Genetic Programming based Iterative Improvement Algorithm for

HW/SW Cosynthesis of Distributted Embedded Systems

Adam Górski and Maciej Ogorzałek

Department of Information Technologies, Jagiellonian University in Cracow,

Prof. Stanisława Łojasiewicza 11, Cracow, Poland

Keywords: Embedded Systems, Architecture, Hardware/Software Co-Synthesis, Genetic Programming.

Abstract: In this work we present a novel genetic programming based iterative improvement approach for

hardware/software cosynthesis of distributed embedded systems. The approach starts from a ready solution

which is an embryo of a genotype. Other nodes in the genotypes are chromosomes. The chromosomes contain

system refinement options. The final solution is obtained after evolution process and mapping genotype to

phenotype. Unlike existing genetic programming iterative improvement methodologies our algorithm starts

from randomly generated system. Therefore the search space is not constrained by any initial condition. It is

also easier for the algorithm to escape local minima of optimizing parameters.

1 INTRODUCTION

Today a lot of devices can be solved as embedded

systems or Internet of Things solutions (IoT).

Examples of such devices can be: cars (Srovnal,

Machacek, Hercik, Slaby and Srovnal, 2010), drones

(Yoon, Anwar, Rakshit and Raychowdhury 2019),

digital door keys (Beaufour and Bonnet 2004),

parking management system (Tsiropoulou et al.

2017) and many others. Nowadays a lot of such

systems have distributed structure (Martins, Tavares,

Solieri, Bertogna, and Pinto, 2020). The complexity

of such systems is increasing. Therefore it is very

important to provide appropriate methodologies to

design of embedded systems.

According to DeMicheli and Gupta (DeMicheli

and Gupta 1997) exist three basic phases of designing

of embedded systems: modelling, implementation

and validation. Another very important problem

appears when system works in defined structure and

meets unexpected situation (Górski and Ogorzałek

2016).

Cosynthesis of distributed embedded systems

(Yen and Wolf 1997) is a process which concurrently

selects architecture of embedded system, assign tasks

to the resources and schedule the tasks. Most of the

cosynthesis methodologies can be divided on two

groups: constructive algorithms and iterative

improvement approaches.

Constructive algorithms (Srinivasan and Jha,

1995) build systems step by step making separate

assignment decisions for each task. The advantage of

those methodologies is low complexity. However

such methodologies can stop in local minima of

optimizing parameters. In (Dave, Lakshminarayana

and Jha 1997) is presented a constructive algorithm

that can change previously made decision, but the

complexity of the algorithm is increasing.

Iterative improvement methodologies (Oh, Ha,

2002) can escape from local minima. Such an

approaches starts from a suboptimal architecture. The

target architecture is obtained after making local

decisions like allocating or deallocating resources or

changing tasks assignment. However obtained

solutions can still be suboptimal.

Genetic methodologies (Conner, Xie, Kandemir,

Link and Dick, 2005) can also escape from local

minima but often results depends on parameters

(Dick, and Jha, 1998). Very good results obtained

using genetic programming (Deniziak and Górski

2008, Górski and Ogorzałek 2014a) especially

genetic programming based adaptive methodologies

(Górski and Ogorzałek 2014b, Górski and Ogorzałek

2017) which can adapt to the environment during its

work.

In this paper we present a novel iterative

improvement methodology for hardware/software

cosynthesis of distributed embedded systems. The

methodology is based on developmental genetic

120

Górski, A. and Ogorzałek, M.

Genetic Programming based Iterative Improvement Algorithm for HW/SW Cosynthesis of Distributted Embedded Systems.

DOI: 10.5220/0010391501200125

In Proceedings of the 10th International Conference on Sensor Networks (SENSORNETS 2021), pages 120-125

ISBN: 978-989-758-489-3

programming (Koza, Bennett III, Lohn, Dunlap,

Keane and Andre, 1997). Genetic programming

evolves tree based genotypes. In the nodes of the trees

are system creating functions. The final solution is

created after evolution process by making genotype

to phenotype mapping.

2 THE NOTATION

PE – Processing Element

CL – Communication Link

b – CL’s bandwidth.

PP – Programmable Processors

HC – Hardware Core

T – task

E – edge in a task graph

G={T,E} – task graph

– cost of an embedded system

c – cost of a task’s execution

– transmission time between tasks T

and T

dij – amount of data transferred between tasks T

and

t – time of a task’s execution

n – number of tasks

e – number of PEs in given database

Π – number of individuals in each population

α – parameter which controls number of individuals

in each population

β – parameter which controls number of individuals

obtained by reproduction

γ – parameter which controls number of individuals

obtained by crossover

δ – parameter which controls number of individuals

obtained by mutation

Φ – number of individuals obtained by reproduction

Ψ – number of individuals obtained by crossover

Ω – number of individuals obtained by mutation

X – number of tasks in a node

P – probability of selection operator

r – position of an individual in a rank list

p – number of communication links in a solution

m – number of programmable processors in a solution

ε – parameter which controls the stop of an algorithm

3 ASSUMPTIONS

Distributed embedded system is built of two kind of

resources: Processing Elements (PEs) and

Communication Links (CLs). PEs execute the tasks.

CLs provides communication between connected

PEs. Processing elements can be divided into:

Programmable Processors (PPs) and Hardware Cores

(HCs). PPs are universal resources and can execute

more than one task. HCs are resources dedicated to

execute a single task only. The cosynthesis process

consists of: resource allocation, task assignment and

task scheduling. Resource allocation is a phase that

selects number and type of Processing Elements.

Task assignment chooses appropriate PE for each

task. In this paper we use a graph representation of

embedded system G={T, E}. The task graph

describes dependencies between tasks. Each node of

the graph contains a task (T). Each edge (E) presents

amount of data that has to be sent between two

connected tasks. The transmission time is described

by the following formula:



(1)

Where b represents a bandwidth. If two connected

tasks are executed on the same resources, the

transmission time is equal to 0.

On figure 1 an example of a task graph was presented.

T4 T5

24 4

T8 T9

T7 T6

Figure 1: Example of task graph.

The graph consists of ten nodes: T0, T1, T2, T3, T4,

T5, T6, T7, T8 and T9. Tasks T1 and T2 can start their

execution only after execution of T0 is terminated. T3

starts its execution after T1. T4 and T5 can start the

execution after T2 is executed. T7 and T8 can be

executed after T4. T9 can start its execution after T9.

Task T6 can start its execution after terminating

execution of T4 or T3. The following groups of tasks:

T1 and T2, T3, T4 and T5, T6, T7, T8 and T9 can be

executed at the same time using another resources.

There are four possible kinds of processing

elements: two Programmable Processors (PP1 and

PP2) and two Hardware Cores (HC1 and HC2).

Processing elements can be connected by two

Genetic Programming based Iterative Improvement Algorithm for HW/SW Cosynthesis of Distributted Embedded Systems

121

communication links (CL1 and CL2). The costs of all

processing elements (C) and connection of PPs to

communication links (c) are given in the database.

Cost of PP1 is equal to 130. Cost of PP2 is equal to

210. Each task is characterized by the cost (c) and

time (t) of its execution. Cost of dedicated resources

(HC1 and HC2) is added to the cost of tasks

execution. Cost of connecting PP1 to CL1 is equal to

8. Cost of connection of PP1 to CL2 is equal to 18.

Table 1: Example of resource database.

Tas

PP1

C=130

PP2

C=210

HC1 HC2

t c t c t c t c

T0 10 9 20 5 5 300 4 380

T1 35 10 14 15 6 200 9 150

T2 18 3 15 4 10 80 2 110

T3 25 5 20 8 9 250 4 300

T4 81 23 72 16 23 180 18 190

T5 13 8 10 9 4 130 5 140

T6 19 25 16 40 9 180 8 210

T7 49 6 23 19 7 200 8 160

T8 14 8 20 10 5 200 4 300

T9 28 12 32 20 8 180 6 250

CL1,

b=2

c=8 c=2 c=35

CL2,

b=12

c=18 c=12 c=41

If n presents a number of tasks to be executed by

the system described by the graph from table1, m is a

number of programmable processors (PPs) and p

describes number of communication links (CLs) the

cost of a final system (C

) is described by the

following formula:







PCCL

PEf

kzk

ccCC

(2)

The methodology presented in next chapter must

find a system with the lowest C

value which does not

exceed the time constrains.

4 THE APPROACH

The presented methodology is genetic programming

approach. Therefore at the beginning the initial

population of genotypes must be created. The number

of individuals in the population is described by the

formula 3 below:

e*n* =





(3)

n is a number of tasks executed by the system, e is

a number of PEs in the database, α is given by the

designer.

According to genetic programming rules every

individual is a tree. The first node in the tree is an

embryo. In existing constructive GP based algorithms

(Deniziak and Górski, 2008, Górski and Ogorzałek,

2014a) embryo is a random implementation only of

the first task. Others iterative improvement GP based

solutions (Górski Ogorzałek 2014b) stat with the

fastest implementation of all the tasks. Unlike others

methodologies in presented approach embryo is

random implementation of all the tasks. Every next

node include number of tasks taken from the previous

node and system building option for the tasks. Such

an approach is also unique in GP based co-synthesis

algorithms in which so far the number of tasks in the

nodes was constrained (equal to 1) or determined by

genes. The number of reassignment tasks is chosen

randomly from 1 to X -1. The algorithm selects task

by maximum value of the following formula:

t)*max(c=

(4)

X is the number of tasks in predecessor node.

Number of nodes and theirs positions are

generated randomly.

Table 2 include the options for reassignment of

tasks. The options are divided into two groups: for

PEs and for CLs. For PEs the following options are

possible: the cheapest implementation of the tasks,

the fastest implementation of the tasks, min value of

multiply of time and cost, the same as task’s

predecessor and the rarest used PP. For CLs the

algorithm can choose from options: the fastest CL, the

cheapest CL and the rarest used CL. Each option,

either for PEs and CLs, has a probability of being

chosen. Options a and b for PEs have probability

0,15, option c 0,3, and options d and e 0,2. Options

a and b for CLs have probability equal to 0,3, option

c equal to 0,4.

Table 2: Options for building system.

Step Option Probability

PE a. The cheapest implementation

of the tasks

0,15

b. The fastest implementation of

the tasks

0,15

c. min

(

t*c

)

0,3

d. The same as task’s predecessor 0,2

e. The rarest used PP

0,2

CL a. The fastest CL

0,3

. The chea

est CL 0,3

c. The rarest use

0,4

Task

scheduling

list scheduling

SENSORNETS 2021 - 10th International Conference on Sensor Networks

122

If there is more than one task assignment to one

PE list scheduling is used to set the tasks’ order.

Every next generation is created using genetic

operators: mutation, crossover and reproductions.

The evolution process is controlled by the parameters

β, γ, δ which value is given by the designer. Number

of individuals created by the reproduction (Φ),

crossover (Ψ) and mutation (Ω) are given by the

following formulas below:

 Φ = β*П

 Ψ = γ*П

 Ω = δ*П

To have the same number of individuals in every

population the following condition must be satisfied:

β + γ + δ = 1 (5)

Crossover randomly selects Ψ genotypes. Then

the crossing point is chosen randomly different for

every genotypes. Next the subtrees of the genotypes

are substituted.

Mutation randomly chooses Ω individuals and

a node. Next the option in the node is substitutes on

another from table 2.

Reproduction selects Φ genotypes and copies

them to a new population. Every individual has

a probability of being chosen. The probability

depends on a position of the individual in a rank list

(r) as follows:









(6)

The approach stops calculating if in ε next

generations better solution was not found. Parameter

ε is given by the designer of the system.

5 THE EXAMPLE

On figure 2 was presented an example of a genotype.

The genotype was generated randomly for the graph

from figure 1. The genotype consists of 5 nodes. In

every node (except the first one) there is an option

from table 2 for PE and for CL. The nodes consists

also a number of a tasks which are taken from the

previous node.

embryo

c/b

Figure 2: Example of genotype.

The first node in the genotype is an embryo. The

embryo is random implementation of all the tasks.

The task assignment was made as follows: PP1: T0,

T2, T4 and T6, PP2: T1 and T8, HC1: T3 and T9,

HC2: T5 and T7. The cost of the solution was 1335.

and the time of execution of all the tasks 135. In the

next node option a for PEs and option b for CLs for 5

tasks were chosen. In the third node for 4 task option

d for PEs and c for CLs was selected. The next node

include options c (PEs) and a (CLs) for 3 tasks. In the

last one node options c for PEs and b for CLs for 2

tasks were chosen. The final cost of generated

solution is 475 and the time of execution of all tasks

is equal to 225.

6 FIRSTS RESULTS

Firsts results obtained by the algorithm described in

this paper were compared with results obtained using

constructive GP based algorithm for cosynthesis

proposed by Deniziak and Górski (Deniziak and

Górski 2008). The results were made on a benchmark

with 10 nodes (Deniziak and Górski 2008).

Algorithm proposed by Deniziak and Górski was

proved to give better results than algorithm EWA for

cosynthesis (Deniziak, 2004) which is not genetic

approach. In table 3 the results were presented. The

parameters were set as follows:

α = 10, β = 0.2,

δ = 0.1, γ = 0.7 ε = 5.

Table 3: Options for building system.

Algorith

Tmax

c generation

DGP08 1000 956 1063 4

GP21 1000 956 788 6

The maximum time constrained was set to 1000.

Algorithm presented in this paper (GP21) generated

solution with the same time as algorithm DGP08

Genetic Programming based Iterative Improvement Algorithm for HW/SW Cosynthesis of Distributted Embedded Systems

123

(time was equal to 956) but with much lower time

(equal to 788). The figure 3 presents graphical

comparison of the results.

Figure 3: Graphical presentation of obtained results.

As it can be observed for used beenchmark

algorithm GP21 generated result about 26% better

than DGP08.

7 CONCLUSIONS AND FUTURE

WORK

The paper presents a novel iterative improvement,

Genetic Programming based approach for

hardware/software cosynthesis of distributed

embedded systems. Unlike other methodologies the

algorithm starts from randomly generated solution.

Therefore it is easier for the algorithm to escape from

local minima of optimizing parameters. Starting from

randomly generated genotype can also increase the

search space. First results are promising, however the

algorithm needs to be further investigated.

The future work will concentrate on providing

more experimental results and test the behaviour of

proposed approach. We will also provide more

Genetic Programing based algorithms in which we

will investigate another system building options and

test others genetic operators.

REFERENCES

Srovnal V. Jr, Machacek, Z. Hercik, R., Slaby, R., Srovnal,

V., 2010. Intelligent car control and recognition

embedded system. In Proceedings of the International

Multi- conference on Computer Science and

Information Technology, pp. 831–836.

Yoon I., Anwar A., Rakshit T., Raychowdhury A., 2019.

Transfer and online reinforcement learning in STT-

Mram based embedded systems for autonomous

drones. In 2019 Design, Automation & Test in Europe

Conference & exhibition (DATE)., pp.1489-1494, IEEE.

Martins J., Tavares A., Solieri M., Bertogna M., Pinto S.,

2020. Bao: A lightweight static partitioning hypervisor

for modern Multi-Core Embedded Systems. In

Workshop on Next Generation Real-Time Systems

L. Sun, L. Zhang, D., Li, B., Guo, B., Li, S., 2010. Activity

recognition on an accelerome- ter embedded mobile

phone with varying positions and orientations. In Zhi-

wen Yu, Ramiro Liscano, Guanling Chen, Daqing

Zhang, Xingshe Zhou (Eds.), Ubiquitous Intelligence

and Computing, Lecture Notes in Computer Sciences,

6406, pp. 548–562, Springer, Xi’an, China.

Beaufour, A., and Bonnet, P.,2004 Personal Servers as

Digital Keys. In Proceedings of the 2

IEEE

International Conf. of Pervasive Computing and

Communications (PerCom).

Tsiropoulou, E. E., Baras, J. S., Papavassiliou S. and Sinha,

S., 2017. RFID-based smart parking management

system. In Cyber-Physical Systems, Vol.3, pp.22-41.

De Micheli, G., Gupta, R., 1997. Hardware/software

co-design. In Proceedings IEEE 95.3 (Mar). IEEE.

Górski, A., Ogorzałek, M.J., 2016. Assignment of

unexpected tasks in embedded system design process.

Microprocessors and Microsystems, Vol. 44, pp. 17-21,

Elsevier.

Yen, T., Wolf, W., 1995. Sensivity-Driven Co-Synthesis of

Distributed Embedded Systems. In Proceedings of the

International Symposium on System Synthesis.

Srinivasan, S., Jha, N.K., 1995. "Hardware-Software Co-

Synthesis of Fault-Tolerant Real-Time Distributed

Embedded Systems", In Proceedings European Design

Automation Conference. pp. 334-339.

Dave, B., Lakshminarayana, G., Jha, N., 1997. COSYN:

Hardware/software Co-synthesis of Embedded

Systems. In Proceedings of the34th annual Design

Automation Conference (DAC’97).

Oh, H., Ha, S., 2002. Hardware-software cosynthesis of

multi-mode multi-task embedded systems with real-time

constraints. In Proceedings of the International Work-

shop on Hardware/Software Codesign, pp.133–138.

Dick, R., P., Jha, N., K., 1998. MOGAC: a multiobjective

Genetic algorithm for the Co-Synthesis of

Hardware-Software Embedded Systems. In IEEE

Trans. on Computer Aided Design of Integrated

Circiuts and systems, vol. 17, No. 10.

Conner, J., Xie, Y., Kandemir, R., Link, G., Dick, R., 2005.

FD-HGAC: AHybrid Heuristic/Genetic Algorithm

Hardware/Software Co-synthesis Framework with

Fault Detection. In Proceedings of Asia South Pacific

Design Automation Conf. (ASP-DAC), pp. 709-712.

Deniziak, S., Górski, A., 2008. Hardware/Software Co-

Synthesis of Distributed Embedded Systems Using

Genetic programming. In Proceedings of the 8th

International Conf. Evolvable Systems: From Biology

to Hardware, ICES 2008. Lecture Notes in Computer

Science, Vol. 5216. SPRINGER-VERLAG.

Górski, A., Ogorzałek, M.J., 2014a. Adaptive GP-based

algorithm for hardware/software co-design of

distributed embedded systems. In Proceedings of the

500

1000

1500

DGP08 GP21

SENSORNETS 2021 - 10th International Conference on Sensor Networks

124

4th International Conf. on Pervasive and Embedded

Computing and Communication Systems, Portugal.

Górski, A., Ogorzałek, M.J., 2014b. Iterative improvement

methodology for hardware/software co-synthesis of

embedded systems using genetic programming. In

Proceedings of the 11th Conf. on Embedded Software

and Systems (Work in Progress Session), Paris, France.

Górski, A., Ogorzałek, M.J., 2017. Adaptive iterative

improvement GP-based methodology for HW/SW co-

synthesis of embedded systems. In Proceedings of the 7

International Joint Conf. on Pervasive and Embedded

Computing and Communication Systems, Madrid, Spain.

Koza, J., R., Bennett III, F., H., Lohn, j., Dunlap, F., Keane,

M., A., Andre, D., 1997. Automated synthesis of

computational circuits using genetic programming. In

Proceedings of the IEEE Conf. on Evolutionary

Computation. IEEE.

Deniziak, S., 2004. Cost-efficient synthesis of

multiprocessor heterogeneous systems. In Control and

Cybernetics, Vol.33, No. 2, pp.341-355.

Genetic Programming based Iterative Improvement Algorithm for HW/SW Cosynthesis of Distributted Embedded Systems

125