Performance Impact of Resource Provisioning on Workflows

1
Performance Impact of Resource Provisioning on
Workflows

• Gurmeet Singh, Carl Kesselman and Ewa Deelman
• Information Science Institute
• University of Southern California
• Presented by
• Nut Taesombut
• Large-Scale Systems Seminar
• May 9, 2005

2
Outline

• Background and Objective
• Simulation Study
• Experimental Results
• Summary Discussion

3
Background
Schedule
Workflow Application
Grid Resources

• Grid
• Resource sharing across multiple domains
• Workflow application
• A collection of tasks with specified dependencies
  
• How to map a workflow application onto
  distributed resources in such a way to minimize
  its completion time ?

4
Resource Provisioning

• Reserve resources for dedicated use by an
  application for certain timeframe
• Tradeoff between predictability of workflow
  performance and waiting time cost
• Performance predictability enables efficient
  resource selection
• Need to wait until the provisioned resources
  become available

5
Objective

• Study of performance impact of resource
  provisioning for a workflow application
• Two performance metrics
• Completion Time Wait Time Runtime
• Resource Utilization
• Three resource provisioning mechanisms
• No provisioning
• Advance reservation
• Dynamic provisioning
• Two resource scheduling policies
• FIFO (First-in, First-out)
• Variant Fair Share

6
Simulation Study

• Maui simulator
• Simulate the running of jobs on the 890-processor
  cluster with different scheduling policies
• Workload trace from the TeraGrid facility
• 2095 jobs on 10-day collection (2/22/05 3/3/05)
• 102 initial running jobs
• Record both requested and actual runtimes

7
Workload Characteristics
Actual Runtime
Requested Runtime

• More than half of the jobs requested a runtime
  between 16 and 24 hours
• Most of the jobs completed in shorter time than
  they requested

8
Application Workflows

• Original Montage workflow
• Contain 17034 tasks in 7 levels (3.1 second
  average runtime)
• 12 Clustered workflows (generated from the
  original)
• Produce workflow graphs with different levels of
  granularity and structures
• Group the tasks at each level into a fixed number
  of clusters

clustering parameter 2
clustering parameter 8
9
Advance Reservation

• Reserve resources for an application in the
  future
• Use the Maui simulator to determine the earliest
  start time
• Decide based on the requested runtimes of the
  running and queued jobs
• Request selection
• Multiple requests can execute the workflow
• Use the best request to determine the minimum
  completion time of the workflow
• Decide based on the earliest start time and the
  metric to optimize

10
Dynamic Provisioning

• Reserve resources for an application during
  runtime
• Use the Maui simulator to determine the actual
  start time
• Compute based on the actual runtimes of the
  running and queued jobs
• Request selection
• Use the best request as determined by advance
  reservation

11
Experimental Results (1) Varied Requests
FIFO

• Completion time of various provisioning requests
• The use of provisioning can reduce the completion
  time
• For FIFO with 84 processors,
• 65 in advance reservation
• 82 in dynamic provisioning
• Dynamic Provisioning vs. Advance Reservation
• FIFO D.P. always outperforms A.R.
• Fair Share D.P. and A.R. performs equally well

Fair Share
12
Experiment Results (2) Varied Workflows
FIFO

• Completion time of various workflows with their
  best request
• The use of provisioning can reduce the completion
  time
• Large workflows tend to obtain significant
  reduction

Fair Share
13
Experiment Results (3) Resource Utilization
Utilization FIFO

• Completion time and resource utilization of
  various workflows using advance reservation
• Consider utilization in optimization metric
• Resource utilization can be much improved by
  small increase in the completion time

Completion Time FIFO
14
Experiment Results (4) Resource Utilization
Utilization Fair Share

• Completion time and resource utilization of
  various workflows using advance reservation
• Consider utilization in optimization metric
• The requests that optimize resource utilization
  also optimize completion time

Completion Time Fair Share
15
Summary

• Study the performance impact of using resource
  provisioning for a workflow application
• Significant reduction in the completion time for
  both FIFO and Fair Share scheduling policies
• Amount of the reduction depends on scales and
  structures of the workflows
• Resource utilization can be improved with small
  increase in the completion time

16
Discussion

• Is the proposed model practical for Grid
  environments?
• Assume homogeneous resource capability and
  scheduling policies
• Assume deterministic and known workflow
  characteristics (e.g., runtime)
• Assume support for reservation systems