Linux as a Real-Time OS 2 Tuesday, July 9, 13 2 Linux as a - - PowerPoint PPT Presentation

July 9th, 2013

A Comparison of Scheduling Latency in Linux, PREEMPT_RT, and LITMUSRT

Felipe Cerqueira and Björn Brandenburg

1

1 Tuesday, July 9, 13

2

Linux as a Real-Time OS

2 Tuesday, July 9, 13

Optimizing system responsiveness

3

Linux as a Real-Time OS

3 Tuesday, July 9, 13

4

PREEMPT_RT (Linux)

Linux as a Real-Time OS

Optimizing system responsiveness

4 Tuesday, July 9, 13

Linux as a Real-Time OS

5

PREEMPT_RT (Linux)

Optimizing system responsiveness Algorithmic changes based on real-time systems research

5 Tuesday, July 9, 13

6

Linux as a Real-Time OS

PREEMPT_RT (Linux)

Optimizing system responsiveness Algorithmic changes based on real-time systems research

6 Tuesday, July 9, 13

PREEMPT_RT

✤ Main real-time branch of Linux ✤ Goal: decrease scheduling latency through the use of low-level

hacks

✤ Convert in-kernel spinlocks into (preemptable) mutexes ✤ Limit the extent of non-preemptable sections ✤ Commonly evaluated with cyclictest ✤ Single, easy-to-compare measure of scheduling latency as

• utput

7

7 Tuesday, July 9, 13

✤ Testbed for applied real-time systems research ✤ Goal ✤ Allow implementation and evaluation of novel multiprocessor

schedulers and synchronization protocols

✤ NOT to reduce scheduling latency ✤ Evaluated with Feather-Trace ✤ Flexible, fine-grained measurement of different overheads 8

8 Tuesday, July 9, 13

✤ Testbed for applied real-time systems research ✤ Goal ✤ Allow implementation and evaluation of novel multiprocessor

schedulers and synchronization protocols

✤ NOT to reduce scheduling latency ✤ Evaluated with Feather-Trace ✤ Flexible, fine-grained measurement of different overheads 9

How do LITMUSRT and PREEMPT_RT compare?

9 Tuesday, July 9, 13

✤ Testbed for applied real-time systems research ✤ Goal ✤ Allow implementation and evaluation of novel multiprocessor

schedulers and synchronization protocols

✤ NOT to reduce scheduling latency ✤ Evaluated with Feather-Trace ✤ Flexible, fine-grained measurement of different overheads 10

How do LITMUSRT and PREEMPT_RT compare? It is not straightforward to compare them!

10 Tuesday, July 9, 13

Objective

11

PREEMPT_RT (Linux)

Direct comparison of scheduling latency between LITMUSRT and PREEMPT_RT

vs.

11 Tuesday, July 9, 13

12

Background

12 Tuesday, July 9, 13

How is LITMUSRT evaluated?

✤ Evaluated with ✤ Lightweight tracing framework for measuring fine-grained

• verheads (e.g., IPI latency, context-switching overhead,

etc.)

✤ Extensively used (20+ publications) ✤ Suitable for schedulability analysis ✤ Check if a task is going to miss a deadline 13

13 Tuesday, July 9, 13

How is PREEMPT_RT evaluated?

✤ Evaluated with cyclictest ✤ Standard benchmark for

assessing real-time responsiveness

✤ Creator: Thomas Gleixner

Current maintainer: Clark Williams

✤ Reports scheduling latency as a single measure ✤ Treats hardware and OS as a black-box 14

14 Tuesday, July 9, 13

Scheduling Latency

interrupt!

!

ISR called

brake sensor HP task ECU

15

Time until the highest-priority task is scheduled

15 Tuesday, July 9, 13

T

Scheduling Latency

interrupt!

!

ISR called scheduler invoked

s

T T

w a k e u p t a s k

ZZ ZZ

16

Time until the highest-priority task is scheduled

16 Tuesday, July 9, 13

T

Scheduling Latency

interrupt!

!

ISR called scheduler invoked

s

T T

w a k e u p t a s k

task picked

P

17

Time until the highest-priority task is scheduled

17 Tuesday, July 9, 13

T

Scheduling Latency

interrupt!

!

ISR called scheduler invoked

s

T T

w a k e u p t a s k

task picked

P

Perform context switch C switched

18

Time until the highest-priority task is scheduled

18 Tuesday, July 9, 13

T

Scheduling Latency

interrupt!

!

ISR called scheduler invoked

s

T T

w a k e u p t a s k

task picked

P

Perform context switch

19

scheduling latency Time until the highest-priority task is scheduled C switched

19 Tuesday, July 9, 13

20

Init Exit Measure

How does cyclictest measure Scheduling Latency?

20 Tuesday, July 9, 13

21

Init Exit Measure Measuring thread is granted real-time status.

How does cyclictest measure Scheduling Latency?

POSIX sched_setscheduler( )

21 Tuesday, July 9, 13

22

Init Exit Measure

How does cyclictest measure Scheduling Latency?

22 Tuesday, July 9, 13

How does cyclictest measure Scheduling Latency?

23

Init Exit Measure Periodically setup one-shot timers with nanosleep. Calculate delta between the instant the task starts executing and the instant the timer should have fired.

23 Tuesday, July 9, 13

24

Init Exit Measure

How does cyclictest measure Scheduling Latency?

24 Tuesday, July 9, 13

25

Init Exit Measure Measuring thread returns to best-effort status.

How does cyclictest measure Scheduling Latency?

POSIX sched_setscheduler( )

25 Tuesday, July 9, 13

cyclictest on LITMUSRT

26

Init Exit Measure LITMUSRT does not use POSIX API to setup real-time tasks!

26 Tuesday, July 9, 13

cyclictest on LITMUSRT

27

Init Exit Measure cyclictest works, but does not measure what we expect... LITMUSRT does not use POSIX API to setup real-time tasks!

27 Tuesday, July 9, 13

Porting cyclictest to LITMUSRT

28

Init Exit Measure cyclictest works, but does not measure what we expect... LITMUSRT does not use POSIX API to setup real-time tasks! POSIX sched_setscheduler( ) POSIX sched_setscheduler( )

28 Tuesday, July 9, 13

Porting cyclictest to LITMUSRT

29

Init Exit Measure POSIX sched_setscheduler( ) POSIX sched_setscheduler( ) cyclictest works, but does not measure what we expect... LITMUSRT does not use POSIX API to setup real-time tasks!

29 Tuesday, July 9, 13

30

Measure LITMUSRT API task_mode() LITMUSRT API task_mode() LITMUSRT Init LITMUSRT Exit No changes in the measurement phase, no bias.

Porting cyclictest to LITMUSRT

30 Tuesday, July 9, 13

31

Study

31 Tuesday, July 9, 13

Questions that We Address

32

Linux (core) Scheduler + Dispatcher Userspace Stock Linux

32 Tuesday, July 9, 13

Questions that We Address

32

Linux (core) Scheduler + Dispatcher Userspace Stock Linux

32 Tuesday, July 9, 13

The Cost of LITMUSRT

33

Linux (core) Scheduling Policy Plugins Userspace LITMUSRT Dispatcher Linux (core) Scheduler + Dispatcher Userspace Stock Linux

33 Tuesday, July 9, 13

The Cost of LITMUSRT

34

Linux (core) Userspace LITMUSRT Linux (core) Scheduler + Dispatcher Userspace Stock Linux How much latency does the scheduling policy interface add to the system? Question 1 Scheduling Policy Plugins Dispatcher

34 Tuesday, July 9, 13

LITMUSRT vs. PREEMPT_RT

35

Linux (core) Scheduler + Dispatcher Userspace PREEMPT_RT Linux (core) Scheduler + Dispatcher Userspace Stock Linux Linux (core) Userspace LITMUSRT Scheduling Policy Plugins Dispatcher

35 Tuesday, July 9, 13

LITMUSRT vs. PREEMPT_RT

36

Linux (core) Scheduler + Dispatcher Userspace PREEMPT_RT Linux (core) Scheduler + Dispatcher Userspace Stock Linux Linux (core) Userspace LITMUSRT Scheduling Policy Plugins Dispatcher

36 Tuesday, July 9, 13

LITMUSRT vs. PREEMPT_RT

36

Linux (core) Scheduler + Dispatcher Userspace PREEMPT_RT Linux (core) Scheduler + Dispatcher Userspace Stock Linux Linux (core) Userspace LITMUSRT Scheduling Policy Plugins Dispatcher

36 Tuesday, July 9, 13

LITMUSRT vs. PREEMPT_RT

37

What is the penalty for LITMUSRT not being based on PREEMPT_RT?

Question 2 Linux (core) Userspace LITMUSRT Linux (core) Scheduler + Dispatcher Userspace PREEMPT_RT Scheduling Policy Plugins Dispatcher

37 Tuesday, July 9, 13

Evaluation

38

Linux (core) Scheduler + Dispatcher Userspace

38 Tuesday, July 9, 13

Evaluation

38

Linux (core) Scheduler + Dispatcher

Userspace

38 Tuesday, July 9, 13

Evaluation

39

Linux (core) Scheduler + Dispatcher Userspace cyclictest background workload

+

39 Tuesday, July 9, 13

Background Workloads

CPU-bound background tasks I/O-bound background tasks

40

background tasks

NO

40 Tuesday, July 9, 13

Experimental Setup

Different kernels:

1.LITMUSRT (Linux 3.0)

Partitioned Fixed Priority (P-FP), Partitioned EDF with synchronization support (PSN-EDF), Global EDF with synchronization support (GSN-EDF)

2.PREEMPT_RT (Linux 3.8.13) 3.Unpatched Linux 3.0 and Linux 3.8.13

41

SCHED_FIFO

}

41 Tuesday, July 9, 13

Experimental Setup

✤ 16-core Intel Xeon platform ✤ cyclictest’s standard setup: ✤ one real-time task per processor ✤ periods: {1000, 1500, 2000, ...} μs ✤ Duration: 20 minutes per experiment ✤ Almost 6 million samples for each case ✤ Results shown in microseconds 42

42 Tuesday, July 9, 13

background tasks

NO

First Scenario

43

43 Tuesday, July 9, 13

44

No Background Tasks

Scheduling Latency (μs)

5 10 15 20 P-FP (LITMUS^RT) Linux 3.0 PREEMPT_RT 11.2 13.9 15.3 2.7 2.9 3.5

Average (99% conf.) Maximum

44 Tuesday, July 9, 13

5 10 15 20 P-FP (LITMUS^RT) Linux 3.0 PREEMPT_RT 11.2 13.9 15.3 2.7 2.9 3.5

45

No Background Tasks

Average (99% conf.) Maximum

Scheduling Latency (μs)

Similar max. and

• avg. latency for Linux

3.0 and LITMUSRT.

45 Tuesday, July 9, 13

5 10 15 20 P-FP (LITMUS^RT) Linux 3.0 PREEMPT_RT 11.2 13.9 15.3 2.7 2.9 3.5

46

No Background Tasks

Average (99% conf.) Maximum

Scheduling Latency (μs)

Improved max. latency for PREEMPT_RT Similar max. and

• avg. latency for Linux

3.0 and LITMUSRT.

46 Tuesday, July 9, 13

CPU-bound background tasks

Second Scenario

47

✤ Tasks running an infinite loop

accessing memory (read/write)

✤ Working set larger than L2

cache size

47 Tuesday, July 9, 13

CPU-bound background tasks

Second Scenario

48

✤ Tasks running an infinite loop

accessing memory (read/write)

✤ Working set larger than L2

cache size Generates memory traffic and cache contention!

48 Tuesday, July 9, 13

CPU-bound Background Tasks

49

Scheduling Latency (μs)

20 40 60 80 P-FP (LITMUS^RT) Linux 3.0 PREEMPT_RT 17.4 72.7 47.6 3.4 4.2 5.2

Average (99% conf.) Maximum

49 Tuesday, July 9, 13

20 40 60 80 P-FP (LITMUS^RT) Linux 3.0 PREEMPT_RT 17.4 72.7 47.6 3.4 4.2 5.2

50

CPU-bound Background Tasks

Scheduling Latency (μs)

Average (99% conf.) Maximum

Latency under PREEMPT_RT is significantly lower.

50 Tuesday, July 9, 13

20 40 60 80 P-FP (LITMUS^RT) Linux 3.0 PREEMPT_RT 17.4 72.7 47.6 3.4 4.2 5.2

51

Average (99% conf.) Maximum

Scheduling Latency (μs)

Latency under PREEMPT_RT is significantly lower. LITMUSRT’s latency lower than

• n Linux 3.0?

CPU-bound Background Tasks

51 Tuesday, July 9, 13

1 10 100 1000 10000 100000 1e+06 1e+07 10 20 30 40 50 60 70 80 90 number of samples

• verhead in microseconds (bin size = 1.00us)

Linux 3.0: scheduling latency (CPU-bound bg tasks) min=2.04us max=72.73us avg=4.22us median=3.86us stdev=1.37us samples: total=5854711

LITMUSRT vs. Linux 3.0: CPU-bound Background Tasks

52

Linux 3.0 P-FP(LITMUSRT)

avg=4.22μs max=72.73μs avg=5.17μs max=47.59μs

log scale!

Scheduling Latency (μs) (Bin size = 1μs)

52 Tuesday, July 9, 13

1 10 100 1000 10000 100000 1e+06 1e+07 10 20 30 40 50 60 70 80 90 number of samples

• verhead in microseconds (bin size = 1.00us)

Linux 3.0: scheduling latency (CPU-bound bg tasks) min=2.04us max=72.73us avg=4.22us median=3.86us stdev=1.37us samples: total=5854711

LITMUSRT vs. Linux 3.0: CPU-bound Background Tasks

53

Linux 3.0 P-FP(LITMUSRT)

avg=4.22μs max=72.73μs avg=5.17μs max=47.59μs

Scheduling Latency (μs) (Bin size = 1μs)

3 samples out of ~6 million

53 Tuesday, July 9, 13

1 10 100 1000 10000 100000 1e+06 1e+07 10 20 30 40 50 60 70 80 90 number of samples

• verhead in microseconds (bin size = 1.00us)

Linux 3.0: scheduling latency (CPU-bound bg tasks) min=2.04us max=72.73us avg=4.22us median=3.86us stdev=1.37us samples: total=5854711

LITMUSRT vs. Linux 3.0: CPU-bound Background Tasks

54

Linux 3.0 P-FP(LITMUSRT)

avg=4.22μs max=72.73μs avg=5.17μs max=47.59μs

Scheduling Latency (μs) (Bin size = 1μs)

Slightly worse latencies

• n average

54 Tuesday, July 9, 13

1 10 100 1000 10000 100000 1e+06 1e+07 10 20 30 40 50 60 70 80 90 number of samples

• verhead in microseconds (bin size = 1.00us)

Linux 3.0: scheduling latency (CPU-bound bg tasks) min=2.04us max=72.73us avg=4.22us median=3.86us stdev=1.37us samples: total=5854711

LITMUSRT vs. Linux 3.0: CPU-bound Background Tasks

55

Linux 3.0 P-FP(LITMUSRT)

avg=4.22μs max=72.73μs avg=5.17μs max=47.59μs

Scheduling Latency (μs) (Bin size = 1μs)

extra spinlock lack of low-level

• ptimizations

55 Tuesday, July 9, 13

20 40 60 80 P-FP (LITMUS^RT) Linux 3.0 PREEMPT_RT 17.4 72.7 47.6 3.4 4.2 5.2

56

CPU-bound Background Tasks

Average (99% conf.) Maximum

Scheduling Latency (μs)

LITMUSRT incurs slightly more latency than Linux 3.0 on average. Latency under PREEMPT_RT is significantly lower.

56 Tuesday, July 9, 13

I/O-bound background tasks

Third Scenario

57

✤ hackbench: Linux scheduler stress tool

bonnie++: Disk and file system benchmark wget: Network activity

57 Tuesday, July 9, 13

I/O-bound background tasks

Third Scenario

58

✤ hackbench: Linux scheduler stress tool

bonnie++: Disk and file system benchmark wget: Network activity Causes a lot of system calls and interrupts

58 Tuesday, July 9, 13

I/O-bound Background Tasks

59

Scheduling Latency (μs)

1 10 100 1,000 10,000 P-FP (LITMUS^RT) Linux 3.0 PREEMPT_RT 44.2 4300.4 3956.5 4.1 6.4 6.6

Average (99% conf.) Maximum

log scale!

59 Tuesday, July 9, 13

1 10 100 1,000 10,000 P-FP (LITMUS^RT) Linux 3.0 PREEMPT_RT 44.2 4300.4 3956.5 4.1 6.4 6.6

60

I/O-bound Background Tasks

Huge impact on scheduling latency under standard Linux.

Scheduling Latency (μs)

Average (99% conf.) Maximum

60 Tuesday, July 9, 13

1 10 100 1,000 10,000 P-FP (LITMUS^RT) Linux 3.0 PREEMPT_RT 44.2 4300.4 3956.5 4.1 6.4 6.6

61

I/O-bound Background Tasks

Huge impact on scheduling latency under standard Linux.

Scheduling Latency (μs)

Average (99% conf.) Maximum

PREEMPT_RT is not affected by the interrupt load!

61 Tuesday, July 9, 13

Summary

1.Cost of the scheduling plugin layer 2.LITMUSRT vs. PREEMPT_RT

62

62 Tuesday, July 9, 13

Summary

1.Cost of the scheduling plugin layer 2.LITMUSRT vs. PREEMPT_RT

63

The overhead introduced by LITMUSRT is small

63 Tuesday, July 9, 13

Summary

1.Cost of the scheduling plugin layer 2.LITMUSRT vs. PREEMPT_RT

64

The overhead introduced by LITMUSRT is small PREEMPT_RT significantly decreases scheduling latency.

64 Tuesday, July 9, 13

Importance of Feather-Trace

✤ cyclictest was ported to LITMUSRT. ✤ Should it become the standard tool

for evaluating LITMUSRT?

65

65 Tuesday, July 9, 13

Importance of Feather-Trace

✤ cyclictest was ported to LITMUSRT. ✤ Should it become the standard tool

for evaluating LITMUSRT?

66

NO!

66 Tuesday, July 9, 13

T

Interference?

interrupt!

!

ISR called scheduler invoked

s

T T

w a k e u p t a s k

OOPS... a higher priority task

67

67 Tuesday, July 9, 13

T

Interference?

interrupt!

!

ISR called scheduler invoked

s

T T

w a k e u p t a s k

another task picked

?

68

68 Tuesday, July 9, 13

T

Interference?

interrupt!

!

ISR called scheduler invoked

s

T T

w a k e u p t a s k

another task picked

?

task picked

P

...

69

69 Tuesday, July 9, 13

T

Interference?

interrupt!

!

ISR called scheduler invoked

s

T T

w a k e u p t a s k

another task picked

?

task picked

P

...

70

This length of this interval depends

• n the execution of other tasks...

70 Tuesday, July 9, 13

T

Interference?

interrupt!

!

ISR called scheduler invoked

s

T T

w a k e u p t a s k

another task picked

?

task picked

P

...

71

This length of this interval depends

• n the execution of other tasks...

..., which depends on other kinds of overhead, preemptions, context switches, etc.

71 Tuesday, July 9, 13

T

Interference?

interrupt!

!

ISR called scheduler invoked

s

T T

w a k e u p t a s k

another task picked

?

task picked

P

...

72

This length of this interval depends

• n the execution of other tasks...

..., which depends on other kinds of overhead, preemptions, context switches, etc.

Overhead-aware schedulability analysis is required!

72 Tuesday, July 9, 13

cyclictest or Feather-Trace?

73

BOTH!

cyclictest LITMUSRT/Feather-Trace

Practical, easy-to- understand measure Can easily compare responsiveness between kernels. For tasks other than the highest-priority ones, schedulability analysis is necessary. Only with Feather-Trace we obtain the data required for the analysis.

73 Tuesday, July 9, 13

Conclusion

✤

74

LITMUSRT: small overheads in comparison with stock Linux Scheduling latency should not be used as the sole metric for quantifying real-time guarantees PREEMPT_RT is highly necessary for Linux as a RTOS LITMUSRT will be ported to PREEMPT_RT soon

74 Tuesday, July 9, 13

Thank You!

We also have a patch that implements Feather-Trace on top of standard Linux, enabling fine-grained measurements.

75

75 Tuesday, July 9, 13

Appendix

76

76 Tuesday, July 9, 13

77

Linux 3.0 vs. Linux 3.8.13

77 Tuesday, July 9, 13

78

No Background Tasks

Scheduling Latency (μs)

5 10 15 20 25 Linux 3.0 Linux 3.8.13 19.7 13.9 2.9 2.9

Average Maximum

78 Tuesday, July 9, 13

79

No Background Tasks

Scheduling Latency (μs)

5 10 15 20 25 Linux 3.0 Linux 3.8.13 19.7 13.9 2.9 2.9

Average Maximum

Similar averages

79 Tuesday, July 9, 13

80

No Background Tasks

1 10 100 1000 10000 100000 1e+06 1e+07 10 20 30 40 50 60 70 80 90 number of samples

• verhead in microseconds (bin size = 1.00us)

Linux 3.0: scheduling latency (no bg tasks) min=1.87us max=13.89us avg=2.89us median=2.77us stdev=0.51us samples: total=5854779 1 10 100 1000 10000 100000 1e+06 1e+07 10 20 30 40 50 60 70 80 90 number of samples Linux 3.8.13: scheduling latency (no bg tasks) min=1.52us max=19.73us avg=2.89us median=2.58us stdev=0.69us samples: total=5854801

Linux 3.0

avg=2.89μs max=13.89μs

Linux 3.8.13

avg=2.89μs max=19.73μs

Similar shapes

80 Tuesday, July 9, 13

81

CPU-bound Background Tasks

Scheduling Latency (μs)

20 40 60 80 Linux 3.0 Linux 3.8.13 64.5 72.7 4.0 4.2

Average Maximum

81 Tuesday, July 9, 13

82

CPU-bound Background Tasks

Scheduling Latency (μs)

20 40 60 80 Linux 3.0 Linux 3.8.13 64.5 72.7 4.0 4.2

Average Maximum

Similar averages

82 Tuesday, July 9, 13

83

CPU-bound Background Tasks

1 10 100 1000 10000 100000 1e+06 1e+07 10 20 30 40 50 60 70 80 90 number of samples

• verhead in microseconds (bin size = 1.00us)

Linux 3.0: scheduling latency (CPU-bound bg tasks) min=2.04us max=72.73us avg=4.22us median=3.86us stdev=1.37us samples: total=5854711 1 10 100 1000 10000 100000 1e+06 1e+07 10 20 30 40 50 60 70 80 90 number of samples Linux 3.8.13: scheduling latency (CPU-bound bg tasks) min=2.14us max=64.47us avg=4.02us median=3.67us stdev=1.20us samples: total=5854707

Linux 3.0

avg=4.22μs max=72.73μs

Linux 3.8.13

avg=4.02μs max=64.47μs

Similar shapes

83 Tuesday, July 9, 13

I/O-bound Background Tasks

84

Scheduling Latency (μs)

1 10 100 1,000 10,000 Linux 3.0 Linux 3.8.13 5464.1 4300.4 6.2 6.4

Average Maximum

log scale!

84 Tuesday, July 9, 13

I/O-bound Background Tasks

85

Scheduling Latency (μs)

1 10 100 1,000 10,000 Linux 3.0 Linux 3.8.13 5464.1 4300.4 6.2 6.4

Average Maximum

Similar averages

85 Tuesday, July 9, 13

86

I/O-bound Background Tasks

1 10 100 1000 10000 100000 1e+06 1e+07 150 300 450 600 750 900 1050 1200 1350 number of samples

• verhead in microseconds (bin size = 1.00us)

Linux 3.0: scheduling latency (IO-bound bg tasks) min=1.85us max=4300.43us avg=6.39us median=4.98us stdev=13.25us samples: total=5854674 1 10 100 1000 10000 100000 1e+06 1e+07 150 300 450 600 750 900 1050 1200 1350 number of samples Linux 3.8.13: scheduling latency (IO-bound bg tasks) min=1.85us max=5464.07us avg=6.23us median=4.60us stdev=15.91us samples: total=5854773

Linux 3.0

avg=6.39μs max=4300.43μs

Linux 3.8.13

avg=6.23μs max=5464.07μs

Similar shapes

86 Tuesday, July 9, 13

87

LITMUSRT’s plugins

87 Tuesday, July 9, 13

88

No Background Tasks

Scheduling Latency (μs)

7.5 15 22.5 30 P-FP PSN-EDF GSN-EDF 14.3 26.2 15.1 3.1 3.5 3.5

Average Maximum

88 Tuesday, July 9, 13

No Background Tasks

89

1 10 100 1000 10000 100000 1e+06 1e+07 10 20 30 40 50 60 70 80 90 number of samples

• verhead in microseconds (bin size = 1.00us)

LITMUS^RT P-EDF: scheduling latency (no bg tasks) min=1.76us max=26.17us avg=3.45us median=2.87us stdev=1.24us samples: total=5854783 1 10 100 1000 10000 100000 1e+06 1e+07 10 20 30 40 50 60 70 80 90

• verhead in microseconds (bin size = 1.00us)

LITMUS^RT G-EDF: scheduling latency (no bg tasks) min=1.59us max=14.34us avg=3.06us median=2.56us stdev=1.18us samples: total=5854797 1 10 100 1000 10000 100000 1e+06 1e+07 10 20 30 40 50 60 70 80 90 number of samples

• verhead in microseconds (bin size = 1.00us)

LITMUS^RT P-FP: scheduling latency (no bg tasks) min=1.96us max=15.13us avg=3.45us median=3.10us stdev=1.03us samples: total=5854818

Similar shapes

PSN-EDF

avg=3.45μs max=26.17μs

GSN-EDF

avg=3.06μs max=14.34μs

P-FP

avg=3.45μs max=15.13μs

89 Tuesday, July 9, 13

No Background Tasks

90

1 10 100 1000 10000 100000 1e+06 1e+07 10 20 30 40 50 60 70 80 90 number of samples

• verhead in microseconds (bin size = 1.00us)

LITMUS^RT P-EDF: scheduling latency (no bg tasks) min=1.76us max=26.17us avg=3.45us median=2.87us stdev=1.24us samples: total=5854783 1 10 100 1000 10000 100000 1e+06 1e+07 10 20 30 40 50 60 70 80 90

• verhead in microseconds (bin size = 1.00us)

LITMUS^RT G-EDF: scheduling latency (no bg tasks) min=1.59us max=14.34us avg=3.06us median=2.56us stdev=1.18us samples: total=5854797 1 10 100 1000 10000 100000 1e+06 1e+07 10 20 30 40 50 60 70 80 90 number of samples

• verhead in microseconds (bin size = 1.00us)

LITMUS^RT P-FP: scheduling latency (no bg tasks) min=1.96us max=15.13us avg=3.45us median=3.10us stdev=1.03us samples: total=5854818

Similar shapes

PSN-EDF

avg=3.45μs max=26.17μs

GSN-EDF

avg=3.06μs max=14.34μs

P-FP

avg=3.45μs max=15.13μs

2 samples

90 Tuesday, July 9, 13

91

CPU-bound Background Tasks

Scheduling Latency (μs)

20 40 60 80 P-FP PSN-EDF GSN-EDF 60.2 73.3 47.6 5.8 5.1 5.2

Average Maximum

91 Tuesday, July 9, 13

1 10 100 1000 10000 100000 1e+06 1e+07 10 20 30 40 50 60 70 80 90 number of samples

• verhead in microseconds (bin size = 1.00us)

LITMUS^RT P-EDF: scheduling latency (CPU-bound bg tasks) min=2.40us max=73.27us avg=5.14us median=4.21us stdev=2.95us samples: total=5854739 1 10 100 1000 10000 100000 1e+06 1e+07 10 20 30 40 50 60 70 80 90

• verhead in microseconds (bin size = 1.00us)

LITMUS^RT G-EDF: scheduling latency (CPU-bound bg tasks) min=1.91us max=60.20us avg=5.81us median=5.39us stdev=2.51us samples: total=5854728 1 10 100 1000 10000 100000 1e+06 1e+07 10 20 30 40 50 60 70 80 90 number of samples

• verhead in microseconds (bin size = 1.00us)

LITMUS^RT P-FP: scheduling latency (CPU-bound bg tasks) min=2.10us max=47.59us avg=5.17us median=4.37us stdev=2.75us samples: total=5854719

CPU-bound Background Tasks

92

PSN-EDF

avg=5.14μs max=73.27μs

GSN-EDF

avg=5.81μs max=60.20μs

P-FP

avg=5.17μs max=47.59μs

Similar shapes

92 Tuesday, July 9, 13

1 10 100 1000 10000 100000 1e+06 1e+07 10 20 30 40 50 60 70 80 90 number of samples

• verhead in microseconds (bin size = 1.00us)

LITMUS^RT P-EDF: scheduling latency (CPU-bound bg tasks) min=2.40us max=73.27us avg=5.14us median=4.21us stdev=2.95us samples: total=5854739 1 10 100 1000 10000 100000 1e+06 1e+07 10 20 30 40 50 60 70 80 90

• verhead in microseconds (bin size = 1.00us)

LITMUS^RT G-EDF: scheduling latency (CPU-bound bg tasks) min=1.91us max=60.20us avg=5.81us median=5.39us stdev=2.51us samples: total=5854728 1 10 100 1000 10000 100000 1e+06 1e+07 10 20 30 40 50 60 70 80 90 number of samples

• verhead in microseconds (bin size = 1.00us)

LITMUS^RT P-FP: scheduling latency (CPU-bound bg tasks) min=2.10us max=47.59us avg=5.17us median=4.37us stdev=2.75us samples: total=5854719

CPU-bound Background Tasks

93

PSN-EDF

avg=5.14μs max=73.27μs

GSN-EDF

avg=5.81μs max=60.20μs

P-FP

avg=5.17μs max=47.59μs

5 samples

Similar shapes

93 Tuesday, July 9, 13

I/O-bound Background Tasks

94

Scheduling Latency (μs)

1 10 100 1,000 10,000 P-FP PSN-EDF GSN-EDF 3905.8 3875.0 3956.5 11.0 6.6 6.6

Average Maximum

log scale!

94 Tuesday, July 9, 13

I/O-bound Background Tasks

95

Scheduling Latency (μs)

1 10 100 1,000 10,000 P-FP PSN-EDF GSN-EDF 3905.8 3875.0 3956.5 11.0 6.6 6.6

Average Maximum

Similar results for P-FP and PSN-EDF

95 Tuesday, July 9, 13

I/O-bound Background Tasks

96

Scheduling Latency (μs)

1 10 100 1,000 10,000 P-FP PSN-EDF GSN-EDF 3905.8 3875.0 3956.5 11.0 6.6 6.6

Average Maximum

Similar results for P-FP and PSN-EDF Higher average for GSN-EDF

96 Tuesday, July 9, 13

1 10 100 1000 10000 100000 1e+06 1e+07 150 300 450 600 750 900 1050 1200 1350 number of samples

• verhead in microseconds (bin size = 1.00us)

LITMUS^RT P-EDF: scheduling latency (IO-bound bg tasks) min=1.98us max=3874.99us avg=6.56us median=5.11us stdev=12.66us samples: total=5854606 1 10 100 1000 10000 100000 1e+06 1e+07 150 300 450 600 750 900 1050 1200 1350

• verhead in microseconds (bin size = 1.00us)

LITMUS^RT G-EDF: scheduling latency (IO-bound bg tasks) min=2.26us max=3905.79us avg=10.95us median=7.38us stdev=14.11us samples: total=5854793 1 10 100 1000 10000 100000 1e+06 1e+07 150 300 450 600 750 900 1050 1200 1350 number of samples

• verhead in microseconds (bin size = 1.00us)

LITMUS^RT P-FP: scheduling latency (IO-bound bg tasks) min=1.89us max=3956.48us avg=6.60us median=5.17us stdev=12.76us samples: total=5854660

I/O-bound Background Tasks

97

Similar shapes

PSN-EDF

avg=6.56μs max=3874.99μs

GSN-EDF

avg=10.95μs max=3905.79μs

P-FP

avg=6.60μs max=3956.48μs

97 Tuesday, July 9, 13

1 10 100 1000 10000 100000 1e+06 1e+07 150 300 450 600 750 900 1050 1200 1350 number of samples

• verhead in microseconds (bin size = 1.00us)

LITMUS^RT P-EDF: scheduling latency (IO-bound bg tasks) min=1.98us max=3874.99us avg=6.56us median=5.11us stdev=12.66us samples: total=5854606 1 10 100 1000 10000 100000 1e+06 1e+07 150 300 450 600 750 900 1050 1200 1350

• verhead in microseconds (bin size = 1.00us)

LITMUS^RT G-EDF: scheduling latency (IO-bound bg tasks) min=2.26us max=3905.79us avg=10.95us median=7.38us stdev=14.11us samples: total=5854793 1 10 100 1000 10000 100000 1e+06 1e+07 150 300 450 600 750 900 1050 1200 1350 number of samples

• verhead in microseconds (bin size = 1.00us)

LITMUS^RT P-FP: scheduling latency (IO-bound bg tasks) min=1.89us max=3956.48us avg=6.60us median=5.17us stdev=12.76us samples: total=5854660

I/O-bound Background Tasks

98

Similar shapes

PSN-EDF

avg=6.56μs max=3874.99μs

GSN-EDF

avg=10.95μs max=3905.79μs

P-FP

avg=6.60μs max=3956.48μs

Higher average

98 Tuesday, July 9, 13

99

threadirqs in Linux 3.8.13

99 Tuesday, July 9, 13

1 10 100 1000 10000 100000 1e+06 1e+07 10 20 30 40 50 60 70 80 90 number of samples

• verhead in microseconds (bin size = 1.00us)

Linux 3.8.13: scheduling latency (no bg tasks) min=1.60us max=25.15us avg=2.82us median=2.65us stdev=0.47us samples: total=5854778

100

No Background Tasks

1 10 100 1000 10000 100000 1e+06 1e+07 10 20 30 40 50 60 70 80 90 number of samples Linux 3.8.13: scheduling latency (no bg tasks) min=1.52us max=19.73us avg=2.89us median=2.58us stdev=0.69us samples: total=5854801

avg=2.82μs max=25.15μs

Linux 3.8.13

avg=2.89μs max=19.73μs

Linux 3.8.13 threadirqs

100 Tuesday, July 9, 13

1 10 100 1000 10000 100000 1e+06 1e+07 10 20 30 40 50 60 70 80 90 number of samples Linux 3.8.13: scheduling latency (CPU-bound bg tasks) min=2.14us max=64.47us avg=4.02us median=3.67us stdev=1.20us samples: total=5854707 1 10 100 1000 10000 100000 1e+06 1e+07 10 20 30 40 50 60 70 80 90 number of samples

• verhead in microseconds (bin size = 1.00us)

Linux 3.8.13: scheduling latency (CPU-bound bg tasks) min=1.95us max=40.90us avg=3.67us median=3.29us stdev=1.19us samples: total=5854684

101

CPU-bound Background Tasks

avg=3.67μs max=40.90μs

Linux 3.8.13 threadirqs Linux 3.8.13

avg=4.02μs max=64.47μs

101 Tuesday, July 9, 13

1 10 100 1000 10000 100000 1e+06 1e+07 150 300 450 600 750 900 1050 1200 1350 number of samples

• verhead in microseconds (bin size = 1.00us)

Linux 3.8.13: scheduling latency (IO-bound bg tasks) min=1.51us max=5203.38us avg=5.89us median=4.67us stdev=14.33us samples: total=5854724 1 10 100 1000 10000 100000 1e+06 1e+07 150 300 450 600 750 900 1050 1200 1350 number of samples Linux 3.8.13: scheduling latency (IO-bound bg tasks) min=1.85us max=5464.07us avg=6.23us median=4.60us stdev=15.91us samples: total=5854773

102

I/O-bound Background Tasks

avg=5.89μs max=5203.38μs

Linux 3.8.13 threadirqs Linux 3.8.13

avg=6.23μs max=5464.07μs

102 Tuesday, July 9, 13