CN101799838B - Method for seeking mapping scheme between tasks and nodes of network on chip - Google Patents

Abstract

The present invention proposes a method for finding a mapping scheme between tasks and nodes of an on-chip network. The process of the method is as follows: initialize a scheme group J _t including K allocation schemes, and then obtain _each On-chip network energy consumption, response time, and fitness value under the scheme, and then perform priority sorting, grouping, switching and other processes according to the size of the appropriate value to finally obtain the optimal scheme. The present invention separates energy consumption and response time as the priority of the assessment scheme In the process of finding the optimal solution, energy consumption and response time are considered at the same time, so that the solution found can minimize energy consumption and system delay at the same time, so that the optimal task and node-to-node mapping scheme can be designed The performance of the network reaches the best, and compared with other methods for finding the priority scheme of mapping between tasks and nodes under multiple objective functions, it has the characteristics of complexity and simple implementation, and the method convergence speed of the present invention has also been confirmed by simulation quick.

Description

Method for finding task-to-node mapping schemes for a network-on-chip

Field

This patent belongs to the technical field of integrated circuit design, especially for the search process of the multi-task and multi-node mapping scheme of the network on chip.

technical background

With the increasing computational complexity of communication terminals and equipment in the future, the demand for integration scale of real-time complex system chips will increase rapidly. Dozens or hundreds of processing units may be integrated on one chip, and signal processing based on multi-core Platform becomes the development trend of software defined radio. In such an integrated system, designing a reliable, high-speed, low-power high-performance on-chip communication system has become a challenge and an opportunity for the development of a System on Chip (SoC). The bus communication structure of the traditional multi-core SoC cannot meet the data exchange requirements in the process of large-scale real-time complex digital signal processing, so a new generation of on-chip interconnection and communication mechanism - on-chip network is proposed.

The system implemented by the network on chip (each node on it is also a processor) is mainly a complex communication system or a digital signal processing system that requires high real-time performance. Communication and signal processing have real-time and processing delay requirements. Therefore, the NoC mapping scheme must make the system delay as small as possible. In addition, the network on chip is realized on the chip of the integrated circuit. In order to ensure the stable operation of the chip, the temperature of the chip should not be high. The temperature of the chip is determined by the energy consumption of the chip, the greater the energy consumption, the higher the temperature of the chip. Therefore, when choosing a mapping scheme, it is necessary to choose a scheme that minimizes energy consumption and delay.

Through novelty searches and extensive literature collection, we found that the published methods for finding mapping schemes between tasks and nodes on a network-on-chip are as follows:

Document "Zhou Ganmin, Yin Yongsheng, Hu Yonghua, etc. NoC Mapping Based on Ant Colony Optimization Algorithm. Computer Engineering and Applications. 2005, 41(18): 7-10." Using an effective ant colony optimization algorithm to realize on-chip network mapping , while the tasks are mapped to the network nodes on a chip, the communication power consumption of the system is reduced as much as possible. The objective function of this method is power consumption, and the method is roughly as follows:

(1) According to the number of ants, assign tasks to each unassigned node according to the probability (this probability is the probability of assigning a node to a task calculated in a way), and each cycle must make all The nodes are assigned to the corresponding tasks, and then use the two-exchange method to perform local search to find the optimal solution with the minimum power consumption, and update the optimal solution with the power consumption value of the optimal solution pheromone attenuation degree and the current pheromone value. Solve the pheromones of the corresponding ants, and finally select the allocation scheme that minimizes power consumption after iterating the above process for many times.

The mapping scheme obtained by this method has a small communication power consumption, but does not consider the delay performance of the system.

The document "Yang Shengguang, Li Li, Gao Minglun, etc. NoC mapping method for energy consumption and time delay. Acta Electronica Sinica. 2008, 36(5): 937-942" is based on a two-dimensional grid structure NoC platform, and establishes an ant-based A unified objective function of swarm algorithms aimed at optimizing system communication energy consumption and link load variance. The calculation formula of the objective function cost of this method is:

cost=λ×E+(1-λ)×VAR

Among them, λ is the proportional coefficient, E is the energy consumption, and VAR is the link load variance. The method is roughly as follows:

(1) According to the number of ants, assign tasks to each unassigned node according to the probability (probability is the probability of assigning a node to a task calculated in a way) according to the number of ants, and make all nodes They are all assigned to the corresponding tasks, and then use the two-exchange method to perform local search to find the solution with the minimum power consumption, and update the ant corresponding to the optimal solution with the power consumption value pheromone attenuation degree of the optimal solution and the current pheromone value After multiple iterations of the above process, the allocation scheme that minimizes the cost of the objective function is finally selected.

This method does not consider that communication energy consumption and execution time are completely different physical quantities, and the obtained mapping scheme does not necessarily have excellent performance.

Literature Tang lei, Shashi Kumar. A two-step genetic algorithm for mapping task graphs to a network onchip architecture. Proceedings of the Euromicro Symposium on Digital System Design 2003. Antalya, Turkey: IEEE, 2003: 180-187. A two-step genetic algorithm approach to task mapping problems. The method is roughly as follows:

(1) First assume that the delays of all edges in the task flow graph are equal, which is a constant, that is, the average number; assign functional units (such as processors, memories, etc.) to each task to minimize the total delay of the system

(2) Use the edge delay of the real data flow graph to further optimize and assign the functional units to the nodes of the NoC.

This method divides the entire optimization stage into two steps, thereby reducing the computational complexity and shortening the computational time, but the energy consumption and delay of the obtained mapping scheme will not be minimized at the same time.

Literature Wenbiao Zhou, Yan Zhang, Zhigang Mao. Pareto based Multi-objective Mapping IP Coreson to NoC Architectures. IEEE Asia Pacific Conference on Circuits and Systems. 2006: 331-334. and Giuseppe Ascia, Vincenzo Catania, Mapsiobping. Miurizio Paleto for Mesh-based NoCA Architectures. International Conference on Hardware/Software Dodesign and System Synthesis, 2004: 182-187. It is proposed to use a multi-objective evolutionary algorithm to map the tasks on the chip network to the nodes of the network structure on the chip based on the Mesh topology. There are two objective functions of this method, one objective function is the average data packet hop count (hop), and the calculation formula is as follows:

$\mathrm{<span\; class="notranslate">\; hop</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\underset{<span\; class="notranslate">\; \&ForAll;</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; d</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; v</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; d</span>}}}\underset{<span\; class="notranslate">\; \&ForAll;</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; d</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; v</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; d</span>}}\underset{<span\; class="notranslate">\; \&ForAll;</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; sidj</span>}}$

Another objective function is the hot spot temperature (T), the calculation formula is: T=max(R×P), where R is the impedance transfer matrix, which was published in ISCA'03 by K.Skadron, M.R.Stan and W.Huang et al. The thermal model tool HotSpot proposed in the article "Temperature-AwareMicro-architecture" in the conference proceedings was obtained. The method is roughly as follows:

(1) First randomly select some schemes to form an initialization scheme group, and perform non-dominated according to the method proposed by N.Srinivas and K.Deb in "Multi-objective optimization function optimization using non-dominated sorting genetic algorithms" published in the second volume of Evolutionary Computation in 1995 The replication, crossover and mutation operations of the genetic algorithm generate a new program group, and then repeat the above replication, crossover and mutation operations several times to select the program that minimizes the objective function.

The non-dominated sorting operation used in this method has high complexity, and the performance of the solution is not particularly good.

Contents of the invention

This patent proposes a method for finding a mapping scheme between tasks and nodes that minimizes energy consumption and response time for multi-task and multi-node on-chip networks. The characteristics of this method are: the obtained mapping scheme has high performance and fast convergence speed , with low complexity.

For the convenience of description, we use vector X=(x ₁ , x ₂ , ... x _n ..., x _N ) to represent a scheme for mapping N tasks to M nodes of the network on chip, and the nth of the vector X The x _n (0≤x _n ≤M-1) in the first position represents the node number of the nth task mapped to a certain node, that is, x _n indicates that the task numbered n is assigned to the _xnth node .

The steps to find the solution that minimizes energy consumption and response time at the same time are as follows:

1. Randomly generate a scheme group J _t containing K allocation schemes, and set a crossover probability p _i and a mutation probability for each allocation scheme X=(x ₁ , x ₂ ,..., x _N ) q _i , and the total number of iterations T, and make the variable t=0 that records the number of iterations;

2. Calculate the energy consumption and response time of the on-chip network under each scheme in the scheme group J _t ;

3. Adopt one of Deb's non-dominated sorting method, elimination method, banker method, ring competition method, recursive method, quick sorting method or improved quick sorting method (published by Science Press in 2007, the author is Zheng Jinhua's " Introduced in "Multi-objective Evolutionary Algorithm and Its Application") to calculate the fitness value of each scheme in the scheme group J _t , the fitness value reflects the priority of the scheme, the smaller the value, the higher the priority, the so-called priority is based on energy Small power consumption and response time are preferred.

4. Randomly divide the K schemes in the scheme group J _t into program groups, wherein L is the number of programs in each program group, select the most preferred program in each program group to form a priority program group Q _t according to the priority relationship determined in the 3rd step;

5. The schemes in the priority scheme group Q _t are randomly paired, and the pairing refers to grouping according to two schemes, and with the crossover probability p _i set in step 1, the two schemes in each pair of schemes are divided into groups. Exchange the number of any node at the same position on each scheme, and then combine all the exchanged schemes to obtain the scheme group Q _t ′;

6. Use the mutation probability q _i set in step 1 to change the number of nodes at any position in each scheme in the scheme group Q _t ′ to obtain the scheme group G _t

个方案，与方案组G_t的方案合并在一起，构成包含K个方案的新的方案组G_t′7. According to the order of the fitness value calculated in step 3 from small to large, select from the program group J _t

schemes, combined with the schemes of scheme group G _t to form a new scheme group G _t ′ containing K schemes

8. Merge scheme groups J _t and G _t into scheme group R _t

9. Calculate the energy consumption and response time of the on-chip network under each scheme in the scheme group R _t ;

10. Using Deb's article A Fast Elitist Non-dominated Sorting Genetic Algorithm for Multi-objective Optimization published in the sixth volume of the magazine IEEE Transactions on Evolutionary Computation in 2002: The boundary set generation method introduced in NSGA-II will scheme group R The 2K schemes in _t are divided into multiple groups, each group represents a boundary set, and the groups are numbered, and the schemes in the smaller numbered group have higher priority than the schemes in the larger numbered group; the so-called priority level is based on energy Small power consumption and response time are preferred.

11. Initialize a new program group F _t without a program, and then add the programs in the group to F _t in sequence from the groups divided in step 10 in descending order until the program The number of plans in the group F _t exceeds K, and then all the plans of the last joined group are taken out, and the last joined group is recorded as V;

12. Prioritize the programs in each group in step 10 by:

12-1 Let the distance value of each scheme in the group be 0, and the distance value indicates whether the scheme is closely related to other schemes.

12-2 Arrange the schemes in reverse order according to the energy consumption of each scheme calculated in step 9, and the schemes with less energy consumption are ranked in front of the schemes with larger energy consumption.

12-3 Calculate the energy consumption distance of each scheme. The energy consumption distance of each scheme is the value obtained by subtracting the energy consumption of the scheme in front of it from the energy consumption of the scheme behind it.

12-4 Arrange the schemes in reverse order according to the response time of each scheme calculated in step 9, and the scheme with a smaller response time is ranked in front of the scheme with a larger response time.

12-5 Calculate the response time distance of each scheme, the response time distance of each scheme is the value obtained by subtracting the response time of the scheme ahead of it from the response time of the scheme behind it,

12-6 Add the response time distance and energy consumption distance of each scheme to get the distance value of the scheme,

12-7 Put the scheme with the least energy consumption and the scheme with the shortest response time at the top, and then sort the remaining schemes according to the distance, and arrange them behind the scheme with the least energy consumption and the scheme with the shortest response time. The larger scheme is ranked in front of the scheme with a smaller distance value, so that the priority of the scheme in the front is higher than the priority of the scheme in the back,

13. From the group V sorted in the 12th step, select the schemes in order of priority from high to low and add them to the scheme group F _t until the number of schemes in the scheme group F _t is K.

个方案小组，其中L是每方案小组中的方案个数，按照第10步和第12步联合确定的优先关系选出每方案小组中最优先的方案组成一个优先方案组F_t′；14. Randomly divide the K schemes in the scheme group F _t into

program groups, where L is the number of programs in each program group, select the most preferred program in each program group to form a priority program group F _t ' according to the priority relationship jointly determined in the 10th step and the 12th step;

15. Randomly pair the schemes in the priority scheme group F _t ', and exchange the number of any node at the same position on the two schemes in each pair of schemes with the crossover probability p _i set in step 1 , and then combine all the swapped schemes to get the scheme group H _t

16. Change the number of nodes at any position in each scheme in the scheme group H _t with the variation probability q _i set in step 1, and obtain the scheme group H _t ′

个方案，与方案组H_t′的方案合并在一起，构成包含K个方案的新的方案组J_t+1 17. According to the priority order of the programs in the program group F _t jointly determined in the 10th and 12th steps, select from the program group F _t in the order of priority from high to low

schemes, combined with the schemes of the scheme group H _t ′ to form a new scheme group J _t+1 containing K schemes

18. If t<T, then t=t+1, then return to step 2, otherwise go to step 19

19. Calculate the fitness value of each scheme in the scheme group J _T according to the method described in step 3, and use the scheme with the smallest fitness value, that is, the scheme with the smallest energy consumption and response time at the same time, as the task of the network on chip to communicate between nodes. Mapping scheme.

The energy consumption of the present invention can be:

E＝E _PE +E _Net

Among them, E is energy consumption, E _PE is the sum of energy consumption of all processing nodes on the network-on-chip, and E _Net is the energy consumption of the communication network connecting each processing node on the network-on-chip. The energy consumption E _Net of the communication network can use the energy consumption calculation method proposed in the paper "Analysis of Power Consumption on Switch Fabrics in Network Routers" published by Terry Tao Ye, Luca Benini, and Giovanni DeMicheli in the conference proceedings of DAC'02. J. Hu and R. Marculescu published in the article "Energy-and performance-aware mapping for regular NoC architectures" in IEEE Transaction on Computer-Aided Design of Integrated Circuits and Systems, Vol. 24, 2005.

The calculation formula of the response time of the present invention (the longest delay experienced by data from input to output) can be:

D=P+TR+TW

Among them, D is the response time, P is the sum of the processing time of all nodes on the input-to-output path, TR is the sum of the delays of all links on the input-to-output path, and TW is the queuing and sum of data packets on the input-to-output path sum of waiting times.

The Essence and Effect of the Invention

The present invention separates energy consumption and response time as the priority index of the assessment scheme, but considers energy consumption and response time at the same time in the process of finding the optimal scheme, so that the found scheme minimizes energy consumption and system delay at the same time, Therefore, the performance of the network designed by using this optimized scheme is the best, and compared with other methods for finding the priority scheme for mapping between tasks and nodes under multiple objective functions, it has the characteristics of complexity and simplicity, through It can also be concluded from the simulation that the method of the present invention has a fast convergence speed.

Description of drawings

Figure 1 is a task diagram composed of six tasks.

Figure 2 is an on-chip network composed of 16 nodes.

1 is a task, 2 is the traffic between tasks, 3 is a node of the network on chip, and 4 is a router of the network on chip.

Detailed ways

There are 6 tasks in the task graph shown in FIG. 1 , and there are 16 nodes in the network on chip shown in FIG. 2 . This embodiment is to find a mapping scheme that distributes the 6 tasks shown in Figure 1 to the 16 nodes of the on-chip network shown in Figure 2, so that the energy consumption and response time of the on-chip network after mapping can reach the minimum value . For the convenience of description, we use vector X=(x ₁ , x ₂ , x ₃ , x ₄ , x ₅ , x ₆ ) to represent a scheme for mapping 6 tasks to 16 nodes in the network on chip, and the nth of the vector X The x _n (0≤x _n ≤15) in the first position represents the node number that the nth task is mapped to a certain node, that is, x _n indicates that the task numbered _n is assigned to the xnth node.

The steps to find the solution that minimizes energy consumption and response time at the same time are as follows:

1. Randomly generate a scheme group J _t containing 20 allocation schemes. Each allocation scheme in the scheme group is represented by a vector: (11, 1, 15, 9, 0, 15), (4, 1, 0, 8 , 1, 6), (7, 5, 15, 8, 7, 8), (10, 11, 5, 0, 6, 8), (0, 6, 7, 6, 15, 9), (14 , 11, 11, 0, 8, 5), (15, 5, 14, 0, 15, 1), (3, 0, 8, 7, 1, 9), (2, 2, 1, 8, 2 , 8), (3, 13, 8, 15, 5, 8), (8, 8, 9, 12, 10, 2), (14, 0, 3, 2, 15, 5), (5, 2 , 4, 9, 0, 8), (7, 2, 15, 11, 1, 11), (7, 1, 15, 2, 1, 11), (9, 12, 3, 9, 10, 11 ), (11, 3, 12, 14, 14, 6), (11, 12, 4, 7, 1, 9), (4, 8, 12, 12, 9, 4), (1, 15, 6 , 9, 3, 9). The probability of crossover of each scheme is set as 0.4, the probability of variation is 0.01, the total number of iterations is 100, and the variable t=0 for recording the number of iterations;

2. Calculate the energy consumption and response time of the on-chip network under each scheme in the scheme group J _t :

2-1. The calculation formula of energy consumption (E) is: E=E _PE +E _Net , E _PE is the sum of energy consumption of all processing nodes on the network on chip, and E _Net is the communication network connecting each processing node on the network on chip energy consumption. The energy consumption of all processing nodes on the network is equal to the processing time of all tasks in the task graph, and the processing time of each task can be obtained by estimating the calculation amount of each task, that is, while determining the task graph, determine the energy consumption of the task. In this example, the value of E _PE is constant at 1.18 for various schemes. Using the energy consumption calculation method proposed in the article "Analysis of Power Consumption on Switch Fabrics in Network Routers" published by Terry TaoYe, Luca Benini, and Giovanni De Micheli in the conference proceedings of DAC'02. to calculate the communication network under various schemes The energy consumption E _Net is: 54.69, 38.09, 61.69, 45.84, 43.84, 46.39, 75.19, 41.84, 46.99, 48.59, 33.09, 55.24, 38.14, 43.89, 50.09, 40.94, 44.64, 43.89, 29.44, 5

2-2. The calculation formula of response time (D) adopts: D=P+TR+TW, P is the sum of the processing time of all nodes on the input to output path, and TR is the delay of all links on the input to output path The sum of TW is the sum of the queuing and waiting time of data packets on the input to output path. In this example, the response times of each scheme are: 0.16, 0.1601, 0.1601, 0.1601, 0.1601, 0.1601, 0.1602, 0.1601, 0.1601, 0.1601, 0.1601, 0.1601, 0.16, 0.16, 0.1601, 0.1601, 0.1601, 0 0.1601.

3. Using Deb's non-dominated sorting method (published by Science Press in 2007, authored by Zheng Jinhua's "Multi-objective Evolutionary Algorithm and Its Application") to calculate the fitness value of each program in the program group J _t , which is: 0.7895, 1.6316, 0.1579, 1, 1.1579, 0.1579, 0, 1.3684, 0.4737, 0.4737, 1.6316, 0.4737, 1.8947, 1.6316, 0.4737, 1.6316, 1, 1.2632, 2, 0.7895, the value reflects the priority of the scheme The smaller the higher the priority.

4. Divide the 20 schemes in the scheme group J _t into 10 scheme groups at random, each scheme group has 2 schemes, select the scheme with the highest priority in each scheme group according to the priority relationship determined in step 3, and then select from The most preferred scheme selected by all scheme groups forms a new scheme group Q _t , and the schemes in the new scheme group Q _t are: (4, 1, 0, 8, 1, 6), (10, 11, 5, 0, 6, 8), (0, 6, 7, 6, 15, 9), (3, 0, 8, 7, 1, 9), (2, 2, 1, 8, 2, 8) , (8, 8, 9, 12, 10, 2), (5, 2, 4, 9, 0, 8), (9, 12, 3, 9, 10, 11), (11, 12, 4, 7, 1, 9), (4, 8, 12, 12, 9, 3)

5. The schemes in the scheme group Q _t are randomly paired, and the pairing refers to grouping according to two schemes, and with the crossover probability 0.4 set in step 1, pair each pair of schemes at random positions exchange the numbers of the nodes, and then combine all the exchanged schemes to obtain the scheme group Q _t ′. The schemes in the scheme group Q _t ′ are: (10, 1, 0, 8, 1, 6 ), (4, 11, 5, 0, 6, 8), (0, 6, 7, 6, 15, 9), (3, 0, 8, 7, 1, 9), (2, 2, 1 , 8, 2, 8), (8, 8, 9, 12, 10, 2), (5, 2, 4, 9, 0, 8), (9, 12, 3, 9, 10, 11), (11, 12, 4, 7, 1, 9), (4, 8, 12, 12, 9, 3)

6. With the variation probability 0.01 set in step 1, change the numbering of nodes at random positions in each scheme in the scheme group Q _t ′ to obtain the scheme group G _t , and the schemes in the scheme group G _t are: (10, 1 , 0, 8, 1, 6), (4, 11, 5, 0, 6, 8), (0, 6, 7, 6, 15, 9), (3, 0, 8, 7, 1, 9 ), (2, 2, 1, 8, 2, 8), (8, 8, 9, 12, 10, 2), (5, 2, 4, 9, 0, 8), (9, 12, 3 , 9, 10, 11), (11, 12, 4, 7, 1, 9), (4, 8, 12, 12, 9, 3)

7. According to the order of fitness from small to large, select 10 schemes from the scheme group J _t and merge them with the schemes of scheme group G _t to form a new scheme group G _t ′ containing 20 schemes

8. Merge scheme groups J _t and G _t ' into scheme group R _t containing 40 schemes, namely: (11, 1, 15, 9, 0, 15), (4, 1, 0, 8, 1, 6), (7, 5, 15, 8, 7, 8), (10, 11, 5, 0, 6, 8), (0, 6, 7, 6, 15, 9), (14, 11, 11, 0, 8, 5), (15, 5, 14, 0, 15, 1), (3, 0, 8, 7, 1, 9), (2, 2, 1, 8, 2, 8) , (3, 1, 3, 8, 15, 5, 8), (8, 8, 9, 12, 10, 2), (14, 0, 3, 2, 15, 5), (5, 2, 4, 9, 0, 8), (7, 2, 15, 11, 1, 11), (7, 1, 15, 2, 1, 11), (9, 12, 3, 9, 10, 11) , (11, 3, 12, 14, 14, 6), (11, 12, 4, 7, 1, 9), (4, 8, 12, 12, 9, 4), (1, 15, 6, 9, 3, 9), (4, 1, 0, 8, 1, 6), (10, 11, 5, 0, 6, 8), (0, 6, 7, 6, 15, 9), ( 3, 0, 8, 7, 1, 9), (2, 2, 1, 8, 2, 8), (8, 8, 9, 12, 10, 2), (5, 2, 4, 9, 0, 8), (9, 12, 3, 9, 10, 11), (11, 12, 4, 7, 1, 9), (4, 8, 12, 12, 9, 3), (11, 3, 12, 14, 14, 6), (0, 6, 7, 6, 15, 9), (11, 12, 4, 7, 1, 9), (3, 0, 8, 7, 1, 9), (4, 1, 0, 8, 1, 6), (8, 8, 9, 12, 10, 2), (7, 2, 15, 11, 1, 11), (9, 12, 3, 9, 10, 11), (5, 2, 4, 9, 0, 8), (4, 8, 12, 12, 9, 4)

9. Calculate the energy consumption and response time of the on-chip network under each scheme in the scheme group R _t . The energy consumption of each scheme is: 54.69, 38.09, 61.69, 45.84, 43.84, 46.39, 75.19, 41.84, 46.99, 48.59, 33.09, 55.24，38.14，43.89，50.09，40.94，44.64，43.89，29.44，53.74，38.09，45.84，43.84，41.84，46.99，33.09，38.14，40.94，43.89，40.84，44.64，43.84，43.89，41.84，38.09，33.09， 43.89, 40.94, 38.14, 29.44; the response time of each scheme is:

10、0.16，0.1601，0.1601，0.1601，0.1601，0.1601，0.1602，0.1601，0.1601，0.1601，0.1601，0.1601，0.16，0.16，0.1601，0.1601，0.1601，0.1601，0.16，0.1601，0.1601，0.1601，0.1601，0.1601， 0.1601, 0.1601, 0.16, 0.1601, 0.1601, 0.1601, 0.1601, 0.1601, 0.1601, 0.1601, 0.1601, 0.1601, 0.16, 0.1601, 0.16, 0.16.

11. Using Deb's article A Fast Elitist Non-dominated Sorting Genetic Algorithm for Multi-objective Optimization published in Volume 6 of the journal IEEE Transactions on Evolutionary Computation in 2002: The boundary set generation method introduced in NSGA-II to divide the scheme group R _t The 40 schemes in the group are divided into 13 groups, each group represents a boundary set, and the groups are numbered, and the schemes in the smaller numbered group have higher priority than the schemes in the larger numbered group; the so-called priority level is based on energy consumption It is better to have a small response time. The schemes in Group 1 obtained by the above method include Scheme 19, Scheme 40, the schemes in Group 2 include Scheme 13, Scheme 27, and Scheme 39, and the schemes in Group 3 include Scheme 2, Scheme 11, Scheme 21, and Scheme 26. Scheme 35, Scheme 36, the schemes in Group 4 include Scheme 16, Scheme 28, and Scheme 38, the schemes in Group 5 include Scheme 8, Scheme 24, and Scheme 34, and the schemes in Group 6 include Scheme 14, Scheme 37, and Group 7 The plans in it include plan 18, plan 29, and plan 33. The plans in group 8 include plan 5, plan 23, and plan 32. The plans in group 9 include plan 4, plan 17, plan 22, plan 30, plan 31, and the group The plans in group 10 include plan 1 and plan 20, the plans in group 11 include plan 9, plan 10, plan 12, plan 15, and plan 25, the plans in group 12 include plan 3, plan 6, and the plans in group 13 include Option 7.

12. Initialize a new program group F _t without a program, and then from the groups divided in step 10, according to the order of group numbers from small to large, from group 1 to group 13, add the plans in the group to In F _t , until the number of schemes in the scheme group F _t exceeds 20, all the schemes of the last joined group (namely group 7) are taken out at this time.

13. Determine the priority of each program in each group obtained in step 10. The specific method is as follows:

12-1 Let the distance value of each scheme in the group be 0, and the distance value indicates whether the scheme is closely related to other schemes.

12-2 Arrange the schemes in reverse order according to the energy consumption of each scheme calculated in step 9, and the schemes with less energy consumption are ranked in front of the schemes with larger energy consumption.

12-3 Calculate the energy consumption distance of each scheme. The energy consumption distance of each scheme is the value obtained by subtracting the energy consumption of the scheme in front of it from the energy consumption of the scheme behind it.

12-4 Arrange the schemes in reverse order according to the response time of each scheme calculated in step 9, and the scheme with a smaller response time is ranked in front of the scheme with a larger response time

12-5 Calculate the response time distance of each scheme, the response time distance of each scheme is the value obtained by subtracting the response time of the scheme ahead of it from the response time of the scheme behind it,

12-6 Add the response time distance and energy consumption distance of each scheme to get the distance value of the scheme,

12-7 Put the scheme with the least energy consumption and the scheme with the shortest response time at the top, and then sort the other schemes according to the distance, and rank them behind the scheme with the least energy consumption and the scheme with the shortest response time, and the distance value is larger The scheme of the distance value is ranked in front of the scheme with a smaller distance value, so that the priority of the scheme in the front is higher than the priority of the scheme in the back

14. From the group sorted in the 12th step, select the scheme according to the priority level and join the scheme group F _t until the number of schemes in the scheme group F _t is 20. After completion, the schemes in the scheme group F _t are: ( 4, 8, 12, 12, 9, 4), (4, 8, 12, 12, 9, 4), (5, 2, 4, 9, 0, 8), (5, 2, 4, 9, 0, 8), (5, 2, 4, 9, 0, 8), (4, 1, 0, 8, 1, 6), (8, 8, 9, 12, 10, 2), (4, 1, 0, 8, 1, 6), (8, 8, 9, 12, 10, 2), (4, 1, 0, 8, 1, 6), (8, 8, 9, 12, 10, 2), (9, 12, 3, 9, 10, 11), (9, 12, 3, 9, 10, 11), (9, 12, 3, 9, 10, 11), (3, 0, 8, 7, 1, 9), (3, 0, 8, 7, 1, 9), (3, 0, 8, 7, 1, 9), (7, 2, 15, 11, 1, 11) , (7, 2, 15, 11, 1, 11), (11, 12, 4, 7, 1, 9)

15. The 20 programs in the program group F _t are randomly divided into 10 program groups, each program group has 2 programs, and the most preferred program of each program group is selected according to the priority relationship jointly determined in steps 10 and 12, and then A new scheme group F _t ′ will be formed from the most preferred scheme selected from all scheme groups. The schemes in the scheme group F _t ′ include (4, 8, 12, 12, 9, 4), (5, 2 , 4, 9, 0, 8), (5, 2, 4, 9, 0, 8), (8, 8, 9, 12, 10, 2), (8, 8, 9, 12, 10, 2 ), (8, 8, 9, 12, 10, 2), (9, 12, 3, 9, 10, 11), (3, 0, 8, 7, 1, 9), (3, 0, 8 , 7, 1, 9), (7, 2, 15, 11, 1, 11)

16. Randomly pair the schemes in the scheme group F _t ′, exchange the numbers of nodes at random positions between each pair of schemes with the crossover probability 0.4 set in step 1, and then merge all the pairs The scheme group H _t is obtained, namely: (4, 8, 12, 12, 9, 4), (5, 2, 4, 9, 0, 8), (8, 2, 4, 9, 0, 8), (5, 8, 9, 12, 10, 2), (8, 8, 9, 12, 10, 2), (8, 8, 9, 12, 10, 2), (9, 12, 3, 9 , 10, 11), (3, 0, 8, 7, 1, 9), (3, 0, 8, 7, 1, 9), (7, 2, 15, 11, 1, 11)

17. Change the numbering of nodes at random positions in each scheme in the scheme group H _t with the probability 0.01 of the variation set in step 1 to obtain the scheme group H _t ′, namely: (4, 8, 12, 12, 9 , 4), (5, 2, 4, 9, 0, 8), (8, 2, 4, 9, 0, 8), (5, 8, 9, 12, 10, 2), (8, 8 , 9, 12, 10, 2), (8, 8, 9, 12, 10, 2), (9, 12, 3, 9, 10, 11), (3, 0, 8, 7, 1, 9 ), (3, 0, 8, 7, 1, 9), (7, 2, 15, 11, 1, 11)

18. Select 10 schemes from the scheme group F _t according to the order of fitness from small to large, and merge them with the schemes of scheme group H _t ′ to form a new scheme group J _t+1 containing 20 schemes

19. If the number of iterations t is less than 100, then t=t+1, and then return to step 2 to perform a new round of iterative operation, otherwise enter step 19;

20. Calculate the fitness value of each scheme in the scheme group J ₁₀₀ according to the method described in step 3, and use the scheme with the smallest fitness value, that is, scheme (5, 2, 9, 9, 5, 4) as the network on chip The mapping scheme between tasks and nodes.

Claims (5)

1. be used to seek the method for mapping scheme between the task of network-on-chip and node, it is characterized in that: seek energy consumption and the response time task and the node mapping scheme of minimum simultaneously of sening as an envoy to, the searching process is as follows:

One, generates a scheme group J who comprises K allocative decision at random _t, and be each allocative decision X=(x ₁, x ₂..., x _N) set the Probability p of one intersection _iProbability q with variation _i, and total iterations T, and the variable t=0 of order record number of iterations;

Two, obtain scheme group J _tIn network-on-chip energy consumption and response time under every kind of scheme;

Three, calculate scheme group J _tIn the fitness value of each scheme, the priority of this fitness value size reflection scheme, it is high more be worth more little priority, so-called priority level be little by energy consumption and response time be excellent;

Four, with scheme group J _tIn K scheme be divided at random

Individual scheme group, wherein L is the scheme number in the every scheme group, selects in the every scheme group prepreerence scheme according to the precedence relationship of determining in the 3rd step and forms a priority scheme group Q _t

Five, with priority scheme group Q _tIn scheme press random pair, described pairing is meant according to two schemes divides into groups for one group, and with the Probability p of the intersection set in the step 1 _iThe numbering of the node on any one same position on two schemes in every pair of scheme is exchanged, and then all are merged by every pair of scheme after exchanging obtain scheme group Q ' _t

Six, the variation probability q to set in the first step _iChange scheme group Q ' _tIn in each scheme in the numbering of any one locational node, obtain scheme group G _t

Seven, the fitness value order of calculating by the 3rd step from small to large is from scheme group J _tThe middle selection

Individual scheme is with scheme group G _tScheme combine, constitute the new scheme group G ' comprise K scheme _t

Eight, with scheme group J _tAnd G ' _tBe merged into scheme group R _t

Nine, numerical procedure group R _tIn network-on-chip energy consumption and response time under every kind of scheme;

Ten, with scheme group R _tIn 2K scheme be divided into a plurality of groups, each group represents a boundary set, to little group #, the scheme in the less group of numbering is than the priority height of the scheme of the bigger group of numbering; So-called priority level is little of excellent by energy consumption and response time;

11, new scheme group F who does not have scheme of initialization _t, from the good group of the tenth step branch, the numbering order from big to small by group joins F with the scheme in the group successively then _tIn, up to scheme group F _tIn the scheme number surpass K, the scheme of the group that will add is at last all taken out then, and the group that will be added at last is designated as V;

12, determine in the tenth step the priority ranking of each scheme in each group, concrete grammar is as follows:

12-1 makes that the distance value of each scheme in the group is 0, and described distance value represents whether getting in touch of this scheme and other scheme be tight;

12-2 carries out inverted order according to the 9th size that goes on foot the energy consumption of each scheme that calculates to scheme and arranges, and the less scheme of energy consumption comes the front of the bigger scheme of energy consumption,

12-3 calculates the energy consumption distance of each scheme, and the energy consumption of each scheme distance deducts the resulting value of energy consumption of that scheme that comes its front for the energy consumption of that scheme of coming its back,

12-4 carries out the inverted order arrangement according to the response time size of each scheme that the 9th step calculated to scheme, and the response time, less scheme came the front of bigger scheme of response time,

12-5 calculates the response time distance of each scheme, and response time of each scheme distance deducts the resulting value of response time of that scheme that comes its front for response time of that scheme of coming its back,

12-6 obtains the distance value of scheme with the response time distance of each scheme and energy consumption apart from addition,

12-7 comes the foremost to the scheme of the scheme of energy consumption minimum and response time minimum, according to the distance size remaining scheme is sorted then, and come the back of the scheme of the scheme of energy consumption minimum and response time minimum, the scheme that distance value is bigger comes the front of the less scheme of distance value, and the priority ratio that comes the scheme of front like this comes the priority height of the scheme of back;

13, join scheme group F from the select progressively scheme of having passed through in the V of group after the ordering of the 12 step according to priority from high to low _t, up to scheme group F _tIn the scheme number be K;

14, with scheme group F _tIn K scheme be divided at random

Individual scheme group, wherein L is the scheme number in the every scheme group, selects priority scheme group F ' of prepreerence scheme composition in the every scheme group according to the precedence relationship of determining in the tenth step and the 12 step _t

15, with priority scheme group F ' _tIn scheme random pair again, with the Probability p of the intersection set in the first step _iThe numbering of the node on any one same position on two schemes in every pair of scheme is exchanged, and then all are merged by every pair of scheme after exchanging obtain scheme group H _t

16, with the probability q of the variation set in the first step _iChange scheme group H _tIn in each scheme in the numbering of any one locational node, obtain scheme group H ' _t

17, unite definite scheme group F according to the tenth step and the 12 step _tThe priority ranking of middle scheme is according to priority order from high to low, from scheme group F _tThe middle selection

Individual scheme is with scheme group H _tScheme combine, constitute the new scheme group J comprise K scheme _T+1

If 18 t＜T, then t=t+1 turned back to for second step then, otherwise entered for the 19 step;

19, calculate scheme group J _tIn the fitness value of each scheme, with the scheme of fitness value minimum, promptly minimum simultaneously scheme of energy consumption and response time as the task of network-on-chip to internodal mapping scheme;

2. according to the method that is used to seek mapping scheme between the task of network-on-chip and node of claim 1, it is characterized in that: described energy consumption is the energy consumption that the energy consumption sum of all processing nodes on the network-on-chip is added the communication network that connects each processing node.

3. according to the method that is used to seek mapping scheme between the task of network-on-chip and node of claim 1 or claim 2, it is characterized in that: the described response time is that the processing time sum that is input to all nodes on the outgoing route adds that the time-delay sum that is input to all links on the outgoing route adds queuing and the stand-by period sum that is input to packet on the outgoing route.

4. according to the method that is used to seek mapping scheme between the task of network-on-chip and node of claim 1 or claim 2, it is characterized in that: the fitness value that calculates each scheme adopts one of exclusive method, banker method, challenge tournament method, recurrence method, quick sort or improved quick sort.

5. according to the method that is used to seek mapping scheme between the task of network-on-chip and node of claim 3, it is characterized in that: the fitness value that calculates each scheme adopts one of exclusive method, banker method, challenge tournament method, recurrence method, quick sort or improved quick sort.

Images