[PDF] - Day 8 Workflow Cloud Resource Provisioning Todays Agenda PDF Document

SLIDE 1

2/7/2018 1

Day 8 Workflow Cloud Resource Provisioning Todays Agenda

Introduction
What is workflow?
What are major QoS requirements?
How is workflow represented?
Our work in workflow scheduling
Research problems
Questions and conclusions

Introduction

Cloud provides resources necessary to execute

workflow applications based on

– Resources are charged by its type and duration – Pay as you go (e.g., hourly pricing scheme) – Other evolving cloud fee structures spot and reserved pricing)

J1 J1 Jn

Cloud consumer

request

Cloud service provider

CPU time usage
Data transfer cost (billed by GB) include the cost to copy data to/from cloud
ver network This billed by GB
Storage cost (billed by GB) is the cost to store VM images

SLIDE 2

2/7/2018 2

Todays Agenda

Introduction
What is workflow?
What are major QoS requirements?
How is workflow represented?
Our work in workflow scheduling
Research problems
Questions and conclusions

What is workflow?

Simply stated workflow application is divisible

application with large task that has precedence constraints.

Workflow applications are being used in a

range of domains, such as

– astrophysics, – bioinformatics, and – disaster modeling and prediction. – neurosurgical imaging.

Directed Acyclic Graph

Workflow application can be modelled as a a

Directed Acyclic Graph (DAGs).

– the nodes represent the set of workflow tasks – the arcs represent the set of control flow or data dependencies between the tasks. – the dependent tasks require a specific execution order due to the relationship between them. – tasks have very different I/O and computational behavior.

T1 T2 T3 T5 2 T4 T7 T8 T6

SLIDE 3

2/7/2018 3 Microarray gene expression

Microarray gene expression data analysis workflow expression

– Task 1 gene expression data is obtained from a microarray experiments. – Task 2: cluster analysis algorithms are used to identify genes that share similar patterns of gene expression profiles that are then predicted to be co-regulated as part of an interactive biochemical pathway. – … – Task 8: the consensus sequence is fed to the BLAST utility to determine if the gene is a new candidate or for the iteration to continue.

The Montage application created by NASA/IPAC closes together multiple input images to form custom mosaics of the sky.

CyberShake workflow is used by the Southern California Earthquake Center to distinguish earthquake threatening a region. The Epigenomics workflow created by the USC Epigenome Center and the Pegasus framework is used to automate the different

perations in genome

sequence processing. LIGO’s Inspiral Analysis workflow is used to create and analyze gravitational waveforms from data gathered during the coalescing of compact binary systems

Real workflow applications Epigenomics Workflow

Orchestrate complex, multi-stage scientific computations
It is possible to automatically parallelize it on distributed

resources

Split Filter & Convert Map Merge Analyze

Epigenomics Workflow

From Gideon Juve

SLIDE 4

2/7/2018 4 Large-Scale, Data-Intensive Workflows

Montage Galactic Plane Workflow

– 18 million input images (~2.5 TB) – 900 output images (2.5 GB each, 2.4 TB total) – 10.5 million tasks (34,000 CPU hours)

Scientific workflow management systems are designed to

automatically distribute data and computations for these large applications

John Good (Caltech)

From Gideon Juve

Workflow Application Patterns

Montage (astronomy)

– I/O: High (95% of time waiting on I/O) – Memory: Low – CPU: Low

Epigenome (bioinformatics)

– I/O: Low – Memory: Medium – CPU: High (99% of time in CPU)

Broadband (earthquake science)

– I/O: Medium – Memory: High (75% of time tasks use > 1GB) – CPU: Medium

From Gideon Juve

Todays Agenda

Introduction
What is workflow?
What are major QoS requirements?
How is workflow represented?
Our work in workflow scheduling
Research problems
Questions and conclusions

SLIDE 5

2/7/2018 5

Workflow Management System Workflow Scheduling

A process that maps and manages the

execution of inter-dependent tasks on the distributed resources.

It allocates suitable resources to workflow

tasks such that the execution can be completed to satisfy objective functions imposed by users.

Question: How visible is the general scheduler

for scheduling workflows?

Scheduling Structure

Create a schedule that meet the objectives of the

problem

For example

– minimizes the total execution cost of a workflow

– while satisfying a user-defined deadline

It is well known to be an NP-complete problem.
The scheduling processes of the workflow

applications are a multiobjective optimization problem (also known as Pareto optimization),

SLIDE 6

2/7/2018 6

Workflow Scheduling Classification

There are two types of workflow scheduling:

– the best-effort workflow scheduling – the quality of services (QoS) constraint workflow scheduling.

Best-effort workflow scheduling

Focuses on reducing the execution time of the whole

workflow tasks regardless of other factors.

An example best-effort workflow scheduling algorithm

– min-min algorithm – execute the small tasks first and delays the larger tasks for a longer time – max–min algorithm - execute the large tasks first and the small tasks are delayed for a longer time.

Todays Agenda

Introduction
What is workflow?
What are major QoS requirements?
How is workflow represented?
Our work in workflow scheduling
Research problems
Questions and conclusions

SLIDE 7

2/7/2018 7

QoS Requirements

Quality of Service (QoS) is used to measure

the level of satisfaction of a service.

A variety of QoS is discussed in the literature:

makespan, cost, reliability and energy For example

– The deadline of a workflow is defined as the maximum finish time of its last task to be executed. – Budget is defined as the maximum amount that a user wants to pay for executing a workflow application on computing resources.

QoS Requirements

Makespan

– An important metric in workflow scheduling – It is defined as the maximum completion time of the workflow. 𝑛𝑏𝑙𝑓𝑡𝑞𝑏𝑜 = max

(𝑑𝑝𝑛𝑞𝑚𝑓𝑢(𝑢))

Task 𝒰 T1 T2 T3 T4 T5 T6 T7 T8 Completion Time 𝑢 3 8 12 4 6 10 12 8

QoS Requirements

This scheduling problem can be formulated as

follows:

min 𝑛𝑏𝑙𝑓𝑡𝑞𝑏𝑜

s. t. x, = 1 t ∈ 𝒰, r ∈ ℛ
∑

T , x, ≤ 𝒠

x, ∈ 0, 1

SLIDE 8

2/7/2018 8

QoS Requirements

Cost represents the cost related for workflow tasks to

complete all its tasks. The aim is to minimize the total execution cost of a workflow

The scheduling objective can be formulated as follows

minimize T t,r + C t,r

s. t.

T t, r = 𝐹𝑈 t,r − t t − t C t,r = 𝑑 t, r − c c − c

– T t, r : the execution time of task t on r – C t, r : the monetary cost for executing task t on r. – t : the maximum execution time – t : the minimum execution time – c : the maximum monetary cost – c : the minimum monetary cost.

Multi-objective Criteria

The workflow scheduling problem becomes more

challenging when we consider multiple QoS parameters.

Meta-heuristic methods or search-based strategies have

been used to achieve good solutions are quite common.

Research problem: meta-heuristics or search-based

strategies usually need significantly high planning costs in terms of the time consumed to produce good results, which makes them less useful in real platforms that need to obtain map decisions on the fly.

Todays Agenda

Introduction
What is workflow?
What are major QoS requirements?
How is workflow represented?
Our work in workflow scheduling
Research problems
Questions and conclusions

SLIDE 9

2/7/2018 9

Workflow representation

tentry

t2

texit

t1 t3 t4 t5

Directed acyclic graph

dummy node dummy node precedence constraint task

The workflow application tasks are dependent on each
ther,

– the output of some tasks is the input to another. – the order of their execution must be considered when assigning the tasks to VM.

Workflow representation

How do we represent workflow for processing?

Tasks 1 2 3 4 1 1 1 2 1 3 1 4

t1 t2 t3 t4

7 4 9 6

Tasks 1 2 3 4 1 4 7 2 9 3 6 4

Task Interdependence Link cost

The numbers on the link are estimated transfer

time/transfer cost of sending the required data along the link

Todays Agenda

Introduction
What is workflow?
What are major QoS requirements?
How is workflow represented?
Our work in workflow scheduling
Research problems
Questions and conclusions

SLIDE 10

2/7/2018 10

Layered workflow scheduling algorithm

For effective scheduling and execution of data intensive

workflows in cloud

– How many VM we need to execute the workflow?

Structure aware workflow resource estimation and

provisioning approach

– Aim is to overcome the problems of overprovisioning and under provisioning. – The number of VMs required to execute the workflow.

The number of tasks in the workflow and
the arrangement of tasks in the workflow

Aziz, Maslina Abdul, Jemal Abawajy, and Tutut Herawan. "Layered workflow scheduling algorithm." Fuzzy Systems (FUZZ-IEEE), 2015 IEEE International Conference on. IEEE, 2015.

Layered workflow scheduling algorithm

Layer demarcation

– The first step is to determine the number of physical layers (PL) in the workflow – This is done based on the location of tasks in the workflow. – For example, we create a table with entry for each task and its layer based on the tasks distance from the entry task

Layer 1 Layer 2 Layer 3 Layer 4

Aziz, Maslina Abdul, Jemal Abawajy, and Tutut Herawan. "Layered workflow scheduling algorithm." Fuzzy Systems (FUZZ-IEEE), 2015 IEEE International Conference on. IEEE, 2015.

Layered workflow scheduling algorithm

Prioritize tasks based on

their dependencies and computation time.

This is achieved as follows

– We keep a priority list – Forward scan - look at the children tasks – Backward scan - look at the parent task

Layer 1 Layer 2 Layer 3 Layer 4

Aziz, Maslina Abdul, Jemal Abawajy, and Tutut Herawan. "Layered workflow scheduling algorithm." Fuzzy Systems (FUZZ-IEEE), 2015 IEEE International Conference on. IEEE, 2015.

SLIDE 11

2/7/2018 11

Priority Generation

Layer 1 2 3 4 Task / Dependency T1 T2 T3 T4 T5 T6 T7 T8 T9 T10 Comp 20 20 20 15 5 5 10 15 20 20 Comm 8 4 2 8 4 8 8 8 8 Task T1

√ √ √ √ √

T2

× √ √

T3

× √

T4

× √ √

T5

× √

T6

× √

T7

× √

T8

× × × √

T9

× × × √

T10

× × ×

Total 5 3 2 3 2 2 2 4 4 3 New Rank 1 2 4 3 5 6 9 8 7 10 New List T1 T2 T4 T3 T5 T6 T9 T8 T7 T10

In layer one task T

is ranked

first since this is the t. In layer two there are five tasks with different number of dependencies.

T and T have the same

number of dependencies of the

ther three tasks.
Tasks T,T and T also have

the same total dependencies.

For T and T, the same

amount of dependencies and computation time, the communication time is considered Completion or communication time is used as a tie breaker

Layered workflow scheduling algorithm

Layer 1 Layer 2 Layer 3 Layer 4

Aziz, Maslina Abdul, Jemal Abawajy, and Tutut Herawan. "Layered workflow scheduling algorithm." Fuzzy Systems (FUZZ-IEEE), 2015 IEEE International Conference on. IEEE, 2015.

Hire 1 VM Hire 5 VM Hire 3 VM Hire 1 VM It is customary to considered fixed number

f resources to the whole

life time of the workflow application. We choose to follow different approach: resources can be acquired at any time and released when they are idle, which save the total charged cost

Determining Required VMs

The first step is to determine the

number of VMs needed to execute the workflow

Maximum number of VMs = =

Task

R1 R2 R3 Processor List (Pi) T1 R1 T2 20 28 28 R1 T4 40 22 22 R2 T3 40 37 24 R3 T5 40 37 44 R2 T6 40 42 44 R1 T9 45 42 44 R2 T8 45 62 46 R3 T7 45 62 60 R1 T10 55 62 60 R1

1 3 2

1,2

3

2,3

1

1,3

2

1,2,3

The second step is to refine the

number of VMs by applying cost guided compaction process

SLIDE 12

2/7/2018 12

Resource allocation and result

20 40 60 80 R1 R2 R3

T8 T3 T10 T1 T7 T2 T4 T6 T5 T9

20 40 60 80 R1 R2 R3

T10 T2 T5 T7 T9 T3 T6 T8 T4 T1

20 40 60 80 R1 R2 R3

T10 T7 T5 T3 T4 T6 T8 T1 T2 T9

HLFET and ETF with the makespan of 88 mins

MCP with the makespan of 85 mins LWFS with the makespan of 82 mins

Result for small (10 tasks) workflow

Research problems

Are the current workflow scheduling approaches

ideal/sufficient?

– Research problem: Tasks can have very different I/O and computational behavior. – Research problem: Workflows have different deadline and monetary constraints. – Research problem: Users may have various workflow application scenarios.

Research problem: Minimize the monetary cost of workflows,

addressing both the price and performance dynamics in clouds

Research problem: The majority of studies about workflow

scheduling focus on single workflow application scheduling. However, these approaches are not adequate for cloud infrastructures.

Research problems

Are the current workflow scheduling approaches

ideal/sufficient?

– Research problem: Majority of workflow provisioning approaches consider only one on-demand instance type. – Research problem: Majority of workflow provisioning approaches are ignorant to the special features and

ptimization opportunities in workflows

– Research problem: Rarely other components such as storage system are considered – Research problem: Compare cost and performance of different distributed storage systems for sharing data in workflows – Research problem : Existing algorithm do not consider the various loads of the available resources.