context ... to private clouds and a provisioning strategy .... more VMs, it uses the public cloud to host the.
Provisioning  and  Resource   Alloca2on    for  Green  Clouds      

Guilherme  Arthur  Geronimo,  Jorge  Werner,  Carlos   Becker  Westphall,  Carla  Merkle  Westphall,   Leonardo  Defen2     Networks  and  Management  Laboratory   Federal  University  of  Santa  Catarina   JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

1  

Summary   1  -­‐  Introduc2on   2  –  State  of  the  Art   3  –  Model   4  –  Proposal  (Results)   5  -­‐  Conclusions   6  –  Future  Works   7  –  Some  References   JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

2  

(J.  Werner,    G.  A.  Geronimo,  C.  B.  Westphall  et  al.  CLEI  EJ  2012)  

JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

3  

1  Introduc2on   -­‐   The   aim   of   Green   Cloud   Compu2ng   is   to   achieve   a   balance   between   the   resource   consump2on   and   quality  of  service.   -­‐  Dynamic  provisioning  and  alloca2on  strategies  are   needed   to   regulate   the   internal   se_ngs   of   the   cloud  to  address  oscillatory  peaks  of  workload.   -­‐  In  this  context,  we  propose  strategies  to  op2mize   the   use   of   the   cloud   resources   without   decreasing   the  availability.   JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

4  

1  Introduc2on   -­‐   This   work   introduces   two   hybrid   strategies   based   on   a   distributed   system   management   model,   describes   the   base   strategies,   opera2on   principles,  tests,  and  presents  the  results.   -­‐   We   extended   CloudSim   to   simulate   the   organiza2on   model   upon   which   we   were   based   and   to   implement   the   strategies,   using   this   improved  version  to  validate  our  solu2on.     JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

5  

1  Introduc2on   -­‐   We   aim   to   propose   an   alloca2on   strategy   to  private  clouds  and  a  provisioning  strategy   for  Green  Clouds,  which  suits  the  oscillatory   workload  and  unexpected  peaks.       -­‐   We   will   focus   on   finding   a   solu2on   that   consumes   low   power   and   generates   acceptable  request  losses.   JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

6  

1  Introduc2on   Organiza2on  of  this  presenta2on:     -­‐   2.   comments   the   state   of   the   art   based   in   some   references;   -­‐  3.  explains  under  which  model  the  strategies  were   based;   -­‐  4.  presents  the  proposal,  tests,  and  the  results;   -­‐  5.  concludes  this  presenta2on;  and   -­‐  6.  addresses  some  future  works.   JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

7  

2  State  of  the  Art   -­‐   The   reference   [8]   uses   a   Dynamic   Voltage   Frequency   Scaling   (DVFS)   strategy   to   decrease   the   energy   consump2on   in   PMs   used  as  virtualiza2on  hosts.     -­‐   It   adapts   the   clock   frequency   of   the   CPUs   with  the  real  usage  of  the  PMs.  It  decreases   the   frequency   in   idle   nodes   and   increases   when  is  needed.     -­‐     JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

8  

2  State  of  the  Art   -­‐   The   workload   balance   strategy   for   clusters   in   [9],  tries  to  achieve  a  lower  energy  consump2on   unbalancing   the   cluster   workload,   genera2ng   idle  nodes  and  turning  off  them.   -­‐   The   paper   [10]   tries   to   decrease   the   hos2ng   costs  in  public  and/or  federated  clouds  using  the   costs  and  fines  in  contracts  as  metrics  to  bejer   allocate  the  resources.     JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

9  

3  Model   -­‐   Management   Systems   based   on   the   Organiza2on   Theory,   providing   the   means   to   describe   why   /   how   elements   of   the   cloud   environment   should   behave   to   achieve   global   system   objec2ves,   which   are   (among   others):   op2mum   performance,   reduce   opera2ng   costs,   appointment   of   dependence,   service   level   agreements,  and  energy  efficiency.       JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

10  

3  Model   -­‐   Managing   Cloud   through   the   principles   of   the   Organiza2on   Theory   provides   the   possibility   for   an   automa2c   con   figura2on   management   system,   since   adding   a   new   element   (e.g.,   V i r t u a l   M a c h i n e s ,   P h y s i c a l   M a c h i n e s ,   Uninterrupted   Power   Supply,   Air   Condi2oning)   is   just   a   majer   of   adding   a   new   service   on   the   Management  Group.       JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

11  

3  Model   -­‐   The   proposed   strategies   are   based   on   a   proac2ve   management   of   Clouds,   which   is   based   on   the   distribu2on   of   responsibili2es   in   holes,   as   seen   in   next   figure.   The   responsibility   of   management   of   the   cloud   elements   is   distributed   among   several   agents,   separated   in   holes,   and   each   agent   controls   individually,   a   Cloud  element  that  suits  him.           JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

12  

3  Model  

JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

13  

4  Proposal   -­‐   For   the   conscious   resource   provisioning   of   the   data   center,   we   propose   a   hybrid   strategy   that   uses   public   cloud   as   an   external   resource   used   to   mi2gate   probable   Service   level   Agreements   (SLAs)   breaches   due   to   unexpected   workload   peaks.       JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

14  

4  Proposal   -­‐   In   parallel,   to   the   op2mal   use   of   local   resources,   we   propose   a   strategy   of   dynamic   reconfigura2on   of   the   VMs   ajributes,  allocated  in  the  data  center.     -­‐  Given  the  distributed  model  presented  in   the   previous   sec2on,   we   use   the   Cloud   simula2on   tool   CloudSim   to   simulate   the   university  data  center  environment.     JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

15  

4  Proposal  (Alloca2on)   -­‐   The   resource   alloca2on   strategy   is   a   proposal   that   introduces   a   composi2on   of   two   other   approaches:   (1)   the   migra2on   of   VMs,   which   aims   to   focus   on   the   processing   of   cloud,   and   (2)   the   Dynamic   Reconfigura2on   of   VMs,   which   aims   to   relocate  dynamically  the  resources  used  by   the  VMs.     JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

16  

4    Proposal  (Alloca2on)    

(J.  Werner,    G.  A.  Geronimo,  C.  B.  Westphall  et  al.  CLEI  EJ  2012)   JANUARY  29TH,  SEVILLE,  SPAIN        

IARIA  ICN  2013    -­‐  GLOBENET  2013  

17  

4  Proposal  (Alloca2on)        

 

Parameter  

Value  

VM  –  Image  size  

1GB  

VM  -­‐  RAM  

256MB  

PM  -­‐  Engine  

Xen  

PM  -­‐  RAM  

8GB  

PM  -­‐  Frequency  

3.0GHZ  

PM  -­‐  Cores  

2  

PROPOSED  SCENARIO  CHARACTERISTCS     (J.  Werner,    G.  A.  Geronimo,  C.  B.  Westphall  et  al.  CLEI  EJ  2012)  

    JANUARY  29TH,  SEVLLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

18  

4  Proposal  (Alloca2on)  

(J.  Werner,    G.  A.  Geronimo,  C.  B.  Westphall  et  al.  CLEI  EJ  2012)   JANUARY  29TH,  SEVLLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

19  

4  Proposal  (Alloca2on)   1)  VMs  Migra+on  Strategy:  This  strategy  aims  to   reduce  power  consump2on  by  disabling  the  idle   PMs  of  the  Cloud.  To  induce  idleness  in  the  PMs,   the  VMs  are  migrated  and  concentrated  in  a  few   PMs.    2)  VMs  Dynamic  Reconfigura+on  Strategy:   It   adjusts   the   parameters   of   the   VM,   without   migra2ng   it   or   turning   it   off.   For   example,   we   can   increase   or   decrease   the   parameters   of   CPU   and  memory  allocated.    

 

JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

20  

4  Proposal  (Alloca2on)   Four   scenarios   were   simulated   in   order   to   seek   the   compara2ve   analysis   between   ordinary   cloud   (Scenario   1),   the   exis2ng   methods   (Scenarios:   2   and   3),   and   the   proposed   approach   (Scenario   4).   Those   were:   No   strategies;   Migra2ng   VMs   Strategy;   Reconfiguring   the   VMs   Strategy;   Reconfiguring  and  migra2ng  VMs  Strategy.       JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

21  

4  Proposal  (Alloca2on)   Scenario  

Reconf.  Strategy   Migrat.  Strategy   ConsumpMon  

Timeout  

1  

No  

No  

-­‐  

-­‐  

2  

No  

Yes  

84.3  %  

8.0  %  

3  

Yes  

No  

0.4  %  

-­‐  

4  

Yes  

Yes  

87.2  %  

7.3  %  

Table   I   (RESULTS   OF   ALLOCATION’S   SCENARIOS)   shows   the   results   of   the   simula2ons.   It   tells   what   strategies   were   used   in   each   scenario   and   what   percentage   (approximate)   reduc2on   was  obtained,  compared  to  the  scenario  without  strategies.   JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

22  

4  Proposal  (Provisioning)   -­‐  The  hybrid  strategy  is  based  on  the  merge   of   two   other   strategies,   the   On-­‐Demand   strategy   (OD)   and   the   Spare   Resources   strategy  (SR).     -­‐   It   aims   to   present   a   power   consump2on   lower   than   the   SR   strategy   and   a   wider   availability  than  the  OD  strategy.       JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

23  

4  Proposal  (Provisioning)   1)   On-­‐Demand   Strategy:   The   principle   of   OD   strategy   is   to   ac2vate   the   resources   when   they   are   needed.   In   our   case,   when   a   service   reaches   a   satura2on   threshold,   new   VMs   would  be  instan2ated.  When  there  is  no  more   space   to   instan2ate   new   VMs,   new   PMs   would   be   ac2vated   to   host   the   new   VMs.   The   opposite   also   applies;   when   a   threshold   of   idleness  is  reached,  the  idle  VMs  and  PMs  are   disabled.      

JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

24  

4  Proposal  (Provisioning)   On-­‐Demand   Strategy   proved   to   be   very   efficient   energe2cally,   since   it   maintains   a   minimum  amount  of  ac2ve  resources.  But,   it   has   been   shown   ineffec2ve   in   scenarios   that   had   sudden   spikes   in   demand,   because   the   process   to   ac2vate   the   resource   took   too   much   2me,   and   the   requests  ended  up  genera2ng  losses.      

JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

25  

4  Proposal  (Provisioning)   Spare   Resource   Strategy:   To   mi2gate   the   problem  of  requests  2meouts,  originated  by  a   long  ac2va2on  2me  of  resources,  we  adopt  the   strategy   SR,   whose   principle   is   reserve   idle   resources  ready  to  be  used.  In  our  case,  there   was   always   one   idle   VM   ready   to   process   the   incoming   requests   and   one   idle   PM   ready   to   instan2ate   new   VMs.   If   these   resources   were   used,  they  were  no  longer  considered  idle,  and   new  idle  resources  were  ac2vated.     JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

26  

4  Proposal  (Provisioning)   The   Spare   Resource   strategy   has   been   shown   effec2ve   in   remedying   unexpected   peak   demands,  but  it  showed  the  same  behavior  of   OD  strategy  in  cases  where  demand  rose  very   rapidly;   in   other   words,   the   idle   feature   was   not   enough   to   process   the   demand.   Another   nega2ve   point   was   the   energy   consump2on;   since   they   always   had   an   ac2ve   and   idle   resource,   the   consump2on   has   been   greater   than  the  OD  strategy.     JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

27  

4  Proposal  (Provisioning)   3)   Hybrid   Strategy:   Seeking   the   merger   of   the   strengths   of   the   previous   strategies   and   mi2ga2ng   its   shortcomings,   we   propose   a   hybrid   strategy.   This   strategy   aims  to  reduce  the  energy  consump2on  on   private   cloud   and   reduce   the   breakage   of   SLA’s  service  in  general.     JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

28  

4  Proposal  (Provisioning)   As  shown  in  next  figure,  the  cloud  enables  the   VMs   when   the   service   in   ques2on   reaches   its   satura2on   threshold,   just   as   the   OD   strategy.   When   more   PMs   space   is   unable   to   allocate   more  VMs,  it  uses  the  public  cloud  to  host  the   new  VMs  while  the  PM  is  passing  through  the   ac2va2on   process.   This   is   to   fulfill   requests   that   would   be   lost   during   the   ac2va2on   process.       JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

29  

4  Proposal  (Provisioning)  

(J.  Werner,    G.  A.  Geronimo,  C.  B.  Westphall  et  al.  CLEI  EJ  2012)   JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

30  

4  Proposal  (Provisioning)   4)   Tests   Results:   As   previously   men2oned,   we   performed   some   modifica2ons   to   the   CloudSim   code,   in   order   to   enable   the   simula2on  of  scenarios.  Before  we  started   the  simula2on,  we  defined  some  variables   for   the   scenario,   such   as   the   satura2on   threshold  and  idleness,  for  example.  Some   of  these  variables  are  shown  in  Table  II.       JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

31  

4  Proposal  (Provisioning)   Variable  

Value  

Satura2on  Threshold  (Load  1  minute)  

1.0    

Idleness  Threshold  (Load  1  minute)    

0.1    

Ac2va2on  VM  2me  (seconds)    

10    

Ac2va2on  PM  2me  (seconds)    

120    

Size  of  Request  (MI)    

1000  to  2000    

Number  of  PMs    

8  

Maximum  number  of  VMs  per  PMs    

5  

SLA  2meout  threshold  (seconds)    

10  

Table  II  (SIMULATION’S  VARIABLES)   JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

32  

4  Proposal  (Provisioning)   To  get  an  overview  of  how  each  strategy  would   behave   in   different   scenarios,   we   ran   a   series   of   tests   which   varied   (1)   the   amount   of   requests   and   (2)   the   size   of   the   requests.   To   maintain   the   defined   request   distribu2on   (explained   in   the   beginning   of   Sec2on   3),   we   used   mul2pliers   to   increase   the   requests.   Those  mul2pliers  started  from  2  to  20  in  steps   of  2  (2,  4,  6,  etc.).     JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

33  

4  Proposal  (Provisioning)   The   size   of   the   requests   ranged   from   1000   to   2000  MI  (Millions  Instruc2ons),  in  steps  of  100   (1000,   1100,   1200,   etc.).   This   way,   it   performed   a   series   of   100   simula2ons.   This   test   evaluated   the   power   consump2on   of   the   private   cloud   and   the   total   number   of   2meouts.   Next   figures   demonstrates   100   simula2ons   in   two   images,   the   percentage   of   2meouts  (top)  and  the  energy  consump2on  of   the  private  cloud  (bojom)  are  plojed.     JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

34  

4  Proposal  (Provisioning)  

JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

35  

4  Proposal  (Provisioning)  

JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

36  

4  Proposal  (Provisioning)   Table   III   shows   the   results   obtained   in   the   ”worst   case   scenario”,   by   defini2on,   with   the   mul2plier   equal   to   20   and   the   request   size   equal   to   2000   MI.   Regarding   the   results   in   Table  III,  it  took  the  Hybrid  Strategy  as  a  basis   of   comparison.   In   this   case,   the   values   listed   are   for   hybrid   strategy.   For   example,   the   hybrid   strategy   presented   3%   less   requisi2on   2meouts  than  the  OD  strategy.       JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

37  

4  Proposal  (Provisioning)  

Timeouts   Consump2on  

On  demand  

Spare  

-­‐  3  %  

 15  %  

 

  %   -­‐  18  

-­‐  52  %  

  Table  III  (HYBRID  STRATEGY  COMPARED  TO  THE  OTHER   STRATEGIES)       JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

38  

5  Conclusions     Based   on   what   was   presented   in   the   previous   sec2ons,   and   considering   the   objec2ves   set   at   the   beginning   of   this   paper,   we   consider   the   intended   goal   was   achieved.   Two   strategies   for   alloca2on   and   provisioning,   were   proposed;   both   aimed   at   op2mizing   the   energy   resource   without  sacrificing  service  availability.       JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

39  

5  Conclusions     The   alloca2on   strategy   in   private   clouds,   compared   to   a   normal   cloud,   demonstrated   a   87%   reduc2on   in   energy   consump2on.   It   was   observed   that   this   strategy   is   not   effec2ve   in   scenarios   where   the   workload   is   oscilla2ng.   That’s   because   it   ends   up   genera2ng   too   much   unnecessary   reconfigura2ons   and   migra2ons.   Despite   this,   it   s2ll   shows   a   significant   gain   in   energy   savings   when   compared   to   a   cloud   without  any  strategy  deployed.       JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

40  

5  Conclusions     The   hybrid   strategy   for   provisioning   in   green   clouds,   demonstrated   a   52%   consump2on   reduc2on  over  the  SR  strategy,  and  a  2meout  rate   3%  lower  than  the  OD  strategy.  Thus,  we  conclude   that   the   use   of   this   strategy   is   recommended   in   situa2ons  where  the  ac2va2on  2me  of  the  resource   is   expensive   for   the   health   of   SLA.   We   also   iden2fied  that  using  this  is  not  recommended  when   the   public   cloud   should   be   used   sparingly   due   to   their  course  or  other  factors.       JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

41  

6  Future  Works     As  future  work,  we  aim  at  adding  the  strategy  of   Dynamic   Reconfigura2on   of   VMs   in   public   clouds.   This   procedure   was   not   adopted   because,   during   the   development   of   this   work,   this   feature   was   not   a   market   reality.   We   also   intend  to  invest  in  new  simula2ons  of  the  cloud   extending  the  variables  (such  as  DVFS  and  UPS)   and,   if   possible,   explore   some   ar2ficial   intelligence   techniques   such   as   Bayesian   networks.       JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

42  

6  Future  Works     Our   PCMONS   (Private   Cloud   Monitoring   System),   open-­‐ source   solu2ons   for   cloud   monitoring   and   management,   also   will   help   to   manage   green   clouds,   by   automa2ng   the   instan2a2on   of   new   resource   usage.     We   foresee,   in   opposi2on   to   unexpected   peaks   scenarios,   work   with   cloud   management   based   on   prior   knowledge   of   the   behavior  of  hosted  services.  It  is  believed  to  be  necessary   to   develop   a   descrip2on   language   to   represent   the   structure   and   behavior   of   a   service,   enabling   the   exchange   of   informa2on   between   applica2ons   for   planning,  provisioning,  and  managing  the  cloud.     JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

43  

7  Some  References     -­‐  J.  Werner,  G.  A.  Geronimo,  C.  B.  Westphall,  F.  L.  Koch,  R.   R.   Freitas,   and   C.   M.   Westphall,   “Environment,   services   and   network   management   for   green   clouds,”   CLEI   Electronic  Journal,  vol.  15,  no.  2,  p.  2,  2012.     -­‐   R.   Buyya,   A.   Beloglazov,   and   J.   Abawajy,   “Energy-­‐ Efficient  management  of  data  center  resources  for  cloud   compu2ng:   A   vision,   architectural   elements,   and   open   challenges,”   in   Proceedings   of   the   2010   Interna+onal   Conference   on   Parallel   and   Distributed   Processing   Techniques   and   Applica+ons   (PDPTA   2010),   Las   Vegas,   USA,  July  12,  vol.  15,  2010.      

JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

44  

7  Some  References     -­‐   R.   Buyya,   “Modeling   and   simula2on   of   scalable   cloud   compu2ng   environments   and   the   cloudsim   toolkit:   Challenges   and   opportuni2es,”   in   HPCS   2009.   Interna+onal   Conference   on.   IEEE,   2009,   pp.   1–11.     -­‐  G.  von  Laszewski,  L.  Wang,  A.  Younge,  and  X.  He,   “Power   aware   scheduling   of   virtual   machines   in   dvfs   enabled   clusters,”   in   Cluster   Compu+ng   and   Workshops,   2009.   CLUSTER   ’09.   IEEE   Interna+onal   Conference  on,  2009,  pp.  1–10.      

JANUARY  29TH,  SEVILLE,  SPAIN    

IARIA  ICN  2013    -­‐  GLOBENET  2013  

45