1. Deepak Poola Chandrashekar
PhD Completion Seminar
Robust and Fault-Tolerant Scheduling for Scientific Workflows in Cloud Computing Environments
Cloud Computing and Distributed Systems (CLOUDS) Laboratory,
Department of Computing and Information Systems
The University of Melbourne, Australia
Deepak Poola Chandrashekar
Supervisor: Prof. Rajkumar Buyya
Co-Supervisor: Prof. Kotagiri Ramamohanarao

2. Clusters Super Computers Grid Cloud Computing
Web Service oriented architecture
Clusters
Super Computers
Grid
Cloud Computing
Provenance, reliability, results derivation, and sharing.
Shell Scripts and Complex Job-Control Language.
Streaming Workflow, IoT based workflows.
RPC, DFS, databases, Distributed Object Technologies
Specialized worfklows for data intensive, compute-intensive, instance- intensive applications

3. Workflows
Scientific workflow systems aim at automating large complex data analysis to make it easier for scientists.
Workflows are collection of tasks that are data dependent or control dependent. Workflows can be represented as Directed Acyclic Graph
Workflow scheduling maps tasks to resources whilst maintaining dependencies
Jargons
Makespan
Cost
Deadline
Budget
Sample Workflow
With the increase in processing power and computation tools, computation has become a “third branch” in scientific research along theory and experiment. Scientific worklfow aims to automate these complex analysis and facilitates collaborative research. 3

4. Dynamically configured and delivered on demand Highly scalable
Cloud Computing
Cloud Computing
On-Demand
Dynamically configured and delivered on demand
Highly scalable
Provide different levels of services
Driven by market principles
Pay as you go
cloud computing is a model for enabling ubiquitous, convenient, on-demand access to a shared pool of configurable computing resources (e.g., networks, servers, storage, applications and services) that can be rapidly provisioned and released with minimal management effort or service provider interaction. 4

5. Cloud Variants
Image Courtesy: Windows Azure 5

6. Other Inherent Challenges
Cloud Computing
Performance variance among different types of VMs
Resource-level failures
Virtualization overhead
Failures or outages in Clouds
Other Inherent Challenges
Network latencies and failures
Task-level failures
Failures associated with storage and I/O
Failures in workflows management systems

7. Started by Amazon around December 2009
Spot Instances
Started by Amazon around December 2009
Idle or unused datacenter capacity
Spot price is decided in an Auction-like mechanism
Varies with time and instance type
Varies between regions and availability zones
bid should be higher than or equal to the spot price
Offers upto 60% cost reductions

8. Thesis Research Question
How to make workflow scheduling algorithms robust and fault-tolerant for cloud computing environments?

9. Novel scheduling heuristics Resource provisioning policies
Thesis Contributions
Novel scheduling heuristics
Resource provisioning policies
Bidding Strategies
A performance evaluation study on time, cost and fault-tolerance
Metrics to measure robustness and fault-tolerance
Multi-cloud resource plug-in

10. Thesis Organization

11. Derived from Publication
Literature Survey
Derived from Publication
Poola D., Salehi, M. A., Rao K., and Buyya R., A Taxonomy and Survey of Fault-Tolerant Workflow Management Systems in cloud and Distributed Computing Environments, Submitted to the ACM Computing Surveys. Under Review.

12. Components of workflow scheduling

13. Taxonomy of Fault-Tolerant Scheduling Techniques

14. Existing Workflow Management Systems

15. Thesis Organization

16. Robust Scheduling of Scientific Workflows
Derived from Publication
Poola D., Garg S.K., Buyya R., Yang Y., and Rao K., Robust Scheduling of Scientific Workflows with Deadline and Budget Constraints in clouds, Proceeding of the 28th IEEE International Conference on Advanced Information Networking and Applications (AINA-2014), Victoria Canada.

17. Robustness is maximized Makespan and cost is minimized
Objective
Find a schedule to map workflow tasks on heterogeneous Cloud resources such that
Robustness is maximized
Makespan and cost is minimized
Constraints : Deadline and Budget
To the best of our knowledge, there has been no study in workflow scheduling algorithm for Clouds maximizing robustness, and minimizing makespan and cost at the same time. Also there are very few works which schedule
workflow tasks on heterogeneous Cloud resources. This study tries to address these shortcomings. 17

18. Absorb some degree of uncertainty While maintaining a stable solution
Robustness
Absorb some degree of uncertainty
While maintaining a stable solution
Robustness alone is not a metric
It gives an idea of the stability of the solution
Much needed in mission and time critical applications
Achieved with redundancy in time or space 18

19. Uncertainties: two types are considered Task failures
System Model
Cloud Environment : A single datacenter with heterogeneous resources are considered.
Uncertainties: two types are considered
Task failures
Performance variations of VMs
Workflow: Represented as a DAG
Data transfer time between tasks are considered

20. Total cost: sum price of all VMs of different types used.
Definitions
Makespan: is the total elapsed time required to execute the entire workflow
Total cost: sum price of all VMs of different types used.
Robustness Metric:
Robustness probability: is the likelihood of the workflow to finish before the given deadline
Tolerance time: is the amount of time a workflow can be delayed without violating the deadline constraint

21. A workflow is divided into a set of partial critical paths PCP
Scheduling Algorithm
A workflow is divided into a set of partial critical paths PCP
PCP of a node ti is a unassigned critical parents (CP) tp , and critical parent of tp recursively until there are no further unassigned parents
CP is the parent of the node, for whom the combination of start time, execution time and transfer time is the maximum
A suitable resource is selected for this PCP 21

22. 2 1 4 1 5 1 2 4 5 3 3 2 3 2 4 3 5 22

23. 2 1 4 1 5 1 2 4 5 3 3 2 3 2 4 3 5 23

24. 2 1 4 1 5 1 2 4 5 3 3 2 3 2 4 3 5 24

25. 2 1 4 1 5 1 2 4 5 3 3 2 3 2 4 3 5 25

26. 2 1 4 1 5 1 2 4 5 3 3 2 3 2 4 3 5 26

27. 2 1 4 1 5 1 2 4 5 3 3 2 3 2 4 3 5 27

28. Scheduling Algorithm (Contd...)
Robustness Type
No Robustness
Slack
One node failure
Two node failure
Solution Set $ $ $ $ RT RT RT RT VM VM VM VM RT $ RT $ RT $ RT $ VM VM VM VM RT $ RT $ RT $ RT $ VM VM VM VM RT $ RT $ RT $ RT $ VM VM VM VM 28

29. Scheduling Algorithm (Contd...)
PCP assigning policy
Robustness-Cost-Time (RCT)
Robustness-Time-Cost (RTC)
Weighted
Feasible Solution Set $ RT VM RT $ RT $ RT $ VM VM VM $ RT VM $ $ $ RT RT RT VM VM VM
Solution $ $ RT RT VM VM 29

30. Gravitational Physics
Simulation setup
Application Model
Considered 5 workflows as presented by Bharathi et.al (WORKS’08).
Montage:
Astronomy
Epigenomics:
Biology
LIGO:
Gravitational Physics
SIPHT:
Biology
Cybershake:
Earthquake Science 30

31. Considered 10 types of VMs.
Simulation setup
Resource Model
Considered 10 types of VMs.
Fastest VM has 5 time more processing power than the slowest VM, the prices varies proportionately.
VMs differs with respect to memory, CPU cores, MIPS and OS
Virtually unlimited resources is assumed
Data transfer costs are not considered
Storage costs are not considered
The Cloud in our model has only one Datacenter 31

32. Failure Model: Two types of failure models are considered.
Simulation setup
Failure Model: Two types of failure models are considered.
Failure traces: simulated from traces.
Used Condor Grid failure dataset from failure trace archive.
Failure probability: 10% failure probability is considered.
Checkpointing mechanism is assumed
Performance variation (ICSOC’09)
Random variable from a normal distribution is added to the execution time
Standard deviation of this distribution is a function of the execution time 32

33. Reference Algorithms: Two reference algorithms are used.
Simulation setup
Reference Algorithms: Two reference algorithms are used.
ICPCP : (FGCS’12)
Robust Bi-objective genetic algorithm (GA) : (ICoCC’2006)
These algorithms are chosen for their similarity with our approach. ICPCP schedules tasks by grouping them into PCPs, similar to our algorithm and GA tries to maximize the slack time to be robust, which is the approach we adapt as well. 33

34. Results : Effect on robustness
Tolerance Time 34

35. Results : Effect on robustness
Robustness Probability 35

36. Results : Effect on Makespan36

37. Results : Effect on Cost37

38. We presented three resource allocation policies.
Summary
We presented three resource allocation policies.
Objectives being robustness, makespan, cost.
Add slack time judiciously to make the schedule robust.
Tested with 2 failure models, 5 workflows and 2 robustness metrics.
Results indicate our policies provide robust schedules with minimal makespan with slight increase in cost.

39. Thesis Organization

40. Fault-Tolerant Scheduling Using Spot Instances
Derived from Publication
Poola D., Rao K., and Buyya R., Fault-Tolerant Workflow Scheduling Using Spot Instances on Clouds, Proceeding of the 13th International Conference on Computational Science (ICCS-2014), Cairns Australia.

41. Just-in-time and adaptive scheduling heuristic
Objective
Just-in-time and adaptive scheduling heuristic
Using spot and on-demand instances
An intelligent bidding strategy
Minimizes the execution cost
Providing a robust schedule
Satisfying the deadline constraint
To the best of our knowledge, there has been no study in workflow scheduling algorithm for Clouds maximizing robustness, and minimizing makespan and cost at the same time. Also there are very few works which schedule
workflow tasks on heterogeneous Cloud resources. This study tries to address these shortcomings. 41

42. Workflow is represented a DAG Makespan is the total elapsed time
Background
Workflow is represented a DAG
Makespan is the total elapsed time
Pricing models
On-Demand
Spot
Critical Path is the longest path from the start node to the exit node

43. Latest Time to On-Demand (LTO)
It is the latest time the algorithm has to switch to on-demand instances to satisfy the deadline constraint
Start
LTO
Deadline
Spot Instances
On-Demand

44. We use Downey’s analytical model Downey’s model requires:
Runtime Estimation
We use Downey’s analytical model
Downey’s model requires:
task’s average parallelism, A,
coefficient of variance of parallelism, σ,
task length
the number of cores
Cirne et al model to generate A and σ

45. Estimates the failure probability of a particular bid price
Failure Estimator
Estimates the failure probability of a particular bid price
Based on spot price
The history price of one month prior is considered
Total time of the spot price history, HT
And total out of bid time, OBTbidt is measured

46. Scheduling Algorithm

47. Scheduling Algorithm (Contd..)

48. Scheduling Algorithm (Contd..)

49. Two type of Scheduling Algorithms
Conservative: CP and LTO is estimated on the lowest cost instance.
CP is the longest, hence less slack time
Uses spot instances cautiously under relaxed deadlines
Aggressive: CP and LTO is estimated on the highest cost instance.
CP is smallest, hence more slack time
opt on-demand instances that are expensive under failures

50. Intelligent Bidding Strategy Current spot price (pspot)
On-demand price (pOD)
Failure probability (FP) of the previous bid price
LTO
Current time (CT) α β

51. Intelligent Bidding Strategy
α : dictates how much higher the bid value must be above the current spot price
β : determines how fast the bid value reaches the on-demand price
FP of the previous bid is used as a feedback to the current bid price

52. Intelligent Bidding Strategy

53. Other Bidding Strategies
On-Demand Bidding Strategy : uses the on-demand price as the bid price.
Naive Bidding Strategy: uses the current spot price as the bid price for the instance

54. CloudSim was used for simulation
Simulation Setup
CloudSim was used for simulation
LIGO workflow with 1000 tasks was considered
For On-Demand 9 different VMs types were considered
For Spot, 1 VM type was used

55. Results : Comparison between algorithms
Mean execution cost of algorithms with varying deadline (with 95% confidence interval)

56. Results : Comparison between bidding strategies
Mean Execution Cost of bidding strategies with varying deadline (with 95% confidence interval)

57. Results : Task Failures
Mean of task failures due to bidding strategies

58. Results : Checkpointing

59. They minimize the execution cost
Summary
Two scheduling heuristics that map workflow tasks onto spot and on-demand instance are presented
They minimize the execution cost
They are robust and fault-tolerant towards out-of-bid failures and performance variations
A bidding strategy that bids intelligently to minimize the cost is presented
Demonstrates the use of checkpointing, which offers cost savings up to 14%

60. Thesis Organization

61. Workflow Scheduling Using Task Replication
Derived from Publication
Poola D., Rao K., and Buyya R., Enhancing Reliability of Workflow Execution Using Task Replication and Spot Instances, Submitted to the ACM Transactions on Autonomous and Adaptive Systems. Under Review.

62. Just-in-time and adaptive scheduling heuristic
Objective
Just-in-time and adaptive scheduling heuristic
Using spot and on-demand instances
Task replication strategy.
Minimizes the execution cost
Providing a robust schedule
Satisfying the deadline constraint
To the best of our knowledge, there has been no study in workflow scheduling algorithm for Clouds maximizing robustness, and minimizing makespan and cost at the same time. Also there are very few works which schedule
workflow tasks on heterogeneous Cloud resources. This study tries to address these shortcomings. 62

63. Essentially Critical Task

64. Proposed Solution: Heuristics
Scenario 1: Mapping task on already running spot instances
LTO ahead of current time
Map tasks on already running spot instances
Saves costs
Consolidates resource utilization
Among available free slot, the resource which can finish within the tasks’ latest finish time is chosen

65. Proposed Solution: Heuristics
Scenario 2: Mapping task on a new spot instance
If no free slot that honor deadline are found
If there is sufficient time to start a spot instance
Bid price is estimated
New spot instance is created

66. Proposed Solution: Heuristics
Scenario 3: Mapping task to an on-demand instance
LTO is before the current time
Finds suitable on-demand instance
Critical path for all instance types are calculated
First finds a free slot
If no free slot, then find instance that can honor deadline
No instance, then create an on-demand instance

67. Proposed Solution: Heuristics
Scenario 4: Task duplication under short deadline
Critical tasks are replicated to provide fault-tolerance
Two variants are proposed
Essential Critical Path Task Replication (ECPTR)
ESCTs are replicated
Critical Task Replication(CTR)
All critical tasks are replicated

68. Results: Effect on Fault-Tolerance
Failure Probability
Tolerance Time

69. Results: Effect on Makespan
Mean Makespan
Resource Consolidation

70. Results: Effect on Cost
Mean Cost
Resource Consolidation

71. Thesis Organization

72. Multi-Cloud Integration for WFMS
We extend cloudbus workflow management system
The broker is enhanced to provision resources dynamically
Effective fault-tolerant technique is developed
Resource provisioning algorithm for multi-cloud is proposed

73. Cloudbus Workflow Management System
Workflow Portal
I/O Models
Many-to-many, many-to-one, synchronization
Workflow Engine
Workflow submission handler
Workflow scheduler
Event Service
Resource Broker

74. Components of workflow scheduling

75. Motivation for Multi-Cloud
Pricing
Billing Periods
Resource Characteristics
Private Cloud
Regulatory and Legal

76. Multi-cloud resource provisioning policy
Create a pool of spot instances
If free resource allocate
Else, if budget is there, create an on-demand instance
If no budget
Find a suitable resource that will complete the task in the earliest

77. Effect on Makespan under failures Resource Instantiation time
Results
Effect on Makespan under failures
Resource Instantiation time

78. Output

79. Proposed novel fault-tolerant heuristics
Summary
Proposed novel fault-tolerant heuristics
Experimented with various fault-tolerant techniques
Extensive literature survey
Resource allocation policies
Bidding Strategies
Multi-cloud integration for workflow management systems

80. Future Research Directions
Cloud Failure Characteristics
Metrics for fault-tolerance
Cloud Pricing Models
Multiple tasks on a single instance
Workflow specific scheduling
Multi-cloud challenges

81. © Copyright The University of Melbourne 2009