a dynamic critical path algorithm for scheduling scientific workflow applications on global grids
DESCRIPTION
A Dynamic Critical Path Algorithm for Scheduling Scientific Workflow Applications on Global Grids. Report: Wei-Cheng Lee. e-Science IEEE 2007. Abstract. - PowerPoint PPT PresentationTRANSCRIPT
2008-06-04
A Dynamic Critical Path Algorithm for Scheduling Scientific Workflow Applications on Global Grids
Report: Wei-Cheng Leee-Science IEEE 2007
2008-06-04 2
Abstract
Dynamic Critical Path (DCP) based workflow scheduling algorithm that determines efficient mapping of tasks by calculating the critical path in the workflow task graph at every step.
It assigns priority to a task in the critical path which is estimated to complete earlier.
Using simulation, we have compared the performance of our proposed approach with other existing heuristic and meta-heuristic based scheduling strategies for different type and size of workflows.
2008-06-04 3
Outline
The Proposed DCP-G Algorithm2
Performance Evaluation4
Introduction31
Calculation of AEST and ALST in DCP-G33
Results and Observations35
Conclusions4
2008-06-04 4
Introduction
Dynamic Critical Path (DCP)
Dynamic Critical Path for Grids (DCP-G)
Design for mapping tasks on to homogeneous processors.
Map and schedule tasks in a workflow on to heterogeneous resources in a dynamic Grid environment.
Heuristic
Meta-heuristic
Myopic, Min-Min, Max-Min, HEFT.
GRASP, Genetic Algorithm.
2008-06-04 5
The Proposed DCP-G Algorithm
For a task graph, the lower and upper bounds of starting time for a task are denoted:
Absolute Earliest Start Time (AEST)
Absolute Latest Start Time (ALST)
DCP algorithm DCP-G• The initial AEST and ALST values are calculated for the resource which provides the minimum execution time for the task.
• For mapping a task on the critical path, all available resources are considered by DCP-G
• When a task is mapped to a resource, its execution time and data transfer time from the parent node are updated accordingly.
2008-06-04 6
Calculation of AEST and ALST in DCP-G
In DCP-G, the start time of a task is not finalized until it is mapped to a resource.
Absolute Execution Time
Absolute Data Transfer Time
2008-06-04 7
Here, the communication time between two tasks is considered to be zero if they are mapped to the same resource, and equal to the ADTT of parent task, if the child is not mapped yet.
Dynamic Critical Path Length
ALST of a task t in resource R, denoted as ALST(t, R)
Calculation of AEST and ALST in DCP-G
2008-06-04 8
Resource Selection
To select the resource that provides the minimum execution time for that task.
This is discovered by checking all the available resources for one that minimizes the potential start time of the critical child task on the same resource, where the critical child task is the one with the least difference of AEST and ALST among all the child tasks of the critical task.
2008-06-04 9
DCP-G Example
2008-06-04 10
T0`s AET is: 150000MI / 1500MIPS = 100/PerSecond
T0`s ADTT is: 1GByte / 200Mbps = 8Gb/200Mbps
= 40/PerSecond
T2`s AEST is: 0 + 100 + 40 = 140
T2`s ALST is: 420 – 200 – 80 = 140
DCPL is : 840 + 50 = 890
DCP-G Example
2008-06-04 11
T2`s AEST is: 0 + 100 + 0 = 100
T2`s ALST is: 380 – 200 – 80 = 100
If T4 was mapped to R2, the DCPL should equal to 770.
Because the AEST(T4) is: 300+400+20=720
So, DCPL is: 720 + 50 = 770.
Whereas, mapping to R1, AEST(T4) is: 300 + 400 + 0 = 700, and DCPL is: 700 + 50 = 750
DCP-G Example
2008-06-04 12
Performance Evaluation
1. Workflow model.
2008-06-04 13
Performance Evaluation
2. Resource model.
2008-06-04 14
Results and Observations
1. Execution time in static environment.
2008-06-04 15
Results and Observations
2008-06-04 16
Results and Observations
2. Execution time in dynamic environment.
2008-06-04 17
Results and Observations
2008-06-04 18
3. Scheduling time.
GRASP creates Restricted Candidate List (RCL) for each unmapped ready task and then selects resource for the task in random.
Results and Observations
2008-06-04 19
Conclusion
It is evident that among the heuristics- based scheduling techniques, DCP-G can generate better schedule by up to 20% in static environment, especially for random and parallel workflow, irrespective or workflow size.
GA and GRASP can generate more effective schedule than DCP-G for random and fork-join workflow, but they suffer from the problem of higher scheduling time.
In dynamic environment, heuristics based techniques adapt to the dynamic nature of resources and can avoid performance degradation. But metaheuristics based techniques perform worse in this situation due to the unavailability of mapped resources at certain intervals.
2008-06-04
Wei-Cheng Lee