62results about How to "Improve Parallel Computing Efficiency" patented technology

The embodiment of the invention discloses an integrated circuit electromagnetic response calculation method and device based on a multi-level parallel strategy. According to the method, the method comprises the steps of dividing the calculation of electromagnetic response characteristics of each frequency point needing to be calculated in frequency domain simulation of the multi-layer super-large-scale integrated circuit into a plurality of calculation subtasks; and respectively carrying out parallel computing on each computing subtask by adopting the plurality of parallel subparticles, and independently executing the plurality of computing tasks corresponding to the plurality of frequency points by utilizing the plurality of parallel coarse particles to finish multi-process parallel computing of the plurality of frequency points. According to the method, the calculation efficiency of each part in the electromagnetic response characteristic calculation of the frequency points in the frequency domain simulation of the multi-layer super-large-scale integrated circuit can be improved, the parallel calculation efficiency of the electromagnetic response characteristics of a large numberof integrated circuits with different frequency points can also be improved, and the efficient calculation requirement is met.

The invention discloses a method and a system for dispatching distributed parallel computing job. The method includes establishing network connection among a client side and a plurality of servers and dividing a complex computing task into a series of independent computing tasks; creating a plurality of threads on the client side and respectively dispatching the independent computing tasks to the multiple servers and the client side by the aid of the threads; enabling the multiple servers and the client to operate the dispatched computing tasks by a memory-sharing parallel process; enabling the client side to receive operation results fed by the various servers after computation is completed. The method and the system for dispatching the distributed parallel computing job have the advantages that a total computing task set is divided into the series of independent computing tasks, parallel computing is carried out on the server and the client side by the memory-sharing process, accordingly, installation programs of server sides are simplified, time consumption for transmitting large quantities of data via a network is saved, and the parallel computing efficiency is improved.

The invention provides a dynamic load balancing method based on a Linux parallel computing platform, and belongs to the field of parallel computing. The hardware architecture of the dynamic load balancing method comprises multiple computers participating in computing, and each computer is provided with a Linux operating system and an MPI software development kit. In the parallel computing process, general computing tasks are divided into multiple stages with the same execution time to be executed. The routine job scheduling technology in a system is utilized, the current resource utilization rate of each node is firstly read before parallel computing of each time phase begins, calculated performance and computational complexity of each node are combined, dynamic allocation is conducted on the computing tasks of the nodes, the fact that computing time of the nodes in every phase is basically equal is guaranteed, and system synchronization waiting delay is reduced. Through the dynamic adjustment strategy, the general computing tasks can be completed by means of higher resource utilization rates, the dynamic load balancing method breaks through efficiency bottlenecks caused by low configuration computational nodes, the computing time is further saved on the basis of the parallel computing, and computational efficiency is improved.

A neural network processor includes: a storage circuit that stores initial input data and a weight value required for performing a convolution operation; at least one computing circuit including: a data buffer that caches initial input data; the weight buffer caches the weight value; the convolution arithmetic unit is used for carrying out convolution operation in the current layer of convolutionneural network according to the initial input data and the weight values to obtain a plurality of first convolution results, and accumulating the first convolution results with the corresponding relations to obtain a plurality of second convolution results; at the same time, deleting the plurality of first convolution results; and the result buffer is used for caching the plurality of second convolution products as the initial input data of the next layer of convolutional neural network. The invention further provides a convolutional neural network data multiplexing method and device, electronic equipment and a storage medium. According to the invention, through multi-level data multiplexing, the operation speed of the neural network processor is improved, and the power consumption is reduced.

The invention provides an image automatic segmentation method and device based on deep learning edge detection. The method comprises the steps of drawing a certain number of training samples accordingto a high-resolution remote sensing image; training an HED edge detection model through the training sample; performing edge detection on the remote sensing image to be segmented by using the trainedHED edge detection model to generate an edge probability graph; performing post-processing on the edge probability graph to generate a vector polygon; and simplifying the vector polygon to obtain animage segmentation result of the remote sensing image to be segmented. According to the image segmentation method provided by the invention, parameters do not need to be adjusted, the segmentation accuracy is higher, automatic operation can be realized, and the image segmentation efficiency is high.

The invention discloses a GPU thread design method of power flow Jacobian matrix calculation. The method comprises the steps that: power grid data is input, and pre-processing is carried out on a node admittance matrix Y; a CPU calculates position mapping relation tables respectively between sub-matrix non-zero elements and node admittance matrix Y non-zero elements; the calculates position mapping relation tables respectively between sub-matrix non-zero elements and Jacobian matrix non-zero elements; in a GPU, injection power of each node is calculated by an injection power kernel function S; and in the GPU, non-zero elements in sub-matrixes are respectively calculated by Jacobian sub-matrix calculation kernel functions and stored in a Jacobian matrix. According to the invention, the calculation of Jacobian matrix non-zero elements is carried out by the sub-matrixes, and the judgment of branch structures of sub-matrix regional process belonging to the elements, which is required by the direct calculation process using a single kernel function, is avoided, so that the execution efficiency is improved; in addition, the non-zero elements in the sub-matrixes having identical calculation formulas are calculated in a centralized manner, so that the problem of unbalanced thread loads is solved, and the efficiency of parallel calculation is improved.

The invention discloses an automatic starting-stopping and computation task dynamic allocation method for mass parallel coarse particle computation. The method comprises the following steps: defining parallel coarse particles according to a problem computation characteristic; dynamically allocating computation tasks in the parallel coarse particles and input parameters corresponding to the computation tasks to all processes including a master process by the master process according to a file marking technology and a dynamic allocation computation task strategy; dynamically allocating memories to the processes including the computation tasks based on an automatic starting-stopping technology; and after completion of parallel computation of all the parallel coarse particles, collecting output parameters of all the processes by the master process, and combining and integrating the output parameters to obtain a final result of complete running. Through adoption of the method, communications among the processes are reduced to the maximum extent; the hard disk reading-writing bottleneck occurring since a memory peak value is greater than an available physical memory during multi-process parallel computation is avoided; meanwhile, the problem of non-equivalent complexity of computation examples is solved perfectly; and the parallel computation efficiency is increased greatly.

The embodiment of the invention provides a load flow calculation method and device for a large-scale power distribution network, and the method comprises the steps: decomposing a target power distribution network into a plurality of sub-networks, and obtaining an equation for solving a contact branch current value between the sub-networks; calculating constant impedance, constant current and constant power parameters of each subnet ZIP load model; calculating a node admittance matrix of each subnet according to the constant impedance parameters, and improving the node admittance matrix; calculating the node injection current of the current iteration according to the node voltage of the current iteration and the constant current and constant power parameters; obtaining the current iterationcontact branch current according to the improved node admittance matrix, the current iteration node injection current and an equation used for solving the contact branch current between the subnets;obtaining the node voltage phasor of the next iteration according to the contact branch current of the current iteration if the difference value between the contact branch current of the previous iteration and the next iteration is greater than the convergence criterion. According to the embodiment of the invention, the efficiency and flexibility of load flow parallel computing are considered.

The invention discloses a network maximum flow parallel solving method which comprises the follow steps that 1) network surplus flow is promoted in a parallel iteration mode; and 2) the network flow is optimized in a sub-network merging mode. According to the step 1) of the method, a network is divided into a plurality of sub-networks which are distributed to each computing unit, each computing unit divides border top points into three different levels in a marking mode according to different states of border arcs and border top points of the sub-network of the computing unit, the border top points with different levels push the surplus flow in different modes, pushing efficiency of the surplus flow is guaranteed, communication problems are effectively solved, solving speed of the parallel method is quickened, and meanwhile correctness of results is guaranteed through the step 2).

The invention relates to the field of function material design and high-performance calculation, in particular to a coarse particle parallel method and system for electromagnetic function material optimization design. The method comprises the steps of obtaining a serial version by taking a coarse particle as an independent execution module, and obtaining a parallel version by taking the coarse particle as a basic execution unit; further, processing a plurality of calculation tasks by adopting an allocation policy of random sorting, wherein the allocation policy of random sorting can thoroughly disrupt an allocation sequence of all the calculation tasks to obtain a new sequence of the calculation tasks; then, applying for allocating to-be-allocated calculation tasks in the new sequence by each calculation process through adopting a first-application first-allocation policy, and performing calculation until all the calculation tasks are finished; and finally, performing statistics on calculation results, performing targeting processing, comparing a processing result with an expected optimization goal, and judging whether the optimization goal is achieved or not. According to the coarse particle parallel method, the problem of low parallel calculation efficiency in electromagnetic function material optimization design at the present stage is effectively solved.

The invention provides a point cloud feature matching method and system based on an FPGA, and a path planning system. The system comprises a processing system part and a programmable logic part. The processing system part performs periodic point cloud data acquisition, the acquired point cloud data is written into a specified address space of a memory and communicates with the programmable logic part through a register channel, and the programmable logic part acquires the point cloud data from the specified address space frame by frame and performs logic processing. The KNN process of processing feature point cloud matching on a traditional processor CPU is avoided and simplified.

The invention discloses a discrete data uniform quantification algorithm in mobile cloud calculation. A mobile cloud calculation data clustering model and a channel model are constructed, fluctuation discrete data acquisition is carried out, a path is selected according to a Logistic model, an optimal probability density of mobile cloud calculation intercluster fluctuation discrete data uniform quantification is obtained, a Bayes rough set uniform quantification optimization object function is constructed, and thus algorithm optimization is realized. The algorithm employed by the invention can effectively realize feature classification of mobile cloud calculation intercluster fluctuation discrete data, has a quite good fluctuation discrete data quantification effect and improves the parallel calculation efficiency of cloud calculation.

The invention provides a modal parallel computing method and system for a heterogeneous many-core parallel computer. The method comprises the following steps: the step S1: generating a finite elementmodel rigidity matrix and quality matrix data through a finite element program, dividing the generated finite element model rigidity matrix and quality matrix data into N subareas, and independentlystoring the finite element model rigidity matrix and quality matrix data of each area in a file, wherein N is an integral multiple of a single core group; and the step S2, enabling the main cores of the core groups which are subjected to parallel computing to synchronously read the finite element model rigidity matrix and the quality matrix data corresponding to the sub-regions, enabling the coregroups to have no data communication exchange, and enabling the slave cores in the core groups to have no data communication exchange. Layering of the calculation process and data communication is achieved through a layering strategy, a large amount of data communication is limited in each core group, and the advantage that the communication rate in the domestic heterogeneous many-core parallel computer core group is high is fully played.

The invention discloses a frequency domain astronomical image target detection method and system, and the method is realized based on a CPU-GPU heterogeneous processor, and the method comprises the steps of carrying out the preprocessing of a collected original astronomical image through a CPU based on a reference image obtained in advance; respectively carrying out image partitioning on the reference image and the preprocessed astronomical image by adopting an overlapping storage method, multiplying a Gaussian basis function by a polynomial to obtain n groups of basis vectors of a convolution kernel, and inputting the n groups of basis vectors into a GPU; enabling the GPU to perform fitting according to the reference sub-images and the astronomical sub-images in combination with n groups of base vectors to obtain a convolution kernel corresponding to each reference sub-image, using the convolution kernels for performing frequency domain filtering blurring processing on each reference sub-image to obtain a template image, and inputting the template image into the CPU; and discarding the edge of the template image by the CPU, connecting the remaining parts and making a difference between the remaining parts and the original astronomical image to obtain a difference image, thereby realizing target detection of the astronomical image.

The present invention relates to a method of improving WRF parallel computing efficiency, and belongs to the technical field of weather forecast and supercomputing application. The method reduces the number of processes in the X direction so as to improve the WRF parallel computing efficiency. The method specifically comprises the following steps of (S1) configuring the number of total processes NP of WRF parallel computing; (S2) setting the number of processes of the WRF in x and y directions nproc_x and nproc_y, wherein nproc_x<nproc_y; (S3) searching the optimal nproc_x through experiments; and (S4) using the optimal nproc_x in business application to improve the WRF parallel computing efficiency. According to the method, the WRF parallel computing efficiency can be effectively improved, supercomputing energy consumption can be reduced, and the method is easy to use and has good application prospect.

The invention discloses a layered data center resource optimization method and system based on SDN and NFV, and the method comprises the steps: enabling an MANO controller to arrange a request initiated by a client to generate a corresponding service chain, and taking chains with the same source and destination nodes into a feature group; filtering the grouped service chains to obtain a part of which the bandwidth requirement is higher than a set threshold value, marking the highest priority, and arranging the remaining service chains in a descending order according to the requirement of the remaining service chains on the time delay to indicate the priority; calculating the first K shortest paths between the source node and the destination node for each group of service chains; formulating, by the MANO controller, a mapping strategy of each service chain by using a coevolutionary multi-population competition genetic algorithm according to the first K shortest paths; and issuing, by the MANO controller, the formulated mapping strategy to an SDN controller of a basic control layer, and enabling the SDN controller to convert the mapping strategy into a flow table suitable for being processed by a switch and issue the flow table to the switch for execution. The end-to-end time delay of the service chain is small, and the utilization rate of network resources is high.

The invention relates to the field of computer technology, and particularly provideds a GPU parallel compute resource allocation method and apparatus, The methods include: a computing task determinedby a computing process is obtained, Aa layered directed acyclic graph model is obtained by adopting a two-layer parallel computing model to process the computational task under the initial parameters,a layered directional acyclic graph model is used to compute the computational task by using a two-layer parallel computational model, according to the number of a plurality of preset parameters andthe number of calculation elements corresponding to each preset parameter, the thread block is labeled according to the preset parameters and the calculation elements corresponding to each preset parameter, and the calculation elements corresponding to each preset parameter are allocated according to the labels of each thread block, so that each thread block calculates the allocated calculation elements. By using the above method, the parallel computing efficiency can be effectively improved.