comments on “algorithmic aspectsof hardware/software partitioning:1d search algorithms”

2
Comments on “Algorithmic Aspects of Hardware/Software Partitioning: 1D Search Algorithms” Haojun Quan, Tao Zhang, Qiang Liu, Jichang Guo, Xiaochen Wang, and Ruimin Hu Abstract—In this paper, the work in [1] is analyzed. An error in its theoretical description part is pointed out and illustrated by a simple example. A modification suggestion is proposed to make the theoretical description of the work [1] more deliberate and thus being used appropriately. Index Terms—Hardware/software partitioning, 1D search algorithm Ç 1 INTRODUCTION Jigang et al. [1] proposed a 1D search approach for hardware/soft- ware partitioning, which reduces the time complexity compared with a 2D search method developed in literature. However, when we tried to exploit the approach [1], we found an error in the theo- retical description of the paper. The error leads to a mismatch between the proposed approach and the algorithm realizations in [1]. As a result, it is desirable to point out the error and make sug- gestions to make the proposed approach [1] more deliberate and not mislead readers. Therefore, in this comment, the approach and theorem pro- posed in [1] are briefly described and an error is pointed out. Section 2 shows a simple example to illustrate the mentioned error. After that, Section 3 suggests a possible modification. As described in [1], given an undirected graph G ¼ ðV;EÞ;V ¼fv 1 ;v 2 ; ... ;v n g; s; h : V ! R þ , and c : E ! R þ . s i and h i denote the software cost and the hardware cost of node v i , respectively. Let x ¼ x 1 ;x 2 ; ... ;x n f g denote a partition scheme, in which x i ¼ 1ðx i ¼ 0Þ indicates that node v i is assigned to software (hardware). Thus, the software cost and the hardware cost of x can be formulated as SðxÞ¼ P n i¼1 s i x i and HðxÞ¼ P n i¼1 h i ð1 x i Þ, respectively. Meanwhile, the communication cost of x is CðxÞ¼ P n1 i¼1 P n j¼iþ1 c ij jx i x j j. Given a constraint constant R, the hard- ware/software partitioning problem can be formulated as the fol- lowing minimization problem P : P minimize P n i¼1 h i ð1 x i Þ subject to P n i¼1 s i x i þ CðxÞ R: 8 > > < > > : The value of P n i¼1 h i is fixed for a certain G, so the solution of problem P is identical to the solution of problem Q: Q maximize P n i¼1 h i x i subject to P n i¼1 s i x i þ C x ðÞ R: 8 > > < > > : Then a new variable m 0; 1Þ for C x ðÞ¼ mR is defined and the problem Q is transformed to the following problem: Q 0 maximize P n i¼1 h i x i subject to P n i¼1 s i x i ð1 mÞR: 8 > > < > > : Let Q x ðÞ¼ P n i¼1 h i x i , and x m denote the optimal solution of the problem Q 0 with m. Then Theorem 1 is given in [1]. Theorem 1. Qðx m1 Þ Qðx m2 Þ for m 1 m 2 , where x m1 and x m2 are the optimal solutions of the problem Q 0 with m 1 and m 2 , respectively. Then [1] gives the following conclusion: Conclusion 1. Let x m denote the optimal solution of the prob- lem Q 0 for a given m in (0, 1). The main idea derives from the fact that: if x m fulfills P n i¼1 s i x i þ P n1 i¼1 P n j¼iþ1 c ij jx i x j j R; x m definitely is a feasible solution of the problem Q. Hence, x m is qualified to be a candidate partition. Moreover, if x m that fulfills P n i¼1 s i x i þ P n1 i¼1 P n j¼iþ1 c ij jx i x j j R is found, there is no need to further search for the optimal solution of Q 0 with m þ Dm for any Dm value. This is because, according to Theo- rem 1, there does not exist a better feasible solution than x m for larger m values in (0, 1) for the problem Q 0 . However, Conclusion 1 is incorrect, because Theorem 1 derived from problem Q 0 is applied to problem Q without considering the impact of communication cost on partitioning schemes. To illus- trate this, an example will be given in the next section. 2 EXAMPLE As shown in Fig. 1, each node representing a task contains two val- ues in a bracket, where the first value denotes the software cost and the second value denotes the hardware cost. Let R ¼ 7.5. There are eight possible partitioning schemes, as illustrated in Fig. 2. The corresponding analysis is shown in Table 1. According to Conclusion 1 described in the previous section, searching communication cost from 0 to 3, the optimal scheme should be scheme 2 (extreme condition) or 3. However, in fact the optimal scheme is scheme 6 or 7, which are feasible partitioning solutions and have the minimum hardware cost. This observation obviously disagrees with Conclusion 1. This disagreement is actually caused by the fact that Conclu- sion 1 does not consider the communication uncertain varia- tions with different hardware/software partitioning schemes. This effect is actually considered in the algorithm realizations in [1], but not in the theoretical part. Therefore, a possible mod- ification is given in the next section. 3 PROPOSED MODIFICATION Considering the analysis above and the algorithms in [1], we pro- pose to add a rule to the 1D search approach. This will make the proposed approach clear and complete, but less general. To solve problem P , order the task nodes according to hard- ware-to-software ratios: h 1 s 1 h 2 s 2 h n s n : (1) Then the following rule for defining partitioning schemes is given: Rule 1. As for the solutions of problem P , following the method of Alg-greedy in [1], a search starts from the left to the right of (1). The first partitioning scheme (or simply scheme 1) is that node H. Quan, T. Zhang, Q. Liu, and J. Guo are with the School of Electronic and Information Engineering, Tianjin University, Tianjin 300072, China. E-mail: {quanhaojun, zhangtao, qiangliu, jcguo}@tju.edu.cn. X. Wang and R. Hu are with the National Engineering Research Center for Multimedia Software, Wuhan University, Wuhan 430072, China. E-mail: [email protected], [email protected]. Manuscript received 11 Jan. 2012; revised 18 June 2012; accepted 3 July 2012; date of publication 18 July 2012; date of current version 5 Mar. 2014. Recommended for acceptance by E. Macii. For information on obtaining reprints of this article, please send e-mail to: [email protected], and reference the Digital Object Identifier below. Digital Object Identifier no. 10.1109/TC.2012.174 IEEE TRANSACTIONS ON COMPUTERS, VOL. 63, NO. 4, APRIL 2014 1055 0018-9340 ß 2012 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Upload: phungnga

Post on 22-Mar-2017

212 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Comments on “Algorithmic Aspectsof Hardware/Software Partitioning:1D Search Algorithms”

Comments on “Algorithmic Aspectsof Hardware/Software Partitioning:

1D Search Algorithms”

Haojun Quan, Tao Zhang, Qiang Liu, Jichang Guo,Xiaochen Wang, and Ruimin Hu

Abstract—In this paper, the work in [1] is analyzed. An error in its theoretical

description part is pointed out and illustrated by a simple example. A modification

suggestion is proposed to make the theoretical description of the work [1] more

deliberate and thus being used appropriately.

Index Terms—Hardware/software partitioning, 1D search algorithm

Ç

1 INTRODUCTION

Jigang et al. [1] proposed a 1D search approach for hardware/soft-ware partitioning, which reduces the time complexity comparedwith a 2D search method developed in literature. However, whenwe tried to exploit the approach [1], we found an error in the theo-retical description of the paper. The error leads to a mismatchbetween the proposed approach and the algorithm realizations in[1]. As a result, it is desirable to point out the error and make sug-gestions to make the proposed approach [1] more deliberate andnot mislead readers.

Therefore, in this comment, the approach and theorem pro-posed in [1] are briefly described and an error is pointed out.Section 2 shows a simple example to illustrate the mentionederror. After that, Section 3 suggests a possible modification.

As described in [1], given an undirected graph G ¼ðV ;EÞ; V ¼ fv1; v2; . . . ; vng; s; h : V ! Rþ, and c : E ! Rþ. si andhi denote the software cost and the hardware cost of node vi,respectively. Let x ¼ x1; x2; . . . ; xnf g denote a partition scheme, inwhich xi ¼ 1ðxi ¼ 0Þ indicates that node vi is assigned to software(hardware). Thus, the software cost and the hardware cost of x can

be formulated as SðxÞ ¼Pn

i¼1 sixi and HðxÞ ¼Pn

i¼1 hið1� xiÞ,respectively. Meanwhile, the communication cost of x is CðxÞ¼Pn�1

i¼1

Pnj¼iþ1 cijjxi � xjj. Given a constraint constant R, the hard-

ware/software partitioning problem can be formulated as the fol-lowing minimization problem P :

P

minimizePn

i¼1

hið1� xiÞ

subject toPn

i¼1

sixi þ CðxÞ � R:

8>><

>>:

The value ofPn

i¼1 hi is fixed for a certain G, so the solution ofproblem P is identical to the solution of problem Q:

Q

maximizePn

i¼1

hixi

subject toPn

i¼1

sixi þ C xð Þ � R:

8>><

>>:

Then a new variable m 2 ð0; 1Þ for C xð Þ ¼ mR is defined and theproblem Q is transformed to the following problem:

Q0maximize

Pn

i¼1

hixi

subject toPn

i¼1

sixi � ð1� mÞR:

8>><

>>:

Let Q xð Þ ¼Pn

i¼1 hixi, and x�m denote the optimal solution of theproblem Q0 with m. Then Theorem 1 is given in [1].

Theorem 1. Qðx�m1Þ � Qðx�m2Þ for m1 � m2, where x�m1 and x�m2 are theoptimal solutions of the problem Q0 with m1 and m2, respectively.

Then [1] gives the following conclusion:Conclusion 1. Let x�m denote the optimal solution of the prob-

lem Q0 for a given m in (0, 1). The main idea derives from thefact that: if x�m fulfills

Pni¼1 sixi þ

Pn�1i¼1

Pnj¼iþ1 cijjxi � xjj � R; x�m

definitely is a feasible solution of the problem Q. Hence, x�m isqualified to be a candidate partition. Moreover, if x�m that fulfillsPn

i¼1 sixi þPn�1

i¼1

Pnj¼iþ1 cijjxi� xjj � R is found, there is no

need to further search for the optimal solution of Q0 withmþ Dm for any Dm value. This is because, according to Theo-rem 1, there does not exist a better feasible solution than x�m forlarger m values in (0, 1) for the problem Q0.

However, Conclusion 1 is incorrect, because Theorem 1 derivedfrom problem Q0 is applied to problem Q without considering theimpact of communication cost on partitioning schemes. To illus-trate this, an example will be given in the next section.

2 EXAMPLE

As shown in Fig. 1, each node representing a task contains two val-ues in a bracket, where the first value denotes the software costand the second value denotes the hardware cost.

Let R ¼ 7.5. There are eight possible partitioning schemes, asillustrated in Fig. 2. The corresponding analysis is shown inTable 1.

According to Conclusion 1 described in the previous section,searching communication cost from 0 to 3, the optimal schemeshould be scheme 2 (extreme condition) or 3. However, in fact theoptimal scheme is scheme 6 or 7, which are feasible partitioningsolutions and have the minimum hardware cost. This observationobviously disagrees with Conclusion 1.

This disagreement is actually caused by the fact that Conclu-sion 1 does not consider the communication uncertain varia-tions with different hardware/software partitioning schemes.This effect is actually considered in the algorithm realizationsin [1], but not in the theoretical part. Therefore, a possible mod-ification is given in the next section.

3 PROPOSED MODIFICATION

Considering the analysis above and the algorithms in [1], we pro-pose to add a rule to the 1D search approach. This will make theproposed approach clear and complete, but less general.

To solve problem P , order the task nodes according to hard-ware-to-software ratios:

h1

s1� h2

s2� � � � � hn

sn: (1)

Then the following rule for defining partitioning schemes isgiven:

Rule 1. As for the solutions of problem P , following the methodof Alg-greedy in [1], a search starts from the left to the right of (1).The first partitioning scheme (or simply scheme 1) is that node

� H. Quan, T. Zhang, Q. Liu, and J. Guo are with the School of Electronic andInformation Engineering, Tianjin University, Tianjin 300072, China.E-mail: {quanhaojun, zhangtao, qiangliu, jcguo}@tju.edu.cn.

� X. Wang and R. Hu are with the National Engineering Research Center forMultimedia Software, Wuhan University, Wuhan 430072, China.E-mail: [email protected], [email protected].

Manuscript received 11 Jan. 2012; revised 18 June 2012; accepted 3 July 2012;date of publication 18 July 2012; date of current version 5 Mar. 2014.Recommended for acceptance by E. Macii.For information on obtaining reprints of this article, please send e-mail to:[email protected], and reference the Digital Object Identifier below.Digital Object Identifier no. 10.1109/TC.2012.174

IEEE TRANSACTIONS ON COMPUTERS, VOL. 63, NO. 4, APRIL 2014 1055

0018-9340 � 2012 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Page 2: Comments on “Algorithmic Aspectsof Hardware/Software Partitioning:1D Search Algorithms”

v1ðs1; h1Þ is assigned to software and other nodes are assigned tohardware. Scheme 2 is that nodes v1 and v2 are assigned to soft-ware and other nodes are assigned to hardware. Similarly, schemetð1 � t � n� 1Þ means nodes v1 to vt are assigned to software andother nodes are assigned to hardware. Note that only (n� 1) parti-tioning schemes are considered.

Let CðxÞ ¼ mR;m 2 ð0; 1Þ. For a certain value m, the maximumof t can be obtained according to

Pti¼1 si � ð1� mÞR. If the parti-

tioning scheme is a feasible solution of problem P , it is the optimalsolution under CðxÞ ¼ mR, because the search method of Rule 1 isused. Now, let P ðxmÞ ¼

Pni¼tþ1 hi, where P ðxmÞ denotes the total

hardware cost for CðxÞ ¼ mR.

Theorem 1. Under Rule 1, for problem P , if there are two optimal parti-tioning solutions xm1 and xm2 for m1 and m2, respectively, thenP ðxm1Þ � P ðxm2Þ for m1 � m2, where P ðxmÞ denotes the totalhardware cost of the optimal solution of problem P whenCðxÞ ¼ mR.

Proof. Based on the partitioning schemes defined in Rule 1, weuse

Ptm1i¼1 si � ð1� m1ÞR and

Ptm2i¼1 si � ð1� m2ÞR to calculate

the maximum of tm1 and tm2, respectively.

{ m1 � m2

; ð1� m1ÞR � ð1� m2ÞR; tm1 � tm2; where tm1 and tm2 are the corresponding max val-ues. This means when m ¼ m1 more nodes are assigned to soft-ware, compared to m ¼ m2

; P ðxm1Þ � P ðxm2Þ tuAccording to Theorem 1, if an optimal solution of problem P

with a certain m is found, there is no need to further search for theoptimal solution of problem P with mþ Dm for any possible Dm

value. Therefore, under Rule 1, Conclusion 1 is correct, so the algo-rithms proposed in [1] are correct.

ACKNOWLEDGMENTS

This work was supported by the Major National Science and Tech-nology Special Projects (2010ZX03004-003-03). The authors wouldlike to thank the anonymous reviewers for their valuable and con-structive comments.

REFERENCE

[1] W. Jigang, T. Srikanthan, and G. Chen, “Algorithmic Aspects of Hardware/Software Partitioning: 1D Search Algorithms,” IEEE Trans. Computers,vol. 59, no. 4, pp. 532-544, Apr. 2010.

Fig. 2. Possible hardware/software partitioning schemes for the undirected graphshown in Fig. 1.

TABLE 1Analysis of Partitioning Schemes

Fig. 1. A simple undirected graph for hardware/software partitioning.

1056 IEEE TRANSACTIONS ON COMPUTERS, VOL. 63, NO. 4, APRIL 2014