Jun Ma, Sanjay Mehrotra and Huanyuan ShengImpact Solver for Optimization Services, November 8, 2006

On Implementing a Parallel Integer Solver

Using Optimization Services

Jun Ma

Huanyuan (Wayne) Sheng

Joint work with

Sanjay Mehrotra

2

Jun Ma, Sanjay Mehrotra and Huanyuan ShengImpact Solver for Optimization Services, November 8, 2006

Outline

• Brief Introduction– Impact Generalized MIP Solver – Optimization Services (OS)

• Distributed Parallel For Integer Programming Using OS

• Conclusion

3

Jun Ma, Sanjay Mehrotra and Huanyuan ShengImpact Solver for Optimization Services, November 8, 2006

IntroductionImpact GMIP Solver

• IMPACT -- Integrated Mathematical Programming Advanced Computational Tools

• Features– Generalized Mixed Integer Nonlinear Solver (GMIP).– Generalized Hyperplanes based Branch and Bound.– Standalone Solver and Remote Solver Service.– Unified NATIVE Interface with Optimization Services.

4

Jun Ma, Sanjay Mehrotra and Huanyuan ShengImpact Solver for Optimization Services, November 8, 2006

Impact GMIP Solver Features Algorithms

• Research Focus– Mixed Integer Nonlinear Programming– Parallel computing for MINLP

• Algorithm Studies– Heuristics for generalized branch and bound

methods – Optimization Services based distributed parallel,

e.g. communications, load balance handling.

5

Jun Ma, Sanjay Mehrotra and Huanyuan ShengImpact Solver for Optimization Services, November 8, 2006

Impact GMIP Solver Features Algorithms

RootNode

Proper BranchingHyperplane

being u

Leftchild

RightChild

Add u

lT xu 1 l

T xu

Growing Left Growing Right

Generate Children Generate Children

• Starting Node

6

Jun Ma, Sanjay Mehrotra and Huanyuan ShengImpact Solver for Optimization Services, November 8, 2006

Impact GMIP Solver Features Algorithms

• Subsequent Node

ParentNode(Left)

Proper BranchingHyperplane

being u prime

Leftchild

RightChild

Add new u prime

lT xu ''

Growing Left Growing Right

Generate Children Generate Children

SiblingNode

1 lT xu

1'' lT xu

7

Jun Ma, Sanjay Mehrotra and Huanyuan ShengImpact Solver for Optimization Services, November 8, 2006

Impact GMIP Solver Features Algorithms

• Generate Proper Branching Hyperplanes

– Basis Reduction Based (Mehrotra and Li)• LLL• GBR (Generalized Basis Reduction)

– Heuristics (ongoing)

8

Jun Ma, Sanjay Mehrotra and Huanyuan ShengImpact Solver for Optimization Services, November 8, 2006

GenerateNodes

Instance

Number OfNodes > 100

N

NodeQueue

Y

DistributedSystem

DistributedSystem

DistributedSystem

End

9

Jun Ma, Sanjay Mehrotra and Huanyuan ShengImpact Solver for Optimization Services, November 8, 2006

10

Jun Ma, Sanjay Mehrotra and Huanyuan ShengImpact Solver for Optimization Services, November 8, 2006

IMPACT GMIP Parallelization over Distributed Optimization Services

ImpactGMIP

OShL

OShL

OShL

OShL

OShL - hookup

CommunicationgetJobID (String OSoL)

solve (String OSiL, String OSoL)

send (String OSiL, String OSoL)

retrieve (String OSoL)

kill (String OSoL)

knock (String OSpL, String OSoL)

OSiL - instance

OSoL - option

Representation

OSrL - result

OSpL - process

Nodes (OSiL Integer)

OSServer

(Linux)

OSServer

(WinXP)

OSServer

(Mac OS)

Lindo

CPLEX

IMPACT

Call back

(OSrL)

11

Jun Ma, Sanjay Mehrotra and Huanyuan ShengImpact Solver for Optimization Services, November 8, 2006

Conclusion

• Introduced the Generalized MIP and showed it is friendliness for Parallelization

• Showed Optimization Services has a general and high extendable design fit for many derived researches

12

Jun Ma, Sanjay Mehrotra and Huanyuan ShengImpact Solver for Optimization Services, November 8, 2006