Distributed Task Scheduling for Physics Fusion Applications J. - - PowerPoint PPT Presentation

▶
distributed task scheduling for physics fusion
SMART_READER_LITE
LIVE PREVIEW

Distributed Task Scheduling for Physics Fusion Applications J. - - PowerPoint PPT Presentation

Jun 16, 2023 460 likes •632 views

Enabling Grids for E-sciencE Distributed Task Scheduling for Physics Fusion Applications J. Herrera A. Cappa E. Huedo M.A. Tereshchenko R.S. Montero F. Castejn I.M. Llorente www.eu-egee.org EGEE-II INFSO-RI-031688 EGEE and gLite are

slide-1
SLIDE 1

EGEE-II INFSO-RI-031688

Enabling Grids for E-sciencE

www.eu-egee.org

EGEE and gLite are registered trademarks

Distributed Task Scheduling for Physics Fusion Applications

  • J. Herrera
  • E. Huedo

R.S. Montero I.M. Llorente

  • A. Cappa

M.A. Tereshchenko

  • F. Castejón
slide-2
SLIDE 2

2

Enabling Grids for E-sciencE

EGEE-II INFSO-RI-031688

Index

  • Self-schedulers Algorithms
  • Grid Trapezoid Self-Scheduler
  • MARATRA
  • Results
  • Conclusion
slide-3
SLIDE 3

3

Enabling Grids for E-sciencE

EGEE-II INFSO-RI-031688

Self-Schuler Algorithms

  • Static Algorithms

– The chunk size is fixed in compile-time – Advantages: Minimum scheduling overhead – Disadvantages: Load imbalance

  • Dynamic Algorithms

– The chunk size is adjusted in execution time – The most important algorithms:

CSS GSS

TSS

FSS FISS

– Advantages: A good load balance – Disadvantages: No suitable to distributed environments

slide-4
SLIDE 4

4

Enabling Grids for E-sciencE

EGEE-II INFSO-RI-031688

Self-Schuler Algorithms

  • Distributed Algorithms

– Node characteristics: cpu speed, memory size, etc – More important algorithms:

DTSS

DFSS DFISS

DTFSS

– Advantages: Suitable to heterogeneous cluster – Disadvantages: Not fixed to Grid environments

slide-5
SLIDE 5

5

Enabling Grids for E-sciencE

EGEE-II INFSO-RI-031688

Grid Trapezoid Self-Scheduler

  • Main characteristics:

– Grid characteristics: a high degree of heterogeneity, high fault rate, dynamic resource availability, etc. – Distributed and dynamic algorithm – Better workload balancing between all Grid resources – Decrease of the scheduling overhead – Transparent from the user point of view

  • Three pillars:

– Grid Benchmarking Model – User statistics – Trapezoid Self-scheduler (TSS)

slide-6
SLIDE 6

6

Enabling Grids for E-sciencE

EGEE-II INFSO-RI-031688

Grid Benchmarking Model

  • Provides a way to investigate performance properties
  • f Grid environments
  • Two main parameters:

– Asymptotic performance (r): the maximum rate of performance

in task executed per second

– Half-performance length (n1/2): the number of task required to

  • btain the half of asymptotic performance
  • 2 x n1/2 testbed equivalent nodes number
slide-7
SLIDE 7

7

Enabling Grids for E-sciencE

EGEE-II INFSO-RI-031688

Grid Benchmarking Model

slide-8
SLIDE 8

8

Enabling Grids for E-sciencE

EGEE-II INFSO-RI-031688

Grid Benchmarking Model

slide-9
SLIDE 9

9

Enabling Grids for E-sciencE

EGEE-II INFSO-RI-031688

Rmax(i)

  • The execution time ratio factor between the faster node

and the i node

  • User execution statistics
  • Ordered ranking (by time execution) of all testbed

nodes

  • GridTrapezoid Self-Scheduler uses Rmax(i) and n1/2 to
  • btain the first chunk size based on TSS scheduler
slide-10
SLIDE 10

10

Enabling Grids for E-sciencE

EGEE-II INFSO-RI-031688

MARATRA

  • MAssive RAy TRAcing in Fusion Plasmas
  • Parameter Sweep Application
  • Executable:

– Truba (traces 1 ray of the microwave bunch) – 1.8 MB – 9' (Pentium 4 3.20 Ghz) – Input files =~ 70 KB – Output files =~ 549 KB

  • Cluster Environment:

– 2 x 103 rays factor – 5 sites (higher, medium1, medium2, medium-lower, lower) – 10 nodes per site – Experiments with: CSS(40), GSS, TSS, GTSS, FSS, FISS

slide-11
SLIDE 11

11

Enabling Grids for E-sciencE

EGEE-II INFSO-RI-031688

Experiment Wall Time

slide-12
SLIDE 12

12

Enabling Grids for E-sciencE

EGEE-II INFSO-RI-031688

Load Balance

slide-13
SLIDE 13

13

Enabling Grids for E-sciencE

EGEE-II INFSO-RI-031688

Conclusion

  • New Self-scheduler algorithm

– Distributed and dynamic – Transparent from the user point of view – Adaptative – Grid characteristics

  • MARATRA experiment
  • Decrease of wall time execution
  • Load balance between all grid nodes
  • Scheduling overhead reduction
slide-14
SLIDE 14

14

Enabling Grids for E-sciencE

EGEE-II INFSO-RI-031688

Future work

  • Increase the problem size

– Objective: 104 and 105

  • Add GTSS to GridWay metascheduler
slide-15
SLIDE 15

EGEE-II INFSO-RI-031688

Enabling Grids for E-sciencE

www.eu-egee.org

EGEE and gLite are registered trademarks

Distributed Task Scheduling for Physics Fusion Applications

  • J. Herrera
  • E. Huedo

R.S. Montero I.M. Llorente

  • A. Cappa

M.A. Tereshchenko

  • F. Castejón