Integrated Scheduling of EndtoEnd Aperiodic and Period Tasks

1
Integrated Scheduling of End-to-End Aperiodic and
Period Tasks

• Chenyang Lu
• Yuanfang Zhang, Chris Gill
• Distributed Object Computing Center
• Department of Computer Science and Engineering

2
Motivation

• Aperiodic tasks in ARMS systems
• have hard end-to-end deadlines
• are critical to the system
• Existing middleware and tools lack scheduling
  support for aperiodic tasks

3
Goals

• Integrated schedulability analyses for aperiodic
  and periodic tasks
• Offline analysis of predefined scenarios ?
  certification
• Online analysis ? resource allocation and control
• Investigate two approaches
• Aperiodic Utilization Bound (AUB)
• Deferrable Server (DS)

4
Accomplishments

• Implemented both schedulability analyses
• Developed efficient heuristics for tuning
  deferrable servers
• Compared two approaches via numerical study
• Implemented deferrable server in TAOs federated
  event channel new!

5
Aperiodic Utilization BoundSynthetic Utilization

• Task Ti
• Ai Arrival time
• Di End-to-end deadline
• Ci,j Subtask execution time on node j
• Aperiodic or periodic
• Current tasks S(t) Ti Ai lt t lt AiDi
• Tasks that have arrived but whose deadlines have
  not expired
• Synthetic utilization of node j

6
Aperiodic Utilization BoundSchedulability Tests

• End-to-end deadlines of all tasks are guaranteed
  if
• Pipeline
• Arbitrary graph

4
2
3
1
T.F. Abdelzaher, G. Thaker, P. Lardieri, A
Feasible Region for Meeting Aperiodic End-to-end
Deadlines in Resource Pipelines, ICDCS 2004.
7
Deferrable Server

• Incorporate aperiodic tasks into a periodic
  scheduling framework
• Server a periodic task responsible for
  processing aperiodic requests.
• Budget maximum time the server can run in a
  period
• Server is suspended when its budgets runs out
• Bound aperiodic tasks impact on periodic tasks
• Servers budget is replenished in the end of each
  period

8
Deferrable ServerSchedulability Test for
Periodic Tasks
Interference from server

• Subtask Tijs response time
• es,j budget of server on node Pj
• Ps,j period of sever on Pj
• Caj execution time of aperiodic subtask on Pj
  (1am)
• Task Ti is schedulable if
• ni number of subtasks of task Ti

Interference from periodic tasks with higher
priorities
9
Deferrable ServerSchedulability Test for
Aperiodic Tasks

• Divide end-to-end deadline into subdeadlines
• Subdeadline Dij is guaranteed if

10
Deferrable ServerSelect Parameters

• Assign subdeadlines Deadline ? subdeadlines
• Select budget for the server on each processor
• Subdeadlines, budget ? period
• Select period based on schedulability test for
  aperiodic tasks ? all aperiodic tasks are
  schedulable
• Test schedulable for periodic tasks

11
Assign Subdeadlines

• Proportional deadline (PD) Subdeadline
  proportional to execution time
• Even deadline (ED) Evenly divide deadline

12
Select Budget

• Smaller budget ?
• Shorter period ? better response time
• Higher overhead
• Budget should be a common divisor (CD) of the
  total execution time of all aperiodic subtasks on
  the same processor ? reduce response time
• Minimum common divisor
• bounded by the granularity of scheduling
• Maximum common divisor

13
Simulation Settings

• Workload All aperiodic and periodic tasks arrive
  simultaneously
• worst-case scenario
• Offline certification
• Baseline Exhaustive search (ES) of all
  combinations of subdeadline assignments and
  budgets
• Represent the best case
• Exponential complexity

14
Simulation Results
Min CD performs almost as well as ES!
AUB is more pessimistic when used offline
Max CD results in lower schedulability than min CD
15
Summary

• Both DS and AUB provide efficient, integrated
  schedulability analysis for aperiodic/period
  tasks
• Efficient heuristic for DS tuning
• Minimum Common Divisor
• Even Deadline or Proportional Deadline
• Comparison
• DS is less pessimistic than AUB when applied
  offline
• DS requires more complex run-time mechanism

16
Future

• Validate analyses via empirical studies on TAO
• Investigate online case
• Integrate with resource allocation and control
• Integrate with RACE