23 results about "Pipeline scheduling" patented technology

The invention relates to the technical field of artificial intelligence, and provides a model distributed training optimization method, electronic equipment and a storage medium, and the method comprises the steps: respectively constructing a calculation flow and a communication flow based on the calculation operation and the communication operation in a model; in the training iteration of the model, executing staggered pipeline scheduling for at least two micro-batches of tasks; wherein the staggered pipeline scheduling comprises the following steps: scheduling a first micro-batch of calculation tasks to a calculation flow for execution, and scheduling a second micro-batch of communication tasks to a communication flow for execution, or scheduling a first micro-batch of communication tasks to the communication flow for execution, and scheduling the second micro-batch of calculation tasks to the calculation flow for execution, the tasks of the first micro-batch and the tasks of the second micro-batch are processed in parallel. According to the method, the parallel processing of the calculation task and the communication task is realized by executing the staggered assembly line scheduling on the tasks of different micro-batches, and the hardware utilization rate and the model training efficiency are improved.

The invention provides a hybrid pipeline scheduling method and device for distributed deep learning training. The invention relates to the technical field of artificial intelligence, in particular to the technical fields of distributed training, machine learning, deep learning and the like. According to the specific scheme, training data are divided into N continuous data sequences, and each data sequence comprises K micro-batches; performing calculation on the first N-1 data sequences by adopting a forward-reverse interleaving scheduling strategy; performing calculation on the Nth data sequence by adopting an interleaved 1F1B scheduling strategy; wherein the calculation is executed on the Nth data sequence by adopting an interleaved 1F1B scheduling strategy, and the calculation comprises the following steps of: for any micro-batch in the Nth data sequence, immediately triggering reverse calculation after the forward calculation is finished; and for any micro-batch, immediately releasing the video memory occupied by the activation value generated by forward calculation of the micro-batch after the reverse calculation of the micro-batch is completed.

This disclosure provides interleaved pipeline scheduling methods, apparatus, devices, and storage media, particularly relating to the computer technology field, and more particularly to the artificial intelligence technology field, including deep learning, neural networks, and parallel computing. [Solution] A specific solution includes obtaining micro-batch scheduling parameters based on the pipeline splitting dimension, interleave dimension, and cumulative count; determining the transmission interval of a micro-batch requiring cache scheduling based on the micro-batch scheduling parameters; and performing cache scheduling on the computation results of the micro-batch requiring delayed transmission based on the transmission interval and cache units. According to this disclosure, a cache unit, for example, a cache queue, can be created, the computation results of some computation units can be cached in the cache unit, and the computation results can be transmitted with a delay. This allows for delayed transmission of the computation results of some micro-batches, improving the applicability of interleaved pipeline scheduling.

The embodiment of the invention provides a model training method, a model training device, electronic equipment and a storage medium. The model training method comprises the steps that in response to acquired insertion layer configuration corresponding to a first model, insertion layer information is determined according to the insertion layer configuration, the insertion layer configuration comprises parallel strategy information, and the insertion layer information comprises the number of insertion layers, the type of the insertion layers and insertion layer position information; updating the first model according to the insertion layer information to obtain a second model; and performing staggered assembly line parallel training on the second model. According to the method, performance measurement decoupling during operation can be achieved, complex search and analysis are avoided, the availability of staggered assembly line scheduling can be guaranteed in the composition stage, the flexibility and compatibility of the system are improved, support is provided for high-performance training under different model structures and different parallel configurations, and the training efficiency is improved.

The application provides a container pipeline scheduling management method and system based on an ARM+DSP heterogeneous multi-core processor and a medium, and the method comprises the following steps: configuring scheduling related attributes of a container, wherein the scheduling related attributes of the container comprise a container duration time DT and a container interval time IT, and the container comprises an ARM container and a DSP container; according to the scheduling related attributes of the container and a pre-established container linked list, a next container to be run is selected according to a time window static scheduling strategy, and the container pipeline is controlled to run. Compared with the prior art, the application realizes independent running of various containers under the premise that the containers are sequentially arranged, do not affect each other and can interact with each other under the condition of sharing physical resources.

The application provides an MLBM-DEM pipeline scheduling optimization method for a heterogeneous multi-domain processor. The method comprises the following steps: S1, performing data division based on application perception and the heterogeneous multi-domain processor, respectively calculating the calculation amount of the MLBM and the DEM, and the calculation capability proportion of the MLBM and the DEM; S2, allocating the calculation resources of the CPU end and the acceleration domain cluster according to the calculation capability proportion of the MLBM and the DEM, and constructing a pipeline design module for the heterogeneous multi-domain processor; S3, obtaining the remaining calculation resources except the CPU end and the acceleration domain cluster, and performing a Manager-Worker real-time scheduling mechanism. The application realizes MLBM-DEM pipeline scheduling optimization calculation under the condition of ensuring calculation accuracy, and further improves the overall calculation performance.

The application provides a data transmission module, a control method and an optoelectronic hybrid computing system. The data transmission module is applied to a runtime subsystem of the optoelectronic hybrid computing system, and the data transmission module is used for providing a data transmission management interface for the runtime subsystem. The data transmission module comprises a data transmission interface unit, a memory management unit, a direct memory access control unit and a pipeline scheduling unit. The data transmission interface unit is used for receiving a transmission call of the runtime subsystem. The memory management unit is used for managing allocation and release of on-board memory of a field programmable gate array. The direct memory access control unit is used for controlling access transmission between a graphics processor and the field programmable gate array. The pipeline scheduling unit is used for pipeline execution of data transmission and optoelectronic computing. In this way, the data transmission between the GPU and the FPGA can be managed by the data transmission module, and a complete transmission process of GPU-FPGA-GPU can be supported.

The invention provides a Simulink code generation method and device based on instruction pipeline perception, and relates to the technical field of embedded code generation, and the method comprises the steps: carrying out instruction dependency analysis and delay modeling on a target Simulink model, and generating a dependency graph for describing a dependency relationship between modules and an instruction delay model combined with target processor system structure parameters; based on the dependency graph and the instruction delay model, determining an optimal execution sequence of each module through assembly line perception scheduling, and performing assembly line scheduling based on the optimal execution sequence to obtain a scheduling result; and based on the scheduling result, utilizing a local pipeline optimization strategy to generate a target code. According to the method provided by the invention, the data dependence path between the modules is determined by utilizing topology analysis, and the execution sequence is adaptively selected through the minimum penalty priority algorithm to avoid unnecessary waiting and resource conflicts, so that the number of pauses of the instruction assembly line is effectively reduced on the premise of ensuring the semantic consistency of the model.

The invention belongs to the technical field related to processor architectures, and discloses a CPU-GPU efficient collaboration method and system oriented to large-scale graph random walk. The method comprises the following steps: calculating an expected value of a step length required by an original walk task currently to be transmitted to a partition of a GPU to jump out of the partition, and calculating a performance ratio rho between explicit batch transmission and implicit zero-copy transmission of the partition based on the expected value, if rho gt; the method comprises the following steps: 1, selecting an implicit zero-copy transmission mode to transmit the data of the partition to a GPU for random walk, otherwise, adopting an explicit batch transmission mode to transmit the data of the partition to the GPU for random walk. According to the method, the optimal transmission mode can be adaptively selected, redundant transmission is reduced, the throughput is improved, and the execution efficiency of random walk is improved. In addition, the invention further provides the design of a pipeline scheduling mode and a cooperative computing mode, so that the execution efficiency of the system is further improved.

The invention relates to the field of storm forecasting, and discloses a parallel calculation storm surge set numerical forecasting method and device which are used for improving the efficiency, accuracy and robustness of storm surge set numerical forecasting. Through a dynamic resource allocation strategy, computing nodes are migrated in real time according to historical variance priority signals, computing resource utilization is optimized, a double-assembly-line scheduling mechanism is constructed, prediction and correction tasks are alternately executed, assembly line waiting is eliminated in combination with hardware-level synchronization signals, and generation of a random disturbance field is accelerated through an FPGA coprocessor. And the physical rationality of a disturbance field is ensured, an entropy comparator is introduced to detect abnormal data in real time, calculation is rapidly recovered through a process thermal migration technology, a high-precision storm surge water level field time sequence is generated in combination with an MPI parallel framework, and a forecast result is output. According to the method, high efficiency, accuracy and robustness of storm surge forecasting are realized through technical innovation, and powerful technical support is provided for disaster prevention and reduction.

The application discloses a method, device, computer storage medium and terminal for inventory emergency management. The embodiments of the present disclosure determine product oil production and sales constraints and inventory management constraints by using product oil management related information such as node supply and demand, channel transportation, node inventory and pipeline correlation, couple inventory emergency management and production and sales to a product oil logistics daily scheduling model, consider product oil logistics planning and inventory emergency management requirements, and thus can obtain a refinery delivery execution plan, an oil depot demand execution plan, a product oil logistics plan, a pipeline scheduling plan and a node inventory plan that meet various constraint conditions, effectively improve the utilization rate of product oil demand prediction, guarantee the oil depot inventory level, achieve the economic and practical purposes of the product oil logistics plan, and improve the efficiency of the product oil inventory emergency management.

The invention discloses a natural gas pipeline network scheduling method, electronic equipment and a storage medium, relates to the technical field of natural gas pipeline scheduling, and aims to determine a better natural gas pipeline network scheduling scheme. The method comprises the following steps: obtaining a linearized mathematical model; solving the target function based on the flow constraint and the operation flow constraint of the compressor at the supercharging node to obtain an initial optimization result; according to the first search step size, based on the initial optimization result and a penalty sequence linear programming algorithm, solving the linearized mathematical model to obtain a first optimization result; according to a second search step length, based on the first optimization result and a penalty sequence linear programming algorithm, solving the linearized mathematical model to obtain a second optimization result; based on the third search step size, the second optimization result and a penalty sequence linear programming algorithm, solving the linearized mathematical model to obtain a third optimization result; and determining a natural gas pipeline network scheduling scheme based on the third optimization result.

This invention relates to the field of artificial intelligence technology, providing a distributed training optimization method, electronic device, and storage medium for a model. The method includes: constructing computational and communication flows based on computational and communication operations in the model; and performing interleaved pipeline scheduling for at least two micro-batches of tasks during model training iterations. The interleaved pipeline scheduling includes: scheduling computational tasks from a first micro-batch to the computational flow for execution, while simultaneously scheduling communication tasks from a second micro-batch to the communication flow, or scheduling communication tasks from the first micro-batch to the communication flow for execution, while simultaneously scheduling computational tasks from the second micro-batch to the computational flow for execution, so that the tasks from the first and second micro-batches are processed in parallel. This invention achieves parallel processing of computational and communication tasks by performing interleaved pipeline scheduling on tasks from different micro-batches, improving hardware utilization and model training efficiency.

The invention particularly relates to an Internet of Things floating point time sequence data compression transmission and decompression reading method based on a GPU. According to the scheme, an asynchronous pipeline scheduling mechanism driven by a lightweight asynchronous event is adopted in the data batch transmission and compressed data flow return process between a CPU and a GPU, so that the problem of high transmission delay caused by data transmission in an existing GPU compression scheme is solved; on the other hand, an accurate and rapid floating-point number-to-integer conversion method with theoretical guarantee is provided, and the problem of precision loss existing in a traditional rapid conversion method is solved; besides, a bit-plane adaptive coding scheme is also designed, so that an optimal storage strategy can be dynamically selected for different parts of the data, and the negative influence of an abnormal value on the compression ratio is effectively overcome. According to the scheme provided by the invention, on the premise of ensuring that the data is absolutely lossless, the compression and decompression speed obviously superior to those of the existing CPU and GPU schemes can be realized, and the data communication transmission compression ratio and throughput in the Internet of Things system can be obviously improved.

The application relates to the field of storm prediction, and discloses a storm surge ensemble numerical prediction method and device based on parallel computing, which are used for improving the efficiency, accuracy and robustness of storm surge ensemble numerical prediction. Through a dynamic resource allocation strategy, a historical variance priority signal is used to migrate computing nodes in real time, computing resource utilization is optimized, a double-pipeline scheduling mechanism is constructed, prediction and correction tasks are alternately executed, pipeline waiting is eliminated by using a hardware-level synchronization signal, a random disturbance field is generated by using an FPGA coprocessor for acceleration, the physical rationality of the disturbance field is ensured, an entropy value comparator is introduced to detect abnormal data in real time, and computing is quickly recovered through process live migration technology, a high-precision storm surge water level field time series is generated by combining an MPI parallel framework, and a prediction result is output. Through technical innovation, the application realizes the efficiency, accuracy and robustness of storm surge prediction, and provides strong technical support for disaster prevention and reduction.

This invention relates to an automated operation and maintenance method and system based on pipeline scheduling, comprising: configuring a pipeline; generating tasks through a pipeline scheduling service; and executing tasks, including: the pipeline execution service obtaining node tasks from an execution queue and executing automated operation and maintenance operations according to the node tasks; the pipeline execution service sending the results of the automated operation and maintenance operations to the pipeline scheduling service; the pipeline scheduling service receiving the node task execution results; and proceeding to the next step based on the node task execution results and task status. This method and system can greatly expand automated operation and maintenance scenarios, reduce manual operations in various scenarios, and improve the level of automated operation and maintenance.

The application relates to a general-purpose GPU-oriented JPEG decoding batch processing and double-buffer pipeline scheduling method, belongs to the field of heterogeneous computing and image processing, and comprises the following steps: obtaining code stream data of a to-be-decoded JPEG image; performing entropy decoding and inverse quantization on the code stream data by a CPU to obtain a plurality of DCT coefficient blocks; if the size of the to-be-decoded JPEG image is not smaller than a preset size threshold, the DCT coefficient blocks are cached in a ring buffer located in a host memory; the state of the ring buffer is continuously monitored; when a preset batch triggering condition is met and a GPU is detected to be available, the DCT coefficient blocks in the ring buffer are batch copied to GPU display memory; based on the DCT coefficient blocks in the GPU display memory, integer IDCT transformation and color space conversion calculation are performed by the GPU to obtain an original image corresponding to the to-be-decoded JPEG image, and the original image is read back to the host memory; wherein the double-buffer pipeline is adopted to perform batch copying of the DCT coefficient blocks, integer IDCT transformation, color space conversion calculation and reading back of the original image. Efficient real-time decoding is realized.

A method for configuring a pipeline scheduling strategy, a pipeline control terminal, and a pipeline system. The method includes: determining a processing content of a scheduling strategy; the processing content of the scheduling strategy including a target function determined from processing functions provided by a pipeline; determining a target sample applying the scheduling strategy; and configuring the scheduling strategy according to the processing content and the target sample, where the scheduling strategy is used to instruct a processing module with the target function on the pipeline to process the target sample.

Some embodiments provide systems and methods for gas pipeline pathway analysis, control and communication. In some embodiments, a physical pipeline scheduling and control system comprises a server or computer-based computing system configured to receive a user input corresponding to a request for a gaseous resource, automatically obtain data from a plurality of data sources, automatically simulate a plurality of potential pathways of the gas pipeline network, automatically output a pathway recommendation comprising at least a first recommended pathway, receive a user selection of a selected recommended pathway, automatically format a nomination for the selected recommended pathway, and automatically access and control a separate electronic submission system and electronically submit the nomination for the request for the gaseous resource to the electronic submission system for execution.

The application relates to the technical field of data governance, in particular to a distributed data governance system based on a new type of scheduling algorithm of a pipeline. The system comprises a basic framework unit, a pipeline scheduling unit, a unified service unit and a data management unit; the basic framework unit is used for managing a network framework supporting system operation; the pipeline scheduling unit is used for managing scheduling of an inter-process communication pipeline; the unified service unit is used for unified standardized management of data; and the data management unit is used for comprehensive governance of data. The application can realize distributed data governance, improve the fusion degree of the system and enhance the extensibility of the system; can improve the transmission effect of a data exchange process, improve the integrity and security of data; through centralized scheduling management of data, the framework is convenient for node expansion of the system, can support high-concurrency access of a large user, and can realize visual management of a data full life cycle, shorten a data management cycle and reduce cost waste.

The application discloses a method and device for processing transportation information, a method and device for constructing a transportation model, and establishes a logistics submodel and a pipeline scheduling submodel, and the logistics submodel and the pipeline scheduling submodel are coupled by transportation connection relations and associated constraint conditions, a logistics optimization model obtained by the coupling can simultaneously consider a pipeline and a logistics path of product transportation, and transportation plan information of products to be transported is obtained by operation of the logistics optimization model, thereby providing technical support for improving product transportation efficiency.

To provide a pipeline scheduling policy setting method, a pipeline control terminal, and a pipeline system.SOLUTION: The method includes the steps of determining processing content of a scheduling policy that includes a target function determined from processing functions provided by a pipeline; determining a target sample to which the scheduling policy is applied; and setting a scheduling policy that instructs a processing module of the target function of the pipeline to process the target sample based on the processing content and the target sample. According to the method, it is possible to flexibly specify different processing methods on the pipeline for different samples and to increase the degree of intelligence of pipeline scheduling.SELECTED DRAWING: None

Embodiments of the present disclosure provide a model training method, a model training device, an electronic device and a storage medium. The model training method comprises: in response to obtaining an insertion layer configuration corresponding to a first model, determining insertion layer information according to the insertion layer configuration, wherein the insertion layer configuration comprises parallel strategy information, and the insertion layer information comprises an insertion layer number, an insertion layer type and insertion layer position information; updating the first model according to the insertion layer information to obtain a second model; and performing staggered pipeline parallel training on the second model. The method can decouple from runtime performance measurement, avoid complex search and analysis, ensure the availability of staggered pipeline scheduling in the composition stage, improve system flexibility and compatibility, support high-performance training under different model structures and different parallel configurations, and improve training efficiency.