CN102508935A

CN102508935A - On-chip network mapping method based on ant-colony chaos genetic algorithm

Info

Publication number: CN102508935A
Application number: CN2011102831241A
Authority: CN
Inventors: 潘红兵; 易伟; 何书专; 王佳文; 李丽
Original assignee: Nanjing University
Current assignee: Nanjing University
Priority date: 2011-09-22
Filing date: 2011-09-22
Publication date: 2012-06-20

Abstract

An on-chip network mapping method based on ant colony chaotic genetic algorithm. It is based on standard ant colony algorithm and introduces genetic algorithm at the same time. The parameters of each ant are coded with real numbers, and the code is used as the chromosome in the genetic algorithm. In one iteration, genetic algorithm is used to adjust the encoded parameters of ant colony algorithm. During the operation of the algorithm, the present invention also monitors the results of each cycle in the algorithm. If it is detected that the algorithm falls into a local optimal solution, the mutation probability of the genetic algorithm is increased by introducing a chaotic model, and then the The algorithm modifies the parameters of the ant colony algorithm. The invention can effectively improve the search ability of the algorithm for the solution space, avoid stagnation in the local optimal solution, and has good practical value and wide application prospect for solving the large-scale on-chip network mapping problem.

Description

A Network-on-Chip Mapping Method Based on Ant Colony Chaos Genetic Algorithm

技术领域 technical field

本发明涉及片上网络映射方法，具体地说是一种能有效提高映射算法在解空间的搜索效率，避免算法陷入局部最优解的映射方法，为一种基于蚁群混沌遗传算法的片上网络映射方法。The invention relates to an on-chip network mapping method, specifically a mapping method that can effectively improve the search efficiency of the mapping algorithm in the solution space and avoid the algorithm from falling into a local optimal solution, and is an on-chip network mapping based on an ant colony chaotic genetic algorithm method.

背景技术 Background technique

随着半导体工艺技术步入纳米阶段及片上集成度的不断提高，全局连线延时上升至数倍于芯片时钟，传统体系结构已无法满足片上系统(System-on-a-Chip，SoC)通信需求。传统单核处理器芯片虽然可以通过进一步提高集成度的方法来提高性能，然而其开销将过于庞大，各类相关瓶颈问题也越发难以解决。因此采用多个相对简单处理器取代单个处理器的做法被提出，并受到广泛关注。随着片上核数的进一步增加，传统架构，如交叉开关、层次化总线等已经无法满足多核间通信需求，为此针对多核技术的片上网络(Network-on-Chip，NoC)架构被提出，并因为其在可扩展性、可重用性等方面具有无可争议的优势，业已成为解决多核片上系统通信问题最具潜力的方案之一。With the semiconductor process technology entering the nanometer stage and the continuous improvement of on-chip integration, the global connection delay has risen to several times the chip clock, and the traditional architecture has been unable to meet the System-on-a-Chip (SoC) communication. need. Although the performance of traditional single-core processor chips can be improved by further increasing the integration level, the overhead will be too large, and various related bottleneck problems will become increasingly difficult to solve. Therefore, the method of replacing a single processor with multiple relatively simple processors has been proposed and has received widespread attention. With the further increase of the number of cores on a chip, traditional architectures, such as crossbar switches and hierarchical buses, cannot meet the communication requirements between multi-cores. For this reason, the Network-on-Chip (NoC) architecture for multi-core technology was proposed, and Because of its indisputable advantages in scalability, reusability, etc., it has become one of the most potential solutions to solve multi-core system-on-chip communication problems.

在NoC系统的设计过程中，当系统拓扑及路由策略等基本通信架构确定以后，需要面对的首要问题即片上网络映射问题。良好的映射方法可以使得系统整体在能耗、延时、容错性等各方面的性能均有所改善，因此多年来，无论业界还是学界，都对该问题表现出极大的研究热情。In the design process of the NoC system, after the basic communication architecture such as the system topology and routing strategy is determined, the primary problem that needs to be faced is the network on chip mapping problem. A good mapping method can improve the overall performance of the system in terms of energy consumption, delay, and fault tolerance. Therefore, over the years, both the industry and the academic community have shown great research enthusiasm for this issue.

具体而言，映射过程是指在已知NoC体系结构和IP核间通信量的基础上，按某种方法将各IP核分配到NoC中各资源节点上，以实现特定应用与NoC体系结构相互对应的过程，映射结果的优劣则是通过比较目标函数而得出的。以二维网格状NoC为例，NoC映射过程如图1所示。Specifically, the mapping process refers to assigning each IP core to each resource node in the NoC in a certain way on the basis of the known NoC architecture and the traffic between IP cores, so as to realize the interaction between specific applications and the NoC architecture. In the corresponding process, the pros and cons of the mapping results are obtained by comparing the objective functions. Taking a two-dimensional grid-like NoC as an example, the NoC mapping process is shown in Figure 1.

用数学方式准确描述NoC映射过程则需要首先给出两个定义。To accurately describe the NoC mapping process mathematically requires two definitions.

定义1：给定应用特征图G(V，E)为有向非循环加权图，图中顶点vi∈V，表示一个执行特定任务的IP核；有向弧ei，j∈E，表示顶点vi与vj之间的通信关系，其权重wi，j表示vi与vj之间的通信量，bi，j则为vi与vj之间通信带宽要求。Definition 1: A given application feature graph G(V, E) is a directed acyclic weighted graph, in which a vertex vi ∈ V represents an IP core performing a specific task; a directed arc ei, j ∈ E represents a vertex vi For the communication relationship with vj, its weight wi, j represents the communication volume between vi and vj, and bi, j is the communication bandwidth requirement between vi and vj.

定义2：给定NoC结构特征图P(R，P)为有向图。图中顶点ri∈R，表示NoC中的一个资源节点；有向弧pi，j∈P表示从顶点ri到rj的路径；hi，j表示从ri到rj之间的曼哈顿距离；Bi，j为路径pi，j能提供的最大通信带宽。Definition 2: The given NoC structure feature graph P(R, P) is a directed graph. The vertex ri∈R in the figure represents a resource node in the NoC; the directed arc pi, j∈P represents the path from the vertex ri to rj; hi, j represents the Manhattan distance between ri and rj; Bi, j is The maximum communication bandwidth that path pi, j can provide.

则NoC映射过程为：给定G和P，寻找G→P的映射函数map()，要求目标函数尽可能优化，并同时满足如下约束：Then the NoC mapping process is: given G and P, find the mapping function map() of G → P, and require the objective function to be optimized as much as possible, and at the same time satisfy the following constraints:

$&ForAll; &ForAll; {v v}_{i i} &Element; &Element; V V &DoubleRightArrow; &DoubleRightArrow; map map (({v v}_{i i})) &Element; &Element; R R - - - - - - ((11))$

$&ForAll; &ForAll; {v v}_{i i} &NotEqual; &NotEqual; {v v}_{j j} &DoubleLeftRightArrow; &DoubleLeftRightArrow; map map (({v v}_{i i})) &NotEqual; &NotEqual; map map (({v v}_{j j})) - - - - - - ((22))$

size(G)≤size(P) (3)size(G)≤size(P) (3)

$&ForAll; &ForAll; {b b}_{i i,, j j} \leq \leq {B B}_{i i,, j j} - - - - - - ((44))$

其中，式(1)、(2)用以保证IP核与资源节点的一一对应，式(3)、(4)则保证规模和带宽要求。Among them, equations (1) and (2) are used to ensure the one-to-one correspondence between IP cores and resource nodes, and equations (3) and (4) are used to ensure the scale and bandwidth requirements.

发明内容 Contents of the invention

本发明要解决的问题是：为了有效提高NoC系统整体性能，需要增强映射算法的搜索能力。The problem to be solved by the present invention is: in order to effectively improve the overall performance of the NoC system, it is necessary to enhance the search capability of the mapping algorithm.

本发明的技术方案为：一种基于蚁群混沌遗传算法的片上网络映射方法，通过蚁群混沌遗传算法将各IP核分配到片上网络NoC中各资源节点上，实现片上网络映射，所述蚁群混沌遗传算法为：以标准蚁群算法为基础，对每只蚂蚁的参数采用实数编码，并以该编码为遗传算法中的染色体，在蚁群算法每一轮迭代中使用遗传算法对编码过的蚁群算法参数进行调整，即通过遗传算法中的选择及交叉操作对染色体进行调整，更新蚁群算法的参数；同时，对每一轮迭代的结果进行监测，如果监测到蚁群算法陷入局部最优解，则引入混沌模型加大遗传算法的突变概率，进而再通过遗传算法修改蚁群算法参数，直至蚁群算法的最优解满足实际系统芯片设计需要，即NoC的功耗和延时最低，迭代结束，根据最优解完成片上网络映射，所述监测到算法陷入局部最优解是指算法本轮最优解和上一轮最优解相等。The technical solution of the present invention is: a network-on-chip mapping method based on an ant colony chaotic genetic algorithm, which distributes each IP core to each resource node in the network-on-chip NoC through the ant colony chaotic genetic algorithm, and realizes network-on-chip mapping. The swarm chaos genetic algorithm is: based on the standard ant colony algorithm, the parameters of each ant are coded with real numbers, and the code is used as the chromosome in the genetic algorithm. Adjust the parameters of the ant colony algorithm, that is, adjust the chromosomes through the selection and crossover operations in the genetic algorithm, and update the parameters of the ant colony algorithm; at the same time, monitor the results of each round of iteration. To obtain the optimal solution, introduce the chaos model to increase the mutation probability of the genetic algorithm, and then modify the parameters of the ant colony algorithm through the genetic algorithm until the optimal solution of the ant colony algorithm meets the actual system chip design requirements, that is, the power consumption and delay of the NoC At the lowest point, the iteration ends, and the on-chip network mapping is completed according to the optimal solution. The detection of the algorithm falling into a local optimal solution means that the optimal solution of the current round of the algorithm is equal to the optimal solution of the previous round.

具体步骤为：The specific steps are:

1)设置初始化参数并完成对片上系统的初始化过程：设置蚁群算法的最大循环次数，信息启发式因子，期望启发式因子，并设置蚂蚁数，然后将每只蚂蚁放置于各自的起始位置，根据前面的初始化参数生成标准蚁群算法下的初始解；1) Set the initialization parameters and complete the initialization process of the system on chip: set the maximum number of cycles of the ant colony algorithm, the information heuristic factor, the expected heuristic factor, and the number of ants, and then place each ant in its own starting position , generate the initial solution under the standard ant colony algorithm according to the previous initialization parameters;

2)构造迭代解：在构造解的过程中，假设第k只蚂蚁在第t次循环时以概率将IP核Pi分配到资源节点Rj上：2) Constructing an iterative solution: In the process of constructing the solution, it is assumed that the k-th ant is in the t-th cycle with probability Assign the IP core Pi to the resource node Rj:

${p p}_{i i,, j j}^{k k} ((t t)) = = \{\begin{matrix} \frac{{[[{τ τ}_{i i,, j j} ((t t))]]}^{α α} \times \times {[[{η η}_{i i,, j j} ((t t))]]}^{β β}}{\underset{j j &NotElement; &NotElement; {tabu tabu}_{k k}}{Σ Σ} {[[{τ τ}_{i i,, j j} ((t t))]]}^{α α} \times \times {[[{η η}_{i i,, j j} ((t t))]]}^{β β}},, j j &NotElement; &NotElement; {tabu tabu}_{k k} \\ 00,, j j &Element; &Element; {tabu tabu}_{k k} \end{matrix} - - - - - - ((55))$

集合tabu_k(k＝1，2，......，M)，M指IP核的总数，tabu_k用于记录蚂蚁k已经分配过的IP核，τ_i，j(t)表示在第t次循环时，将IP核Pi分配到资源节点Rj上的信息素强度，η_i，j(t)是指IP核Pi分配到资源节点Rj上的启发信息，式中α，β分别为信息启发式因子与期望启发式因子，α，β在第一次迭代时由步骤1)设置初值；Set tabu _k (k=1, 2, ..., M), M refers to the total number of IP cores, tabu _k is used to record the IP cores that ant k has allocated, τ _{i, j} (t) is expressed in During the tth cycle, assign the IP core Pi to the pheromone intensity on the resource node Rj, η _{i, j} (t) refers to the heuristic information that the IP core Pi is assigned to the resource node Rj, where α and β are respectively Information heuristic factor and expected heuristic factor, α, β are initially set by step 1) in the first iteration;

3)信息素更新：用参数ρ表示信息素的持久程度(0＜ρ＜1)，Δτ_i，j为信息素增量：3) Pheromone update: the parameter ρ is used to indicate the persistence degree of the pheromone (0<ρ<1), and Δτ _{i, j} is the pheromone increment:

$Δ Δ {τ τ}_{i i,, j j} = = {Σ Σ}_{k k = = 11}^{M m} Δ Δ {τ τ}_{i i,, j j}^{k k} - - - - - - ((99))$

表示本次循环中蚂蚁k在分配路径(Pi→Rj)上留下的信息量，计算公式为： Indicates the amount of information left by ant k on the distribution path (Pi→Rj) in this cycle, and the calculation formula is:

$Δ Δ {τ τ}_{i i,, j j}^{k k} = = \{\begin{matrix} \frac{11}{cos cos t t ((k k))},, map map ((k k)) includes includes (({P P}_{i i} &RightArrow; &Right Arrow; {R R}_{j j})) \\ 00,, else else \end{matrix} - - - - - - ((1010))$

式(10)中，cost(k)为蚂蚁k根据步骤2)得到的分配方案的成本，最优解为具有最小成本的解，所述cost函数的定义根据不同的目标需求而有所不同，使得最优解对信息量的贡献最大，当针对目标函数最大化的优化问题时，则式(10)中的1/cost(k)变化为cost(k)，此时最优解为具有最大成本的解，在所有蚂蚁完成一次循环以后，根据下式对各分配路径(Pi→Rj)上的信息量作更新：In formula (10), cost(k) is the cost of the allocation plan obtained by ant k according to step 2), the optimal solution is the solution with the minimum cost, and the definition of the cost function varies according to different target requirements, Make the optimal solution contribute the most to the amount of information. When aiming at the optimization problem of maximizing the objective function, the 1/cost(k) in formula (10) changes to cost(k). At this time, the optimal solution has the maximum For the cost solution, after all ants complete a cycle, update the amount of information on each distribution path (Pi→Rj) according to the following formula:

τ_i，j(t+1)＝ρ×τ_i，j(t)+Δτ_i，j (11)τ _i,j (t+1)=ρ×τ _i,j (t)+Δτ _i,j (11)

4)判断最优解：蚁群算法一次循环中迭代解的求解为：4) Judgment of the optimal solution: The solution to the iterative solution in one cycle of the ant colony algorithm is:

41)从IP核集合P中按概率

选择一个未分配的IP核Pi分配到Rj上，并将该核添加到tabu_k中；41) From the IP core set P according to the probability

Select an unassigned IP core Pi to assign to Rj, and add the core to tabu _k ;

42)重复执行N步，直到所有的IP核都分配到相应的资源上，tabu_k满；42) Repeat N steps until all IP cores are allocated to corresponding resources and the tabu _k is full;

一次蚁群迭代循环完成后，在所有的蚂蚁中选择最优解，如果最优解满足条件，迭代结束，如果不满足，则进行步骤5)进入下一次循环迭代；After an ant colony iteration cycle is completed, select the optimal solution among all ants, if the optimal solution satisfies the condition, the iteration ends, if not, proceed to step 5) to enter the next cycle iteration;

5)用遗传算法更新蚁群算法参数，并混沌算法避免陷入局部最优解：当蚁群算法得到的解不能满足要求时，使用遗传算法更新蚁群算法的参数，使用实数α，β，Q对各个蚂蚁进行基因编码，每个蚂蚁的染色体即用(α，β，Q)表示：5) Use the genetic algorithm to update the parameters of the ant colony algorithm, and the chaotic algorithm to avoid falling into the local optimal solution: when the solution obtained by the ant colony algorithm cannot meet the requirements, use the genetic algorithm to update the parameters of the ant colony algorithm, and use the real numbers α, β, Q Genetic coding is carried out for each ant, and the chromosome of each ant is represented by (α, β, Q):

α＝x×α_f+(1-x)×α_m (12)α＝x×α _f +(1-x)×α _m (12)

β＝x×β_f+(1-x)×β_m (13)β=x×β _f +(1-x)×β _m (13)

Q＝x×Q_f+(1-x)×Q_m (14)Q＝x×Q _f +(1-x)×Q _m (14)

通过轮盘概率实现优胜劣汰，选择两个种群，将它们的参数(α，β，Q)按比例进行杂交，然后再按照一定的概率对它们的参数进行变异，这里的变异概率由混沌模型来进行调整，将结果保存，重复执行M/2次后，所有蚂蚁的参数都被更新，然后进入下一轮蚁群算法的迭代过程，即回到步骤1)；Survival of the fittest is achieved through roulette probability. Two populations are selected, and their parameters (α, β, Q) are crossed in proportion, and then their parameters are mutated according to a certain probability. The mutation probability here is performed by the chaotic model. Adjust, save the result, repeat the execution M/2 times, all the parameters of the ants are updated, and then enter the iterative process of the next round of ant colony algorithm, that is, return to step 1);

所述杂交比例系数x的初始值设为0.5，当检测到算法陷入局部最优的时候，采用混沌模型对x值进行调整：The initial value of the hybridization ratio coefficient x is set to 0.5, and when it is detected that the algorithm falls into a local optimum, the chaotic model is used to adjust the value of x:

${cx cx}_{n no}^{m m + + 11} = = 44 {cx cx}_{n no}^{m m} ((11 - - {x x}_{n no}^{m m})) - - - - - - ((1515))$

式中，

表示m次迭代得到的第n个混沌变量，

且

按照式(15)对遗传算法的杂交比例系数x进行更新。In the formula,

Indicates the nth chaotic variable obtained by m iterations,

and

According to formula (15), the hybridization ratio coefficient x of the genetic algorithm is updated.

对每只蚂蚁的参数编码包括：对信息启发式因子，期望启发式因子以及信息素强度进行编码。The parameter coding of each ant includes: coding information heuristic factor, expectation heuristic factor and pheromone intensity.

本发明以标准蚁群算法为基础，同时引入遗传算法，对每只蚂蚁的参数采用实数编码，并以该编码为遗传算法中的染色体，在每一轮迭代中使用遗传算法对编码过的蚁群算法参数进行调整。在算法运行过程中，本发明还对算法中每一轮循环的结果进行监测，如果监测到算法陷入局部最优解，则通过引入混沌模型的方法加大遗传算法的突变概率，进而再通过遗传算法修改蚁群算法参数。本发明可以有效提高算法对解空间的搜索能力，避免其停滞于局部最优解，对于大规模片上网络映射问题的求解有着良好的实用价值和广泛的应用前景。The present invention is based on the standard ant colony algorithm, introduces the genetic algorithm at the same time, adopts real number codes for the parameters of each ant, and uses the codes as the chromosomes in the genetic algorithm, uses the genetic algorithm in each round of iteration to code the coded ant Adjust the parameters of the group algorithm. During the operation of the algorithm, the present invention also monitors the results of each cycle in the algorithm. If it is detected that the algorithm falls into a local optimal solution, the mutation probability of the genetic algorithm is increased by introducing a chaotic model, and then the The algorithm modifies the parameters of the ant colony algorithm. The invention can effectively improve the search ability of the algorithm for the solution space, avoid stagnation in the local optimal solution, and has good practical value and wide application prospect for solving the large-scale on-chip network mapping problem.

附图说明 Description of drawings

图1是NoC映射过程示意图。Figure 1 is a schematic diagram of the NoC mapping process.

图2是本发明蚁群混沌遗传算法流程示意图。Fig. 2 is a flow diagram of the ant colony chaotic genetic algorithm of the present invention.

图3是MPEG4解码器任务分解示意图。Fig. 3 is a schematic diagram of MPEG4 decoder task decomposition.

图4是MPEG4解码器不同映射结果示意图。Fig. 4 is a schematic diagram of different mapping results of an MPEG4 decoder.

图5是本发明对MPEG4解码器映射结果示意图。Fig. 5 is a schematic diagram of the mapping result of the MPEG4 decoder according to the present invention.

图6是本发明与标准蚁群算法在随机任务图下映射结果比较示意图。Fig. 6 is a schematic diagram of the comparison between the present invention and the standard ant colony algorithm in the mapping results under the random task graph.

具体实施方式 Detailed ways

基于遗传的蚁群算法主要是为了解决传统蚁群算法对初始化参数的依赖，通过选择合适种群的参数进行杂交，并以一定的概率变异从而不断对参数进行更新。为了避免算法陷入局部最优，还引入混沌模型，加大突变概率。本发明的主要流程如图2所示。The genetic-based ant colony algorithm is mainly to solve the dependence of the traditional ant colony algorithm on the initialization parameters, by selecting the parameters of the appropriate population for hybridization, and mutating with a certain probability to continuously update the parameters. In order to prevent the algorithm from falling into local optimum, a chaotic model is also introduced to increase the mutation probability. The main flow of the present invention is shown in Figure 2.

设置初始化参数并完成对系统的初始化过程。具体为设置最大循环次数，信息启发式因子，期望启发式因子，并设置蚂蚁数，然后将每只蚂蚁放置于各自的起始位置，进而形成算法最初的解。Set initialization parameters and complete the initialization process of the system. Specifically, the maximum number of cycles, the information heuristic factor, the expected heuristic factor, and the number of ants are set, and then each ant is placed in its own starting position to form the initial solution of the algorithm.

接下来是对新一轮解的构造过程。在构造解的过程中，假设第k只蚂蚁在第t次循环时以概率

将IP核i分配到资源节点j上。Next is the process of constructing a new round of solutions. In the process of constructing the solution, it is assumed that the k-th ant is in the t-th cycle with probability

Assign IP core i to resource node j.

tabu_k(k＝1，2，......，M)用于记录蚂蚁k已经分配过的IP核，M指IP核的总数。τ_i，j(t)表示在第t次循环时，将IP核Pi分配到资源节点Rj上的信息素强度，信息素的更新将在后面详细介绍。η_i，j(t)是指IP核Pi分配到资源节点Rj上的启发信息。式中α，β分别为信息启发式因子与期望启发式因子。α，β在第一次迭代时由初始化系统设置初值，以后在蚁群算法每执行一次迭代后由遗传算法更新。tabu _k (k=1, 2, . . . , M) is used to record the IP cores allocated by ant k, and M refers to the total number of IP cores. τ _{i, j} (t) represents the pheromone strength of the IP core Pi assigned to the resource node Rj in the t-th cycle, and the update of pheromone will be introduced in detail later. η _i,j (t) refers to the heuristic information assigned by the IP core Pi to the resource node Rj. where α and β are information heuristic factor and expectation heuristic factor respectively. The initial values of α and β are set by the initialization system in the first iteration, and are updated by the genetic algorithm after each iteration of the ant colony algorithm.

一个好的映射总是希望把最重要的IP核分配到通信能力最强的资源节点上。因此，定义启发信息如下所示：A good mapping always hopes to allocate the most important IP cores to the resource nodes with the strongest communication capabilities. Therefore, the definition heuristic looks like this:

η_i，j(t)＝VIP(i)/Comm(j) (6)η _i,j (t) = VIP(i)/Comm(j) (6)

η_i，j(t)体现出将IP核Pi分配给资源节点Rj的合理程度。Comm(j)表示Rj在NoC中的通信能力，VIP(i)代表Pi在应用特征图中的重要程度。η _i,j (t) reflects the rationality of allocating IP core Pi to resource node Rj. Comm(j) represents the communication capability of Rj in the NoC, and VIP(i) represents the importance of Pi in the application feature map.

$Comm Comm ((j j)) = = {Σ Σ}_{i i = = 11}^{N N} {h h}_{i i,, j j} - - - - - - ((77))$

$VIP VIP ((i i)) = = {Σ Σ}_{j j = = 11}^{N N} {w w}_{i i,, j j} + + {Σ Σ}_{j j = = 11}^{N N} {w w}_{j j,, i i} - - - - - - ((88))$

式中w_i，j和w_j，i为通信量，N为节点数，h_i，j表示从节点ri到节点rj经过的跳数(hop)。In the formula, w _{i, j} and w _{j, i} are traffic, N is the number of nodes, h _{i, j} represents the number of hops (hops) from node ri to node rj.

求解步骤如下：The solution steps are as follows:

1)从IP核集合P中按概率

选择一个未分配的IP核Pi分配到资源节点Rj上，并将该核添加到tabu_k中；1) From the IP core set P according to the probability

Select an unallocated IP core Pi to allocate to resource node Rj, and add the core to tabu _k ;

2)重复执行N步，直到所有的核都分配到相应的资源上，tabu_k满。2) Repeat N steps until all cores are allocated to corresponding resources and tabu _k is full.

一次蚁群迭代循环完成后，在所有的蚂蚁中选择最优解，然后进入下一次循环迭代。After an ant colony iteration cycle is completed, select the optimal solution among all ants, and then enter the next cycle iteration.

接下来是信息素更新过程。随着程序运行，以前留下的信息素逐渐消逝，同时在每只蚂蚁选择的过程中又有新的信息素加入。用参数ρ表示信息素的持久程度(0＜ρ＜1)，Δτ_i，j为信息素增量。Next is the pheromone update process. As the program runs, the pheromones left before gradually disappear, and at the same time, new pheromones are added during the selection process of each ant. The parameter ρ is used to represent the persistence degree of the pheromone (0<ρ<1), and Δτ _i,j is the increment of the pheromone.

表示本次循环中蚂蚁k在分配路径(Pj→Ri)上留下的信息量，计算公式为：

Indicates the amount of information left by ant k on the distribution path (Pj→Ri) in this cycle, and the calculation formula is:

式中，cost(k)为蚂蚁k根据上面步骤完成的分配方案的成本。最优解具有最小的成本，所以对信息量的贡献最大。至于cost函数的定义，根据不同的目标需求，其定义也会有所不同。当然，如果是针对目标函数最大化的优化问题，则式(10)中的1/cost(k)，应当变化为cost(k)，因为此时最优解为具有最大成本的解，所以应当做此调整以使得其对信息量的贡献最大。在所有蚂蚁完成一次循环以后，根据下式对各分配路径(Pi→Rj)上的信息量作更新。In the formula, cost(k) is the cost of the allocation plan completed by ant k according to the above steps. The optimal solution has the smallest cost, so it contributes the most to the amount of information. As for the definition of the cost function, its definition will be different according to different target requirements. Of course, if it is an optimization problem aimed at maximizing the objective function, then 1/cost(k) in formula (10) should be changed to cost(k), because the optimal solution at this time is the solution with the maximum cost, so it should be This adjustment is made to maximize its contribution to the amount of information. After all ants complete a cycle, update the amount of information on each distribution path (Pi→Rj) according to the following formula.

然后是用遗传算法更新蚁群算法的步骤，也是本发明不同于传统蚁群算法的最关键的步骤之一，对每只蚂蚁的参数采用实数编码，并以该编码为遗传算法中的染色体，在蚁群算法每一轮迭代中使用遗传算法对编码过的蚁群算法参数进行调整，即通过遗传算法中的选择及交叉操作对染色体进行调整，以解决蚁群算法本身对于参数设置过于敏感的问题。当蚁群算法得到的解不能满足要求时，使用遗传算法更新蚁群算法的参数，可以保证蚁群算法的进行。为了避免遗传算法复杂的编码解码过程，这里使用实数α，β，Q对各个蚂蚁进行基因编码。那么每个蚂蚁的染色体即用(α，β，Q)表示。Then be the step of updating the ant colony algorithm with the genetic algorithm, which is also one of the most critical steps that the present invention is different from the traditional ant colony algorithm. The parameters of each ant are coded with real numbers, and the code is used as the chromosome in the genetic algorithm, In each iteration of the ant colony algorithm, the genetic algorithm is used to adjust the parameters of the encoded ant colony algorithm, that is, the chromosomes are adjusted through the selection and crossover operations in the genetic algorithm to solve the problem that the ant colony algorithm itself is too sensitive to parameter settings. question. When the solution obtained by the ant colony algorithm cannot meet the requirements, the genetic algorithm is used to update the parameters of the ant colony algorithm, which can ensure the progress of the ant colony algorithm. In order to avoid the complex encoding and decoding process of the genetic algorithm, the real numbers α, β, Q are used to encode the genes of each ant. Then the chromosome of each ant is represented by (α, β, Q).

α＝x×α_f+(1-x)×α_m (12)α＝x×α _f +(1-x)×α _m (12)

β＝x×β_f+(1-x)×β_m (13)β=x×β _f +(1-x)×β _m (13)

Q＝x×Q_f+(1-x)×Q_m (14)Q＝x×Q _f +(1-x)×Q _m (14)

具体而言，是通过轮盘概率(Roulette)来实现优胜劣汰的。选择两个合适的种群，将它们的参数(α，β，Q)按照一定的比例进行杂交(Crossover)。然后再按照一定的概率对它们的参数进行变异(Mutate)，将结果保存。重复执行M/2次后，所有蚂蚁的参数都被更新，然后进入下一轮蚁群算法的迭代过程。Specifically, the survival of the fittest is achieved through Roulette. Select two suitable populations, and crossover their parameters (α, β, Q) according to a certain ratio. Then mutate their parameters according to a certain probability, and save the result. After repeating M/2 times, the parameters of all ants are updated, and then enter the next round of ant colony algorithm iteration process.

式(12)-(14)中，杂交比例系数x的初始值均设置为0.5。当然，随着程序的运行，当检测到算法陷入局部最优的时候，将根据接下来的步骤采用混沌模型对x值进行调整，具体而言：In formulas (12)-(14), the initial value of the hybridization ratio coefficient x is set to 0.5. Of course, as the program runs, when it is detected that the algorithm falls into a local optimum, the value of x will be adjusted according to the next steps using the chaotic model, specifically:

当本次的最优解与上一次的最优解相等时，意味着可能陷入了局部最优，为了避免算法停滞，利用混沌对初值变化极度敏感的特性，将此时的系数进行混沌操作如下所示。When the optimal solution of this time is equal to the optimal solution of the last time, it means that it may fall into a local optimum. In order to avoid algorithm stagnation, the coefficients at this time are subjected to chaotic operations by using the characteristics of chaos that is extremely sensitive to initial value changes. As follows.

式中，表示m次迭代得到的第n个混沌变量，

且

按照式(15)对遗传算法的杂交比例系数x进行更新。In the formula, Indicates the nth chaotic variable obtained by m iterations,

and

下面通过具体实施例说明本发明的实施。The implementation of the present invention is illustrated below by specific examples.

实施例1Example 1

为验证本发明，将算法应用在MPEG4解码器的映射问题上。MPEG4解码器可以分解为12个任务，如图3所示，然后把它们交给12个IP核分别执行，此时的映射问题即如何将这12个IP核放置到一个规模为3x4的NoC上去。To verify the invention, the algorithm is applied to the mapping problem of MPEG4 decoder. The MPEG4 decoder can be decomposed into 12 tasks, as shown in Figure 3, and then handed over to 12 IP cores for execution respectively. The mapping problem at this time is how to place these 12 IP cores on a NoC with a scale of 3x4 .

在正式应用本发明解决该问题之前，我们需要先定义目标函数cost的数学表达式。Before formally applying the present invention to solve this problem, we need to define the mathematical expression of the objective function cost.

首先，系统能耗定义为：First, the system energy consumption is defined as:

$E E. ((C C)) = = {Σ Σ}_{i i = = 11}^{N N} {Σ Σ}_{j j = = 11}^{N N} {w w}_{i i,, j j} \times \times {h h}_{i i,, j j} - - - - - - ((1616))$

式中w_i，j为通信量，N为节点数，h_i，j表示从节点ri到节点rj经过的跳数(hop)。为了简化设计模型，本文假设h_i，j为两节点之间的曼哈顿距离(|x_ri-x_rj|+|y_ri-y_rj|)。所以优化能耗的目标是最小化各节点间的加权曼哈顿距离之和，就是将通信任务较重的几个IP核分配到紧邻的节点，实现通信近邻化。In the formula, w _{i, j} is the traffic, N is the number of nodes, h _{i, j} represents the number of hops (hops) from node ri to node rj. In order to simplify the design model, this paper assumes that h _{i, j} is the Manhattan distance between two nodes (|x _ri -x _rj |+|y _ri -y _rj |). Therefore, the goal of optimizing energy consumption is to minimize the sum of the weighted Manhattan distances between nodes, that is, to allocate several IP cores with heavy communication tasks to adjacent nodes to realize communication proximity.

接下来考虑延时影响，出从源节点i到目标节点j的延时如式(17)所示：Next, considering the effect of delay, the delay from source node i to target node j is shown in formula (17):

T_i，j＝(T_b+T_w)×h_i，j+T_b(B-1) (17)T _{i, j} = (T _b + T _w ) × h _{i, j} + T _b (B-1) (17)

式中，Tb为无拥塞时一帧数据通过一个开关和一条链路所需要的时间，Tw为存在拥塞时包头在开关节点处平均等待时间，B为数据包中包含的帧数。其中Tb和B为常系数。由(17)式可以看出，Ti，j依赖于参数Tw和h_i，j。在带宽约束下，减小数据的传输延时Tw可以通过缓解数据拥塞实现，而缓解拥塞的关键就在于平衡链路负载。平衡链路负载就是最小化链路负载方差。所以优化延时以链路负载方差(VAR(L))作为指标：In the formula, Tb is the time required for a frame of data to pass through a switch and a link when there is no congestion, Tw is the average waiting time of the packet header at the switch node when there is congestion, and B is the number of frames contained in the data packet. Where Tb and B are constant coefficients. It can be seen from formula (17) that Ti, j depends on parameters Tw and h _{i, j} . Under bandwidth constraints, reducing data transmission delay Tw can be achieved by alleviating data congestion, and the key to alleviating congestion is to balance link loads. To balance link load is to minimize link load variance. Therefore, the optimized delay takes the link load variance (VAR(L)) as an indicator:

$VAR VAR ((L L)) = = {Σ Σ}_{i i = = 11}^{M m} {[[Load load (({l l}_{i i})) - - Load load (({l l}_{avg avg}))]]}^{22} / / M m - - - - - - ((1818))$

式中M为链路总数，Load(li)为链路li的负载量，Load(lavg)为平均链路负载量。显然，通信延时优化就是将通信任务均匀分布。In the formula, M is the total number of links, Load(li) is the load of link li, and Load(lavg) is the average link load. Obviously, communication delay optimization is to distribute communication tasks evenly.

从(16)、(18)式可以看出，我们一方面希望优化系统能耗E(C)，另一方面希望平衡链路负载VAR(L)。为了达到联合优化的目的，我们定义目标函数定义如下：It can be seen from (16) and (18) that we hope to optimize the system energy consumption E(C) on the one hand, and balance the link load VAR(L) on the other hand. In order to achieve the purpose of joint optimization, we define the objective function as follows:

cost＝λ×E(C)+(1-λ)VAR(L) (19)cost＝λ×E(C)+(1-λ)VAR(L) (19)

式中，λ是比例系数，用于调节通信能耗和延时在成本函数中的比重，取值范围为(0，1)。当λ＝1时，优化通信能耗；λ＝0时，优化通信延时。在实际使用中需要通过具体需要调整λ值，比如在通信能耗的优化更重要时，λ取值(0.5，1]；通信延时更重要时，λ取值[0，0.5)，λ的具体取值根据实际需要进行调整。In the formula, λ is a proportional coefficient, which is used to adjust the proportion of communication energy consumption and delay in the cost function, and the value range is (0, 1). When λ=1, communication energy consumption is optimized; when λ=0, communication delay is optimized. In actual use, the value of λ needs to be adjusted according to specific needs. For example, when the optimization of communication energy consumption is more important, the value of λ is (0.5, 1]; when the communication delay is more important, the value of λ is [0, 0.5), and the value of λ is The specific value is adjusted according to actual needs.

最终试验结果如图4所示。从图4可以看出，当λ＝0.5时，对通信能耗的优化明显弱于λ＝0时的通信能耗优化，而对链路负载的优化弱于λ＝1时的链路负载优化。λ＝0.5的优化在于联合考虑通信能耗和链路负载平衡。也就是说可以通过对λ值进行调整，在满足通信能耗要求的情况下，优化负载平衡。或者是在满足带宽要求的情况下，优化通信能耗。图5为λ＝0.5时的映射结果。和随机放置相比，在通信能耗上降低了31％，链路负载上优化了56％。The final test results are shown in Figure 4. It can be seen from Figure 4 that when λ=0.5, the optimization of communication energy consumption is obviously weaker than that of communication energy consumption when λ=0, and the optimization of link load is weaker than that of link load optimization when λ=1 . The optimization of λ=0.5 is to jointly consider communication energy consumption and link load balance. That is to say, by adjusting the lambda value, the load balance can be optimized under the condition of meeting the communication energy consumption requirement. Or optimize communication energy consumption while meeting bandwidth requirements. Fig. 5 is the mapping result when λ=0.5. Compared with random placement, the communication energy consumption is reduced by 31%, and the link load is optimized by 56%.

实施例2Example 2

为充分体现本发明的优势，生成了一系列的随机任务图，并侧重不同的优化目标(λ＝1，0.5，0)对NoC进行了映射优化，然后其将映射结果同标准蚁群算法进行了比较。本实施例中，目标函数的定义仍然采用实施例1中的(19)式。最终结果如图6所示。图6显示了λ＝1，0.5，0时各映射方案的通信能耗和链路负载方差与标准蚁群算法的对比结果。从图可以看出本发明明显优于传统蚁群算法。λ＝1时映射方案比参考方案cost降低11％，λ＝0.5时映射方案比参考方案降低4％，λ＝0时映射方案比参考方案降低1％。In order to fully reflect the advantages of the present invention, a series of random task graphs are generated, and the NoC is mapped and optimized with emphasis on different optimization objectives (λ=1, 0.5, 0), and then the mapping results are carried out with the standard ant colony algorithm compared. In this embodiment, the definition of the objective function still adopts the formula (19) in Embodiment 1. The final result is shown in Figure 6. Figure 6 shows the comparison results of the communication energy consumption and link load variance of each mapping scheme with the standard ant colony algorithm when λ=1, 0.5, and 0. It can be seen from the figure that the present invention is obviously superior to the traditional ant colony algorithm. When λ=1, the cost of the mapping scheme is 11% lower than that of the reference scheme; when λ=0.5, the cost of the mapping scheme is 4% lower than that of the reference scheme; when λ=0, the cost of the mapping scheme is 1% lower than that of the reference scheme.

本发明可以有效提高映射算法对解空间的搜索能力，避免其陷入局部最优解，对改善片上网络系统整体性能，降低通信开销，减少通信延时，有着积极良好的应用价值。The invention can effectively improve the search ability of the mapping algorithm for the solution space, prevent it from falling into a local optimal solution, improve the overall performance of the on-chip network system, reduce communication overhead, and reduce communication delay, and has positive and good application value.

Claims

1. a network-on-chip mapping method based on ant colony chaos genetic algorithm, it is characterized in that each IP core is assigned to each resource node in network-on-chip NoC by ant colony chaos genetic algorithm, realizes network-on-chip mapping, said ant colony chaos The genetic algorithm is: based on the standard ant colony algorithm, the parameters of each ant are coded with real numbers, and the code is used as the chromosome in the genetic algorithm. Adjust the parameters of the swarm algorithm, that is, adjust the chromosomes through the selection and crossover operations in the genetic algorithm, and update the parameters of the ant colony algorithm; at the same time, monitor the results of each round of iterations. solution, introduce the chaos model to increase the mutation probability of the genetic algorithm, and then modify the parameters of the ant colony algorithm through the genetic algorithm until the optimal solution of the ant colony algorithm meets the actual system chip design requirements, that is, the power consumption and delay of the NoC are the lowest, At the end of the iteration, the network-on-chip mapping is completed according to the optimal solution. The detection of the algorithm falling into a local optimal solution means that the optimal solution of the current round of the algorithm is equal to the optimal solution of the previous round.

2. a kind of network-on-chip mapping method based on ant colony chaos genetic algorithm according to claim 1, is characterized in that concrete steps are:

1) Set the initialization parameters and complete the initialization process of the system on chip: set the maximum number of cycles of the ant colony algorithm, the information heuristic factor, the expected heuristic factor, and the number of ants, and then place each ant in its own starting position , generate the initial solution under the standard ant colony algorithm according to the previous initialization parameters;

2) Constructing an iterative solution: In the process of constructing the solution, it is assumed that the k-th ant is in the t-th cycle with probability

Assign the IP core Pi to the resource node Rj:

{p p}_{i i,, j j}^{k k} ((t t)) = = \{\begin{matrix} \frac{{[[{τ τ}_{i i,, j j} ((t t))]]}^{α α} \times \times {[[{η η}_{i i,, j j} ((t t))]]}^{β β}}{\underset{j j &NotElement; &NotElement; {tabu tabu}_{k k}}{Σ Σ} {[[{τ τ}_{i i,, j j} ((t t))]]}^{α α} \times \times {[[{η η}_{i i,, j j} ((t t))]]}^{β β}},, j j &NotElement; &NotElement; {tabu tabu}_{k k} \\ 00,, j j &Element; &Element; {tabu tabu}_{k k} \end{matrix} - - - - - - ((55))

Set tabu _k (k=1, 2, ..., M), M refers to the total number of IP cores, tabu _k is used to record the IP cores that ant k has allocated, τ _{i, j} (t) is expressed in During the tth cycle, assign the IP core Pi to the pheromone intensity on the resource node Rj, η _{i, j} (t) refers to the heuristic information that the IP core Pi is assigned to the resource node Rj, where α and β are respectively Information heuristic factor and expected heuristic factor, α, β are initially set by step 1) in the first iteration;

3) Pheromone update: the parameter ρ is used to indicate the persistence degree of the pheromone (0<ρ<1), and Δτ _{i, j} is the pheromone increment:

Δ Δ {τ τ}_{i i,, j j} = = {Σ Σ}_{k k = = 11}^{M m} Δ Δ {τ τ}_{i i,, j j}^{k k} - - - - - - ((99))

Indicates the amount of information left by ant k on the distribution path (Pi→Rj) in this cycle, and the calculation formula is:

Δ Δ {τ τ}_{i i,, j j}^{k k} = = \{\begin{matrix} \frac{11}{cos cos t t ((k k))},, map map ((k k)) includes includes (({P P}_{i i} &RightArrow; &Right Arrow; {R R}_{j j})) \\ 00,, else else \end{matrix} - - - - - - ((1010))

In formula (10), cost(k) is the cost of the allocation scheme obtained by ant k according to step 2), the optimal solution is the solution with the minimum cost, and the definition of the cost function varies according to different target requirements, Make the optimal solution contribute the most to the amount of information. When aiming at the optimization problem of maximizing the objective function, the 1/cost(k) in the formula (10) changes to cost(k). At this time, the optimal solution has the maximum For the solution of cost, after all ants complete a cycle, update the amount of information on each distribution path (Pi→Rj) according to the following formula:

τ _i,j (t+1)=ρ×τ _i,j (t)+Δτ _i,j (11)

4) Judgment of the optimal solution: The solution to the iterative solution in one cycle of the ant colony algorithm is:

41) From the IP core set P according to the probability

Select an unassigned IP core Pi to assign to Rj, and add the core to tabu _k ;

42) Repeat N steps until all IP cores are allocated to corresponding resources and the tabu _k is full;

After an ant colony iterative cycle is completed, select the optimal solution among all ants, if the optimal solution satisfies the condition, the iteration ends, if not, proceed to step 5) to enter the next cycle iteration;

5) Use the genetic algorithm to update the parameters of the ant colony algorithm, and the chaotic algorithm to avoid falling into the local optimal solution: when the solution obtained by the ant colony algorithm cannot meet the requirements, use the genetic algorithm to update the parameters of the ant colony algorithm, and use the real numbers α, β, Q Genetic coding is carried out for each ant, and the chromosome of each ant is represented by (α, β, Q):

α＝x×α _f +(1-x)×α _m (12)

β=x×β _f +(1-x)×β _m (13)

Q＝x×Q _f +(1-x)×Q _m (14)

Survival of the fittest is achieved through roulette probability. Two populations are selected, and their parameters (α, β, Q) are crossed in proportion, and then their parameters are mutated according to a certain probability. The mutation probability here is carried out by the chaotic model. Adjust, save the result, repeat the execution M/2 times, all the parameters of the ants are updated, and then enter the iterative process of the next round of ant colony algorithm, that is, return to step 1);

The initial value of the hybridization ratio coefficient x is set to 0.5, and when it is detected that the algorithm falls into a local optimum, the chaotic model is used to adjust the value of x:

{cx cx}_{n no}^{m m + + 11} = = 44 {cx cx}_{n no}^{m m} ((11 - - {x x}_{n no}^{m m})) - - - - - - ((1515))

In the formula,

Indicates the nth chaotic variable obtained by m iterations,

and

3. a kind of on-chip network mapping method based on ant colony chaotic genetic algorithm according to claim 1 or 2, is characterized in that the parameter coding to each ant comprises: to information heuristic factor, expectation heuristic factor and pheromone Intensity is encoded.