Software Dynamics: A New Method of Evaluating Real-Time Performance of Distributed Systems Janusz Zalewski Computer Science Florida Gulf Coast University.

Software Dynamics:A New Method of Evaluating Real-Time

Performance of Distributed Systems

Janusz Zalewski Computer Science

Florida Gulf Coast University

Ft. Myers, FL 33965-6565

http://www.fgcu.edu/zalewski/

FALSE2002, Nashville, Nov. 14-15, 2002

Talk Outline• RT Software Architecture

• Evaluating S/W Architectures

• Timeliness & S/W Dynamics

• Conclusion


Feedback Control System


Generic Real-Time

Software Architecture


• Sensor/Actuator component

• User Interface component• Communication Link component• Database component• Processing component• Timing component.

Basic Components of Real-TimeBasic Components of Real-TimeSoftware ArchitectureSoftware Architecture


Air-Traffic Control System

Physical Diagram


Air-Traffic Control System

Context Diagram


The idea of grouping I/O information into different categories, which later determine the software architecture follows the fundamental software engineering principle of separation of concerns (Parnas, 1970s).


Model of a Distributed

Embedded Simulation


We are missing good (any) measures to characterize Behavioral Properties of a software module (its dynamics).


Interrupt Latency

The time interval between the occurrence of an external event and start of the first instruction of the interrupt service routine.


• H/W logic processing

• Interrupt disable time

• Handling higher H/W priorities

• Switching to handler code.

Interrupt Latency Involves


Real-Time System

Responsiveness


Dispatch Latency

The time interval between the end of the interrupt handler code and the first instruction of the process activated (made runnable) by this interrupt.


Dispatch Latency Involves• OS decision time to reschedule (non-preemptive kernel state)

• context switch time

• return from OS call.


Real-Time Properties

* Responsiveness

* Timeliness

* Schedulability

* Predictability


How to measure these properties?* Responsiveness - just outlined* Timeliness - proposed below* Schedulability - rate monotonic and deadline monotonic analyses.


Two measures of timeliness:* Overall time deadlines are missed (by a task)* Number of times deadlines are missed by X percent


5-task Benchmark


Overall time the deadlines are missed for 100 experiments.


The number of times the deadlines are missed by 2%.

Overall time the deadlines are missed for 100 experiments (CORBA).


The number of times the deadlines are missed by 2% (CORBA).


ATCS: Software Components

Communicating via CORBA


Overall time (in milliseconds) deadlines are missed for 20 aircraft (in 100 experiments).


Number of times deadlines are missed by more than 20% for 20 aircraft (in 100 experiments).




FSI TLM Dish MicrodyneCombiner

3200

MicrodyneRx'er 1200

Internet Gateway

WorkstationWorkstation

DB Server

MUX/DEMUX

Workstation

SGI Origin 2000

Workstation Workstation

FutureExpansion

IEEE 1394 Splitter

Bit Sync

VME Bus

Analog I/O(AIO) Subsystem

Frontend(FEP) Subsystem

Realtime Processing(RTP) Subsystem

Satellite Ground Control Station


SGCS Implementation


SGCS Physical Architecture



One DB Client - 100 requests

0

100

200

300

400

500

600

700

800

0 20 40 60 80 100

Request Number

Tim

e t

o p

roc

es

s r

eq

ue

st

(ms

ec

)

Single DB Client Request Processing Time.


% Deadlines Missed (one DB Client)

0

20

40

60

80

100

120

0 100 200 300 400 500 600 700 800 900

Deadline set (msec)

% D

ead

lin

es m

isse

d

Percent of deadlines missed for one DB Client.


5 DB Clients - 100 requests

600

800

1000

1200

1400

1600

1800

2000

0 20 40 60 80 100

Request Number

Tim

e to

pro

cess

req

ues

t (m

sec)

Client #1

Client #2

Client #3

Client #4

Client #5

Five DB Clients Request Processing Time.


% Deadlines Missed (5 DB Clients)

0

20

40

60

80

100

120

700 900 1100 1300 1500 1700 1900

Deadline set (msec)

% D

ead

lin

es M

isse

d

Client #1

Client #2

Client #3

Client #4

Client #5

Percent of deadlines missed for five DB Clients.


Sensitivity:a measure of the magnitude of system’s response to changes.


Sensitivity: (y1 – y0)/[(y1 + y0)/2] (x1 – x0)/[(x1 + x0)/2]


Sensitivity (Database)

0

20

40

60

80

100

120

0 100 200 300 400 500 600 700 800 900

Deadline set (msec)

% D

ead

lin

es m

isse

dSensitivity = 1.73


Sensitivity (Telemetry)

0

20

40

60

80

100

120

0 50 100 150 200 250 300 350 400 450

Deadline Set (msec)

% D

ead

lin

es M

isse

dSensitivity = 1.00


Sensitivity (GPS)

0

20

40

60

80

100

120

0 20 40 60 80 100

Deadline Set (msec)

% D

ead

lines

Mis

sed

Sensitivity = 1.64


First Order Dynamics

G(s) = K / (*s + 1)


Time constant - :a measure of the speed of system’s response to changes.


• Settling Time:

time when curve reaches 2% max

• Time Constant = 0.25 * Settling Time


% Deadlines Missed (Database)

0

20

40

60

80

100

120

0 200 400 600 800 1000

Deadline Set (msec)

% D

ead

lin

es M

isse

d

Time Constant

Original

= 165 ms


% Deadlines Missed (TDA)

0

20

40

60

80

100

120

0 100 200 300 400 500

Deadline Set (msec)

% D

ead

lin

es M

isse

d

Time Constant

Original

= 87.5 ms


% Deadlines Missed (GPS)

0

20

40

60

80

100

120

0 20 40 60 80 100

Deadline set (msec)

% D

ea

dlin

es

Mis

se

d

Time Constant

Original

= 15 ms


Applet Interface

IGVxWorks I

ITSVxWorks II

SMCSES/workbench

CGFObjecTime

Distributed Embedded Simulation Architecture


Statistical measures of timeliness: * Round-trip time stability * Service time effect


0100020003000400050006000700080009000

10000

0.01 0.1 1 10

Mean Service Time

Qu

eue

Len

gth

Queue Length for NonPeriodic Traffic with Inter-Arrival Time = 0.1

Queue Length for NonPeriodic Traffic with Inter-Arrival Time = 1.75

Queue Length for PeriodicTraffic Only

Service time effect for a specific architecture


0

0.5

1

1.5

2

2.5

3

0 500 1000 1500 2000 2500 3000

Time Units to run simulation

Tim

e U

nit

s

Sending to SES fromVxWorks 1

Sending to SES fromVxWorks 2

Sending to ObjecTime fromSES

Round-trip message time for 5-task simulation


Conclusion• Behavioral Properties are crucial for successful software development

• Sensitivity is one important property

• Software Dynamics seems to be a measurable property as well


Software Dynamics: A New Method of Evaluating Real-Time Performance of Distributed Systems Janusz Zalewski Computer Science Florida Gulf Coast University.

Documents

overall time deadlines

software dynamics

number of times deadlines

os decision time

milliseconds deadlines

software module

interrupt handler code

measures of timeliness