Upload
others
View
6
Download
0
Embed Size (px)
Citation preview
Dynamic Resource Allocation for Priority Processing– Master Project –
Martijn van den [email protected]
Systems Architecture and Networking (SAN)Department of Mathematics and Computer Science
Eindhoven University of Technologythe Netherlands
23 July 2009
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 1 / 34
Outline
1 IntroductionScalable Video AlgorithmsPlatformReal-time SystemsPriority Processing
2 Resource Management MechanismsPreliminary TerminationProcessor Allocation
3 Simulation Environment
4 Simulation ResultsPreliminary TerminationProcessor Allocation
5 Conclusions
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 2 / 34
Scalable Video Algorithms Motivation
Trade-off: Quality versus Resources
• Reuse softwaremodules
• Cost-effective
• Time-to-market(Porting to newplatforms)
By courtesy of Hentschel et. al.
Demo ...
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 3 / 34
Scalable Video Algorithms Motivation
Trade-off: Quality versus Resources
• Reuse softwaremodules
• Cost-effective
• Time-to-market(Porting to newplatforms)
By courtesy of Hentschel et. al.
Demo ...Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 3 / 34
Platform Definition
Definition
A platform, which is capable to run software components, is defined by:
1 hardware (resources)
2 programming language
3 operating system (OS)
4 runtime-libraries (provided by - or built on top of - OS)
Note: A resource can also be a software entity!(We especially consider the processor)
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 4 / 34
Platform Definition
Definition
A platform, which is capable to run software components, is defined by:
1 hardware (resources)
2 programming language
3 operating system (OS)
4 runtime-libraries (provided by - or built on top of - OS)
Note: A resource can also be a software entity!(We especially consider the processor)
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 4 / 34
Hardware Architectures Overview
Flexibilityefficiency
General Purpose DSP ASP
Telephone chips
SignalProcessing
Acceleration
Desktop PC /Laptop
By example of: H. Corporaal, B.Mesman
Current trend
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 5 / 34
Operating System Support
1 Abstraction
provide generic concepts;handle complexity.
2 Virtualization
same abstraction for multiple underlying systems;each user is provided a dedicated platform (sharing of platform).
3 Resource management
sharing, protection of resources;optimize performance of resource usage.
Research subject: Resource Management Mechanisms
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 6 / 34
Operating System Support
1 Abstraction
provide generic concepts;handle complexity.
2 Virtualization
same abstraction for multiple underlying systems;each user is provided a dedicated platform (sharing of platform).
3 Resource management
sharing, protection of resources;optimize performance of resource usage.
Research subject: Resource Management Mechanisms
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 6 / 34
Real-time Systems
Definition
A real-time system is a system which has to fulfill:
1 Functional correctness
2 Timeliness correctness
Requires Analysis!
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 7 / 34
Real-time Systems
Definition
A real-time system is a system which has to fulfill:
1 Functional correctness
2 Timeliness correctness
Requires Analysis!
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 7 / 34
Real-time Task Model
Definition
A task is a sequence of actions that must be performed in reaction to anevent
Definition
A job is an instantiation of a task
Properties of a real-time task:
1 Timing parameters known upfront,(i.e.: computation time, deadline, period);
2 Assume worst-case execution time (WCET) near to average-caseexecution time;
3 priority assignment.
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 8 / 34
Real-time Task Model
Definition
A task is a sequence of actions that must be performed in reaction to anevent
Definition
A job is an instantiation of a task
Properties of a real-time task:
1 Timing parameters known upfront,(i.e.: computation time, deadline, period);
2 Assume worst-case execution time (WCET) near to average-caseexecution time;
3 priority assignment.
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 8 / 34
Policy versus Mechanism
Definition
A policy defines how a system should behave (abstract).
Definition
A mechanism provides the instrument to implement the policy.
Example: processor sharing (scheduling)
Policy:
time
. . .. . .task 1 task 2task 0
Round RobinAssign time-slots, ∆ts, alternating over tasks
∆ts
Mechanism: priority based scheduling and pre-emption
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 9 / 34
Policy versus Mechanism
Definition
A policy defines how a system should behave (abstract).
Definition
A mechanism provides the instrument to implement the policy.
Example: processor sharing (scheduling)
Policy:
time
. . .. . .task 1 task 2task 0
Round RobinAssign time-slots, ∆ts, alternating over tasks
∆ts
Mechanism: priority based scheduling and pre-emption
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 9 / 34
Fixed Priority Scheduling
Definition
Fixed Priority Pre-emptive Scheduling means:
1 highest priority ready task will always execute
2 during runtime the scheduler can pre-empt a lower priority task infavor of a higher priority task
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 10 / 34
Video Processing Characteristics
Periodic behavior with corresponding deadlines
Data-dependent, highly Fluctuating Load
0.06
0.07
0.08
0.09
0.1
0.11
0.12
0.13
0 200 400 600 800 1000
Tim
e(se
cond
s)
Number of processed frames
Time for deinterlacing a frame (576x720)
VQEG source 6
WCET NOT near to average case
Classical real-time approach leads to under-utilization of (processor)resources.
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 11 / 34
Priority Processing Concept
Basic Analyse Enhancetime
Qua
lity
BasicQuality(0%)
100%
1. Basic: simple and fast output atlow quality;
2. Analysis: Sort video contentin order of importance;
3. Enhance: Process video contentaccording to sorted order;
Termination is allowed after a basic output is available.
prel. termination
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 12 / 34
Priority Processing Concept
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 13 / 34
Priority Processing - Overload
Definition
Overload: A task’s computation time exceeds the deadline
time (t)
Period (T )D = T
ta td
basic scalable
preliminary termination
Required: Mechanism for preliminary termination
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 14 / 34
Priority Processing - Overload
Definition
Overload: A task’s computation time exceeds the deadline
time (t)
Period (T )D = T
ta td
basic scalable
preliminary termination
Required: Mechanism for preliminary termination
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 14 / 34
Priority Processing - Gain-time
Definition
Gain-time: Assigned, but unused processor resources becoming availablefor other tasks
time (t)
Period (T )D = T
ta td
basic scalable
Gain-time
Desired: Mechanism allowing gain-time consumption
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 15 / 34
Priority Processing - Gain-time
Definition
Gain-time: Assigned, but unused processor resources becoming availablefor other tasks
time (t)
Period (T )D = T
ta td
basic scalable
Gain-time
Desired: Mechanism allowing gain-time consumption
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 15 / 34
Priority Processing - Competing Algorithms
Multiple independent algorithms share a single processor
Divide period in fixed-size quanta,e.g. time-slots ∆ts
Decision Scheduler divides time-slots:
time(t)
time(t)
Algorithm 1
Algorithm 2
∆ts
ta td
period(T )
Hence: Dynamic Allocation
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 16 / 34
Priority Processing - Competing Algorithms
Multiple independent algorithms share a single processor
Divide period in fixed-size quanta,e.g. time-slots ∆ts
Decision Scheduler divides time-slots:
time(t)
time(t)
Algorithm 1
Algorithm 2
∆ts
ta td
period(T )
Hence: Dynamic Allocation
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 16 / 34
Priority Processing - Competing Algorithms
Multiple independent algorithms share a single processor
Divide period in fixed-size quanta,e.g. time-slots ∆ts
Decision Scheduler divides time-slots:
time(t)
time(t)
Algorithm 1
Algorithm 2
∆ts
ta td
period(T )
Required: Monitoring of time and progress on time-slot scale
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 17 / 34
Outline
1 IntroductionScalable Video AlgorithmsPlatformReal-time SystemsPriority Processing
2 Resource Management MechanismsPreliminary TerminationProcessor Allocation
3 Simulation Environment
4 Simulation ResultsPreliminary TerminationProcessor Allocation
5 Conclusions
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 18 / 34
Priority Processing Architecture (1/2)
PlatformSyst
emSu
ppor
t
Decision Scheduler
Scalable PriorityProcessing Algorithm 1
Scalable PriorityProcessing Algorithm 2
budgetprogress budgetprogress
App
licat
ion
Lev
el
Resource Management
uses
Decision scheduler aims at maximizing total progress of algorithms
Decision Scheduler defines a scheduling policy,e.g. based on Reinforcement Learning or Round Robin
Focus is on mechanisms to support (optimize) the decision scheduler
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 19 / 34
Priority Processing Architecture (2/2)
PlatformSyst
emSu
ppor
t
Decision Scheduler
Scalable PriorityProcessing Algorithm 1
Scalable PriorityProcessing Algorithm 2
budgetprogress budgetprogress
App
licat
ion
Lev
el
Resource Management
uses
Distribution of Responsibilities:
Application Specific (Policy) System Support (Mechanism)
Roll-forward when deadline reached Preliminary termination
Resource distribution Allocation of processor(Reinforcement Learning policy) (Scheduling)
Monitor Progress values Account consumed time(time-slots)
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 20 / 34
Resource Management Mechanisms
Goal:
Share processor resources among competing priority processing algorithmsas efficient as possible.
Compare performance of different implementations for the mechanisms:
1 Preliminary Termination
2 Processor Allocation
Note: Monitoring is missing and left for discussion.
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 21 / 34
Resource Management Mechanisms
Goal:
Share processor resources among competing priority processing algorithmsas efficient as possible.
Compare performance of different implementations for the mechanisms:
1 Preliminary Termination
2 Processor Allocation
Note: Monitoring is missing and left for discussion.
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 21 / 34
Preliminary Termination
Terminate all jobs when deadline is reached:
1 Cooperative termination (polling):
Decision Scheduler sets a flag when the deadline expires;Algorithms check on regular intervals for deadline expiration.
1 whi le ( ! endOfFrame and ! d e a d l i n e E x p i r e d )2 p r o c e s s N e x t B l o c k ( ) ;
2 A-synchronous signalling:
Decision Scheduler sends a signal causing all algorithms to preliminaryterminate jobs
Alg. 1
Signal handlerterminateJob();
proceed with next frame
ControlSignal
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 22 / 34
Preliminary Termination
Terminate all jobs when deadline is reached:
1 Cooperative termination (polling):
Decision Scheduler sets a flag when the deadline expires;Algorithms check on regular intervals for deadline expiration.
1 whi le ( ! endOfFrame and ! d e a d l i n e E x p i r e d )2 p r o c e s s N e x t B l o c k ( ) ;
2 A-synchronous signalling:
Decision Scheduler sends a signal causing all algorithms to preliminaryterminate jobs
Alg. 1
Signal handlerterminateJob();
proceed with next frame
ControlSignal
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 22 / 34
Processor Allocation
Dynamically Allocate Processor Resources:
1 Suspend-resume tasks:Suspend only in favor of others!
Resume
Suspend
time(t)
tactivate tdeadline
period(T )∆ts
2 Manipulate priorities:Substitute:
suspend → assign low priorityresume → assign high priority
Low priority tasks can use gain-time.
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 23 / 34
Processor Allocation
Dynamically Allocate Processor Resources:
1 Suspend-resume tasks:Suspend only in favor of others!
Resume
Suspend
time(t)
tactivate tdeadline
period(T )∆ts
2 Manipulate priorities:Substitute:
suspend → assign low priorityresume → assign high priority
Low priority tasks can use gain-time.
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 23 / 34
Outline
1 IntroductionScalable Video AlgorithmsPlatformReal-time SystemsPriority Processing
2 Resource Management MechanismsPreliminary TerminationProcessor Allocation
3 Simulation Environment
4 Simulation ResultsPreliminary TerminationProcessor Allocation
5 Conclusions
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 24 / 34
Simulation Environment
Priority Processing Applications are prototyped using:
Matlab-Simulink (version 2007b)
Microsoft Windows XP:
use Fixed Priority Scheduler (FPS)administrator privileges required
General Purpose Multi-core machine:
Scalable Video AlgorithmsDecision SchedulerMatlab environment(extract results from simulation)
Video sequences from VQEG
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 25 / 34
Simulation Environment
Priority Processing Applications are prototyped using:
Matlab-Simulink (version 2007b)
Microsoft Windows XP:
use Fixed Priority Scheduler (FPS)administrator privileges required
General Purpose Multi-core machine:
Scalable Video AlgorithmsDecision SchedulerMatlab environment(extract results from simulation)
Video sequences from VQEG
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 25 / 34
Simulation Environment
Priority Processing Applications are prototyped using:
Matlab-Simulink (version 2007b)
Microsoft Windows XP:
use Fixed Priority Scheduler (FPS)administrator privileges required
General Purpose Multi-core machine:
Scalable Video AlgorithmsDecision SchedulerMatlab environment(extract results from simulation)
Video sequences from VQEG
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 25 / 34
Outline
1 IntroductionScalable Video AlgorithmsPlatformReal-time SystemsPriority Processing
2 Resource Management MechanismsPreliminary TerminationProcessor Allocation
3 Simulation Environment
4 Simulation ResultsPreliminary TerminationProcessor Allocation
5 Conclusions
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 26 / 34
Preliminary Termination - Overhead
Computational Overhead of polling-based mechanisms:
0
1
2
3
4
5
6
50 100 150 200
Rel
ativ
e C
ompu
tatio
n T
ime
(% a
dd. t
ime)
Number of processed frames
Polling overhead for scalable deinterlacer (VQEG source 6)
Pixel-based pollingBlock-based polling
Conclusion: Polling causes small overhead on general purpose machines.
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 27 / 34
Preliminary Termination - Overhead
Computational Overhead of polling-based mechanisms:
0
1
2
3
4
5
6
50 100 150 200
Rel
ativ
e C
ompu
tatio
n T
ime
(% a
dd. t
ime)
Number of processed frames
Polling overhead for scalable deinterlacer (VQEG source 6)
Pixel-based pollingBlock-based polling
Conclusion: Polling causes small overhead on general purpose machines.
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 27 / 34
Preliminary Termination - Latency (1/2)
Polling based termination latencies:
40
42
44
46
48
50
52
54
56
58
60
0 50 100 150 200 250 300 350 400 450
Late
ncy(
µs)
Number of processed frames
Termination latency for deinterlacer (VQEG src6)
Latency (dl = 60 ms)
40
42
44
46
48
50
52
54
56
58
60
0 50 100 150 200 250 300 350 400 450
Late
ncy(
µs)
Number of processed frames
Termination Latency for deinterlacer (VQEG src6)
Latency (dl = 60 ms)
Block-based polling Pixel-based polling
Conclusion: Lowerbound on latency due to OS overhead.
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 28 / 34
Preliminary Termination - Latency (1/2)
Polling based termination latencies:
40
42
44
46
48
50
52
54
56
58
60
0 50 100 150 200 250 300 350 400 450
Late
ncy(
µs)
Number of processed frames
Termination latency for deinterlacer (VQEG src6)
Latency (dl = 60 ms)
40
42
44
46
48
50
52
54
56
58
60
0 50 100 150 200 250 300 350 400 450
Late
ncy(
µs)
Number of processed frames
Termination Latency for deinterlacer (VQEG src6)
Latency (dl = 60 ms)
Block-based polling Pixel-based polling
Conclusion: Lowerbound on latency due to OS overhead.
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 28 / 34
Preliminary Termination - Latency (2/2)
Signalling based termination latencies:
Implemented in Linux (not in Windows)
Unpredictable latencies
Conclusion: Polling is the most general solution.
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 29 / 34
Processor Allocation - Gain-time
How suitable is Priority Processing for gain-time consumption?1 Use Round Robin policy within Decision Scheduler2 Non-optimal scheduling (leaving gain-time)
0
10
20
30
40
50
60
40 50 60 70 80 90 100 110 120
Rel
ativ
e G
ain
(% p
roce
ssed
blo
cks)
Period(ms)
Priority Manipulation vs. Suspend/Resume using RR
Enhancement FilterDeinterlacer
Conclusion: Priority manipulation allows gain-time consumption
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 30 / 34
Processor Allocation - Gain-time
How suitable is Priority Processing for gain-time consumption?1 Use Round Robin policy within Decision Scheduler2 Non-optimal scheduling (leaving gain-time)
0
10
20
30
40
50
60
40 50 60 70 80 90 100 110 120
Rel
ativ
e G
ain
(% p
roce
ssed
blo
cks)
Period(ms)
Priority Manipulation vs. Suspend/Resume using RR
Enhancement FilterDeinterlacer
Conclusion: Priority manipulation allows gain-time consumption
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 30 / 34
Processor Allocation - Gain-time
How suitable is Priority Processing for gain-time consumption?1 Use Round Robin policy within Decision Scheduler2 Non-optimal scheduling (leaving gain-time)
0
10
20
30
40
50
60
40 50 60 70 80 90 100 110 120
Rel
ativ
e G
ain
(% p
roce
ssed
blo
cks)
Period(ms)
Priority Manipulation vs. Suspend/Resume using RR
Enhancement FilterDeinterlacer
Conclusion: Priority manipulation allows gain-time consumption
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 30 / 34
Optimized Control
1 Use block-based polling
2 Use Windows FPS
3 Reduced context-switching
4 Allow gain-time consumption
0
2
4
6
8
10
12
14
30 40 50 60 70 80 90 100 110 120
Rel
ativ
e G
ain
(% p
roce
ssed
blo
cks)
Period(ms)
Priority Manipulation vs. Suspend/Resume using RL
Enhancement FilterDeinterlacer
Conclusion: Significant improvement by reducing control overhead
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 31 / 34
Optimized Control
1 Use block-based polling
2 Use Windows FPS
3 Reduced context-switching
4 Allow gain-time consumption
0
2
4
6
8
10
12
14
30 40 50 60 70 80 90 100 110 120
Rel
ativ
e G
ain
(% p
roce
ssed
blo
cks)
Period(ms)
Priority Manipulation vs. Suspend/Resume using RL
Enhancement FilterDeinterlacer
Conclusion: Significant improvement by reducing control overhead
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 31 / 34
Outline
1 IntroductionScalable Video AlgorithmsPlatformReal-time SystemsPriority Processing
2 Resource Management MechanismsPreliminary TerminationProcessor Allocation
3 Simulation Environment
4 Simulation ResultsPreliminary TerminationProcessor Allocation
5 Conclusions
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 32 / 34
Conclusions
Identification of 3 mechanisms for dynamic resource allocation:1 Preliminary termination:
Most platforms, including Windows, do not support signallingPolling is preferred (trade-off granularity)
2 Processor allocation:
Reduced context-switchingPriority manipulation allows gain-time consumption
3 Monitoring:
On a time-slot scaleRelies on availability of high-resolution timers
Open challenges:1 Optimize processor utilization:
Overhead caused by interference of simulation environmentOptimize memory management
2 Mapping on embedded platform (real-time environment)
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 33 / 34
Conclusions
Identification of 3 mechanisms for dynamic resource allocation:1 Preliminary termination:
Most platforms, including Windows, do not support signallingPolling is preferred (trade-off granularity)
2 Processor allocation:
Reduced context-switchingPriority manipulation allows gain-time consumption
3 Monitoring:
On a time-slot scaleRelies on availability of high-resolution timers
Open challenges:1 Optimize processor utilization:
Overhead caused by interference of simulation environmentOptimize memory management
2 Mapping on embedded platform (real-time environment)
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 33 / 34
Conclusions
Identification of 3 mechanisms for dynamic resource allocation:1 Preliminary termination:
Most platforms, including Windows, do not support signallingPolling is preferred (trade-off granularity)
2 Processor allocation:
Reduced context-switchingPriority manipulation allows gain-time consumption
3 Monitoring:
On a time-slot scaleRelies on availability of high-resolution timers
Open challenges:1 Optimize processor utilization:
Overhead caused by interference of simulation environmentOptimize memory management
2 Mapping on embedded platform (real-time environment)
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 33 / 34
Conclusions
Identification of 3 mechanisms for dynamic resource allocation:1 Preliminary termination:
Most platforms, including Windows, do not support signallingPolling is preferred (trade-off granularity)
2 Processor allocation:
Reduced context-switchingPriority manipulation allows gain-time consumption
3 Monitoring:
On a time-slot scaleRelies on availability of high-resolution timers
Open challenges:1 Optimize processor utilization:
Overhead caused by interference of simulation environmentOptimize memory management
2 Mapping on embedded platform (real-time environment)
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 33 / 34
References
L. F. Bic and A. C. Shaw.Operating Systems Principles.Prentice-Hall, Inc., Upper Saddle River, NJ, USA, 2002.
C. Hentschel and S. Schiemenz.Priority-processing for optimized real-time performance with limitedprocessing resources.International Conference on Consumer Electronics (ICCE), 2008.Digest of Technical Papers., Jan. 2008.
S. Schiemenz.Echtzeitsteuerung von skalierbaren Priority-Processing Algorithmen.Tagungsband ITG Fachtagung - Elektronische Medien, pages 108 –113, 17–18 March 2009.
Martijn van den Heuvel (TU/e, SAN) Dynamic Resource Allocation 23 July 2009 34 / 34