Scheduling of Parallel Jobs In a Heterogeneous MultiSite Environment

1
Scheduling of Parallel Jobs In a Heterogeneous
Multi-Site Environment

• By Gerald Sabin from Ohio State
• Reviewed by Shengchao Yu
• 02/2005

2
Overview job scheduling

• Most research focused on homogeneous multi-site
  environment
• Limited research on heterogeneous systems
• Each job mapped to a single processor
• Only addressed the scheduling of independent
  sequential jobs or precedence constrained task
  graphs where each task is sequential

3
Problem of the Paper

• Target system
• A heterogeneous multi-site environment with each
  site as a homogeneous cluster of processor, but
  processors at different sites having different
  speeds
• A stream of parallel jobs submitted to a
  metascheduler

4
Goal Approach

• Goal
• Find an effective schedule to achieve optimized
  average turnaround time of the jobs.
• Basic Approach
• Extension of proven back-filling based parallel
  scheduling in single-site scheduling
• Multiple (or K) simultaneous requests (MR)
• Completion-based conservative backfilling
• Use of Effective Utilization (Efficacy), instead
  of raw utilization

5
Extend Backfilling with MR

• Backfilling conservative, aggressive
• Move forward smaller jobs to fill idle process
  cycles without delaying any jobs with future
  reservation
• Multiple Request (MR)
• Submit each job to multiple sites, and cancel
  redundant submissions once the job started on one
  of the sites

6
Extend Backfilling with MR
7
Adapt Backfilling to Heter Sys

• Characteristics on Heter. Sys.
• The same application performs differently on
  different sites
• No site is the fastest on all applications
• Simple greedy and MR schemes fail
• Introduction of Efficacy and Effective
  Utilization
• Adoption of completion-based conservative
  backfilling

8
Introduction of Efficacy and Effective Utilization
9
Aggressive Vs. Conservative

• Prefer completion time rather than start time
• Conservative backfilling chosen
• Higher Effective Utilization
• Smaller Average Turnaround time
• Fig. 6, 7, 8

10
Aggressive Vs. Conservative
11
Aggressive Vs. Conservative
12
Efficacy Based Scheduling

• Replace Greedy Policy (FCFS) with Efficacy
• Use efficacy as the priority order for the jobs
  in the queue, so that jobs with higher efficacies
  will attempt to backfill before jobs with lower
  efficacies
• In another word, a job will have a greater chance
  to run on its faster machines.

13
Efficacy Based Scheduling
14
Restricted MR

• Communication Overhead due to
• Deciding earliest completion time
• Data transfer for a job
• Select K sites with best completion time
• All about compromise
• Fig 11, 12

15
Restricted MR
16
Restricted MR
17
Conclusion

• Extension of backfilling with the application of
  MR, Completion time based conservative scheme and
  Efficacy in a orthogonal fashion
• A good attempt to optimization across a
  collection of independent jobs for GRID
  environment with the characteristics of the
  heterogeneous system in mind
• A Heuristic Approach
• Scalability and Fault-tolerance
• Sensitivity to the type of applications