Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo...

26
Paper Presentation Amo Guangmo Tong University of Taxes at Dallas [email protected] February 11, 2014 Amo Guangmo Tong (UTD) February 11, 2014 1 / 26

Transcript of Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo...

Page 1: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

Paper Presentation

Amo Guangmo Tong

University of Taxes at Dallas

[email protected]

February 11, 2014

Amo Guangmo Tong (UTD) February 11, 2014 1 / 26

Page 2: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

Overview

1 Techniques for Multiprocessor Global Schedulability Analysis

2 New Response Time Bounds for Fixed Priority MultiprocessorScheduling

Amo Guangmo Tong (UTD) February 11, 2014 2 / 26

Page 3: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

Techniques for Multiprocessor Global SchedulabilityAnalysis

Techniques for Multiprocessor Global Schedulability Analysis

Sanjoy Baruah

Amo Guangmo Tong (UTD) February 11, 2014 3 / 26

Page 4: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

System Model

Hard real-time system T

n sporadic tasks Tn = (en,Dn, pn)

m identical processors ,m > 1

minimum inter-arrival time or period, pi > 0

execution time ei < pi

relative deadline Di ≤ pi (constrained)

utilization ui = ei/pi

Amo Guangmo Tong (UTD) February 11, 2014 4 / 26

Page 5: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

Schedule Algorithms

Partitioned FJP(fixed job priority) algorithmsAssign tasks to processors. Schedule tasks on each uniprocessor.

Example:

Global FJP algorithmsIn this schedule algorithm, tasks can execute and resume on anyprocessors.

Schedulable and Schedulability testAmo Guangmo Tong (UTD) February 11, 2014 5 / 26

Page 6: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

Contributions

Global FJP and partitioned FJP are incomparable.

Provide a schedulability test for global EDF.

Amo Guangmo Tong (UTD) February 11, 2014 6 / 26

Page 7: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

Global and partitioned FJP scheduling are incomparable.

Step 1

Sometimes global FJP scheduling is better than partitioned FJPscheduling.

Example:

τ1 = (2, 3), τ2 = (3, 4), τ3 = (5, 12); u1 = 2/3, u2 = 3/4, u3 = 5/12.Because u1 + u2 > 1,u2 + u3 > 1,u1 + u3 > 1, partitioned FJP dosenot work.

The following figure shows global FJP works.

1 2 3 4 5 6 7 8 9 10 11 12

Amo Guangmo Tong (UTD) February 11, 2014 7 / 26

Page 8: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

Global and partitioned FJP scheduling are incomparable.

Step 2

Sometimes partitioned FJP scheduling is better than global FJPscheduling .Example:

τ1 = (2, 3), τ2 = (3, 4), τ3 = (3, 12), τ4 = (4, 12); u1 = 2/3, u2 =3/4, u3 = 3/12, u4 = 4/12. Because u1 + u4 = 1,u2 + u3 = 1,partitioned FJP works.However no global FJP works. (If we assume τ3,1 has higher prioritythan τ4,1, then τ4,1 misses its deadline.)

1 2 3 4 5 6 7 8 9 10 11 12

Miss deadline

Amo Guangmo Tong (UTD) February 11, 2014 8 / 26

Page 9: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

A schedulability test for global EDF

Let focus on τi ,j .

ri ,j = ta, di ,j = td , Di = td − ta, to is the latest time instant before taat which at least one processor is idle.

L: the union of intervals in [ta, td) during which all m processors areexecuting jobs other than τi ,j .

Some proc. Is idle

Amo Guangmo Tong (UTD) February 11, 2014 9 / 26

Page 10: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

A schedulability test for global EDF

If τi ,j misses its deadline, then |L| > Di − ei . So there exists anL∗ ⊆ L where |L∗| = Di − ei . And the total execution of all tasksdone in [to , ta)

⋃L∗ is m(Ai + Di − ei ).

Some proc. Is idle

Amo Guangmo Tong (UTD) February 11, 2014 10 / 26

Page 11: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

A schedulability test for global EDF

Let I (τk , τi ,Ai ) denote the upper bound of the workload of τk thatcould be done in [to , ta)

⋃L′

where L′

is any subset of [ta, td) withlength Di − ei .I (τk , τi ,Ai ) = I (ek ,Dk ,Ai , ei ,Di )

If, for any value of Ai ,∑n

k=1 I (τk , τi ,Ai ) < m(Ai + Di − ei ), then anyjob τi ,j of τi cannot miss its deadline.

Some proc. Is idle

Amo Guangmo Tong (UTD) February 11, 2014 11 / 26

Page 12: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

Test

If for any value A,∑nk=1 I (τk , τ1,A) < m(A + D1 − e1), jobs in τ1 will meet deadline.∑nk=1 I (τk , τ2,A) < m(A + D2 − e2), jobs in τ2 will meet deadline.

.

.

.∑nk=1 I (τk , τn,A) < m(A + Dn − en), jobs in τn will meet deadline.

Thus, if for all i from 1 to n, and any value Ai ,∑nk=1 I (τk , τi ,Ai ) < m(Ai + Di − ei ), then no job in τ will miss deadline.

Amo Guangmo Tong (UTD) February 11, 2014 12 / 26

Page 13: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

I (τk , τi)

I (τk , τi ,Ai ) denote the upper bound of the workload of τk that could bedone in [to , ta)

⋃L∗ where L∗ is any subset of [ta, td) with length Di − ei .

I (τk , τi ,Ai ) = I (ek ,Dk ,Ai , ei ,Di )

Some proc. Is idle

Carry in job

No carry in job

Amo Guangmo Tong (UTD) February 11, 2014 13 / 26

Page 14: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

Pseudo-polynomial

Previous : O(P(n))

Now:O(P(m, ei , di ))

Other weakness of improvement.

Amo Guangmo Tong (UTD) February 11, 2014 14 / 26

Page 15: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

New Response Time Bounds for Fixed PriorityMultiprocessor Scheduling

New Response Time Bounds for Fixed Priority MultiprocessorScheduling

Nan Guan, Martin Stigge, Wang Yi and Ge Yu

Amo Guangmo Tong (UTD) February 11, 2014 15 / 26

Page 16: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

System Model

sporadic real-time system T

n sporadic tasks Tn = (en, dn, pn)

m identical processors ,m > 1

minimum inter-arrival time or period, pi > 0

execution time ei < pi

relative deadline Di

Di ≤ pi (constrained-deadline)Di and Pi have no relationship(arbitrary-deadline)

utilization ui = ei/pi

fi ,j denotes the finish time of τi ,j , define response time of τi ,j asfi ,j − ri ,j

Amo Guangmo Tong (UTD) February 11, 2014 16 / 26

Page 17: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

Content

Response time bound for constrained-deadline system

Response time bound for arbitrary-deadline system

Amo Guangmo Tong (UTD) February 11, 2014 17 / 26

Page 18: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

Constrained-deadline system

To analyze deadline miss

Some proc. Is idle

To analyze response time bound

Some proc. Is idle

Amo Guangmo Tong (UTD) February 11, 2014 18 / 26

Page 19: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

Constrained-deadline system

When we analyze τi ,j :Let Ωi (x) be the maximum value of the sum of all higher-priority tasks’interference among all possible cases.

Carry-in job

x

Amo Guangmo Tong (UTD) February 11, 2014 19 / 26

Page 20: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

Constrained-deadline system

Then we solve the equation:

x = bΩi (x)m c+ ei

x

Remember: Ωi(x) be the maximum value of the sum ofall higher-priority tasks’ interference against τi ,j .

Amo Guangmo Tong (UTD) February 11, 2014 20 / 26

Page 21: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

Constrained-deadline system

Then x is an upper bound of response time of τi ,j

x

Response time

Amo Guangmo Tong (UTD) February 11, 2014 21 / 26

Page 22: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

Arbitrary-deadline system

For constrained-deadline system, we analyze :

Some proc. Is idle

For arbitrary-deadline system, we analyze :

x

Amo Guangmo Tong (UTD) February 11, 2014 22 / 26

Page 23: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

Arbitrary-deadline system

Similarly, we solve the equation:

x = bΩi (x ,h)m c+ h · ei

x

Remember : Ωi(x , h) be the maximum value of the sumof all higher-priority tasks’ interference against τi ,j .

Amo Guangmo Tong (UTD) February 11, 2014 23 / 26

Page 24: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

Constrained-deadline system

Then an upper bound of response time of τi ,j is

x − (h − 1) · pi

x

Response time of h jobs

Amo Guangmo Tong (UTD) February 11, 2014 24 / 26

Page 25: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

Weakness or improvement

Amo Guangmo Tong (UTD) February 11, 2014 25 / 26

Page 26: Amo Guangmo Tong - University of Texas at Dallascxl137330/courses/spring14/AdvRTS/... · Amo Guangmo Tong University of Taxes at Dallas gxt140030@utdallas.edu February 11, ... tasks

The End

Amo Guangmo Tong (UTD) February 11, 2014 26 / 26