Upload
sendash-pangambam
View
45
Download
6
Tags:
Embed Size (px)
Citation preview
DEADLINE BASED RESOURCE PROVISIONING AND SCHEDULING ALGORITHM FOR
SCIENTIFIC WORKFLOWS ON CLOUDS(BASED ON PSO)
Maria Alejandra Rodrigues and Rajkumar Buyya
*** This work was developed by Maria Alejandra Rodrigues and Rajkumar Buyya. I am just presenting slides on their work.
Objective:
To design a new PSO based resource provisioning and scheduling strategy for scientific workflows on cloud
Minimization of total execution cost Deadline constrained
Workflow:
Automation of process during which inputs or/and outputs are passed from one task to other task(s) according to some rules
Used to model large-scale scientific problems in areas such as bioinformatics, astronomy, physics, etc.
Scientific workflows have ever-growing data and computing requirements
Demand a high-performance computing environment.
Structure of some scientific workflows:
Montage(Used in
Astronomy)
LIGO(Gravitational
waves)
SIPHT(Gene encoding)
CyberShake(Earthquake hazard
characterization)
Stages of workflow execution:
1. Resource provisioning phase Computing resources that will be used to
run the tasks are selected and provisioned
2. Schedule generation phase A schedule is generated and tasks are
assigned to the best suited resource.
Optimization is done in both stages so that user defined QOS are met.
Problems:
Parent tasks and child tasks dependency rules.
Scheduling: NP-hard problem Impossible to find optimal solution in
polynomial time; focus on near optimal ones
Transfer time affects execution time
Problem Formulation:
Workflow: W = (T,E) T = {t1, t2, …, tn} E = {eij};
Edge eij : There is data dependency between task ti and tj; tj is child of ti
User defined deadline of the workflow: w Each VMi is associated with processing capacity (PVMi) and
cost per unit time (CVMi) VM usage cost is quantized.
Problem Definition:
To find a schedule to execute the workflow that minimizes TEC subject to TET ≤ w
Particle Swarm Optimization: (1)
Population based stochastic optimization technique
Inspired by social behaviour of birds and fish
Shares similarities between other evolutionary computing techniques like GA
Developed by Dr. Eberhart and Dr. Kennedy in 1995
Particle Swarm Optimization: (2)
Particle: An individual (candidate solution) that has the ability to move iteratively through the problem space
Each particle has an associated position and velocity pbest and gbest based on fitness function In each iteration, velocity of each particle updated
towards pbest and gbest locations
Iterated until some stopping criterion is met.
PSO Modelling: (1)
KEY STEPS1. Encoding of the problem; representation of
candidate solutions2. Defining the fitness function or objective
function
Particle: Workflow and its associated tasks Dimension of a particle: Number of tasks in the
workflow; n-dimensional Fitness function: TEC
PSO Modelling: (2)
Encoding Example:
Some other strategies are also used for convenience
Makespan, TEC and deadline are evaluated and analysed for different types of workflows