252 results about "Memory pool" patented technology

The invention discloses a compatible operation method for an Android camera HAL in a container based on memory access virtualization, and the method comprises the steps: taking a DMA-Buf memory heap of a Linux kernel host system as a target memory heap, taking an ION memory heap of an Android container system as a source memory heap, building a process exclusive memory management context, an FD cache table and a synchronous fence pool, and carrying out the execution of the process exclusive memory management context, the FD cache table and the synchronous fence pool; kernel registration and node binding of virtual ION equipment, pre-allocation of a first memory pool and access hook registration of kernel layer equipment are completed, and when an ION file descriptor is obtained in an HAL process, legality of a container process is verified, and context binding is initialized to the file descriptor; intercepting a memory allocation request of the HAL process, analyzing and adapting parameters, and preferentially multiplexing first memory pool resources to obtain an ION handle; when a data sharing request is processed, matching an FD cache or generating a new FD through an ION handle; when the HAL process releases the memory, resources are recycled according to a memory source, and when the HAL process exits, the context is cached or destroyed, so that cross-system memory operation compatibility is realized.

A real-time target detection method and system based on RTSP flow and NPU collaborative optimization belong to the technical field of computer vision and artificial intelligence, and are characterized by comprising the following steps: adopting dynamic memory optimization management of a hybrid pipeline architecture, and performing single-time continuous copying through a CPU (Central Processing Unit); through deep integration of innovative technologies such as DMA direct transmission NPU continuous memory pool management, dynamic batch processing scheduling, hybrid assembly line processing, intelligent equipment load balancing, parallel preprocessing optimization and vectorization post-processing, RTSP flow collaborative optimization and the like, the NPU utilization rate is improved to 85% or above, the overall average FPS is improved by 200% or above, the assembly line parallelism degree achieves three times of performance gain, and the production efficiency is greatly improved. The data transmission delay is reduced by 80%, the system stability is remarkably improved, performance degradation is avoided after long-time operation, and the method is suitable for various scenes such as edge calculation and cloud reasoning.

The invention relates to the technical field of memory expansion, and discloses a memory access system and method for a memory pool, the system is connected with PC I e equipment and CXL equipment, the PC I e equipment is used for sending an access request to a host so as to access the CXL memory pool through the host; the CXL equipment is used for sending a read-write command to the CXL switch so as to access the CXL memory pool through the CXL switch, the system comprises a host which is respectively connected with the PC I e equipment and the CXL switch, and the host is used for receiving an access request sent by the PC I e equipment, converting the access request into the read-write command and sending the read-write command to the CXL switch so as to access the CXL memory pool through the CXL switch, cXL equipment and PC I e equipment which are short in memory are expanded through the CXL memory pool, the memory access system of the memory pool can be compatible with the CXL equipment and the PC I e equipment at the same time, unified memory distribution and memory management are provided for the CXL equipment and the PC I e equipment, memory sharing between the CXL equipment and the PC I e equipment is achieved, memories of the CXL equipment and the PC I e equipment are expanded, and therefore the memory access efficiency and the memory utilization rate are improved.

The invention provides an RDMA-based big data transmission and storage system and method and a semiconductor detection device.The big data transmission and storage system comprises a hardware layer and a software layer, the hardware layer comprises an NVMe SSD storage array, an NVMe RAID acceleration card based on a GPU architecture and an RDMA network card, and the software layer comprises an NVMe SSD storage array, an NVMe RAID acceleration card based on the GPU architecture and an RDMA network card; the software layer comprises a multi-thread transmission module, a self-adaptive memory pool module, a transmission control module, a data verification and retransmission module, a scheduling and optimizing module and a disk falling and storage management module. According to the method, multi-channel transmission parallelization is achieved through data direct connection, zero copy transmission, RAID acceleration and multi-thread parallel RDMA transmission in a hardware link, data is directly connected and transmitted from a memory to the NVMe SSD storage array, data receiving and disk falling thread decoupling directly access the NVMe SSD storage array through asynchronous writing, network overload or storage blockage is avoided, and the data transmission efficiency is improved. The data transmission efficiency and reliability are high, and popularization and application in scenes needing large-scale image transmission processing such as semiconductor detection are facilitated.

Datacenter workloads demand flexible memory architectures spanning from rack-level to pod-scale deployments. Embodiments herein disclose systems enabling CXL memory semantics over physical layers based on IEEE 802.3 PMA, facilitating memory disaggregation across datacenter fabric infrastructures. The embodiments comprise processing cores with coherent interconnects, MMUs for address translation, and memory channels supporting substantial memory capacities. Resource Provisioning Units (RPUs) translate between CXL data, optionally encapsulated, transmitted via physical layers based on IEEE 802.3 PMA, and CXL requests, enabling external entities to access memory across different physical address spaces. This architecture provides memory pooling using datacenter network infrastructure, supporting intra-rack memory sharing, inter-pod memory access, distributed AI training across datacenter resources, and elastic memory provisioning for cloud-native applications, overcoming physical layer limitations of traditional CXL implementations while maintaining protocol coherency suitable for GenAL, LLM inference, and HPC workloads.

The invention discloses a multi-target tracking method and system for unmanned aerial vehicle small targets, and belongs to the technical field of computer vision. The method comprises the steps of obtaining multi-level feature maps based on a video frame image, and then unifying the multi-level feature maps into an initial fusion feature map; calculating to obtain a two-dimensional enhanced feature map based on the initial fusion feature map and the original image; calculating to obtain a spatial enhanced feature map based on the two-dimensional enhanced feature maps of the front frame and the rear frame; after splicing the two-dimensional enhanced feature maps of the front frame and the rear frame, sequentially performing feature alignment processing and memory pool interaction processing to obtain a time enhanced feature map of the current frame, and updating the memory pool; and fusing the spatial enhanced feature map and the time enhanced feature map, performing down-sampling and splicing, and further combining an output result with a ByteTrack tracking strategy to realize multi-target tracking. According to the method, the accuracy and identity consistency are remarkably improved, and meanwhile, efficient and real-time reasoning performance is kept.

The present disclosure provides a description of exemplary systems and methods for superseding pending transactions in a distributed ledger. The methods and systems may include a processor which may receive a first transaction associated with a digital asset in a memory pool of a distributed ledger network, the first transaction having a first priority. The processor may receive a second transaction associated with the digital asset in the memory pool of the distributed ledger network, the second transaction having a second priority. The processor may analyze the first priority of the first transaction and the second priority of the second transaction to determine a higher priority. In response to determining that the first priority of the first transaction is less than the second priority of the second transaction, the processor may validate the second transaction having the second priority.

The application discloses a memory allocation method without locking among multiple cores, comprising the following steps: dividing the memory into multiple local memory pools according to the number of CPU cores, wherein the number of the local memory pools is the same as the number of the CPU cores, and each local memory pool corresponds to a different CPU core, and each CPU core does not need to be locked when applying for memory resources in the multiple local memory pools; if the required memory resources of a second CPU core are less than the memory resources in a second local memory pool, the second CPU core applies for the memory resources from the second local memory pool; otherwise, the second CPU core continues to execute; if the required memory resources of the second CPU core are greater than the memory resources in the second local memory pool corresponding to the second CPU core, the second CPU core applies for and uses the memory resources from a first local memory pool corresponding to a first CPU core through multiple shared buffers. The application further discloses corresponding devices, electronic equipment and computer readable storage media, wherein in most scenarios, each CPU core fully uses the memory resources, and the performance is optimal; when cross-core application is required, the memory allocation is realized through the shared buffer mode without locking.

The invention provides a data processing method and system, and the method comprises the steps: responding to an access request carrying a target resource identifier, and querying a virtual address mapping relation corresponding to the target resource identifier; under the condition that the virtual address mapping relationship is not queried, determining a target resource type according to the target resource identifier, and determining an allocated target virtual address from a logic memory pool corresponding to the target resource type, the logic memory pool being obtained by performing resource type division in a virtual address space; a target memory allocation interface is called, a target resource corresponding to the target resource identifier is loaded to a target physical memory, a mapping relation between the target physical memory and an allocated target virtual address is established, a virtual address handle corresponding to the target virtual address is generated, and the target memory allocation interface integrates memory allocation interfaces of all host platforms; accessing the target resource according to the virtual address handle; through the design of virtualization and abstract interfaces, the cross-platform performance consistency can be ensured, and the development complexity can be reduced.

The invention relates to a UDP communication fault diagnosis and recovery method based on protocol stack health degree multi-dimensional perception, and belongs to the technical field of embedded systems and computer network communication. According to the method, aiming at a difficult fault that Ping communication is carried out between devices but UDP applications cannot carry out communication, an internal state of a network protocol stack is actively sensed through a built-in diagnosis agent module, and a traditional external packet capturing means is replaced. The diagnosis process comprises the steps of link layer reachability verification, protocol stack load channel detection, Socket interface state auditing and memory pool and task health degree inspection. The system constructs a protocol stack health degree vector through multi-dimensional sensing data, and performs fault mode matching based on a preset rule base. Once the fault root cause is positioned, the system can automatically trigger a repair action or report an accurate alarm. According to the invention, through automatic and intelligent diagnosis and self-healing embedded in the system, the reliability and maintainability of the embedded network equipment are greatly improved.

The present disclosure provides a video frame filtering method and device, electronic equipment and storage medium, belonging to the technical field of computer. The method comprises: for any video frame in a video frame queue, determining a plurality of reference video frames of the video frame from the video frame queue; in the case that the number of idle memory in the idle memory pool is not less than the first number, covering the data in the first memory in the idle memory pool with the video frame and at least one reference video frame, the idle memory in the idle memory pool being used to store the encoded video frames in the video frame queue in a pyramid structure; filtering the video frame based on the video frame in the first memory and the video frame in the second memory, the second memory being used to store at least one reference video frame in a pyramid structure. The above scheme can reduce the additional memory overhead by multiplexing the idle memory in the idle memory pool, avoiding frequent memory allocation, and thus improving the encoding efficiency of the video.

The invention discloses a Direct3D stream output virtual implementation method based on kernel enhancement, which comprises the following steps: in a DXVK compiling stage, constructing stream data for a shader, establishing a format mapping table and an application scene library of D3D, DXVK and a kernel, initializing a shared memory pool when the kernel is started, after the DXVK is started, establishing an abstraction layer consistent with VKEXTTransformFEDback, analyzing parameters when an application calls a stream output related API (Application Program Interface), and implementing the kernel enhancement based Direct3D stream output virtual implementation. The buffer binding type API triggers a kernel to distribute partitions for streams, the stream control type API triggers a corresponding shader to rewrite, the query operation type API extracts query related information, the execution shader writes data into a temporary delivery buffer, data distribution is completed when a distribution condition is met, a distribution channel of an abstraction layer is packaged into a Vulkan secondary command buffer, and the Vulkan secondary command buffer is used for sending the data to the execution shader. And reusing a DXVK rendering thread command pool and submitting the command pool to the GPU in batches to finish rendering.

This invention discloses a communication method, system, device, and computer-readable storage medium, applied in the field of network communication technology. The method includes: an I / O thread pool receiving messages sent by an external terminal device and loading the messages into a memory pool using zero-copy technology; sending the messages to an independent service thread pool using a serialization executor corresponding to the current communication connection; the independent service thread pool being a collection of task processing threads independent of I / O scheduling; the independent service thread pool performing protocol parsing on the messages, constructing the corresponding communication messages based on the parsing results, and distributing the communication messages to the corresponding lock-free circular queues; and the I / O thread pool communicating based on the priority of the communication messages in the lock-free circular queues using an atomic operation mechanism based on hardware primitives. This invention improves communication efficiency by introducing atomic variables based on hardware primitives to replace traditional mutex locks, achieving asynchronous serial processing within a single device and fully parallel processing between multiple devices.

The invention discloses a shared memory dynamic management method and device, equipment, a medium and a product. The method comprises the following steps: monitoring transmission parameters during service data transmission of a shared memory pool configured between a CPU (Central Processing Unit) and at least one heterogeneous computing power processor; and adjusting the memory size of the shared memory pool according to the transmission parameters. According to the scheme, the shared memory pool is configured, load-aware dynamic memory pool size adjustment is performed according to the transmission parameters during service data transmission, uplink and downlink data size difference and memory allocation under the actual working condition are considered, resource migration control in the base station can be realized, the memory use efficiency is greatly improved, and the resource utilization rate of the base station is improved. The system working performance is improved.

The application provides a charging facility fault early warning model construction method based on multi-source data perception, and belongs to the technical field of charging facility fault early warning.The application forms a multi-source time series data set by collecting charging pile operation state data and environmental data and performing time series alignment, expands fault samples by using an adversarial generative network and a semi-supervised clustering algorithm to form a balanced sample data set, constructs a fault early warning model containing a time series memory pool encoding layer and a frequency domain convolution feature extraction layer, cooperatively optimizes early warning accuracy and response time delay by using a double-layer game optimization framework to obtain an optimal parameter combination, deploys the optimized model to an edge computing module to realize real-time fault early warning, establishes an online incremental learning mechanism based on sliding time window detection data distribution offset, and uses an elastic weight consolidation technology to update model parameters and retain historical knowledge, thereby solving the problem of early warning performance degradation of charging facility fault early warning in a data distribution evolution environment.

The invention discloses an illumination rendering method and device, a storage medium and electronic equipment. The method comprises the following steps: in response to the change of a target scene, obtaining a target bounding box needing to be updated in the target scene, and dividing the target bounding box into sparse blocks to obtain a group of sparse blocks; allocating a memory for each sparse block in the group of sparse blocks, determining a target sparse block index, and creating an effective surface mapping texture in a memory pool based on the target sparse block index; compressing and storing texture data corresponding to the effective surface of each voxel based on the effective surface mapping texture to obtain an effective surface image set; and performing illumination calculation on the target sparse block based on the light source framing in the target scene to obtain illumination information, and storing the illumination information in an effective surface image set based on the target sparse block index and the effective surface mapping texture. According to the method and the device, the technical problem that the utilization rate of a storage space is relatively low due to the fact that idle sparse blocks and voxels need to be stored in an illumination rendering process is solved.

A method, according to one embodiment, includes obtaining information generated by a hypervisor, where the generated information details a classification process. The classification process includes classifying physical memory blocks as residing on a first server or a second server, and associating the physical memory blocks with logical memory blocks of logical partitions of the first server and logical partitions of the second server. The method further includes providing the generated information to an operating system used to manage the logical partitions of the first server and the logical partitions of the second server. The operating system is caused to use the information to classify memory pools to a home memory cluster of the first server or to a remote memory cluster of the second server, and memory pages are caused to be dynamically stored within allocations of the memory pools based on predetermined usage statistics.

Techniques are directed to performing a file system check. Such technique involve determining a plurality of unreleased memory pools for a plurality of phases of a file system check by parsing a resource file indicative of an allocation situation of a memory. Such technique further involve determining a list of phases based on the plurality of unreleased memory pools. Such technique further involve performing the file system check based on the list of phases. Accordingly, it is possible to pre-determine an unreleased memory pool based on a resource file and perform a file system check according to the information of the memory pool, which avoids the problems of memory fragmentation and consecutive out-of-memory caused by the unreleased memory pool, thereby solving the out-of-memory error during the execution of the file system check.

The application discloses a multi-agent collaborative reasoning method and system based on memory sharing and conflict resolution, comprising: constructing a hybrid structured memory pool containing a vector database for semantic recall and a graph database for storing logical structure; deploying functionally heterogeneous multi-agent to normalize original multi-source data into observation nodes and write into the graph database; using the graph database trigger mechanism to listen to the writing event, and automatically triggering the reasoning to generate the hypothesis node and establish the deduction edge when matching the preset rule; detecting whether there is a logical conflict for the hypothesis node of the same subject, and triggering dynamic arbitration if there is; the arbitration mechanism calculates the weighted score based on the confidence, data source credibility and context consistency calculated based on KL divergence, and finally outputs the reasoning conclusion and evidence chain path. The application breaks the information island through the hybrid memory pool, realizes the automatic reasoning by using the trigger, and significantly improves the accuracy, logical consistency and explainability of reasoning in complex scenarios.

Provided are systems, methods, and apparatuses for stream processing architecture for memory pools. In one or more examples, the systems, devices, and methods include receiving a memory transaction request that includes a request header and an input data stream; determining that the request header indicates a function to execute based on the input data stream; determining a location to execute the function based on a location indicator of the request header, the location indicator indicating a source device of the input data stream, a destination device of an output data stream, or a switch that processes data communication between the source device of the input data stream and the destination device of the output data stream; and generating the output data stream based on executing the function in relation to the input data stream.

This application provides a model loading method and a terminal device, relating to the field of terminal technology. The method includes: loading at least a portion of the model files of a preset language model into a first memory pool when no preset language model is running; receiving a first user instruction to trigger a voice assistant function of a first application; and, in response to the first user instruction, running the preset language model while all model files of the preset language model are loaded into the first memory pool, generating a semantic reasoning result for the first user instruction using the preset language model.

The application discloses a memory management method and device, relates to the technical field of heterogeneous computing, and comprises the following steps: obtaining a memory allocation request of a device, and searching for a page-locked memory block corresponding to the memory allocation request to allocate the memory block to the corresponding device; receiving a request for releasing the memory block, identifying a device computing flow associated with the memory block, and recording a synchronization event for each computing flow; storing the recorded synchronization event and metadata of the memory block in a to-be-processed event queue; querying whether an event in the to-be-processed event queue is completed; and when it is queried that a specific event is completed, recycling the corresponding memory block to an idle memory pool, so that the technical problem that in the related art, high-cost memory management function calls are too frequent, performance is low, and it is difficult to cooperate with device asynchronous operation, and data competition and errors are easily caused is solved, and the technical effect that the efficiency of memory allocation and the overall throughput of a system are greatly improved under the premise of ensuring memory safety is achieved.

The invention relates to the technical field of robot model deployment and heterogeneous computing system architecture, in particular to a robot sensing-reasoning inter-node communication system and method based on a page lock memory. The system comprises a sensing node, a shared memory management module and a reasoning node. And the shared memory management module is used for constructing and managing a page lock memory pool and mapping the page lock memory pool to process address spaces of the sensing node and the reasoning node. And the sensing node is used for acquiring the memory block from the pool and directly writing the acquired sensor data into the pool to realize zero-transfer writing. And the reasoning node is used for asynchronously loading the data to the graphics processor video memory through direct memory access operation according to the address information when the data of the target memory block is ready. According to the invention, high-bandwidth, low-delay and high-determinacy data transmission without redundant copy between robot sensing-reasoning nodes is realized.

The invention discloses an infinite context large model processing method and device based on lexical element memory. The method comprises the following steps: firstly, segmenting an ultra-long text into text paragraphs with fixed lengths; secondly, constructing a local attention module based on a decoder type Transform, calculating text paragraphs, and outputting a local attention result and a related vector; and thirdly, constructing a token lexical element memory pool to maintain each lexical element memory unit, and calculating cross-paragraph global attention and outputting a result in combination with the query vector. And fourthly, fusing local and global attention results to realize an infinite context. And 5, based on the correlation vector, incrementally updating the lexical token memory pool and the normalization item, and providing support for subsequent paragraph processing. And 6, constructing a paragraph memory module, and storing the hidden state of the last lexical element of the paragraph for generating the first lexical element of the next text paragraph. The problems that in the prior art, when super-long texts are processed, calculation and storage expenses are large, and semantic information is prone to being lost are solved.

A system for managing non-transient memory in a cloud computing environment, comprising a plurality of data processors configured to cooperatively provide a cloud computing environment, a persistent memory pool system configured to interact with each of the plurality of data processors to identify persistent non-transient data memory devices at each of the data processors and a plurality of memory pools created by the persistent memory pool system, wherein each of the plurality of memory pools has a designated function.

The application discloses a storage engine control method and system supporting large field data and a storage medium, constructs a memory pool for dynamically allocating and releasing memory blocks, the memory pool comprises a plurality of memory blocks stored in a linked list, and the memory block comprises a pointer pointing to a memory position; when large field data needs to be stored, a memory block with a corresponding size is allocated on demand. According to the large field data to be processed, a corresponding table structure and data type are designed; an index structure and a query strategy are determined; the query method comprises insert and find methods, for the large field data type, a B+ tree index structure is used for range query, and a hash index structure is used for point query; a data compression and cache strategy is determined, which is used for reducing memory occupation and improving query performance. The application successfully solves the problem that the memory storage engine in the prior art does not support large field data, improves the SQL penetration performance of the internal and external networks, and has good practicability.

A workload-based network-accessible memory pool provisioning system includes a computing device. A memory pool provisioning subsystem in the computing device identifies a workload type for a workload and, based on the workload type, configures a network-accessible memory pool using network-accessible memory devices coupled to the computing device via a network, and exposes the network-accessible memory pool for use in performing the workload. An operating system provided by the computing device uses the exposed network-accessible memory pool to perform the workload using a first subset of the network-accessible memory devices. A memory orchestration subsystem in the computing device monitors the performance of the workload by the operating system and, in response, modifies the performance of the workload by the operating system such that the operating system performs the workload using a second subset of the network-accessible memory devices that is different than the first subset of the network-accessible memory devices.

The invention relates to the field of memory management, in particular to a concurrent management method of a multi-thread memory pool. A plurality of independent memory pools are maintained, each memory pool corresponds to a preset memory block size, an idle linked list composed of idle memory blocks is maintained in each memory pool, and each idle linked list is provided with a head pointer used for memory block allocation and capacity expansion and a tail pointer used for memory block recovery. Memory block allocation is realized only by operating a head pointer from the head of an idle linked list, and memory block recovery is realized only by inserting the tail of the idle linked list, so that memory block allocation and recovery operations are executed concurrently under a lock-free condition; and setting a first mutual exclusion lock and a second mutual exclusion lock which are separated for each memory pool, wherein the first mutual exclusion lock and the second mutual exclusion lock respectively protect concurrent modification of the head pointer and the tail pointer. The method is simple in structure and low in cost, the number of times of lock competition and the memory fragmentation rate are remarkably reduced, and the memory allocation efficiency and the system throughput under the multi-thread environment are improved.

The invention discloses a zero-copy serialization method and device based on memory mapping, equipment and a medium, and relates to the field of data processing. In the method, a root object to be serialized is received, the root object and sub-objects are analyzed, and a field to be serialized is identified; constructing an object graph containing the reference relationship between the objects; in the object graph, recursively calculating the data size required for serializing the root object and the sub-object to obtain the total serialization size; according to the total serialization size, selecting and distributing a corresponding memory block from a preset layered memory pool, creating a memory mapping file corresponding to the memory block, and establishing a memory mapping buffer area based on the memory mapping file; traversing the object graph, and writing data of the root object and the sub-object into a memory mapping buffer area; the total length of the data is actually written into the memory mapping buffer area, and the total length of the data is written back into the reserved head area of the memory mapping buffer area. By implementing the technical scheme provided by the invention, the data serialization efficiency is improved.

The invention discloses a cloud-edge fusion storage system supporting federated learning, and belongs to the field of cloud-edge cooperative control. An edge client comprises a CPU, a GPU, an RAM and a CSD; the RAM and the CSD are organized into a memory pool based on a CXL protocol; in each round of federated learning, the edge client side executes the following steps: acquiring global update data from the central cloud side and storing the global update data into the memory pool; the global update data comprises an updated global model and a global sample accumulative gradient; training the updated global model by using the local data set to obtain an updated local model, and forming local update data by the updated local model and the local sample cumulative gradient; the importance score sequence of the samples in the local data set does not exceed a preset sample quantity threshold value; and uploading the local update data to a central cloud, so that the central cloud aggregates the local update data of each edge client into global update data. According to the invention, the storage efficiency and the I / O performance of the edge client in a stream data federated learning scene can be improved.