Aperiodic Task Scheduling

Gerhard Fohler
Mälardalen University, Sweden
[email protected]

Real-Time Systems © Gerhard Fohler 2005 1

 Non Periodic Tasks

So far periodic events and tasks

what about others?

• Sporadic (aperiodic, but minimum interarrival time)

– worst case: all sporadic tasks arrive with highest frequency
(with minimum time between arrivals)
– all other arrival patterns less demanding
– if we can schedule worst case, we can schedule all other
– worst case - minimum interarrival time - like periodic task
assume sporadic tasks as periodic for schedulability test

Real-Time Systems © Gerhard Fohler 2005 2

 • Aperiodic (no limitations on arrival times known)
• soft: without deadline
not much to do from scheduling view
• firm: with deadline
(worst case execution time needs to be known as well)
usually “all or nothing” semantic:
when we start task, we want that it runs until completion; else
we don’t start

Real-Time Systems © Gerhard Fohler 2005 3

 Background Services

Fixed priority scheduling, rate monotonic

What is the minimum we can do for aperiodic tasks in a periodically
scheduled system?
Background service: execute aperiodic tasks when no periodic
ones are executing
– no disturbance of periodic tasks (and their feasibility)
– simple run-time mechanisms
• queue for periodics
• queue for aperiodics - FCFS
– no guarantees

Real-Time Systems © Gerhard Fohler 2005 4

 Polling Server
• Background service lives from “left overs” of periodic tasks,
without guarantees
• If enough idle time, ok
• long response times, although faster service possible
• How can we provide that at least a certain amount of processing
goes to aperiodic tasks?

Server task
periodic task, whose purpose is to service aperiodic requests as
soon as possible
– period Ts
– computation time Cs is called capacity of the server

Real-Time Systems © Gerhard Fohler 2005 5

 Polling server algorithm
– at periods Ts server becomes active and serves aperiodic
requests with its capacity Cs
– no aperiodic activities - not execute, waits for next period,
capacity lost
– based on rate monotonic
Lehozcky, Sha, Strosnider, Sprunt 1987, 1989

Real-Time Systems © Gerhard Fohler 2005 6

 Example Polling Server

• two tasks 1 , 2
Ci Ti
1 1 4
2 2 6

• server
Cs=2
TS=5

Real-Time Systems © Gerhard Fohler 2005 7

 server
Cs=2
TS=5

2
c=2 c=1 c=2 c=1
aper. no no
aper
cap

0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 time
Cs
2
1

Real-Time Systems © Gerhard Fohler 2005 8

 Aperiodic Guarantee

Aperiodic guarantee
hard aperiodic task Ta, Ca, Da
worst case:
– aperiodic request misses the server task
– has to wait until next instance
– if Ca <= Cs, aperiodic request completed within two server
periods (one for waiting, one for executing)
2*Ts <= Da
– arbitrary execution times:
Ts + Ca/Cs Ts <= Da

average response time not very good!

Real-Time Systems © Gerhard Fohler 2005 9
 Further FPS Server Algorithms

• Deferrable Server(Lehozcky, Sha, Strosnider 1987, 1995)

– lower bound for periodic tasks
• Priority exchange
– (Lehozcky, Sha, Strosnider 1987)
• Sporadic server
– Sprunt, Sha, Lehozcky 1989
– replenishes capacity only after aperiodic execution
• optimum algorithm
– does not exist!
– Tia, Liu, Shankar 1995
– proof that with static priority assignment, no algorithm exists
to minimize response time

Real-Time Systems © Gerhard Fohler 2005 10

 Dynamic Priority Servers

• EDF based
• Dynamic priority exchange server
– Spuri, Buttazzo 1994, 1996
– like rate monotonic priority exchange, but for EDF
• Dynamic sporadic server
– Spuri, Buttazzo 1994, 1996
• Earliest deadline late server
– Chetto, Chetto 1989

Real-Time Systems © Gerhard Fohler 2005 11

 Total bandwidth server

• Spuri, Buttazzo 1994, 1996

• response time dependent on server period:
– shorter periods have shorter response times
– but higher overhead
• how else shorter response times?
– change the deadline of the aperiodic to earlier time
(its EDF here, so it will get serviced earlier)
– but make sure that total load of aperiodics does not exceed
maximum value (bandwidth) Us

Real-Time Systems © Gerhard Fohler 2005 12

How can we calculate minimum deadline for US?
assume we have all CPU for us:
dl = C
U
1

C
t t+C
U

1 US
0.5

C
t t+2*C
Real-Time Systems © Gerhard Fohler 2005 13
kth aperiodic request
• arrival time rk
• computation time ck
• deadline dk
• server utilization US

dk = max( rk, dk-1) + Ck/Us

d0 = 0

• uses all bandwidth of server

• very simple run-time mechanism
• no extra server task

schedulability
Up + Us <= 1
Sum of periodic load and bandwidth of server less or equal 1.
Real-Time Systems © Gerhard Fohler 2005 14
 Example Total Bandwidth Server
• periodic tasks 1 (3,6), 2 (2,8)
• TBS Us = 1 - Up = 0.25

2 1 dl1= 3+1/0.25=7 1 dl3= max(14,17)+1/0.25=21

2
aper.

0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 time
dl2= 9+2/0.25=17
Real-Time Systems © Gerhard Fohler 2005 15
 Total Bandwith Server - Comments

• based on
– Us not actual periodic load
– worst case c

Real-Time Systems © Gerhard Fohler 2005 16

 Total Bandwith Server - Comments

• TBS assigns deadlines based on maximum Us (not actual load)

dk = max( rk, dk-1) + Ck/Us, d0 = 0

Us=0.125
1

dl1= 6+1/0.125=14
aper.

0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 time

Real-Time Systems © Gerhard Fohler 2005 17

 TB*

• Buttazzo, Sensini - 1997

• assigns deadlines dk first according to TBS
• then shortens, as much as periodics allow
– new d’k= fk …finishing time according to EDF schedule,
including periodics
– apply recursively
– maintains schedulability, since order maintained
• complexity, many steps

Real-Time Systems © Gerhard Fohler 2005 18

 Constant Bandwidth Server

• Abeni and Buttazzo, 1998

• designed for multimedia applications
– sporadic (hard) tasks
– soft tasks: mean execution, interarrival times, not fixed
– periodic tasks
• assign maximum bandwidth of CPU to each soft task
• handles overload of aperiodics
– limited by assigned bandwidth
– might slow down, but not impair effect other tasks
• EDF based

Real-Time Systems © Gerhard Fohler 2005 19

 CBS Definitions

• task i
– sequence of jobs Ji,j
– ri,j … request, arrival time of the jth job of task I

• hard task
– (Ci,Ti)
• Ci worst case execution time
• Ti minimum interarrival time
• deadline equal to next period: di,j = ri,j + Ti
• soft task
– (Ci,Ti)
• Ci mean execution time
• Ti desired interarrival time
• soft deadline equal to next period: di,j = ri,j + Ti
Real-Time Systems © Gerhard Fohler 2005 20
• cS … budget
• (QS,TS)
– QS … maximum budget
– TS … period of server
• US = QS/TS … server bandwidth
• dS,k … deadline associated to server
– initial dS,0 =0

• job Ji,j comes in, is served, assigned dynamic deadline di,j equal
to current server deadline dS,k
– job executes, server budget cs decreased

Real-Time Systems © Gerhard Fohler 2005 21

• cS=0:
– budget recharged to maximum QS
– new server deadline: ds,k+1=ds,k+TS

• Ji,j arrives, CBS active (jobs pending): put in queue

• Ji,j arrives, CBS idle:
– cS (ds,k - ri,j)*US:
• new deadline ds,k+1 = ri,j + TS
• cS recharged to QS
– else
• job served with last server deadline ds,k
• job finishes: next job in queue
• at any time, job assigned last deadline generated by server

Real-Time Systems © Gerhard Fohler 2005 22

 Example CBS
1 (2,3)
hard

cs (ds,0 - r1)*US:ds,1=r1+TS=9
c1=3 c2=2 cs=1<(23-17)*2/7=12/7 no new deadline
2
server active - queue c3=1
soft

r1 r2 c = 0:d =d +T =16r3
S S,2 s,1 S
cS= 0:dS,3=ds,2+TS=23
CBS
(2,7)

0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 time

Real-Time Systems © Gerhard Fohler 2005 23

 • limits impact “harm” by ill behaved aperiodics, e.g., exec time
overrun
• various improvements
– several servers
– capacity exchange
– feedback control
– …..

Real-Time Systems © Gerhard Fohler 2005 24

 Articles

• TBS:
Spuri, Buttazzo
“Efficient Aperiodic Service under Earliest Deadline Scheduling”
Proceedings of the 15th IEEE Real-Time System Symposium
(RTSS 94), Portorico, pp. 2-21, December 1994
• CBS:
L. Abeni and G. Buttazzo, "Integrating Multimedia Applications in
Hard Real-Time Systems", Proceedings of the IEEE Real-Time
Systems Symposium, Madrid, Spain, pp. 4-13, December 1998.

Real-Time Systems © Gerhard Fohler 2005 25

 Schedulability Analysis

First show that aperiodic load executed not exceeds US of server

Lemma: In each interval of time [t1,t2], if Cape is the total execution

time demanded by aperiodic requests arrived at t1 or later and
served with deadlines less than or equal to t2, then
Cape (t2 - t1) US
Proof: by definition:

Cape Ck
t1 rk , d k t 2

Real-Time Systems © Gerhard Fohler 2005 26

 • TB* uses periodic interference…can now calculate it

• (formulae for completeness only)

If(t, dks) =
n
d ks next _ ri ( t )
max 0, 1 Ci
i 1 Ti

next_ri(t)…time at which next instance of I after t starts

Real-Time Systems © Gerhard Fohler 2005 27

TBS assigns deadlines in increasing order,
therefore there must exist two aperiodic requests with indeces
k1 and k2 such that k 2

Ck Ck
t 1 rk , d k t 2 k k1
k2 k2
Cape Ck [ dk max( r k , d k 1 )] * U S
k k1 k k1

[d k2 max(rk1 , d k1 1 )] *U S

 (t 2 t1 ) *U S

Real-Time Systems © Gerhard Fohler 2005 28

Proof main result:
Theorem: Given a set of n periodic tasks with processor utilization
Up and a TBS with processor utilization of US, the whole set is
schedulable by EDF if and only if
Up + US 1
Proof: If:
• assume Up + US 1 plus overflow at time t
• overflow preceded by continuous utilization
• from a point t’ on (t’< t), only instances of tasks ready at t’ or later
and having deadlines less than or equal to t are run
• C total execution time demanded by these instances
• since there is overflow at t: t - t’ < C

Real-Time Systems © Gerhard Fohler 2005 29

• we also know that

n
t t'
C * Ci Cape
i 1 Ti
n
t t'
* Ci (t t ' ) * US
i 1 Ti
(t t ' ) * (U p U S )

it follows: Up+US > 1 …# contradiction

Real-Time Systems © Gerhard Fohler 2005 30

• only if:
• assume aperiodic request enters periodically with period TS and
execution time CS=TSUS, then server behaves like periodic task
• total utilization of processor is then Up+US
• if task set schedulable: UP+US 1


Real-Time Systems © Gerhard Fohler 2005 31