Ravi Chidansh Hema

1
A Grid system with different scheduling
strategies and dynamically choosing an
appropriate scheduling strategy

• Ravi Chidansh Hema

2
Agenda

• Introduction
• Problem statement
• Our approach
• Scheduling algorithms identified
• Proposed Architecture
• Literature Survey

3
Key messages

• Relevance of this work
• Key scheduling algorithms / methods
• One good paper on each algorithm

4
Introduction

• What is Scheduling?
• Allocation of tasks (computational
  communicational) to resources within a time frame
• Why Scheduling?
• Optimize the allocation of tasks to resources to
  achieve performance
• How different is Grid Scheduling from general
  scheduling?
• Resources highly shared
• Heterogeneous distributed resources
• Dynamic arrival of jobs
• Dynamic nature of resources
• Different resource policies

5
Problem Statement

• A Grid system with different scheduling
  strategies and dynamically choosing an
  appropriate scheduling strategy
• Different set of applications
• Different scheduling algorithms available
• Best algorithm for a given application
• Choose subset of resources depending on the
  selected algorithm

6
Our Approach

• Identify and choose appropriate algorithms
• Identify the criteria and scenarios for
  evaluation
• Evaluate the criteria for the algorithms
• Benchmark the algorithms
• Understanding the application requirements
• Get optimal algorithm based on the benchmark
  information
• Return the subset of resources

7
Criteria For Evaluation of Algorithms

• Interference
• considers the impact of job scheduling on
  currently running jobs
• Computation cycle more tasks are assigned to the
  same processor
• Communication bandwidth same cluster and share
  communication bandwidth.
• Three interference-based scheduling policies
• Minimize Total Interference (MTI)
• Minimize Completion Time (MCT)
• Minimize Num Interference (MNI)

8
Criteria For Evaluation of Algorithms(contd)

• Parameter Sweep
• These applications are typically structured as
  sets of experiments,each of which is executed
  with a distinct set of parameters.
• Although parameter sweep experiments are
  independent (i.e. do not communicate), many PSAs
  are structured so that distinct experiments share
  large input files, and produce large output
  files.
• Breakdown, proximity, affinity,etc.

9
Scheduling algorithms identified

• Min-Min
• Iteratively schedules the tasks with the minimum
  completion time on the host that completes it the
  fastest
• Max-Min
• Iteratively schedules the tasks with the maximum
  completion time on the host that completes it the
  fastest Simulated Annealing
• Sufferage
• The application will suffer if it is not
  allocated to the particular resource
• Simulated Annealing
• Global optimization search criteria

10
Scheduling algorithms identified(Contd)

• Meta scheduling
• Takes into consideration the computational and
  communication tasks to compute the total
  completion time
• Linear Programming
• Linear objective function with linear constraints
• APST
• Takes into account the resource performance
  estimates to generate a plan
• AppLeS
• Adaptively selecting the schedule for the
  application
• Nimrod
• Economies of scale

11
Proposed Architecture

App 1
App 2
App 3
App n

• Algorithms
• Min-Min
• Max-Min
• Sufferage

Application Information
Optimal subset Of Resources For the given
application
Main Scheduler
Benchmark information Of Algorithms
Ranking Daemon

• Criteria
• Interference
• Affinity
• Parameter
• Sweep

Resource Information

NWS
MDS
12
The Benchmark Table
1- Worst Case and 4- Best Case
13

• Literature Survey

14

• Meta Scheduling
• Interference Paradigm for Network Job Scheduling
• AppLeS
• Simulated Annealing

15
AppLes Architecture
16
Metascheduling

• 3 categories of applications concurrent,
  pipelined and parallel
• Completion Time TCT computational time
  Communication time
• TCT comp time (ci , Sk) Commu_time (ci, cj,
  si, sj)

App Res info
MS
Application Instantiation
Cost models
17
Simulated Annealing

• Analogy Annealing of metals
• A technique that can be applied to any
  minimization / learning process based on
  successive update steps
• Update steps deterministic / random
• Length of the steps proportional to a parameter
  temperature)
• By a heuristic search it allows a non improving
  move to a neighbor with a probability that
  decreases with time. Rate of decrease known as
  the cooling schedule
• Ex. Cauchy annealing and Boltzmann annealing

18
References

• A meta scheduler for the Grid Sathish S.
  Vadhiyar and Jack J. Dongarra
• The Interference Paradigm for network job
  scheduling J.B. Weissman
• Adaptive Computing on the Grid using AppLeS F.
  Berman, R. Wolski, H. Casanova
• Metascheduling A Scheduling Model for
  Metacomputing Systems - J.B. Weissman
• Adaptive Simulated Annealing (ASA) L. Ingber
• Grid Resource Management State of the Art and
  Future Trends by J. Nabrzyski et. al, Kluwer
  Academic Publishers
• Economic Paradigm for Resource Management and
  Scheduling in Grid Computing Rajkumar B, et al.
• Heuristics for Scheduling Parameter Sweep
  Applications in Grid Environments H Casanova,
  et al.

19

• Questions?

20
Thank You