Flexible Scheduling of Software with Logical Execution Time Constraints* Stefan Resmerita and Patricia Derler University of Salzburg, Austria *UC Berkeley,

Flexible Scheduling of Software with Logical Execution Time Constraints*

Stefan Resmerita and Patricia DerlerUniversity of Salzburg, Austria

*UC Berkeley, USA

October 14, 2009 2RePP: Flexible Scheduling of LET Software

Introduction

• Scope: Embedded software aplications– Set of periodic tasks with predictable timing behavior

– Preemptive scheduling

– Event-triggered tasks

• Problem: Predictability is achieved by restricting the set of feasible schedules

• Aim: Relax scheduling restrictions while preserving predictability


The LET Programming Model

• Specification of logical execution times for tasks– Giotto, TDL, HTL, xGiotto, FTOS

• Implementation – Dedicated runtime system


Main Runtime Operations

1. Update outputs at LET end

2. Invoke task at LET starta) Update inputs

b) Release task for execution


Scheduling

• High-level: scheduling of operations– Static schedule compiled into a „timing program“

– Platform independent

• Low-level: scheduling of task executions– Platform dependent

– May use any policy (e.g., FPS, EDF)

– Schedulability test uses WCET information

• What if the system is not schedulable?


Trade-offs

• Increased predictability– Separation of timing from functionality– Separation of reactivity from scheduling

• Platform independence– Portable timing program

• Performance costs– Application performance (response time)– Platform requirements (memory/time)– Processor utilization (idle time)


This Work

• Provides a methodology for obtaining more flexible high-level schedules

• Keep predictability• Increase processor utilization

• Cost: Portability is reduced– Provide tool support


Enlarging Scheduling Margins

• Use more information about– Execution times of tasks

– Predictability of inputs

– Low-level scheduling


Common Task Structures

• Shared memory• Internal dispatching

• Offsets

• Example (Two tasks)– Periods: 4 and 8

– Offsets: 2 and 40 104 86 122

PowerOn 8ms_A 8ms_B 16ms4ms


Case 1: Reading from Sensors

• Internal port p is connected to a sensor

• The variable p is updated at tLs(T)

• δ(T,p): minimun execution time of T

up to accessing p

• Task T can be started at timetr = tLs(T) – δ(T,p)


Case 2: Reading from LET-Based Tasks

• Ports p1 and p2 are updated from tasks T1 and T2, respectively, at the end of their LETs

tr =max{tLe(T1) – δ(T,p1), tLe(T2) – δ(T,p2)}


Modified Operational Requirements

(O1) Update task outputs at LET end

(O2) Update inputs connected to sensors at LET start

(O3) Update input ports connected to LET tasks at the end of the source task‘s LET

(O4) Release task T at time tr tLs(T) such that• No input port connected to a sensor is accessed before

tLs(T)

• Every port that is accessed before tLs(T) has a constant value between the moment of the access and tLs(T)


Computation of Early Release Times

• Formally:

(1)


Main Result


Schedulability

• Assumption:

The system with classical release times is schedulable.

• Question:

Is the system with release margins schedulable?• Answer:

Depends on the underlying scheduling algorithm


Scheduling with Earliest Deadline First


Fixed-Priority Scheduling

• Counter-example (T1 has higher priority):

• Conservative solution: use only the minimum margin

Classical case: No missed deadline

Early release of T1 leads to a missed deadline for T2


Our Approach: Dual-Priority Scheduling

• Assign a dual priority to each LET-based task– All dual priorities are lower than all nominal ones

• A task is scheduled by FPS:– With nominal priority inside its LET

– With dual priority outside its LET

• Effect: a task is executed outside its LET only if the CPU would be otherwise idle!

• DP scheduling is as predictable as FPS


Mixing Events in

• DPS can be used in systems containing also event-triggered tasks

• Event-triggered tasks are always scheduled with nominal priorities

Theorem 3: If event triggered tasks have lower priorities than LET-based tasks, then their response times remain the same or decrease when using release margins with DPS instead of classical release times with FPS.


Application Example

Inverted pendulum:


Evaluation of DPS


Conclusions

• Approach for relaxed scheduling contraints• Usage of execution time information beyond just

WCET• Employ timing predictability offered by LET to

improve scheduling of the application• Static scheduling, fully automatic• Further work: dynamic scheduling, evaluation

Thank you!

Flexible Scheduling of Software with Logical Execution Time Constraints* Stefan Resmerita and Patricia Derler University of Salzburg, Austria *UC Berkeley,

Documents

t ls t t

tasks t

time t r t ls t

p task t

scheduling highlevel

scheduling restrictions

p slide

lets t r