Transcript
Page 1: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Bin  Packing  with  Linear  Usage  Costs    An  applica)on  to  Energy  Management  in  Data  

Centres  

Hadrien  Cambazard  G-­‐SCOP,  Université  de  Grenoble  

Deepak  Mehta,  Barry  O’Sullivan,  Helmut  Simonis  Cork  Constraint  ComputaAon  Centre,  Cork,  Ireland  

Page 2: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Outline  1.  The  EnergeTIC  Project  (Problem  FormulaAon)  

2.  A  key  sub-­‐problem  BPUC  (Bin  Packing  with  Usage  Costs)  –  Lower  bounds  based  on  LP  –  A  Constraint  Programming  view-­‐point  

3.  Back  to  the  applicaAon  –  A  Lower  bound  by  column  generaAon  relying  on  BPUC  –  Upper  bound  via  Large  Neighborhood  Search  

4.  Overview  of  experimental  results  

Page 3: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

The  EnergeTIC  Project  

•  Interest  increased  by  the  Roadef  Challenge  with  google  •  EnergeTIC  is  a  project  located  in  Grenoble  •  Companies  involved:    

Schneider,  Bull,  Eolas  Business  &  Decision,  UXP      

Design  of  energy  efficient  data  centres  

•  Permanent  increase  of  energy  price  (50  %  by  2020)  •  Growing  market  for  cloud  compuAng  

Page 4: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

The  EnergeTIC  Project  

Problem  formulaAon  Assign  virtual  machines  to  servers  

over  mulAple  Ame-­‐periods  

Energy  models    Characterize    

each  component  

Demand  predicAon    Forecast  the    cpu  needs  

Energy  indicators  -­‐  Usage  -­‐  Energy  

ExploitaAon  constraints  

Solver  Plan  

Page 5: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

The  EnergeTIC  Project  

Problem  formulaAon  Assign  virtual  machines  to  servers  

over  mulAple  Ame-­‐periods  

Energy  models    Characterize    

each  component  

Demand  predicAon    Forecast  the    cpu  needs  

Energy  indicators  -­‐  Usage  -­‐  Energy  

ExploitaAon  constraints  

Solver  Plan  

Page 6: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

EnergeTIC:  Modeling  Equipment  

•  Characterize  energy  consumpAon  of  heterogeneous  servers  – Measures  performed  on  Bull  equipment  –  Linear  model  of  energy  consumpAon    

Fixed  cost  Usage  cost  

Page 7: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

The  EnergeTIC  Project  

Problem  formulaAon  Assign  virtual  machines  to  servers  

over  mulAple  Ame-­‐periods  

Energy  models    Characterize    

each  component  

Demand  predicAon    Forecast  the    cpu  needs  

Energy  indicators  -­‐  Usage  -­‐  Energy  

ExploitaAon  constraints  

Solver  Plan  

Page 8: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Problem  formulaAon  

1 migration

   

t-1 t t+1

STAND  BY  

   

STAND  BY  

STAND  BY    

 

Uit

MiUmaxj

Mmaxj

•  A  server  has  a  CPU,  MEMORY  and  CARDINALITY  capaciAes  •  A  virtual  machine  has  a  CPU  consumpAon  changing  over  Ame  and  a  

fixed  MEMORY  consumpAon  •  The  number  of  migraAons  is  limited  from  one  period  to  the  next  •  ObjecAve  :  state  (ON/STAND_BY)  and  load  of  servers  

             changes  of  state                migraAons  

Page 9: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Problem  formulaAon  

1 migration

   

… …

t-1 t t+1

STAND  BY  

   

STAND  BY  

STAND  BY    

 

Uit

MiUmaxj

Mmaxj

A  series  of  consecuAve  “cost  aware”  mulAdimensional  Bin-­‐Packing  problems  (linked  by  a  limited  number  of  changes)  

Page 10: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Problem  formulaAon  

1 migration

   

… …

t-1 t t+1

STAND  BY  

   

STAND  BY  

STAND  BY    

 

Uit

MiUmaxj

Mmaxj

A  series  of  consecuAve  “cost  aware”  mulAdimensional  Bin-­‐Packing  problems  (linked  by  a  limited  number  of  changes)  

Coupling  constraints  

Page 11: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Problem  formulaAon  

1 migration

   

… …

t-1 t t+1

STAND  BY  

   

STAND  BY  

STAND  BY    

 

Uit

MiUmaxj

Mmaxj

A  series  of  consecuAve  “cost  aware”  mulAdimensional  Bin-­‐Packing  problems  (linked  by  a  limited  number  of  changes)  

 Coupling  constraints  

Page 12: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

•  A  set  of  items  

•  A  set  of  bins  

Bin  Packing  with  Usage  Costs  S = {w1, . . . , wn}

B = {{C1, f1, c1}, . . . , {Cm, fm, cm}}

w1 w2. . .

Cj

Fixed  cost  for  opening  a  bin  

Usage  cost    depending    on  the  load  

Cost  

Load  

Load  

{Cj , fj , cj }

cjfj

•  Minimize    

costj = fj + Loadjcj

Pm

j=1|Loadj>0 costj

Page 13: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Bin  Packing  with  Usage  Costs  

3

9

{9,0,1} {3,0,2} {3,0,2} {3,0,2} {3,0,2}3

2

Items   Bins  

•  A  set  of  items  •  A  set  of  bins  •  Minimize  the  sum  of  the  costs  of  the  used  bins    

S = {w1, . . . , wn}B = {{C1, f1, c1}, . . . , {Cm, fm, cm}}

Page 14: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Bin  Packing  with  Usage  Costs  •  A  set  of  items  •  A  set  of  bins  •  Minimize  the  sum  of  the  costs  of  the  used  bins    

S = {w1, . . . , wn}B = {{C1, f1, c1}, . . . , {Cm, fm, cm}}

3

2

3

9

{9,0,1} {3,0,2} {3,0,2} {3,0,2} {3,0,2}

(P1) (P2)

{9,0,1} {3,0,2} {3,0,2} {3,0,2} {3,0,2}

8x1  +  3x2  +  3x2  +  3x2  =  26   9x1  +  2x2  +  2x2  +  2x2  +  2x2  =  25  

Page 15: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Lower  bounds  for  BPUC  

Bin  j  is  open  Item  i  is  placed  on  bin  j   Load  of  bin  j  

“Standard”  LP  formulaCon:  

Page 16: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Lower  bounds  for  BPUC  

Bin  j  is  open  Item  i  is  placed  on  bin  j   Load  of  bin  j  

“Standard”  LP  formulaCon:  

Linear  relaxaCon  easy  to  characterize:  •  A  unit  of  space  on  bin  j  will  cost  at  least:  •  Sort  the  bins  by  non-­‐decreasing  •  Fill  the  “cheapest”  bins  first  

rj = fj/Cj + cjrj

Page 17: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Lower  bounds  for  BPUC  

rj = fj/Cj + cj

9

f1 =c1 =

9  5  

r1 = 6  

14  3  

7

f3 =c3 =r3 = 5  

1  

5

f4 =c4 =10  r4 =10.2  

3

f2 =c2 =r2 =

1  5  5.33  

12

f5 =c5 =r5 =

12  10  11  

Cj =

3

5

Page 18: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Lower  bounds  for  BPUC  

rj = fj/Cj + cj

9

f1 =c1 =

9  5  

r1 = 6  

14  3  

7

f3 =c3 =r3 = 5  

1  

5

f4 =c4 =10  r4 =10.2  

3

f2 =c2 =r2 =

1  5  5.33  

12

f5 =c5 =r5 =

12  10  11  

9

r1 = 6  

7

r3 = 5  

5

r4 =10.2  

3

r2 =5.33  

12

r5 = 11  

3

5

Page 19: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

3  

Lower  bounds  for  BPUC  

rj = fj/Cj + cj

9

f1 =c1 =

9  5  

r1 = 6  

14  3  

7

f3 =c3 =r3 = 5  

1  

5

f4 =c4 =10  r4 =10.2  

3

f2 =c2 =r2 =

1  5  5.33  

12

f5 =c5 =r5 =

12  10  11  

Total  Load  =  18  Lower  bound  =  7x5  +  3x5.33  +  6x8  =  99  

r1 = 6  r3 = 5   r4 =10.2  r2 =5.33   r5 = 11  O(mlog(m) + n)

7   8  

3

5

Page 20: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

3  

Lower  bounds  for  BPUC  

Total  Load  =  18  

Lower  bound  =  99  

r1 = 6  r3 = 5   r4 =10.2  r2 =5.33   r5 = 11  

OpAmal  soluAon  

Opt  =  129  

7   8  

3

5

Page 21: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Lower  bounds  for  BPUC  based  on  LP  

2 3 4 5 70

F

0 5 9 7 6 15 18 2400 5 12 7

S = {2, 2, 3, 5} B = {{3,1,2},{4,3,1},{7,3,3}}

z3:  Arc-­‐Flow  Model  of  Carvalho  

z1:  Standard  LP  formulaAon   z2:  Gilmore  and  Gomory    

i-­‐th  cuing  pajern  of  bin  j  

C C z2  z3  z1  

Page 22: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Solving  BPUC  exactly  

Arc-­‐Flow  Model  Standard  LP  formulaAon  

Gilmore  and  Gomory    

BPUC  in  the  applicaAon  domain  goes  with:  1.  Large  scale    2.  Side  constraints  (ex:  cardinality  constraints,  2D,  …)  

MIP  

Page 23: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Solving  BPUC  exactly  

Arc-­‐Flow  Model  Standard  LP  formulaAon  

Gilmore  and  Gomory    

BPUC  in  the  applicaAon  domain  goes  with:  1.  Large  scale    2.  Side  constraints  (ex:  cardinality  constraints,  2D,  …)  

MIP  

Not  so  easy  to  extend    with  side  constraints  

Page 24: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Solving  BPUC  exactly  

Arc-­‐Flow  Model  Standard  LP  formulaAon  

Gilmore  and  Gomory    

BPUC  in  the  applicaAon  domain  goes  with:  1.  Large  scale    2.  Side  constraints  (ex:  cardinality  constraints,  2D,  …)  

MIP  

Page 25: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Solving  BPUC  exactly  

Arc-­‐Flow  Model  Standard  LP  formulaAon  

Gilmore  and  Gomory    

Bin-­‐Packing  global  constraint    +  

Cost  propagaAon  using  bound  of  LP  

 

MIP   CP  

BPUC  in  the  applicaAon  domain  goes  with:  1.  Large  scale    2.  Side  constraints  (ex:  cardinality  constraints,  2D,  …)  

Page 26: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Cost  propagaAon  for  BPUC  

r1 = 6  r3 = 5   r4 =10.2  r2 =5.33   r5 = 11  

7  3  

8  

Idea:  Use  LP  bound  to  compute  the  minimum/maximum  possible  filling  of  each  bin  due  to  cost  

LB  =  99    

Page 27: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Cost  propagaAon  for  BPUC  

r1 = 6  r3 = 5   r4 =10.2  r2 =5.33   r5 = 11  

Suppose  a  UB=130,  load  on  bin  3  (cheapest  bin)  must  be  >=  1  (if  not  LB  =  132  >  130)    

7  3  

8  1  

Idea:  Use  LP  bound  to  compute  the  minimum/maximum  possible  filling  of  each  bin  due  to  cost  

LB  =  99    

Page 28: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Cost  propagaAon  for  BPUC  

r1 = 6  r3 = 5   r4 =10.2  r2 =5.33   r5 = 11  

7  3  

8  1  

Idea:  Use  LP  bound  to  compute  the  minimum/maximum  possible  filling  of  each  bin  due  to  cost  

1  

UB  =  130    

We  know  now  that  3  and  1  must  be  open  so:  •  Count  their  fixed  costs  in  the  bound  •  Apply  the  same  reasoning  on  restricted  problem  unAl  a  fixed  point  is  reached  

Page 29: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Cost  propagaAon  for  BPUC  

r1 = 6  r3 = 5   r4 =10.2  r2 =5.33   r5 = 11  

7  3  

8  3  

0  

3  3  

5  

Idea:  Use  LP  bound  to  compute  the  minimum/maximum  possible  filling  of  each  bin  due  to  cost  

UB  =  130    

9  

Update  upper  bounds  Update  lower  bounds  

O(m)

Page 30: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Cost  propagaAon  for  BPUC  

r1 = 6  r3 = 5   r4 = 10.2  r2 = 5.33   r5 = 11  

7  

3  

8  3  

0  

3  3  

5  

UB  =  130     9  

Update  upper  bounds  Update  lower  bounds  

•  Lower  bound  in                      and  filtering  in                for  each  bin  •  Upper/lower  bounds  can  also  be  Aghtened  by  Dynamic  Programming  

(during  search  for  CP,  iniAally  for  MIP)  •  On  the  example,  the  bound  at  the  root  node  (with  UB  =  130):  

–  LP  Bound:  114.4  –  CP  Bound:  119.66  

•  UpdaAng  load  bounds  triggers  the  filtering  of  the  Bin-­‐Packing  global  constraint  commiing/forbidding  items  to  bins  

O(mlog(m) + n) O(m)

Page 31: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Results  on  BPUC  

•  CP  is  good  at  handling  large  BPUC  problems  •  Need  to  implement  branch  and  price  for  the  cuing-­‐

stock  formulaAon  to  compare  with  CP  

Lower  bounds  

Page 32: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Back  to  the  applicaAon  

•  Upper  bounds:    Large  Neighborhood  Search  designed  for  the  ROADEF  Challenge  2012  (Mehta,  O’Sullivan,  Simonis)  

•  Lower  bounds:  Column  generaAon  relying  on  BPUC  

1 migration

   

t-1 t t+1

STAND  BY  

   

STAND  BY  

STAND  BY    

 

Uit

MiUmaxj

Mmaxj

Page 33: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Back  to  the  applicaAon  (Lower  bound)  

•  A  column  is  a  Bin-­‐Packing  (one  Ame  period)  …  •  The  pricing  problem  can  be  well  modeled  with  3  global  

constraints:    –  BinPackingWithUsageCost  (for  cpu  cost  pertubated  by  dual  variables)  –  BinPackingWithUsageCost  (for  cardinality  cost  due  to  dual  variables)  –  BinPacking  (for  memory  limitaAons)  –  GlobalCardinality  (for  limits  on  the  number  of  items  on  each  bin)  

1 migration

   

t-1 t t+1

STAND  BY  

   

STAND  BY  

STAND  BY    

 

Uit

MiUmaxj

Mmaxj

Page 34: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Back  to  the  applicaAon  (Lower  bound)  

•  A  column  is  a  Bin-­‐Packing  (one  Ame  period)  …  

•  The  pricing  problem  can  be  well  modeled  with  3  global  constraints:    –  BinPackingWithUsageCost  (for  cpu  cost  pertubated  by  dual  variables)  –  BinPackingWithUsageCost  (for  cardinality  cost  due  to  dual  variables)  –  BinPacking  (for  memory  limitaAons)  –  GlobalCardinality  (for  limits  on  the  number  of  items  on  each  bin)  

 •  Pricing  is  intractable  but  we  are  only  looking  for  a  lower  bound:  

–  Use  a  lower  bound  of  the  best  reduced  costs  to  get  the  dual  bound  of  the  column  generaAon  

–  Problem  reduced  to  one  Ame  period  are  relaAvely  easy  in  pracAce  

Page 35: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Results  on  the  applicaAon  •  Benchmark:  74  instances  with  at  most  242  virtual  machines,  20  servers,  

287  Ame-­‐periods.  Time-­‐limit  600s.  •  CG  lower  bound  outperforms  the  LP  and  MIP  lower  bounds    

(avg:0.3%,  med:0.1%,  max:7%  of  best  known  upper  bounds)  •  LNS  scales  very  well  in  quality  and  size  

 (avg:0.5%,  med:0%,  max:4.5%  of  the  best  known  lower  bounds)  •  Zoom  on  some  instances:  

Lower  bound   Upper  bound  

Page 36: Bin$Packing$with$Linear$Usage$Costs$$€¦ · Outline$ 1. The$EnergeTIC$Project(Problem$Formulaon)$ 2. A$key$sub5problem$BPUC$(Bin$Packing$with$Usage$Costs)$ – Lower$bounds$based$on$LP$

Conclusion  •  Study  of  a  key  variant  of  Bin-­‐Packing  (BPUC):  

–  Characterize  LP  bound  +  Cost  based  propagaAon  –  Propose  an  extension  of  the  Bin-­‐Packing  global  constraint  with  a  useful  cost-­‐model  for  this  applicaAon  domain  

–  CP  gives  an  effecAve  exact   solver   for  BPUC   specially   to  handle  large-­‐scale  problems  and  side  constraints  

•  Minimizing  energy  in  data  centres:  –  Design   a   lower   bound   based   on   column   generaAon   that   has  some  generality  for  the  applicaAon  domain    

–  Assert  the  quality  of  LNS  on  real  benchmark  –  We  are  currently  tesAng  the  scalability  on  problems  where  the  bin-­‐packings  are  4  Ames  larger:    

1000  vms,  80  servers,  300  Ame-­‐periods  


Recommended