36results about How to "Reduce data transfer overhead" patented technology

A VLAN tagging unit associated with a WAN is used to add VLAN tags based upon logical interfaces. In one example, data received across the logical interface is tagged with an associated VLAN ID by the unit. In another embodiment, flow-based VLAN tagging is done in which in addition to the logical interface other indications are used in the VLAN tagging.

The invention discloses a method for deploying component services for a data-intensive service cooperation system. By the method, the component services are optimally deployed by the aid of a multi-object optimization algorithm on the basis of negative selection. The method includes mapping the component service deployment on the basis of data-intensive service cooperation into a negative selection algorithm, mapping a single deployment scheme of each single component service into a gene, mapping a deployment scheme of all the component services as an antibody and creating all possible genes to form a gene library; sequentially iteratively matching antibodies on the basis of a negative selection process and circularly iteratively matching the antibodies repeatedly to finally obtain antigens, and using the deployment schemes corresponding to the certain antibodies as the optimal deployment schemes. The method has the advantages that in each iteration procedure, a certain quantity of antibodies are generated by means of gene recombination at first, antibody groups are formed, then low-quality antibodies in the antibody groups are eliminated by means of negative selection, and accordingly search space is reduced; a gene warehouse is updated according to an iteration result after each iteration procedure is completed, so that optimal antibodies can be assuredly generated by means of gene recombination in each follow-up iteration procedure.

The invention discloses a check code packet deployment method and device. The method comprises the following steps of pre-storing at least one secondary package in the target file storage space; obtaining an uploaded check packet, and storing the uploaded check packet in the target file storage space; determining a secondary package on which the uploaded check packet depends; and creating a class loader corresponding to the uploaded check packet in the memory, and loading the class in the uploaded check packet and the class in the secondary package on which the uploaded check packet depends into the class loader. According to the scheme, the check packet in the check code packet is decoupled from the secondary package, and the secondary package is stored and managed in a centralized manner, so that the storage overhead and the data transmission overhead can be greatly reduced, and the centralized management of the secondary package in the check code packet is facilitated; and moreover, by loading the classes of the check package and the relied secondary package into the same class loader, the efficient operation of the check code package is ensured.

The invention discloses a thread data dependency group mapping method for a many-core system. The method comprises the steps that application program thread internal data locality is analyzed by calculating data reuse distances, and data dependencies among different threads are measured with a thread dependency matrix; according to characteristics of the data dependencies of application programs and many-core processor construction, application threads are divided into data dependency logical groups capable of reflecting different threads by means of a minimal spanning tree. Reasonable mapping from the threads to different processing core hardware threads of a specific processor is finally achieved, shared storage access conflicts among the threads are reduced, additional transmission overhead is reduced, and an effective solution is provided for fully utilizing many-core system computing resources and promoting the program calculation performance.

The invention provides a training system of back propagation neural network DNN (Deep Neural Network). The training system of the back propagation neural network DNN comprises a first graphics processor assembly, a second graphics processor assembly and a controller assembly, wherein the first graphics processor assembly is used for performing DNN forward calculation and weight update calculation; the second graphics processor assembly is used for perform DNN forward calculation and DNN back calculation; the controller assembly is used for controlling the first graphics processor assembly and the second graphics processor assembly to perform Nth-layer DNN forward calculation respectively according to respective input data, and after the completion of the forward calculation, controlling the first graphics processor assembly to perform the weight update calculation and controlling the second graphics processor assembly to perform the DNN back calculation; and N is a positive integer. The training system provided by the invention has the advantages of high training speed and low data transmission cost, so that the training speed of the back propagation neural network DNN is promoted.

The present invention relates to the technical field of video processing, in particular to a method, device, medium and system for identifying video content. Aiming at the problem that in the existing security monitoring system, when the video content is automatically recognized in real time, the resource occupancy rate of the video recognition terminal is high or the recognition accuracy is low, the embodiment of the present invention provides a real-time video processing on the edge side, based on the edge device Redundant resources and computing power to cooperate to complete the real-time video recognition scheme. This solution uses existing edge devices (such as set-top boxes, smart home gateways, etc.) to form an edge computing network, and completes real-time video processing on the edge side through a series of methods such as video task allocation, scheduling, and prediction of available resources.

The application discloses a data processing optimization method, device, electronic equipment, and readable storage medium, which are applied in the field of computer technology. The data processing optimization method includes: detecting whether the data to be processed stored in the shared buffer of the CPU satisfies the preset startup GPU processing condition; if satisfied, the data to be processed is synchronously copied from the shared buffer to the shared video memory, wherein the shared video memory is a video memory shared by all GPUs; at least one data processing GPU is scheduled by the CPU, Perform multi-GPU accelerated data processing on the data to be processed. The present application solves the technical problem of low efficiency when the CPU performs data processing in the prior art.

The invention provides data decoding method and equipment with uplink coordinated multiple points. The data decoding method comprises the following steps: a main district receives constellation point information from a cooperation district, wherein the data quantity of the constellation point information is smaller than that of uplink data received by the cooperation district; the main district carries out weighted combination on the constellation point information to obtain combined constellation point information; and the main district carries out demodulation and decoding on the combined constellation point information to obtain original information. With the adoption of the technical scheme of the invention, the data transmission cost between the main district and the cooperation district is reduced, the problem that the gain is lowered caused by decoding soft bits after being combined is solved.

The invention provides a data configuration method, a data configuration device and a server. The data configuration method comprises the following steps: according to a position attribute, determining variable position data sets from all data sets; calculating the relevance of each two data sets in all the data sets and constructing a relevance matrix according to the relevance of each two data sets; sequentially traversing data sets in the variable position data sets and searching the data set with the highest relevance with the currently transversed data set from the relevance matrix; and according to a data center of the data set with the highest relevance, placing the currently transversed data set. The invention can implement the effect of placing the data set at the data center according to the relevance between the data sets so as to reduce the cost of data transmission across the data center and improve the integral efficiency of a system.

The invention discloses a data management method based on rendering application data access characteristics. Different from a disk array storage mode used by the existing rendering application, the technical scheme adopts different storage designs for different data categories by summarizing the access data characteristics of the rendering application. Under the situation that the rendering application is enabled to operate normally, the memory of a rendering node is effectively used, and the data access time of the traditional rendering storage is reduced. Besides, data pre-access is taken into consideration aiming at the characteristics of frame data, and data required by rendering can be stored in almost all of the memory in the rendering operation process, so that the utilization rate of memory is improved. Meanwhile, the data management method is convenient to use as the design in the invention is flexible and feasible, the flexibility is embodied on memory allocation, and the available memory space is allocated for each node according to the size of a scene file; the feasibility is embodied in the aspect that the technical scheme in the invention is still available when the number of nodes is increased or reduced in the distributed environment.

The present invention includes a plurality of CPUs using memory as main memory, another function block using memory as a buffer, a CPU interface which controls access transfer from the plurality of CPUs to memory, and a DRAM controller for performing arbitration of the access transfer to the memory. Therein, the CPU interface causes access requests from the plurality of CPUs to wait, and receives and stores the address, data transfer mode and data size of each access, notifies the DRAM controller of the access requests, and then, upon receiving grant signals for the access requests, sends information to the DRAM controller according to the grant signals, whereupon the DRAM controller receives the grant signals, and on the basis of the access arbitration, specifies CPUs for which transfers have been granted so as to send the grant signals to the CPU interface.

The invention discloses a cooperative spectrum sensing method utilizing space diversity. The method comprises the steps that sensing nodes participating in cooperative spectrum sensing calculate power magnitudes of received signals, the sensing nodes participating in cooperative spectrum sensing send the power magnitudes of the received signals to a fusion center, the fusion center calculates an inspection statistics magnitude according to the power magnitudes of the signals received by the sensing nodes participating in cooperative spectrum sensing, and finally, the fusion center judges whether an authorized user signal exists in the signals received by the sensing nodes to achieve spectrum sensing by comparing the inspection statistics magnitude with a judgment threshold. The method has the advantages that the noise power magnitude does not need to be known, and the performance is close to that of the energy inspection method based on soft combination with the noise power magnitude known.

The invention discloses a data acquisition method and device based on a Spark computing framework. The method includes: after receiving the table object access request, obtaining the computing resource information of Spark and the data distribution information of the data table to be accessed in the MPP cluster; generating multiple Partitions according to the computing resource information and data distribution information; wherein, each A Partition corresponds to part of the data in the data table; through the generated multiple Partitions, the data table is obtained from the MPP cluster. The present invention makes full use of the data storage characteristics of the MPP cluster, and directly and quickly acquires data sets from the storage nodes of the MPP through multiple Partitions. Furthermore, in the case of sufficient computing resources, the data table of the storage node can be further split to achieve the purpose of increasing parallelism and improving data import performance. According to the data distribution of the MPP cluster, data can be obtained from local storage first, reducing data transmission overhead, saving network bandwidth, reducing network delay, and improving computing performance.