Task Planning on the Grid

Slide 1

Slide 1 text

TASK PLANNING ON THE GRID Chad Taylor

Slide 2

Slide 2 text

Beware: OPINIONS ahead!

Slide 3

Slide 3 text

This is interesting!

Slide 4

Slide 4 text

Interesting = Ninjas ⋀ Pirates ⋀ Zombies

Slide 5

Slide 5 text

Let’s pretend...

Slide 6

Slide 6 text

No content

Slide 7

Slide 7 text

No content

Slide 8

Slide 8 text

One day...

Slide 9

Slide 9 text

This image is licensed under the Creative Commons Attribution 3.0 Unported license. Attribution: Laura Poitras / Praxis Films

Slide 10

Slide 10 text

No content

Slide 11

Slide 11 text

No content

Slide 12

Slide 12 text

No content

Slide 13

Slide 13 text

No content

Slide 14

Slide 14 text

No content

Slide 15

Slide 15 text

No content

Slide 16

Slide 16 text

...

Slide 17

Slide 17 text

...

Slide 18

Slide 18 text

... I like turtles!

Slide 19

Slide 19 text

No content

Slide 20

Slide 20 text

Motivation? Check.

Slide 21

Slide 21 text

Do you even CLOUD, bro?

Slide 22

Slide 22 text

Why cloud?

Slide 23

Slide 23 text

Why cloud? Homogeneity

Slide 24

Slide 24 text

Why cloud? Homogeneity Reliability

Slide 25

Slide 25 text

Why not cloud?

Slide 26

Slide 26 text

Why not cloud? $

Slide 27

Slide 27 text

What percentage of time is your computer running 100% CPU usage?

Slide 28

Slide 28 text

What percentage of time is your organization’s computers running 100% CPU usage?

Slide 29

Slide 29 text

The Grid

Slide 30

Slide 30 text

Why grid?

Slide 31

Slide 31 text

Why grid? Use existing resources

Slide 32

Slide 32 text

Why grid? Use existing resources $

Slide 33

Slide 33 text

Challenges?

Slide 34

Slide 34 text

Challenges? Heterogeneous resources

Slide 35

Slide 35 text

Challenges? Heterogeneous resources Dynamic nature

Slide 36

Slide 36 text

Planner DAG GRID Services Executor (Yu, et al. 2007) (Blythe, et al. 2005)

Slide 37

Slide 37 text

Planner DAG GRID Services Executor (Yu, et al. 2007) (Blythe, et al. 2005)

Slide 38

Slide 38 text

Planner Performance History Repository Predictor Scheduler Scheduler (Yu, et al. 2007)

Slide 39

Slide 39 text

Performance History Repository

Slide 40

Slide 40 text

Stores: * Task type * Input task parameters * Resource characteristics * Measured fitness values (Burton, et al. 1997)

Slide 41

Slide 41 text

NoSQL?

Slide 42

Slide 42 text

Predictor

Slide 43

Slide 43 text

Learn task performance characteristics and make accurate predictions for future task executions Goal:

Slide 44

Slide 44 text

Artificial neural network with on-line learning Algorithm:

Slide 45

Slide 45 text

Implementation: * One neural network per task type * Train in real-time with data in repository * Skip expensive backtracking costs * Custom hardware? (Burton, et al. 1997)

Slide 46

Slide 46 text

* How many nodes are required for each ANN? * How to prevent unlearning? Challenges:

Slide 47

Slide 47 text

Scheduler

Slide 48

Slide 48 text

Create accurate, optimal schedules Goal:

Slide 49

Slide 49 text

What is a schedule?

Slide 50

Slide 50 text

What is a schedule? “Assignment of tasks to specific time intervals of resources” (Fibich, et al. 2005)

Slide 51

Slide 51 text

A schedule with 5 tasks and 3 resources

Slide 52

Slide 52 text

What is an optimal schedule?

Slide 53

Slide 53 text

What is an optimal schedule? A schedule that “minimizes a given optimality criterion” (Fibich, et al. 2005)

Slide 54

Slide 54 text

* makespan * maximum lateness * # tardy jobs Optimality criteria: (Fibich, et al. 2005)

Slide 55

Slide 55 text

Scheduler Performance History Repository Predictor Symbiosis!

Slide 56

Slide 56 text

Scheduling is NP-Complete (García-Galán, et al. 2012)

Slide 57

Slide 57 text

Scheduling is NP-Complete (García-Galán, et al. 2012) :’-(

Slide 58

Slide 58 text

Multi-objective Particle Swarm Optimization (MOPSO) Algorithm:

Slide 59

Slide 59 text

PSO that minimizes more than one optimization function

Slide 60

Slide 60 text

Demo

Slide 61

Slide 61 text

What next?

Slide 62

Slide 62 text

No content

Slide 63

Slide 63 text

Compare MOPSO against other algorithms (e.g., multi-objective GAs)

Slide 64

Slide 64 text

What about pirates?

Slide 65

Slide 65 text

No content

Slide 66

Slide 66 text

Questions? https://github.com/tessellator/mopso @tessellator

Slide 67

Slide 67 text

References Blythe, James, et al. "Task scheduling strategies for workflow-based applications in grids." Cluster Computing and the Grid, 2005. CCGrid 2005. IEEE International Symposium on. Vol. 2. IEEE, 2005. Burton, Bruce, et al. "Identification and control of induction motor stator currents using fast on-line random training of a neural network." Industry Applications, IEEE Transactions on 33.3 (1997): 697- 704. Coello Coello, Carlos A., and Maximino Salazar Lechuga. "MOPSO: A proposal for multiple objective particle swarm optimization." Evolutionary Computation, 2002. CEC'02. Proceedings of the 2002 Congress on. Vol. 2. IEEE, 2002. Fibich, Pavel, Ludek Matyska, and Hana Rudová. "Model of grid scheduling problem." Exploring Planning and Scheduling for Web Services, Grid and Autonomic Computing (2005): 17-24. García-Galán, S., R. P. Prado, and J. E. Muñoz Expósito. "Fuzzy scheduling with swarm intelligence- based knowledge acquisition for grid computing." Engineering Applications of Artificial Intelligence 25.2 (2012): 359-375. Reyes-Sierra, Margarita, and CA Coello Coello. "Multi-objective particle swarm optimizers: A survey of the state-of-the-art." International Journal of Computational Intelligence Research 2.3 (2006): 287- 308. Schopf, Jennifer M. "A general architecture for scheduling on the grid." Special issue of JPDC on Grid Computing 4 (2002). Yu, Zhifeng, and Weisong Shi. "An adaptive rescheduling strategy for grid workflow applications." Parallel and Distributed Processing Symposium, 2007. IPDPS 2007. IEEE International. IEEE, 2007.