HeteroPar 2013 Optimization of a Cloud Resource Management Problem from a Consumer Perspective...

HeteroPar 2013

Optimization of a Cloud Resource Management Problem from a Consumer Perspective

Rafaelli de C. Coutinho, Lucia M. A. Drummond andYuri Frota

Fluminense Federal University – UFF Brazil

Sponsorship:

Outline

Motivation

Definition and Formulation

GRASP

Results and Conclusions

Outline

oMotivation

oProblem Definition and Mathematical

Formulation

oGRASP for Resource Selection in Cloud

Environments

oPreliminary Results and Conclusion

Outline

Outline

Motivation


GRASP


Motivation

Cloud Computing

a large-scale distributed computing paradigm in which computing

resources are available to consumers via Internet.

Motivation

Outline

Motivation


GRASP


Motivation

Cloud Computing

a large-scale distributed computing paradigm in which computing

resources are available to consumers via Internet.

It delivers infrastructure, platform and software as services by signing service-level agreements (SLAs) with consumers.

Motivation

Outline

Motivation


GRASP


Motivation The provider should offer resource-economic services.

New pricing models based on the pay-as-you-go policy are necessary to address the highly variable demand for cloud resources.

Motivation

Outline

Motivation


GRASP


Motivation The provider should offer resource-economic services.

New pricing models based on the pay-as-you-go policy are necessary to address the highly variable demand for cloud resources.

For example, cloud service consumers might have an SLA with a cloud service provider concerning how much bandwidth, CPU, and memory the consumer can use at any given time throughout the day.

Motivation

Outline

Motivation


GRASP


Motivation Application schedulers have different policies that vary according to the objective function.

Independently of the objective function, resource allocation problems in clouds are NP-hard problems.

Motivation

Outline

Motivation


GRASP


Motivation

Cloud computing becomes more widespread

andComputing resources grow

More researches have been conducted on

Resource Allocation Problem in Clouds.

Motivation

Outline

Motivation


GRASP


Definition and FormulationA resource management problem that aims to reduce the payment cost and the execution time of the user application is defined.

An integer programming formulation and a heuristic based on Greedy Randomized Adaptive Search Procedure are proposed.


Outline

Motivation


GRASP


Definition Problem Infrastructure-as-a-Service (Iaas) model:

• cloud consumers request computing resources as processing power, disk storage, memory and architecture type, for a period of time and pays only what he/she uses.

• Cloud providers have a wide variety of virtual machines.

• Cloud consumers have to decide which packages should purchase in order to minimize a specific objective as execution time or payment cost.


Outline

Motivation


GRASP




Cloud providers have a wide variety of package types.

• Cloud consumers have to decide which packages should purchase in order to minimize a specific objective as execution time or payment cost.


Outline

Motivation


GRASP




Cloud providers have a wide variety of package types.

Cloud consumers have to decide which packages should purchase in order to minimize a specific objective as execution time or payment cost.


Outline

Motivation


GRASP


Modeling the problem … Let be the set of packages types offered by a cloud

provider during a set of time periods.


Package

Package

Package

Outline

Motivation


GRASP


Modeling the problem … Let be the set of packages types offered by a cloud

provider during a set of time periods.

Each package type has: • cost• computing resources:

• disk storage • memory capacity • processing power ofGflop per period of time


Package

Package

Package

𝑐𝑝𝑑𝑝𝑚𝑝

Package

𝑔𝑝

Outline

Motivation


GRASP


Modeling the problem … The set of consumer requirements is defined by:• maximum cost • maximum time • disk storage • memory capacity • processing demand of Gflop.


, ,

Cloud Provider

Outline

Motivation


GRASP


Modeling the problem … The set of consumer requirements is defined by:• maximum cost • maximum time • disk storage • memory capacity • processing demand of Gflop.

Let , the maximum limit of packages that can be purchased for each consumer in each period of time.


, ,

Cloud Provider

Outline

Motivation


GRASP


Modeling the problem … Let , a binary variable for each , and such that if and only if packageof type is purchased at time ; otherwise

Let , the variable that defines the last time period that a package was purchased by the consumer.


Outline

Motivation


GRASP


The Mathematical Formulation CC-IP

Subject to:1. The paid costs do not exceed the maximum recommended cost. Definition and

Formulation

Outline

Motivation


GRASP



Subject to:2. The storage and memory capacity should be sufficient large to meet the requirements in each period of time.


Outline

Motivation


GRASP



Subject to:3. The processing power is at least large enough to satisfy the total demand.

4. The number of selected packages does not exceed the cloud providers limit.


Outline

Motivation


GRASP



Subject to:5. Guarantee the correct interpretation of variable

6. There will be no gaps of time in feasible solutions.


Outline

Motivation


GRASP



Subject to:7. An order between the packages should be established and the symmetry should be eliminate.

8. The binary and integrality requirements of the variables.


Outline

Motivation


GRASP


GRASP for Resource Selection in Cloud Exact procedures have often proved incapable of finding solutions as they are extremely time-consuming. • Heuristics and metaheuristics provide sub-

optimal solutions in a reasonable time.

GRASP

Outline

Motivation


GRASP




A Greedy Randomized Adaptive Search Procedure (GRASP) is proposed. GRASP

Outline

Motivation


GRASP




A Greedy Randomized Adaptive Search Procedure (GRASP) is proposed.

The proposed heuristic GraspCC is composed of two phases: • a construction phase coCC• a local search phase lsCC

GRASP

Outline

Motivation


GRASP


GraspCCAdittional Definitions• A solution as a set of 3-tuples representing that package of type was purchased on period .

• Let be the family of all possible solutions.

• Let be the last period that a package was selected.

GRASP

Outline

Motivation


GRASP


GraspCCAdittional Definitions• Define a cost function which will measure the quality of the solution:

GRASP

Outline

Motivation


GRASP


Construction Phase - coCCInitialization• is defined, as the set of packages such that its elements appears in descending order of cost and processing power.

Meeting Disk and Memory RequirementsChoose a package among the first packages to be added to the first period of time while the solution does not satisfy the disk and memory requirements.

Meeting Processing RequirementsThe chosen packages are replicated in future periods until the processing demand is satisfied.

GRASP

Outline

Motivation


GRASP


Local Search Phase - lsCCInitialization• The neighborhood is defined as the family of all solutions obtained by exchanging tuples in with another tuples not in .

• lsCC starts with the solution provided by the construction phase.

Improving solution• It iteratively replaces the current solution by that with minimum cost functionwithin its neighborhood.

Termination• The local search halts when no better solution is found in the neighborhood of the current solution.

GRASP

Outline

Motivation


GRASP


GraspCC

GraspCC consists to perform the coCC following by the lsCC until the maximum number of iterations without improvement in the best solution found is satisfied.

GRASP

Outline

Motivation


GRASP


Preliminary Results Implemented in:

• C • CPLEX 12.4, for mathematical formulation.

Executed on a computer with processor Intel Core i7 3.4Hz and 12Gb of RAM under Linux (Ubuntu 12.04) operating system.

Tested over a set of instances constructed from the requirements of real applications combined with the sets of virtual machine packages available in commercial clouds.


Outline

Motivation


GRASP


Preliminary Results• These instances use requirements of five real applications:• an algorithm for the Quadratic Assignment Problem

that can be solved by a branch-and-bound algorithm;• three applications related to manipulation of biologic

sequences: RAXML, ModelGenerator and Segemehl;• a typical analysis user job for the CMS experiment

running at the LHC/CERN grid.

• They use packages of two commercial clouds:• Amazon EC2, with package groups of high

performance computing (HPC) clusters • Google Cloud Platform, with packages from the

Google Compute Engine


Outline

Motivation


GRASP


Preliminary Results Accomplished experiments:

1. An evaluation of the CC-IP mathematical formulation

2. An evaluation of the GraspCC heuristic


Outline

Motivation


GRASP






For all experiments, these values for and were considered:

The other used values were: and , equal to the maximum number of packages offered by each cloud provider and .

Outline

Motivation


GRASP






The objectives of the cost function were normalized due to their distinct range values. - They share the same minimum and maximum values (0 and 1).

- The payment cost was divided by the most expensive package cost times the maximum time informed by the cloud consumer.

- The execution time was divided by the maximum time informed by the cloud consumer.

Outline

Motivation


GRASP


Preliminary Results

Analyzing the different goals in the proposed cost function:(i) the cloud consumer is only concerned with

payment cost ( and )(ii) he only prioritizes the execution time ( and

)


Outline

Motivation


GRASP


Preliminary Results

Analyzing the different goals in the proposed cost function:(i) the cloud consumer is only concerned with

payment cost ( and )(ii) he only prioritizes the execution time ( and

)


For example, nug24-cbb with packages of Amazon(i) 3 time units for execution time and $46.20 for payment cost. Packages set: 8 packages of type 2 (ii) 2 time units for execution time and $84.00 for payment cost. Packages set: 20 packages of type 2

Outline

Motivation


GRASP


Preliminary ResultsComparison of the GraspCC with the CC-IP

Instances

GraspCC CC-IP

Function

cost

Solution value

TotalTime

Gap (%)

Function

cost

Solution Value

Total Time

Time valu

e

Payment

Time Value

Payment

nug22-sbb_am

12 0.0456 1 4.5 0.01 0.0000 0.0456 1 4.5 0.59

nug28-sbb_am

72 0.0128 1 40.8 0.07 0.0000 0.0128 1 40.8 27.23

nug22-cbb_am

12 0.0519 1 11.8 0.03 0.0000 0.0519 1 11.8 0.77

nug28-cbb_am

72 0.0123 1 36.9 0.07 0.0000 0.0123 1 36.9 21.20

nug30-cbb_am

84 0.2652 23 1034.7 3.63 0.0000 0.2652 23 1034.7 118.42

segemehl_am

4 0.1688 1 16.8 0.07 0.0000 0.1688 1 16.8 0.26

cms-1500_am

24 0.1625 4 410.4 0.53 0.0000 0.1625 4 410.4 32.83

nug22-sbb_go

12 0.0497 1 4.9 5.36 0.0066 0.0497 1 4.9 77.72

nug28-sbb_go

72 0.0655 6 87.4 133.66 0.0000 0.0655 6 87.435278.

98nug22-cbb_go

12 0.1137 2 18.6 17.97 0.0000 0.1137 2 18.6 246.79

nug28-cbb_go

72 0.0541 5 70.7 60.00 0.0859 0.0552 5 70.686400.

00nug30-cbb_go

84 0.8502 81 1573.22987.7

90.0000 0.8502 81 1573.2

5085.60

segemehl_go4 0.5038 3 26.2 52.40 0.0000 0.5038 3 26.2 23.52

cms-1500_go24 0.4796 13 573.5 53.27 0.0000 0.4797 13 573.986400.

00


Outline

Motivation


GRASP




Instances

GraspCC CC-IP

Function

cost

Solution value

TotalTime

Gap (%)

Function

cost

Solution Value

Total Time

Time valu

e

Payment

Time Value

Payment

nug22-sbb_am

12 0.0456 1 4.5 0.01 0.0000 0.0456 1 4.5 0.59

nug28-sbb_am

72 0.0128 1 40.8 0.07 0.0000 0.0128 1 40.8 27.23

nug22-cbb_am

12 0.0519 1 11.8 0.03 0.0000 0.0519 1 11.8 0.77

nug28-cbb_am

72 0.0123 1 36.9 0.07 0.0000 0.0123 1 36.9 21.20

nug30-cbb_am

84 0.2652 23 1034.7 3.63 0.0000 0.2652 23 1034.7 118.42

segemehl_am

4 0.1688 1 16.8 0.07 0.0000 0.1688 1 16.8 0.26

cms-1500_am

24 0.1625 4 410.4 0.53 0.0000 0.1625 4 410.4 32.83

nug22-sbb_go

12 0.0497 1 4.9 5.36 0.0066 0.0497 1 4.9 77.72

nug28-sbb_go

72 0.0655 6 87.4 133.66 0.0000 0.0655 6 87.435278.

98nug22-cbb_go

12 0.1137 2 18.6 17.97 0.0000 0.1137 2 18.6 246.79

nug28-cbb_go

72 0.0541 5 70.7 60.00 0.0859 0.0552 5 70.686400.

00nug30-cbb_go

84 0.8502 81 1573.22987.7

90.0000 0.8502 81 1573.2

5085.60

segemehl_go4 0.5038 3 26.2 52.40 0.0000 0.5038 3 26.2 23.52

cms-1500_go24 0.4796 13 573.5 53.27 0.0000 0.4797 13 573.986400.

00

Outline

Motivation


GRASP


Instances

GraspCC CC-IP

Function

cost

Solution value

TotalTime

Gap (%)

Function

cost

Solution Value

Total Time

Time valu

e

Payment

Time Value

Payment

nug22-sbb_am

12 0.0456 1 4.5 0.01 0.0000 0.0456 1 4.5 0.59

nug28-sbb_am

72 0.0128 1 40.8 0.07 0.0000 0.0128 1 40.8 27.23

nug22-cbb_am

12 0.0519 1 11.8 0.03 0.0000 0.0519 1 11.8 0.77

nug28-cbb_am

72 0.0123 1 36.9 0.07 0.0000 0.0123 1 36.9 21.20

nug30-cbb_am

84 0.2652 23 1034.7 3.63 0.0000 0.2652 23 1034.7 118.42

segemehl_am

4 0.1688 1 16.8 0.07 0.0000 0.1688 1 16.8 0.26

cms-1500_am

24 0.1625 4 410.4 0.53 0.0000 0.1625 4 410.4 32.83

nug22-sbb_go

12 0.0497 1 4.9 5.36 0.0066 0.0497 1 4.9 77.72

nug28-sbb_go

72 0.0655 6 87.4 133.66 0.0000 0.0655 6 87.435278.

98nug22-cbb_go

12 0.1137 2 18.6 17.97 0.0000 0.1137 2 18.6 246.79

nug28-cbb_go

72 0.0541 5 70.7 60.00 0.0859 0.0552 5 70.686400.

00nug30-cbb_go

84 0.8502 81 1573.22987.7

90.0000 0.8502 81 1573.2

5085.60

segemehl_go4 0.5038 3 26.2 52.40 0.0000 0.5038 3 26.2 23.52

cms-1500_go24 0.4796 13 573.5 53.27 0.0000 0.4797 13 573.986400.

00



GraspCC presented an outstanding improvement of the execution time, in

average 99% less than the execution time of CC-IP.

HeteroPar 2013

Optimization of a Cloud Resource Management Problem from a Consumer Perspective

Rafaelli de C. Coutinho, Lucia M. A. Drummond andYuri Frota

Fluminense Federal University – UFF Brazil

Sponsorship:

Outline

Motivation


GRASP


Local Search Phase - lsCC

GRASP

HeteroPar 2013 Optimization of a Cloud Resource Management Problem from a Consumer Perspective...

Documents

Transcript of HeteroPar 2013 Optimization of a Cloud Resource Management Problem from a Consumer Perspective...