46 results about "Massively parallel computation" patented technology

A method and system for performing network transfer adaptive optimization in a large-scale parallel computing system. The method of network transfer adaptive optimization includes forming a message to be transferred through the network based on obtained information related to a task executed by computing nodes, wherein the message includes identification information of the computing nodes to perform data transfer and corresponding sink nodes, and an amount of data the computing nodes to transfer to the corresponding sink nodes; transferring the message to a network layer; and forming a new data transfer pattern for data transfer between the computing nodes and the corresponding sink nodes in accordance with the received message.

The invention discloses a hypersonic transverse flow transition prediction method considering a surface roughness effect. The hypersonic transverse flow transition prediction method is based on the fact that a critical transverse flow Reynolds number and surface roughness meet a logarithmic relationship, and based on hypersonic wind tunnel experimental data, transition critical momentum thicknessReynolds numbers of different roughness under a hypersonic condition are obtained through CFD layer solution, a least square method is adopted to solve a relation coefficient, a criterion relational expression is obtained, and prediction of transverse flow transition is achieved by adding a transverse flow source item DSCF into an existing gamma-Retheta t transition model momentum thickness transport equation. Due to the fact that a gamma-Retheta t transition model is a localization model, the transverse flow criterion ReSCF is obtained through iteration, the momentum thickness Reynolds numberdoes not need to be solved through integration, and localization is achieved. Therefore, the method does not relate to calculation or calling of non-local amount, the prediction technology achieves complete localization, and the method is suitable for large-scale parallel calculation.

The invention relates to the field of parallel computation, and discloses a parallel computation dispatch method in the heterogeneous environment. According to the method, a plurality of JVM task slots having different internal storages, and idle task slot sets are built, tasks in the parallel computation are divided into I/O intensities and CPU intensities, the tasks are appointed to suitable task slots for calculation, and parallel calculation efficiency in the heterogeneous environment is optimized. The parallel computation dispatch method has the advantages that the sizes and the types of the internal storages required by the tasks are determined dynamically, resource using efficiency of heterogeneous clusters is improved, total operation time of the parallel calculation work is shortened, and the situation that the internal storages overflow in a task operation process is avoided.

This invention discloses a large scale parallel computer system sectionalizing starting method for solving the technology problem of how to reducing starting time based on large scale of parallel computer system with CC_NUMA structure. Technology project is characterized by the following steps: dividing all nodes into groups, setting a sort of node controller to control them to choose BSP in parallel and by stages in Firmware and their initialization and allocation, including three steps of start namely choosing BSP, LBSP and GBSP. The invention has the advantages of reducing system starting time, improving system usability and getting conducive to expanding the system structure.

A method and apparatus for selectively parking routers used for routing traffic in mesh interconnects. Various router parking (RP) algorithms are disclosed, including an aggressive RP algorithm where a minimum number of routers are kept active to ensure adequate network connectivity between active nodes and/or intercommunicating nodes, leading to a maximum reduction in static power consumption, and a conservative RP algorithm that favors network latency considerations over static power consumption while also reducing power. An adaptive RP algorithm is also disclosed that implements aspects of the aggressive and conservative RP algorithms to balance power consumption and latency considerations in response to ongoing node utilization and associated traffic. The techniques may be implemented in internal network structures, such as for single chip computers, as well as external network structures, such as computing clusters and massively parallel computer architectures. Performance modeling has demonstrated substantial power reduction may be obtained using the router parking techniques while maintaining Quality of Service performance objectives.

The invention belongs to the field of integrated circuit design, and in particular relates to a thermal-expandability three-dimensional parallel cooling integration method. The three-dimensional vertical integration technology proposed at present has disability in thermal expansion. The invention provides a thermal-expandability three-dimensional parallel cooling integration method. In the figure 1 described in the specification, all device layers are parallel with a cooling direction, each device layer is in a shape of a strip, short sides of the strip are parallel to the cooling device, and long sides of the strip are vertical to the cooling device, therefore each device layer is ensured to obtain an independent and shorter cooling channel by virtue of a high-thermal-conductivity silicon substrate (instead of thermal-conducting through holes) installed on the device layer and the fact that the highest temperature of a thermal-expandability three-dimensional parallel cooling integrated chip has no relation with the number of overlapped device layers is ensured. The invention also provides an analysis model for calculating the highest substrate temperature of the thermal-expandability three-dimensional parallel cooling integrated chip, and an analytic expression of the highest substrate temperature of a three-dimensional chip is derived, therefore the thermal expandability of the method disclosed by the invention can be proved theoretically. The method can be widely applied to a three-dimensional integration scheme-based massively parallel computation on-chip system needing good cooling performance urgently.

A memory management mechanism a nodal having multiple processors in a massively parallel computer system dynamically configures nodal memory on demand. A respective variable-sized subdivision of nodal memory is associated with each processor in the node. A processor may request additional memory, and the other processor(s) may grant or veto the request. If granted, the requested memory is added to the subdivision of the requesting processor. A processor can only access memory within its own subdivision. Preferably, each subdivision contains a daemon which monitors memory usage and generates requests for additional memory.

The invention provides a method and device for achieving OLAP analysis based on ClickHouse. The method comprises the steps of receiving kafka data in real time through a kafka engine table of the ClickHouse; carrying out data partitioning on the kafak data; carrying out TTL management on the partitions; carrying out data fragmentation; indexing a primary key; generating dynamic code. Kafka data isreceived in real time through a kafka engine table of ClickHouse, big data processing is conducted in a database mode, the method is simpler and easier to use, maintenance is easy, the query speed ishigh, through data partition TTL management, the problem of data storage resource wasting is solved, through data fragmentation, the large-scale parallel computing capacity of a cluster is improved,and a query result is quickly returned. And moreover, transverse linear expansion can be realized, and a large-scale distributed cluster is constructed, so that the capacity of processing mass data isachieved.

A high-speed parallel true random number generator with extra-strong scalability is characterized in that a pulse sequence is inputted to an array waveguide grate by a super-continuous spectrum entropy source, is cut into N narrow-band sub optical pulse sequences, enters a dispersion compensation fiber array generation time domain for spreading, is further inputted to a photoelectric detector array through another N polarization-preserving fibers, is transformed into corresponding electric pulse sequences, finally enters an analog-digital converter array and is quantized into N independent high-speed parallel true random number sequences. The single-channel code rate can reach +Gbps magnitude, at least N=1000 independent parallel true random numbers can be simultaneously outputted, the scalability of an existing parallel true random number generator is increased by at least three to four orders of magnitude, and requirements of modern large-scale parallel computing and high-speed secure network communication can be greatly met.

The invention discloses global medium-term numerical forecast GRAPES_GFS. The global medium-term numerical forecast GRAPES_GFS is characterized in that in the aspect of large-scale parallel computing,the improvement of an existing framework is proposed for the shortcomings of an existing semi-implicit semi-Lagrangian integration scheme in terms of computational efficiency and scalability, and a solving algorithm with high efficiency and high performance is proposed based on the semi-implicit integral scheme and linearization system. The non-equidistant difference is adopted to improve the calculation accuracy of the background temperature profile and solve the problem of large errors of the background temperature profile at the level of drastic change of model vertical stratification thickness. The digital filtering module is reconstructed and optimized to improve the stability and computational efficiency of digital filtering. A 3D-Var integrated single-point test system is established.

The invention relates to a method for acquiring non-local variable information of a boundary layer of an unstructured grid, and belongs to the technical field of boundary layer transition prediction in fluid mechanics. According to the method, firstly, grids are preprocessed, all calculations for acquiring boundary layer information are completed in the w(i, j) array, and the i representing wall grid units in the w(i, j) array is independent from one another, so that large-scale parallel calculation can be performed, and the method is suitable for a modern CFD calculation method. The nearest wall surface distance is searched by adopting a circulating box method, and the reordering of the grids is completed on the basis, so that the calculation efficiency is relatively high. As the w(i, j) array is basically along the normal direction of the wall surface, and wall surface distance correction, boundary outer edge and boundary layer non-local variable parameter correction are carried out, the obtained result has high precision.

A tbps all-optical parallel true random number producing method having ultra-strong scalability comprises the steps of utilizing an optical device to produce an ultra-short pulse sequence, conducting spectrum cutting on the obtained ultra-short pulse sequence to obtain N narrow-band ultra-short pulse sequences, then conducting power adjustment on the N narrow-band ultra-short pulse sequences to enable average powers of the sequences to equal, converting randomly fluctuating peak powers into different pulse states, encoding the peak powers as 1 if the peak powers are larger than pulse output of the average powers, encoding the peak powers as 0 if the peak powers are lower than pulse-free output of the average powers, achieving repeated-frequency N all-optical parallel true random number production and filtering out random numbers through N optical band pass filter elements. According to the tbps all-optical parallel true random number producing method, the limitation of inherent periodicity of a parallel pseudo random number generator is overcome, meanwhile at least 10000 independent parallel true random numbers are output, the scalability of an existing parallel true random number generator is improved by 3 to 4 magnitudes, and the demands for large-scale parallel computing and high-speed secure network communication are met.

The invention discloses a parallel sequence pattern mining method based on a Spark cloud computing platform. For the problems that the existing serialization sequence pattern mining algorithm has low-effective algorithm capacity during processing mass data and the existing parallel sequence pattern mining algorithm based on Hadoop has high IO overhead and imbalanced loading, the invention designs a reasonable projection sequence database splitting strategy, so that the problem of imbalanced loading can be solved to maximum. On the basis, parallelization on an original PrefixSpan algorithm can be realized according to the characteristic of a MapReduce programming frame, so that the mass data sequence pattern mining efficiency can be improved by utilizing the massive parallel computing capacity of the Spark cloud computing platform. The technical scheme provided by the invention has the characteristics of easiness, and rapidness and can better improve the efficiency of sequence pattern mining.

The invention provides an automatic uncertainty analysis platform and method based on large-scale CFD parallel computing software. The automatic uncertainty analysis platform comprises a user interaction module, a database, a server, a CFD test module, an NIPC, a post-processing module and a monitoring module. The database is connected with the user interaction module and the server, the server module is connected with and controls the CFD test module, the monitoring module and the NIPC and post-processing module, the CFD test module comprises sample generation, batch processing and automatic calculation, and finally a plurality of calculation results are output to the NIPC and post-processing module. And after uncertainty analysis and result post-processing, an analysis result and original data are returned to a database and finally returned to the user interaction module. According to the method, automatic uncertainty analysis of large-scale CFD software is completed through the processes of network access, automatic example and sample loading, cluster calculation submission, real-time progress monitoring, NIPC uncertainty analysis, automatic post-processing and the like, and experimental comparison diagrams, analysis reports, summary reports and the like are provided.

The invention discloses a financial product picture information identification method based on domain analysis. The method comprises the following steps: data preparation work, summarizing financial market data from each channel to perform the persistence in a database, and establishing a data model for the original data; accurately describing a domain model of the financial product by using a formalized method, and establishing an association relation between the model attribute in the domain model and the data field in the financial market database; identifying the continuous numerical type variable of all positions in the financial product picture provided by the user; analyzing the large-scale parallel computing capacity by using the big data, and computing the matching degree of the output result and the domain model attribute according to the domain model, thereby performing interpretation on the financial product picture information. By using the method disclosed by the invention, the threshold of understanding the financial product by the normal user is lowered, and the identification prediction of the picture information related to the financial product is improved.

The invention provides a CPU-GPU (Central Processing Unit-Graphics Processing Unit) architecture-based business process model discovery method, which comprises the following steps of: copying an event log into a GPU from a CPU; dividing the event log into a plurality of sub-logs; one thread in the GPU calls one sub-log, and the sub-log is converted into a sub direct following graph; combining all sub direct following graphs in the thread into a final direct following graph; the final direct following graph is copied to a CPU from the GPU, and the CPU converts the direct following graph into a process model. According to the method, the characteristic of large-scale parallel computing of the GPU is fully utilized, high-parallel execution of process model discovery is realized, and the speed of business process model discovery is increased. Compared with a distributed system, the architecture occupies smaller computer resources.