1(18)
Using Constraint Programming for Embedded System
Radoslaw Szymanek
Department of Computer Science
Lund Institute of Technology
Sweden
Email: Radoslaw.Szymanek@cs.lth.se
2(18)
Embedded System
performs computation but it is not a general purpose computer
usually performs computation to achieve given goal/functionality
mobile phones, home appliances, security and safety systems
Radoslaw Szymanek
January 13, 2004
3(18)
Issuses within Embedded Systems
minimal cost and maximal functionality
memory issues - increased audio/video requirements
early good optimization can achieve substanstial gains
Radoslaw Szymanek
January 13, 2004
4(18)
Scheduling Problem
heterogeneous multiprocessor system
parallel computation and communication
heterogeneous constraints - time, memory constraints
Radoslaw Szymanek
January 13, 2004
5(18)
Constraint description of Scheduling Problem
diffn, cumulative, element constraints
arithmetic constraints
resource constrained scheduling
Radoslaw Szymanek
January 13, 2004
6(18)
Optimization
requires search heuristic
resource based criteria in optimization function
reneweable (data memory) and non reneweable resources (time and
code memory)
Radoslaw Szymanek
January 13, 2004
7(18)
Result
easy to extend the model
easy to answer ad-hoc questions, what if?
guided search methods
Radoslaw Szymanek
January 13, 2004
8(18)
Memory Intricacies
current memory systems allow fast access under given conditions
high penalties if conditions not respected
high number of memory modules to increase effectiveness
Radoslaw Szymanek
January 13, 2004
9(18)
SDRAM memory
high bandwidth can be obtained if used with care
U=2
P=2
B0
B1
P0
P1
S=4
S1
Radoslaw Szymanek
S2
January 13, 2004
S3
S4
t
10(18)
SDRAM memory
page buffers requires special access patterns
U=2
P=2
B0
B1
P0
P1
S=4
S1
Radoslaw Szymanek
S2
January 13, 2004
S3
S4
t
11(18)
SDRAM memory
interleaved data access possible from multiple open page buffers
U=2
P=2
B0
B1
P0
P1
S=4
S1
Radoslaw Szymanek
S2
January 13, 2004
S3
S4
t
12(18)
Memory Bandwidth Modeling
main bottleneck in current applications
diffn constraints for page reservation during data access
cumulative constraints for size constraints
cumulative for memory bandwidth constraints
Radoslaw Szymanek
January 13, 2004
13(18)
Application Execution Flexibility
each task can execute in fixed number of possible ways
each execution option indicates optimal operation scheduling within
a task
operation scheduling freedom restricted to few representative
choices
application execution flexibility obtained by task execution flexibility
Radoslaw Szymanek
January 13, 2004
14(18)
Time and Energy Pareto Application Diagram
Hierarchical composition of application task Pareto Diagrams
T2
TN
time
T1
App
time
TN−1
time
energy
energy
energy
energy
energy
Energy consumption is a sum of execution consumption of tasks
time
time
Task energy consumption depends on the option chosen
Execution time of the application depends on scheduling and task
execution option
Radoslaw Szymanek
January 13, 2004
15(18)
Conflict Graph Constraints
Specifies relation between task data which needs to be satisfied
Memory conflict constraint requires two data to be stored in
different memory
Page conflict constraint requires two data to be stored in such a
way that they are accessed from different pages
Memory compatibility specifies that two data must be stored in the
same memory
Bank compatibility can be used to facilitates page sharing during
data access
Radoslaw Szymanek
January 13, 2004
16(18)
Application execution optimization
Find for each task appropriate execution option
Find best assignment from bandwidth and size point of view
Find best scheduling for given tasks execution options and data
assignment
Compute constraints for assignment improvement
Find better assignment and reschedule application
Radoslaw Szymanek
January 13, 2004
17(18)
Memory conclusions
Problem divided into 3 subproblems
Estimates used for the unknown parameters
Solution to each step gives possibility to improve results of the
previous stage
Global constraints and tailored search methods
Radoslaw Szymanek
January 13, 2004
18(18)
Conclusions
Heterogeneous problems and specific search methods
Adaptability
Global constraints and search methods
JaCoP - our own constraint solver written in Java
Radoslaw Szymanek
January 13, 2004