on the ineffectiveness of 1/m-based interference bounds in...

11
On the Ineffectiveness of 1/m-based Interference Bounds in the Analysis of Global EDF and FIFO Scheduling Alessandro Biondi and Youcheng Sun Scuola Superiore Sant’Anna, Pisa, Italy University of Oxford, Oxford, UK + + * * Journal paper published in Real-Time Systems, July 2018, Volume 54, Issue 3, pp 515–536

Upload: others

Post on 02-Jun-2020

2 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: On the Ineffectiveness of 1/m-based Interference Bounds in ...2018.rtss.org/wp-content/uploads/2018/12/bp_01.pdf · G-EDF P-EDF, first-fit, decreasing utilizations utilization y o

On the Ineffectiveness of 1/m-based Interference Bounds in the

Analysis of Global EDF and FIFO Scheduling

Alessandro Biondi and Youcheng Sun

Scuola Superiore Sant’Anna, Pisa, Italy

University of Oxford, Oxford, UK

+

+

*

*

Journal paper published in Real-Time Systems, July 2018, Volume 54, Issue 3, pp 515–536

Page 2: On the Ineffectiveness of 1/m-based Interference Bounds in ...2018.rtss.org/wp-content/uploads/2018/12/bp_01.pdf · G-EDF P-EDF, first-fit, decreasing utilizations utilization y o

2A. Biondi

Response-time analysis for sporadic tasks under global scheduling

• Most popular proposal due to Bertogna and Cirenei (RTSS’07)

• Extended to refine carry-in/carry-out interference

• The same analysis principle has been widely adopted as a building

block to analyze

RTA for Global Scheduling

locking protocols hierarchical scheduling

Vestal’s mixed-criticality schedulers

processor affinities

parallel tasks cache-related delays

priority assignments heterogeneous multiprocessors etc…

Page 3: On the Ineffectiveness of 1/m-based Interference Bounds in ...2018.rtss.org/wp-content/uploads/2018/12/bp_01.pdf · G-EDF P-EDF, first-fit, decreasing utilizations utilization y o

3A. Biondi

Response-time analysis for sporadic tasks under global scheduling

• Most popular proposal due to Bertogna and Cirenei (RTSS’07)

• Extended to refine carry-in/carry-out interference

• The same analysis principle has been widely adopted as a building

block to analyze

RTA for Global Scheduling

𝑅𝑖UB = 𝐶𝑖 +

𝟏

𝒎

othertasks 𝜏𝑗

<bound on interfering workload of 𝜏𝑗>

number of processors

Working principle: exploit work-conserving property to bound interference

locking protocols hierarchical scheduling

Vestal’s mixed-criticality schedulers

processor affinities

parallel tasks cache-related delays

priority assignments heterogeneous multiprocessors etc…

Page 4: On the Ineffectiveness of 1/m-based Interference Bounds in ...2018.rtss.org/wp-content/uploads/2018/12/bp_01.pdf · G-EDF P-EDF, first-fit, decreasing utilizations utilization y o

4A. Biondi

…it’s enough to sort the tasks by decreasing relative deadlines, and thenassign them to processors with any heuristic (e.g., first-fit)

Main results

If a set of sporadic real-time tasks is schedulable under

global EDF according to the Bertogna and Cirenei’s

analysis, then it is also schedulable under partitioned EDF

1

Page 5: On the Ineffectiveness of 1/m-based Interference Bounds in ...2018.rtss.org/wp-content/uploads/2018/12/bp_01.pdf · G-EDF P-EDF, first-fit, decreasing utilizations utilization y o

5A. Biondi

…it’s enough to sort the tasks by decreasing relative deadlines, and thenassign them to processors with any heuristic (e.g., first-fit)

Main results

If a set of sporadic real-time tasks is schedulable under

global EDF according to the Bertogna and Cirenei’s

analysis, then it is also schedulable under partitioned EDF

The same holds for global FIFO scheduling when analyzed

with a 1/m-based interference bound

1

2

E.g., relevant for multi-unit FIFO-based locking protocols

Page 6: On the Ineffectiveness of 1/m-based Interference Bounds in ...2018.rtss.org/wp-content/uploads/2018/12/bp_01.pdf · G-EDF P-EDF, first-fit, decreasing utilizations utilization y o

6A. Biondi

…it’s enough to sort the tasks by decreasing relative deadlines, and thenassign them to processors with any heuristic (e.g., first-fit)

Main results

If a set of sporadic real-time tasks is schedulable under

global EDF according to the Bertogna and Cirenei’s

analysis, then it is also schedulable under partitioned EDF

The same holds for global FIFO scheduling when analyzed

with a 1/m-based interference bound

Y. Sun and M. Di Natale, “Assessing the pessimism of current multicore global

fixed-priority schedulability analysis”, In Proceedings of the 33rd Annual ACM

Symposium on Applied Computing (SAC), April 9-13, 2018, Pau, France

…and also for global fixed-priority scheduling

1

2

E.g., relevant for multi-unit FIFO-based locking protocols

Page 7: On the Ineffectiveness of 1/m-based Interference Bounds in ...2018.rtss.org/wp-content/uploads/2018/12/bp_01.pdf · G-EDF P-EDF, first-fit, decreasing utilizations utilization y o

7A. Biondi

Take-away messages

G-EDF

P-EDF, first-fit, decreasing utilizations

utilization

sch

edu

lab

ility

rati

o• Besides the analytical

dominance, experimental

results showed that P-EDF

outperforms G-EDF withtoday’s analysis (30 heuristics

have been tested)

• The same emerged for FIFO

scheduling

(latest RTA RTSS’15)

Page 8: On the Ineffectiveness of 1/m-based Interference Bounds in ...2018.rtss.org/wp-content/uploads/2018/12/bp_01.pdf · G-EDF P-EDF, first-fit, decreasing utilizations utilization y o

8A. Biondi

Take-away messages

G-EDF

P-EDF, first-fit, decreasing utilizations

utilization

sch

edu

lab

ility

rati

o• Besides the analytical

dominance, experimental

results showed that P-EDF

outperforms G-EDF withtoday’s analysis (30 heuristics

have been tested)

• The same emerged for FIFO

scheduling

1/m-based analysis techniques for global scheduling are ineffective

with respect to much simpler and predictable partitioned schedulers

• Main source of pessimism is not related to carry-in and carry-out!

• What about other task models? What’s the impact on previous work?

(latest RTA RTSS’15)

Page 9: On the Ineffectiveness of 1/m-based Interference Bounds in ...2018.rtss.org/wp-content/uploads/2018/12/bp_01.pdf · G-EDF P-EDF, first-fit, decreasing utilizations utilization y o

9A. Biondi

Take-away messages

G-EDF

P-EDF, first-fit, decreasing utilizations

utilization

sch

edu

lab

ility

rati

o• Besides the analytical

dominance, experimental

results showed that P-EDF

outperforms G-EDF withtoday’s analysis (30 heuristics

have been tested)

• The same emerged for FIFO

scheduling

1/m-based analysis techniques for global scheduling are ineffective

with respect to much simpler and predictable partitioned schedulers

• Main source of pessimism is not related to carry-in and carry-out!

• What about other task models? What’s the impact on previous work?

Future analyses for global scheduling should always compare against

partitioned approaches

(latest RTA RTSS’15)

Page 10: On the Ineffectiveness of 1/m-based Interference Bounds in ...2018.rtss.org/wp-content/uploads/2018/12/bp_01.pdf · G-EDF P-EDF, first-fit, decreasing utilizations utilization y o

10A. Biondi

Take-away messages

G-EDF

P-EDF, first-fit, decreasing utilizations

utilization

sch

edu

lab

ility

rati

o• Besides the analytical

dominance, experimental

results showed that P-EDF

outperforms G-EDF withtoday’s analysis (30 heuristics

have been tested)

• The same emerged for FIFO

scheduling

1/m-based analysis techniques for global scheduling are ineffective

with respect to much simpler and predictable partitioned schedulers

• Main source of pessimism is not related to carry-in and carry-out!

• What about other task models? What’s the impact on previous work?

Future analyses for global scheduling should always compare against

partitioned approaches

Is global scheduling really the way to go? Anyway, we need new

analysis techniques…

(latest RTA RTSS’15)

Page 11: On the Ineffectiveness of 1/m-based Interference Bounds in ...2018.rtss.org/wp-content/uploads/2018/12/bp_01.pdf · G-EDF P-EDF, first-fit, decreasing utilizations utilization y o

Thank you!Alessandro Biondi

[email protected]

𝟏

𝒎

Do you want to know more about this work?

Come visiting me during the poster session!