9results about How to "Eliminate overhead" patented technology

This invention relates to a method and system for multi-source information knowledge fusion and intelligent retrieval. The method includes: acquiring multi-source heterogeneous data and performing unified knowledge modeling on the multi-source heterogeneous data to obtain a unified knowledge framework. Based on the unified knowledge framework, a large language model is invoked to perform structured preprocessing on the multi-source heterogeneous data to generate structured knowledge triples. The structured knowledge triples are then structured using attribute graphs, and the associated data of each entity is vectorized and stored using an embedding model to obtain a fused knowledge body. User query requests are received and responded to by performing deep analysis on the user input data to obtain the user's query intent and query elements. A condition graph is constructed based on the user's query intent and query elements, and the condition graph is mapped to a knowledge network for entity localization and condition graph matching to generate structured retrieval results. This reduces network overhead and scheduling latency, and improves retrieval response speed and accuracy.

The application relates to the field of data calculation, and relates to a neural network operation method and device, a chip, electronic equipment and a storage medium. The neural network operation method comprises the following steps: acquiring input data and Wk*Hk sub-convolution kernel groups of neural network operation, and entering an operation step; N Wk*Hk*C convolution kernels of the neural network operation are split to obtain N*Wk*Hk 1*1*C sub-convolution kernels, and the N*Wk*Hk 1*1*C sub-convolution kernels are divided into Wk*Hk sub-convolution kernel groups; the operation step comprises the following steps: rearranging the input data according to a data rearrangement mode corresponding to each sub-convolution kernel group to obtain rearranged input data corresponding to each sub-convolution kernel group; the sub-convolution kernel groups and the rearranged input data corresponding to the sub-convolution kernel groups are convolved to obtain convolution results of the sub-convolution kernel groups; the convolution results of the sub-convolution kernel groups are accumulated to obtain an accumulated result, and data located at an effective position in the accumulated result is taken as an output result of the neural network operation. The hardware design overhead, the increase of data memory access and the increase of dynamic power consumption caused by img2col are eliminated.

The invention discloses a social network link prediction-oriented time sequence diagram network parallel training acceleration method. The method comprises the following steps: firstly, calculating a redundancy score and an old score for each user interaction edge in a training set, and calculating a comprehensive information loss score according to the redundancy score and the old score; secondly, according to a preset core set retention proportion alpha, selecting all user interaction side comprehensive information loss scores and upper alpha quantiles of repetitiveness as threshold values, and discarding user interaction with the comprehensive information loss scores lower than the threshold values, so that a simplified core set is obtained through single-time preprocessing; then carrying out adaptive batch division on the obtained core set, and dynamically determining an acceptable maximum user interaction number in each batch according to a redundancy and old comprehensive information loss score; and finally, training the time sequence diagram neural network based on the divided batches to realize social network link prediction. The method not only improves the training efficiency, but also greatly improves the precision of the model.

The invention discloses a server-free remote memory access performance optimization method based on cross-process memory tracking, and belongs to the technical field of computer memory management. The method comprises the following steps: designing a memory state table as a unified perception layer for remote memory access of homologous server-free containers, and realizing cross-process tracking of memory page states among different containers; a memory state table query process is embedded into hardware page table traversal, page state query is completed by hardware acceleration, and extra overhead on a memory access key path is eliminated; the remote memory access is merged and optimized based on the memory state table, cached local memory pages are reused in the container creation stage, the same remote memory access requests are merged in the operation stage, and the page missing exception frequency and the network bandwidth contention are reduced. According to the method, extra overhead caused by remote memory access in a server-free environment is effectively reduced, page missing abnormity and network bandwidth contention in a concurrent scene are greatly reduced, and the running performance and the system expandability of memory-intensive server-free applications are remarkably improved.

Several examples of using RF SerDes components for chip-to-chip interconnect multibus replicators are disclosed. In one example, the multibus replicator includes: an initiating write bus replicator unit for receiving write commands and write data from multiple write buses and transmitting write commands and write data using one or more RF SerDes transmitters; and a destination write bus replicator unit for receiving write commands and write data from RF SerDes transmitters and providing write commands and write data on multiple write buses.

The embodiment of the invention provides a message processing method and device, electronic equipment and a storage medium, and relates to the technical field of networks, the method comprises the following steps: applying for a memory space for an original message, reserving an extension head area in the memory space, storing data of the original message in the memory space, and storing the data of the original message in the extension head area; the initial position of data storage is identified by a data initial pointer, and the extended header area is located in front of the initial position pointed by the data initial pointer, then adding an outer-layer message header to the original message in response to the need, and if the extended header area meets the length requirement of the outer-layer message header, adding the outer-layer message header to the original message. If not, moving the data start pointer towards the start direction of the memory space by the length of the outer-layer message header, filling the outer-layer message header in the expanded header area vacated after moving, and packaging the original message into the target message, thereby remarkably reducing memory access operation, ensuring that enough space is available in the whole life cycle of the message, and improving the message access efficiency. And the certainty and efficiency of the message processing flow are improved.

This application provides a QoS control method and system based on IO demand prediction. The method includes: capturing the IO system call trajectory of a target AI task and extracting an enhanced IO feature vector, where the enhanced IO feature vector is a feature vector containing system call features and file metadata features; matching a pre-trained enhanced IO fingerprint library to identify the current execution stage; predicting the IO demand of the next stage based on a Hidden Markov Model (HMM), where HMM is a probabilistic model used for sequence state prediction; and directly calling the kernel's native block device layer QoS interface to update the token bucket parameters of the target AI task based on the prediction results, where the block device layer QoS interface refers to the token bucket parameter update interface provided by the Linux kernel's rq-qos module. This application achieves non-intrusive prediction of AI task IO demands and ultra-low latency dynamic control of block device layer bandwidth without modifying application code or kernel code.

This invention discloses an SSL VPN system and packet processing method based on a layered architecture. This solution establishes a completely decoupled architecture between the control plane and the data plane, configuring the SSL VPN process specifically for processing encrypted control layer packets, while configuring raw data layer packets to be directly delivered to an independent high-performance forwarding engine via a bypass mechanism. This invention provides multi-core support for high-volume, high-concurrency data packet scenarios while preserving SSL VPN security. Furthermore, it relies on the shared memory channel of the forwarding engine to avoid performance losses associated with CPU kernel-mode / user-mode context switching and packet copying processes in traditional forwarding models.