First-principles computing and hybrid communication methods and devices for heterogeneous acceleration platforms
By distinguishing between streaming and recomputable data on heterogeneous acceleration platforms and utilizing shared memory windows and a unified set communication interface layer, the problem of high fine-grained communication overhead is solved, the relationship between video memory and communication overhead is optimized, parallel scalability and hardware adaptation flexibility are improved, and the consistency of computation results is ensured.
Patent Information
- Application Number
- CN202611590791.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-14
- Publication Date
- 2026-11-17
AI Technical Summary
In first-principles calculations on heterogeneous acceleration platforms, existing technologies suffer from high fine-grained cross-compute node communication overhead, making parallel expansion difficult. Furthermore, they lack a comprehensive assessment of data recalculation and transmission costs, have a single shared memory window usage pattern, and exhibit a contradiction between memory usage and communication overhead, resulting in insufficient flexibility in cross-hardware platform adaptation.
A hybrid communication approach for heterogeneous acceleration platforms is adopted. By creating a shared memory window within the computing node, streaming data and recomputable data are distinguished. Streaming data is transmitted through a set communication plane, while recomputable data is independently and redundantly recomputed and shared within the node. A unified set communication interface layer is set up to adapt to different communication backends and dynamically manage the capacity of the shared memory window.
It reduces cross-compute node transmission, optimizes the trade-off between memory usage and communication overhead, improves parallel scalability, reduces communication overhead, enhances hardware platform adaptability and flexibility, and ensures consistency of computation results.
Smart Images

Figure CN123115041A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of high-performance computing, and more specifically to a hybrid communication method and apparatus for first-principles computing on heterogeneous acceleration platforms. Background Technology
[0002] First-principles calculations using plane-wave pseudopotential density functional theory are core numerical simulation methods for conducting electronic structure simulations, molecular dynamics simulations, and time-dependent evolution calculations of materials. They are widely used in scientific research and engineering scenarios such as new material composition screening, analysis of material microscopic property mechanisms, study of crystal defect evolution behavior, and interface reaction simulation. As the scale of the materials to be simulated continues to increase, the computing power and memory resources of a single server node are no longer sufficient to meet the time and storage requirements of simulating large atomic systems. Heterogeneous acceleration clusters, composed of general-purpose central processing units (CPUs) and multiple accelerated computing devices (GPUs / DCUs), are gradually becoming the mainstream hardware platform for this type of simulation. In this type of heterogeneous parallel simulation, a two-dimensional process topology is often used, namely, a process group topology consisting of a NODE1 reciprocal space G-sphere and an FFT real space grid multiplied by a NODE2 k-point process group. Core computational steps such as three-dimensional fast Fourier transform and matrix reduction rely on cross-node communication for data interaction, and the communication overhead significantly impacts the parallel scalability of the simulation.
[0003] In existing engineering implementation schemes, each round of self-consistent field iteration or molecular dynamics iteration generates two types of computational data with significantly different characteristics. The first type is streaming data, which consists of large data blocks with regular memory layout, making it suitable for cross-computing node transmission via aggregated communication. The second type of data is strongly bound to individual atoms, requiring recalculation with each update of atom coordinates, and is broken down into numerous fine-grained data entries.
[0004] Existing processing solutions for this type of fine-grained atomic data mostly adopt the approach of global fragmentation generation combined with cross-computing node collection and distribution: all atomic fragments are allocated to various processes within the MPI communication domain, and each process only calculates the data corresponding to its assigned atoms; after the calculation is completed, the full data distribution is completed by broadcasting across computing nodes one atom at a time; when force and stress calculation scenarios are enabled, a single atom will generate multiple cross-computing node broadcasts until all computing nodes obtain the full content of the corresponding data.
[0005] Existing implementations have revealed several limitations in practical use. First, the communication overhead generated by fine-grained small message broadcasting is difficult to amortize with parallel scaling. The number of broadcast executions increases linearly with the total number of atoms in the system, and the communication traffic of a single broadcast message changes synchronously with the number of receiving computing nodes. Increasing the number of computing nodes can only distribute the computational load, and it is difficult to reduce this part of the cross-computing node communication overhead; after the number of nodes increases, the proportion of this communication time in the total time of a single iteration will increase, and the speedup benefits brought by parallelism are easily offset by communication overhead. It is difficult to achieve the ideal parallel scaling effect for large-scale molecular dynamics simulations by simply adding cluster nodes.
[0006] Secondly, existing implementations lack a comprehensive evaluation mechanism for the cost of data recalculation and transmission. Some fine-grained data related to atoms do not have computational dependencies between individual atomic entries, and can be regenerated locally on the computing node based on atomic coordinates and pseudopotential parameters. However, existing solutions rarely consider the cost trade-off between local recalculation and network transmission. Even if the cost of local recalculation is relatively lower, the network transmission path is still preferred.
[0007] Secondly, the usage pattern of shared memory windows within compute nodes is relatively simple. In existing solutions, shared memory windows are generally used to forward data within a node after cross-compute node broadcasting is completed; data needs to be transmitted through the cross-compute node network to reach the local machine first, and then forwarded to other processes within the node via shared memory windows.
[0008] Furthermore, there is a trade-off between memory usage and communication overhead. The full content of the fine-grained atomic data needs to reside in the video memory of the accelerated computing devices within each computing node for subsequent nonlocal term, force, and stress solutions. If a fragmented generation mode is adopted to control video memory usage, it will introduce a large number of cross-computing node broadcast collection operations, increasing communication overhead; if the goal is to reduce communication overhead, each node needs to store the complete data, which will increase video memory usage.
[0009] Meanwhile, the lack of capacity management logic for shared memory windows leads to additional system call overhead. Existing implementations generally do not perform storage capacity pre-assessment based on pseudopotential cutoff radius and real space grid density; during simulation phases where the cell and atomic sphere grid sizes remain unchanged, shared memory windows are repeatedly destroyed and rebuilt, generating numerous operating system calls and increasing simulation time.
[0010] In addition, the business logic of aggregated communication is often deeply bound to a specific communication backend, making porting across heterogeneous hardware platforms a significant undertaking. When simulation tasks are migrated to different acceleration hardware clusters, targeted modifications to the upper-layer scheduling code are often required. Furthermore, most existing implementations lack configurable runtime mode switching capabilities; if it is necessary to compare simulation results of different data processing logics, it is generally necessary to modify the underlying source code and recompile, resulting in insufficient adaptability and flexibility. Summary of the Invention
[0011] This invention addresses the problem of excessive fine-grained cross-computation node communication overhead caused by some iterative data in existing first-principles heterogeneous parallel communication implementations by providing a hybrid communication method and apparatus for first-principles computing on heterogeneous acceleration platforms.
[0012] According to a first aspect of the present invention, a hybrid communication method for first-principles calculations on heterogeneous acceleration platforms is provided, applied to a heterogeneous acceleration platform containing multiple computing nodes, for performing first-principles calculations based on plane-wave pseudopotential density functional theory. The method includes: obtaining a cross-node aggregated communication plane based on the topological partitioning of the same computing node, and creating shared memory windows within each computing node; in each round of self-consistent field iteration or molecular dynamics iteration, dividing the data used in each iteration into streaming data and recomputable data based on whether the data used in each iteration is updated with changes in atomic positions, whether it can be independently and redundantly recalculated within a single computing node, and whether a globally consistent view needs to be maintained across computing nodes; transmitting the streaming data between computing nodes through the cross-node aggregated communication plane; each computing node independently and redundantly recalculating the complete recomputable data, and sharing the recomputable data with multiple processes within the same computing node using the shared memory windows within each computing node; wherein, in each round of self-consistent field iteration or molecular dynamics iteration, the transmission process of the streaming data and the sharing process of the recomputable data are carried out collaboratively.
[0013] Furthermore, the recomputable data is data that is updated as the atomic position changes, can be independently and redundantly recomputed within a single computing node, and does not require maintaining a globally consistent view across computing nodes.
[0014] Furthermore, the step of sharing the recomputable data with multiple processes within the same computing node using a shared memory window within each computing node includes: writing the complete recomputable data into the shared memory window through a process that performs independent redundant recomputation; and reading the recomputable data from the shared memory window after each process completes synchronization within the computing node.
[0015] Furthermore, the following management steps are performed on the shared memory window: The upper bound of the shared memory window's capacity is determined based on the upper limit of the number of atoms within a single computing node and the upper limit of the atomic sphere grid size; during each round of self-consistent field iteration or molecular dynamics iteration, it is determined whether the shared memory window needs to be rebuilt based on the actual capacity requirement and the allocated capacity; if the actual capacity requirement exceeds the allocated capacity, the shared memory window is rebuilt; if the actual capacity requirement does not exceed the allocated capacity, the already created shared memory window is reused, and the newly obtained recalculated data from this iteration is overwritten into the shared memory window.
[0016] Furthermore, in the first-principles calculation process based on plane-wave pseudopotential density functional theory, the recalcible data includes at least one of the following: nonlocal pseudopotential β-projection operator, real space grid values corresponding to the nonlocal pseudopotential β-projection operator, spatial derivative of the nonlocal pseudopotential β-projection operator with respect to atomic coordinates, and atomic sphere grid index.
[0017] Furthermore, a unified set communication interface layer is set up for the streaming data. The unified set communication interface layer provides standard set communication primitives to the caller and is compatible with various set communication backends. The target set communication backend is called through the unified set communication interface layer, and the streaming data is transmitted between computing nodes based on the cross-computing node set communication plane.
[0018] Furthermore, the method further includes: receiving a mode selection instruction to indicate whether to enable a first operating mode; when the mode selection instruction indicates that the first operating mode is enabled, performing the independent redundant recalculation on the recalculable data, and sharing the recalculable data with multiple processes within the same computing node through the shared memory window; when the mode selection instruction indicates that the first operating mode is not enabled, transmitting the recalculable data as streaming data through a cross-computing node aggregated communication plane.
[0019] According to a second aspect of the present invention, a hybrid communication device for first-principles calculations on a heterogeneous acceleration platform is provided, applied to a heterogeneous acceleration platform containing multiple computing nodes, for performing first-principles calculations based on plane-wave pseudopotential density functional theory, comprising: a communication carrier construction module, used to obtain a cross-computational communication plane based on the topological partitioning of the same computing node, and to create shared memory windows within each computing node; and a data partitioning module, used to, in each round of self-consistent field iteration or molecular dynamics iteration, update the data used in each iteration based on whether the data changes with atomic positions, whether independent redundant recalculation can be performed within a single computing node, and... Whether it is necessary to maintain a globally consistent view across computing nodes, the data used in each iteration is divided into streaming data and recalcible data; a streaming data transmission module is used to transmit the streaming data between computing nodes through the cross-computing node aggregate communication plane; a recalcible data processing module is used for each computing node to independently and redundantly recalculate the complete recalcible data, and to share the recalcible data with multiple processes within the same computing node using the shared memory window within each computing node; wherein, in each round of self-consistent field iteration or molecular dynamics iteration, the streaming data transmission process and the recalcible data sharing process are carried out in concert.
[0020] According to a third aspect of the present invention, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method.
[0021] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method.
[0022] Compared with the prior art, the present invention has the following beneficial effects.
[0023] 1. In each iteration, this invention differentiates the data used in the iteration based on whether the data is updated with changes in atomic position, whether it can be independently recalculated within a single computing node, and whether a globally consistent view needs to be maintained across computing nodes: streaming data suitable for cross-computing node transmission is transmitted between nodes using cross-computing node set communication; recalcible data is recalculated locally on the computing node and then shared among processes within the computing node using a shared memory window within the computing node, which can reduce the cross-computing node transmission behavior corresponding to this type of data.
[0024] 2. This invention reduces cross-node transmission by sacrificing redundant recalculation overhead within the computing node. The data storage space required for redundant recalculation originates from the storage allocated by the business process itself for storing recalculated data, eliminating the need for additional persistent video memory allocation independent of the business process. This design is suitable for hardware characteristics of heterogeneous cluster nodes with strong computing power within nodes and relatively limited network bandwidth between nodes, creating a new trade-off between video memory usage and cross-node communication overhead.
[0025] 3. This invention sets capacity control logic for shared memory windows: the upper limit of storage capacity is pre-evaluated based on the pseudopotential cutoff radius and real space grid density; when the size of the atomic sphere grid changes and causes a change in storage requirements, window reconstruction is triggered; when the atomic position moves but the grid parameters remain unchanged, the existing window is reused and an overwrite is performed; during the simulation stage where the cell scale and real space grid parameters remain unchanged, the window can be reused after initial allocation, which can reduce the overhead caused by the operating system repeatedly creating and destroying windows.
[0026] 4. The cross-compute node transmission of streaming data adopts a unified set communication interface layer. This interface layer provides standard set communication primitives at the top and adapts to multiple types of set communication backends at the bottom. The upper-layer iterative scheduling logic is decoupled from the lower-layer communication implementation. When switching set communication backends, the upper-layer business logic does not need to make specific modifications to the communication backend, which facilitates migration to different acceleration hardware clusters.
[0027] 5. The cross-component communication plane for streaming data and the shared memory plane within the compute nodes for recomputable data reuse the same set of two-dimensional process topologies for compute nodes and execute collaboratively within the same iteration step, which can reduce the additional overhead caused by maintaining multiple sets of parallel communication domains separately. Under this topology, the cross-component network links mainly carry the large-block regular communication traffic corresponding to streaming data, which is adapted to the processing characteristics of large-block regular communication by the aggregated communication hardware.
[0028] 6. The present invention provides a selectable operating mode switch: it can switch between a first operating mode and a first operating mode that is not enabled. The first operating mode adopts the processing path of the present invention. When the first operating mode is not enabled, the recalculated data is transmitted as streaming data through cross-computing node set communication, which facilitates the comparison of simulation results of the two processing paths under the same hardware and system scale.
[0029] 7. Using the method provided by this invention, the downstream non-local terms, force, and stress solutions can directly read the recalcible data obtained through sharing within this computing node, and the data access interface is consistent with the reading method of the downstream computing kernel. When switching between enabling the first running mode and not enabling the first running mode, the difference in the calculation results mainly comes from the floating-point rounding introduced by the different order of parallel floating-point reduction summation. This difference is within the normal consistency tolerance range of first-principles parallel computing. Attached Figure Description
[0030] Figure 1 This is a flowchart illustrating the first-principles computing hybrid communication method for heterogeneous acceleration platforms provided by the present invention.
[0031] Figure 2 A comparison curve showing the time consumption of the wmask generation stage as a function of the number of nodes, provided by this invention.
[0032] Figure 3 The present invention provides a structural block diagram of a first-principles computing hybrid communication device for heterogeneous acceleration platforms.
[0033] Figure 4 This is a schematic diagram of the electronic device structure provided by the present invention. Detailed Implementation
[0034] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.
[0035] Example 1 Reference Appendix Figure 1 This application provides a hybrid communication method for first-principles calculations on heterogeneous acceleration platforms, applicable to multi-node distributed heterogeneous acceleration platforms. Each computing node employs a heterogeneous hardware combination of a central processing unit (CPU) and at least one acceleration computing device, which can be a graphics processing unit (GPU) or a deep computing unit (DCU). Multiple computing nodes are interconnected through a high-speed interconnect network to perform self-consistent field iterations or molecular dynamics iterations in first-principles calculations based on plane-wave pseudopotential density functional theory. The method specifically includes the following steps.
[0036] S1. Based on the topology partitioning of the same computing node, a cross-computing node set communication plane is obtained, and a shared memory window is created inside each computing node.
[0037] This step is executed once after the simulation task starts. The topology can be partitioned using a two-dimensional process topology consisting of a NODE1 reciprocal space G-sphere and an FFT real-space mesh multiplied by NODE2's k-point process groups. Based on the partitioning results, a cross-computation node aggregate communication plane is instantiated. Simultaneously, a shared memory window is created within each computation node, granting access to all parallel processes within the same computation node for high-speed data sharing within the node.
[0038] S2. In each round of self-consistent field iteration or molecular dynamics iteration, the data used in each round of iteration is divided into streaming data and recalcible data based on whether the data used in each round of iteration is updated with changes in atomic position, whether it can be independently and redundantly recalculated within a single computing node, and whether it is necessary to maintain a globally consistent view across computing nodes.
[0039] In this embodiment, when classifying the data involved in the iteration, the following factors are considered: (1) whether the data is updated as the atomic position changes; (2) whether the data can be independently and redundantly recalculated within a single computing node; and (3) whether the data needs to maintain a globally consistent view through cross-computing node communication.
[0040] Data that updates with changes in atomic position, can be independently and redundantly recalculated within a single computing node, and does not require maintaining a unified distribution view across computing nodes is classified as recalcible data. Other data that needs to maintain consistency through a cross-node aggregated communication plane is classified as streaming data.
[0041] In first-principles calculations based on plane-wave pseudopotential density functional theory, recalcible data may include at least one of the following: nonlocal pseudopotential β-projection operator, real-space grid values corresponding to the nonlocal pseudopotential β-projection operator, spatial derivatives of the nonlocal pseudopotential β-projection operator with respect to atomic coordinates, and atomic sphere grid indices. Streaming data may include electron wavefunctions, charge densities, fast Fourier transform-related transpose data, and reduction matrices, as well as other data that needs to be processed consistently across multiple computational nodes.
[0042] In one embodiment of this application, the overhead associated with independent redundancy recalculation within a computing node can be further compared with the transmission overhead across computing nodes to identify data processed using the independent redundancy recalculation method within the computing node. For example, when the size of single-atom related data is constrained by the size of the atomic sphere mesh, the accelerated computing devices within the computing node can perform local calculations in parallel, and the fine-grained message transmission across computing nodes is greatly affected by network latency, it is preferable to process this type of data as recalcible data.
[0043] S3. The streaming data is transmitted between computing nodes through the cross-computing node aggregate communication plane.
[0044] In one embodiment of this application, a unified set communication interface layer is configured for streaming data. This unified set communication interface layer provides standard set communication primitives to the caller and adapts to three different types of set communication backends: a set communication library for accelerated computing devices, a message passing interface supporting direct communication between the graphics memory and the accelerated computing device, and a host-side message passing interface. This invention selects the set communication backend through a pluggable backend selection rule; it reads the backend identifier parameter from the simulation input and directly loads the corresponding set communication backend. Even when switching between different set communication backends, the calling format of the upper-layer standard set communication primitives remains consistent, and the transmission process of streaming data across the set communication plane of computing nodes is completed through the currently selected set communication backend.
[0045] The communication business logic is decoupled from the underlying aggregated communication backend. The scheduling timing and primitive calling logic of streaming data remain unchanged. Only the downward adaptation implementation of the interface layer changes according to the selected target aggregated communication backend. All recalculated data is recalculated and the publishing process runs within a single computing node. This part of the processing is not affected by the selected aggregated communication backend. Regardless of the type of aggregated communication backend used, the processing steps of independent redundant recalculation, shared memory window writing, and internal synchronous reading within the node remain consistent.
[0046] This step only performs cross-compute node aggregation communication on large blocks of streaming data with regular arrangement rules. This type of business scenario is an advantageous scenario for the backend hardware processing of various aggregation communication.
[0047] S4. Each computing node independently and redundantly recalculates to obtain the complete recalculable data, and uses the shared memory window within each computing node to share the recalculable data with multiple processes within the same computing node.
[0048] In one embodiment of this application, sharing recomputable data with multiple processes within the same computing node using a shared memory window within each computing node includes: writing the complete recomputable data into the shared memory window through a process performing independent redundant recomputation; and reading the recomputable data from the shared memory window after each process completes internal synchronization within the computing node. During the processing of recomputable data, there is no cross-computing node byte transmission carrying business payloads; only internal integer consistency reduction within the computing node and barrier synchronization operations without data payloads are retained.
[0049] The following management steps are performed on the shared memory window: The upper bound of the shared memory window's capacity is determined based on the upper limit of the number of atoms within a single computing node and the upper limit of the atomic sphere grid size; during each round of self-consistent field iteration or molecular dynamics iteration, it is determined whether the shared memory window needs to be rebuilt based on the actual capacity requirement and the allocated capacity; if the actual capacity requirement exceeds the allocated capacity, the shared memory window is rebuilt; if the actual capacity requirement does not exceed the allocated capacity, the already created shared memory window is reused, and the newly obtained recalculated data from this iteration is overwritten into the shared memory window.
[0050] Downstream nonlocal term calculations, force solutions, and stress solutions directly consume the complete recalculated data generated by this computing node, without needing to perform cross-computation node index alignment operations.
[0051] Furthermore, this method also includes: S5, in each round of self-consistent field iteration or molecular dynamics iteration, the transmission process of streaming data and the sharing process of recalculated data are carried out in a coordinated manner.
[0052] In one embodiment of this application, the two processes are executed collaboratively as follows: the results of the same set of computing node topology partitioning already constructed in S1 are reused, and the two processes are supported by the results of this computing node topology partitioning, without the need to maintain multiple independent parallel communication domains. During the same round of self-consistent field iteration or molecular dynamics iteration, the streaming data transmission process and the recalculable data sharing process are carried out collaboratively, and the two processes are executed by multiple parallel processes within each computing node. The cross-computing node network, i.e., the high-speed interconnect communication link between different computing nodes, only carries large, regular messages corresponding to streaming data. Fine-grained atomic-level cross-computing node communication, which generates short messages at the unit of a single atom, is completely eliminated. Before the computing kernel reads the two types of data, the streaming data completes the cross-computing node transmission, and the recalculable data completes the intra-node sharing. The two types of data converge at the synchronization point of the same iteration step and are provided for subsequent simulation calculations.
[0053] Furthermore, this method also includes: S6, receiving a mode selection instruction, which is used to indicate whether to enable the first operating mode in order to control the processing path of recalculated data.
[0054] In one embodiment of this application, the specific execution method of mode switching is as follows: when the mode selection instruction indicates that the first running mode is enabled, the independent redundant recalculation is performed on the recalcible data, and the recalcible data is shared with multiple processes in the same computing node through a shared memory window; when the mode selection instruction indicates that the first running mode is not enabled, the recalcible data is transmitted as streaming data through a cross-computing node aggregated communication plane.
[0055] The two operating modes are based on the same nonlocal pseudopotential β projection operator mathematical operation rules, with the only difference being the data fragmentation and allocation method and the data distribution and transmission path. The slight difference in the simulation results comes from the floating-point rounding caused by the different summation order of the parallel floating-point reduction operation. The simulation output results of the two operating modes meet the engineering acceptance requirements for the consistency of the parallel results of first-principles calculations.
[0056] Furthermore, this method also includes: S7, completing the internal calculation of a single iteration, updating the atomic positions, and repeating S2-S6 until the self-consistent field convergence condition or the molecular dynamics simulation termination condition is met, outputting simulation results such as total energy, atomic forces, electronic band structure, and simulation trajectory.
[0057] The following examples, using molecular dynamics simulations of multi-atom materials, provide a complete and detailed explanation of the invention. These examples are for illustrative purposes only and are not intended to limit the scope of protection of the invention.
[0058] The test hardware environment is a multi-node distributed heterogeneous acceleration platform, with each computing node configured with 4 accelerated computing devices; the simulation object is a 319-atom material system, performing plane wave pseudopotential density functional theory molecular dynamics simulation; the simulation input configuration includes cell parameters, pseudopotential parameters, and k-point grid parameters, and can receive mode selection commands to select whether to enable or disable the first running mode; the aggregate communication backend is configured as an aggregate communication library for accelerated computing devices.
[0059] Step 1: The simulation task is started. Execute S1 of this embodiment to construct the two-dimensional process topology partition of NODE1×NODE2. Based on the same set of computing node topology partitions, a cross-computing node collective communication plane is obtained, and shared memory windows are created inside each computing node. Initialize the unified collective communication interface layer and select the collective communication library for accelerated computing devices. Receive the mode selection instruction carrying the first running mode identifier and enter the molecular dynamics iterative loop.
[0060] Step 2: Execute S2 of this embodiment to classify the data used in this iteration according to the preset classification rules determined when the simulation task is started (the classification results can be verified in each iteration); in this case, the nonlocal pseudopotential β projection operator, real space grid values, first / second order spatial derivatives, and atomic sphere grid index are classified as recalcible data; the electron wave function, charge density, FFT transpose plane, and reduced small matrix are classified as streaming data.
[0061] Step 3: Execute S3 of this embodiment, call the standard collection communication primitive of the unified collection communication interface layer, and complete the cross-computing node transmission of streaming data through the selected collection communication backend using the cross-computing node collection communication plane; the communication business logic is decoupled from the underlying collection communication backend.
[0062] Step 4: Execute S4 of this embodiment. Each computing node independently and redundantly calculates the full content of the recalcible data. The process that performs independent redundant recalculation writes the complete recalcible data into the shared memory window. After completing the internal synchronization of the computing node, each process reads the recalcible data from the shared memory window. During the processing of recalcible data, there is no cross-computing node byte transmission carrying business payload.
[0063] The shared memory window management logic is executed as follows: Based on the upper limit of the number of atoms within a single computation node and the upper limit of the atomic sphere mesh size, the upper bound of the shared memory window's capacity is determined. During each iteration of the simulation, the actual capacity requirement of the shared memory window is compared with the allocated capacity to determine whether to rebuild the shared memory window. When the actual capacity requirement exceeds the allocated capacity, the shared memory window is rebuilt; when the actual capacity requirement does not exceed the allocated capacity, the already created shared memory window is reused, and the newly obtained recalculated data from this iteration is overwritten into the shared memory window. Downstream nonlocal terms and force and stress solution stages directly use the complete recalculated data of this computation node.
[0064] Step 5: Execute S5 of this embodiment. The transmission process of streaming data and the sharing process of recalculated data are carried out in tandem in the current molecular dynamics iteration step. The simulation kernel consumes streaming data and recalculated data respectively to complete the three-dimensional FFT transformation and charge density construction. It consumes recalculated data to complete the calculation of nonlocal pseudopotential terms and the solution of atomic forces and stresses.
[0065] Reference Appendix Figure 2 The timing and error indicators described below were obtained from the simulation test of this application and are only used to illustrate the technical effects of this application, and are not intended to limit all simulation scenarios.
[0066] According to the actual performance test results: In the first running mode, the time taken for the wmask generation stage is 0.893s for 1 node, 0.893s for 2 nodes, and 0.901s for 4 nodes. The stage time is basically decoupled from the number of computing nodes. The stage time does not increase much when the number of computing nodes doubles.
[0067] In another embodiment of this application, a mode selection instruction carrying an identifier indicating that the first operating mode is not enabled is received; the processing logic for not enabling the first operating mode in S6 is executed, and the recalculated data is incorporated into the streaming data and transmitted via the cross-computing node set communication plane; the traditional processing flow of global fragmentation generation combined with cross-computing node collection is executed; in the same simulation example, under the first operating mode not enabled, the wmask generation stage takes 1.104s for 1 node, 1.522s for 2 nodes, and 1.870s for 4 nodes. The stage time increases continuously with the increase in the number of computing nodes, and doubling the number of nodes brings a significant increase in time consumption.
[0068] As can be seen from the comparison, without enabling the first operating mode, the data processing method is consistent with existing technologies: recomputable data and streaming data are transmitted together through cross-component node communication. The overhead in this stage cannot be diluted by the number of parallel nodes; the more nodes there are, the higher the proportion of communication overhead caused by message interaction between cross-component nodes. The first operating mode of this invention can eliminate all cross-component node message interactions carrying business payloads corresponding to recomputable data, and the overhead in this stage is basically decoupled from the node scale. Even in a single-node scenario, this invention achieves time savings due to the optimization of memory allocation logic.
[0069] The two operating modes are based on the same nonlocal pseudopotential β projection operator mathematical operation rules, with the only difference being the data fragmentation and allocation method and the data distribution and transmission path. The slight difference in the simulation results comes from the floating-point rounding caused by the summation order of the parallel floating-point reduction operation. The simulation output results of the two operating modes meet the engineering acceptance requirements for the consistency of the parallel results of first-principles calculations.
[0070] To further illustrate the advantages of this invention over existing technologies in terms of recomputable data communication overhead, a quantitative comparison of the cross-computation node message exchange volume in two operating modes within a single molecular dynamics iteration step is presented below. Existing solutions require multiple cross-computation node broadcast operations for recomputable data within each molecular dynamics iteration step. These broadcast operations are initiated sequentially by atom, with each message carrying small blocks of data related to the atomic sphere grid and projection. The total cross-computation node message exchange volume reaches the order of magnitude of the product of the number of atoms, the size of the atomic sphere grid, and the number of projection channels. The number of broadcast executions increases linearly with the number of atoms, and the data size of a single message increases synchronously with the number of receiving computation nodes. This invention employs an independent redundant recomputation computation within each computation node, coupled with shared memory distribution. In each molecular dynamics iteration step, recomputable data no longer generates cross-computation node messages carrying service payloads; only integer consistency reduction within the computation node (i.e., reduction operations at the atom count level) and fence synchronization operations without data payloads are retained. The high-speed interconnect links between different computing nodes no longer bear the business data transmission task of recomputable data. The corresponding cost is to increase the independent redundant recomputation calculation workload locally on each computing node, thereby eliminating the communication bottleneck that continues to increase as the number of computing nodes expands.
[0071] The above-mentioned comparative simulation test used a material system with 319 atoms as the test sample, with each computing node configured with 4 accelerated computing devices. The test timing object was the generation stage of wmask, i.e., the nonlocal pseudopotential β projection operator and its derivative. For the same calculation case, only the running mode was switched, and the average value was taken from multiple tests. Compared with not using the first running mode, the time reduction of the first running mode of the present invention under different node scales is as follows: the time reduction is about 19% for the 1-node scenario, about 41% for the 2-node scenario, and about 52% for the 4-node scenario. In the prior art, the time of this stage continuously increases with the number of nodes. The time curve of the first running mode of the present invention tends to be flat, and the bottleneck of parallel expansion is eliminated. The single-node scenario can also achieve performance improvement. This improvement comes from the memory allocation optimization brought about by the dynamic management of shared memory windows, proving that the gain of the technical solution of the present invention does not simply rely on reducing cross-computing node communication.
[0072] The simulation results from multiple sets of comparisons in this application show that the total energy difference between the first operating mode and the mode without the first operating mode is less than 1×10⁻⁶. -6 eV, the difference in atomic force is less than 5 × 10⁻⁶ -3 eV / Å, meeting the accuracy standard for first-principles material simulation of plane wave pseudopotential density functional theory.
[0073] The unified aggregate communication interface layer set up in this invention can simultaneously adapt to both graphics processor-accelerated backends and general-purpose accelerated computing backends, with both backends using completely identical aggregate communication calling logic. The graphics processor-accelerated backend loads an aggregate communication library oriented towards the graphics processor, while the general-purpose accelerated computing backend loads an aggregate communication library with a matching interface format. Under both types of accelerated backends, the process of streaming data transmission across computing nodes and the process of sharing recomputable data within nodes remain unchanged; only the underlying aggregate communication library differs. The technical solution is not bound to a specific acceleration hardware architecture and can adapt to various heterogeneous acceleration platforms.
[0074] The following is an example of the complete calculation process of the present invention from input parsing to result output, which corresponds one-to-one with steps S1 to S7 of the aforementioned embodiment.
[0075] Step 1, Input Analysis: Read the simulation input file and structure file, determine the system to be simulated, functional type, k-point parameters, and set communication backend selection parameters, and analyze the mode selection instructions used to identify the first operating mode or not to enable the first operating mode.
[0076] Step 2: Parallel topology establishment, execute Example S1: Establish a two-dimensional process topology of the NODE1 reciprocal space G sphere and the FFT real space grid partitioned by NODE2 k-point process group; obtain the cross-computing node set communication plane based on the topology partitioning result of the same computing node, and create shared memory windows in the internal of each computing node; establish the internal communication domain of the computing node, and initialize the selected set communication backend.
[0077] Step 3: Data Classification, Execution Example S2: Based on whether the iterative data is updated with changes in atomic position, whether it can be independently and redundantly recalculated within a single computing node, and whether it needs to maintain a globally consistent view across computing nodes, the iterative data is divided into streaming data and recalcible data.
[0078] Step 4: Streaming data transmission, executing Example S3: Streaming data is transmitted between computing nodes through a cross-computing node aggregated communication plane.
[0079] Step 5: Recomputable data processing. Execute Example S4: Each computing node independently and redundantly recomputes to obtain complete recomputable data, and uses the shared memory window within each computing node to share the recomputable data with multiple processes within the same computing node; if the first running mode is not enabled, the traditional processing flow of global sharding generation combined with cross-computing node collection is executed.
[0080] Step Six: Dual-process collaborative scheduling, executing Example S5: The transmission process of streaming data and the sharing process of recalculated data are carried out collaboratively.
[0081] Step 7, Mode Selection Control, Execution Example S6: Receive mode selection instruction, which includes a first operating mode or an operating mode identifier that does not enable the first operating mode, and control the processing path of recalculated data; when in the first operating mode, perform independent redundant recalculation on the recalculated data and share the recalculated data through a shared memory window; when the first operating mode is not enabled, transmit the recalculated data as streaming data through a cross-compute node aggregated communication plane.
[0082] Step 8: Iterative calculation and atomic position update. Execute Example S7: Complete all calculations within a single iteration, update atomic positions, and repeat S2 to S6 until the self-consistent field convergence condition or the molecular dynamics simulation termination condition is met. Output simulation results such as total energy, atomic force, and charge density.
[0083] The entire methodology performs all computational steps locally on a multi-node distributed heterogeneous acceleration platform, eliminating the need for external independent computing nodes or remote computing power services. A hybrid communication strategy designs separate transmission paths for streaming and recomputable data, dynamically reconstructing or reusing windows based on shared memory window capacity requirements. It switches between the two data processing paths using mode selection instructions. Each step is executed sequentially: topology initialization, data classification, cross-node communication, intra-node sharing, and iterative looping. This constructs a hybrid communication process for first-principles simulation of plane wave pseudopotentials that balances parallel simulation accuracy with optimized cross-node communication overhead.
[0084] Example 2 Please see Figure 3 This invention also provides a hybrid communication device for first-principles calculations on heterogeneous acceleration platforms, used to execute the hybrid communication method for first-principles calculations on heterogeneous acceleration platforms provided in Embodiment 1. This device is deployed on a multi-node distributed heterogeneous acceleration platform and is used to perform first-principles self-consistent field iterative calculations or molecular dynamics iterative calculations based on plane-wave pseudopotential density functional theory. The device consists of a communication carrier construction module, a data partitioning module, a streaming data transmission module, and a recomputable data processing module. These functional modules cooperate and are decoupled to jointly realize the parallel scheduling function of hybrid communication in a heterogeneous acceleration environment. The specific working methods of each module are as follows.
[0085] The communication carrier construction module is used to obtain a cross-node aggregate communication plane based on the topology partitioning of the same computing node, and to create shared memory windows within each computing node. After the device starts, the communication carrier construction module completes the global parallel topology initialization in one go, and completes the topology partitioning of the reciprocal space grid and k-point process groups using a two-dimensional process topology approach. It instantiates the cross-node aggregate communication plane based on a unified topology structure, providing a unified communication domain support for distributed communication across multiple computing nodes. Simultaneously, the communication carrier construction module independently creates a shared memory window within each computing node. The shared memory window has access permissions to all parallel processes within the computing node, providing a hardware carrier for high-speed sharing of recomputable data within nodes. The topology structure and communication carrier are generated only during the simulation task initialization phase and do not need to be reconstructed during iterations.
[0086] The data partitioning module classifies data based on its dynamic and parallel processing characteristics during each self-consistent field or molecular dynamics iteration. It distinguishes data according to three criteria: whether the data dynamically changes with atomic position updates, whether redundant recalculation can be performed independently within a single computation node, and whether a globally consistent view needs to be maintained across computation nodes. Data that updates with atomic position, supports independent redundant recalculation within a node, and does not require global cross-node synchronization is classified as recalcible data. This includes data such as nonlocal pseudopotential β-projection operators, real-space grid values, spatial derivatives, and atomic sphere grid indices. Data that updates frequently and requires multi-node collaboration to maintain global consistency is classified as streaming data. This includes data such as electron wave functions, charge densities, FFT transpose planes, and reduced small matrices. The data partitioning module performs dynamic classification before each iteration, adapting to the data distribution characteristics of different atomic configurations and simulation systems.
[0087] The streaming data transmission module is used to transmit streaming data across computing nodes via a cross-node aggregated communication plane. The module is configured with a unified aggregated communication interface layer, which can adaptively adapt to various types of aggregated communication backends. It uses standardized aggregated communication primitives to distribute, collect, and synchronize large blocks of regular streaming data. The streaming data transmission module performs cross-node aggregated communication only for streaming data, completely avoiding fine-grained, small-message atomic cross-node communication behavior, significantly reducing communication latency and message interaction overhead in distributed parallel scenarios. The communication scheduling logic of the streaming data transmission module is decoupled from the underlying hardware communication backend, maintaining consistent communication logic across different heterogeneous acceleration hardware platforms.
[0088] The recomputable data processing module controls each computing node to independently and redundantly recompile complete recomputable data and utilizes a local shared memory window to achieve multi-process data sharing within the node. During each iteration, the module drives a single computing node to complete redundant computation of all recomputable data without relying on data synchronization or message transmission from other computing nodes. After computation, the module writes the complete recomputable data into the shared memory window of the current node, and parallel processes within the node read the data from the shared memory window through an internal synchronization mechanism. Simultaneously, the module possesses dynamic shared memory window management capabilities, determining window capacity requirements based on the upper limits of the number of atoms and the grid size, and adaptively reconstructing or reusing overwrites to ensure efficient and stable sharing of recomputable data within the node. The module generates no cross-component data transmission carrying business payloads throughout the process, completely eliminating distributed communication bottlenecks associated with recomputable data.
[0089] During operation, the streaming data transmission process across computing nodes and the intra-node sharing process of recomputable data are executed in parallel. Each module reuses the same set of global topology resources, eliminating the need to maintain multiple independent communication domains. Both types of data processing proceed synchronously through intra-node parallel processes. Before the computing kernel performs iterative computation, the streaming data transmission module completes cross-node data transmission, and the recomputable data processing module completes intra-node data sharing. The two types of data converge at the same iteration synchronization point and are then uniformly supplied to the subsequent simulation computation kernel.
[0090] Based on the above modular architecture design, this device uses a hybrid communication mechanism that combines streaming data communication across computing node sets, redundant recalculation within recalculated data nodes, and local sharing to differentiate between the two types of iterative data. While ensuring that the accuracy of first-principles simulation calculations remains unchanged, it significantly reduces the increase in communication overhead caused by parallel scaling, effectively improving the parallel scalability and operating efficiency of molecular dynamics simulations and self-consistent field iterative calculations on a multi-node heterogeneous acceleration platform.
[0091] Example 3 Based on Embodiment 1 described above, this embodiment also provides an electronic device, please refer to the appendix. Figure 4 , Figure 4 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this disclosure.
[0092] like Figure 4 As shown, an electronic device may include a processing unit (such as a central processing unit, graphics processing unit, etc.) that can perform various appropriate actions and processes based on a program stored in read-only memory (ROM) or a program loaded from a storage device into random access memory (RAM). The RAM also stores various programs and data required for the operation of the electronic device. The processing unit, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.
[0093] Typically, the following devices can be connected to an I / O interface: input devices such as touchscreens, touchpads, keyboards, mice, and cameras; output devices such as liquid crystal displays (LCDs) and speakers; storage devices such as magnetic tapes and hard drives; and communication devices. Communication devices allow electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 4 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or have instead. Figure 4 Each box shown can represent a device or multiple devices as needed.
[0094] In particular, according to some embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, some embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a ROM. When the computer program is executed by a processing device, it performs the functions defined above in the methods of some embodiments of this disclosure.
[0095] Example 4 Based on Embodiment 1 above, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above method.
[0096] It should be noted that, in some embodiments of this disclosure, the computer-readable medium described above may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In some embodiments of this disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In some embodiments of this disclosure, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium may be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), or any suitable combination thereof.
[0097] In some embodiments, the client and server may communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and may interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.
[0098] The aforementioned computer-readable storage medium may be included in the aforementioned electronic device or may exist independently without being assembled into the electronic device. The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to: obtain a cross-component communication plane based on the same computing node topology partitioning, and create shared memory windows within each computing node; in each round of self-consistent field iteration or molecular dynamics iteration, divide the data used in each iteration into streaming data and recomputable data based on whether the data used in each iteration is updated with changes in atomic positions, whether it can be independently and redundantly recalculated within a single computing node, and whether a globally consistent view needs to be maintained across computing nodes; transmit the streaming data between computing nodes through the cross-component communication plane; each computing node independently and redundantly recalculates the complete recomputable data, and shares the recomputable data with multiple processes within the same computing node using the shared memory windows within each computing node; wherein, in each round of self-consistent field iteration or molecular dynamics iteration, the transmission process of the streaming data and the sharing process of the recomputable data are carried out collaboratively.
[0099] Computer program code for performing operations of some embodiments of this disclosure can be written in one or more programming languages or a combination thereof, including object-oriented programming languages such as Python and C++, and conventional procedural programming languages such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).
[0100] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.
[0101] The units described in some embodiments of this disclosure can be implemented in software or hardware. The described units can also be housed in a processor; for example, a processor may be described as including a communication carrier construction unit, a data partitioning unit, a streaming data transmission unit, and a recomputable data processing unit. The names of these units do not necessarily limit the specific unit itself; for example, the communication carrier construction unit may also be described as "a unit that obtains a cross-computational node aggregate communication plane based on the topology partitioning of the same computing node, and creates shared memory windows within each computing node."
[0102] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), etc.
[0103] Obviously, those skilled in the art will understand that the various steps of the present invention described above can be performed in a manner different from that described above, and the simulation methods and experimental equipment include, but are not limited to, the above description. The steps of the present invention described above can be performed in a different order in certain circumstances, and the steps shown or described above can be performed separately. Therefore, the present invention is not limited to any particular combination of hardware and software.
[0104] The above description, in conjunction with specific embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such deductions or substitutions should be considered within the scope of protection of the present invention.
Claims
1. A hybrid communication method for first-principles calculations on heterogeneous acceleration platforms, applied to heterogeneous acceleration platforms containing multiple computing nodes, for performing first-principles calculations based on plane-wave pseudopotential density functional theory, characterized in that... The method includes: A cross-computing node collective communication plane is obtained based on the topology partitioning of the same computing node, and a shared memory window is created within each computing node; In each round of self-consistent field iteration or molecular dynamics iteration, the data used in each round of iteration is divided into streaming data and recalcible data based on whether the data used in each round of iteration is updated with changes in atomic position, whether it can be independently and redundantly recalculated within a single computing node, and whether it is necessary to maintain a globally consistent view across computing nodes. The streaming data is transmitted between computing nodes via the cross-computing node aggregate communication plane; Each computing node independently and redundantly recalculates to obtain the complete recalculable data, and uses the shared memory window within each computing node to share the recalculable data with multiple processes within the same computing node. In each round of self-consistent field iteration or molecular dynamics iteration, the transmission process of the streaming data and the sharing process of the recalculated data are carried out in concert.
2. The first-principles computing hybrid communication method for heterogeneous acceleration platforms according to claim 1, characterized in that, The recomputable data is data that is updated as the atomic position changes, can be independently and redundantly recomputed within a single computing node, and does not require maintaining a globally consistent view across computing nodes.
3. The first-principles computing hybrid communication method for heterogeneous acceleration platforms according to claim 1, characterized in that, The method of sharing the recomputable data with multiple processes within the same computing node using shared memory windows within each computing node includes: The complete recalculated data is written into the shared memory window by executing an independent redundant recalculation process. After each process completes internal synchronization within the computing node, it reads the recomputable data from the shared memory window.
4. The first-principles computing hybrid communication method for heterogeneous acceleration platforms according to claim 3, characterized in that, The following management steps are performed on the shared memory window: The upper bound of the capacity of the shared memory window is determined based on the upper limit of the number of atoms in a single computing node and the upper limit of the size of the atomic sphere grid. During each round of self-consistent field iteration or molecular dynamics iteration, it is determined whether the shared memory window needs to be rebuilt based on the actual capacity requirement and the allocated capacity of the shared memory window. If the actual capacity requirement exceeds the allocated capacity, then the shared memory window is rebuilt; If the actual capacity requirement does not exceed the allocated capacity, the created shared memory window is reused, and the newly obtained recalculated data in this iteration is overwritten into the shared memory window.
5. The first-principles computing hybrid communication method for heterogeneous acceleration platforms according to claim 1, characterized in that, In the first-principles calculations based on plane-wave pseudopotential density functional theory, the recalcifiable data includes at least one of the following: nonlocal pseudopotential β-projection operator, real-space grid values corresponding to the nonlocal pseudopotential β-projection operator, spatial derivatives of the nonlocal pseudopotential β-projection operator with respect to atomic coordinates, and atomic sphere grid index.
6. The first-principles computing hybrid communication method for heterogeneous acceleration platforms according to claim 1, characterized in that, A unified collection communication interface layer is set up for the streaming data. The unified collection communication interface layer provides standard collection communication primitives to the caller and is compatible with various collection communication backends. The target set communication backend is invoked through the unified set communication interface layer, and the streaming data is transmitted between each computing node based on the set communication plane across computing nodes.
7. The first-principles computing hybrid communication method for heterogeneous acceleration platforms according to claim 1, characterized in that, The method further includes: receiving a mode selection instruction, the mode selection instruction being used to indicate whether to enable a first operating mode; When the mode selection instruction indicates that the first running mode is enabled, the independent redundant recalculation is performed on the recalcible data, and the recalcible data is shared with multiple processes in the same computing node through the shared memory window. When the mode selection instruction indicates that the first operating mode is not enabled, the recomputable data is transmitted as streaming data through the aggregated communication plane across computing nodes.
8. A hybrid communication device for first-principles calculations on heterogeneous acceleration platforms, applied to a heterogeneous acceleration platform containing multiple computing nodes, for performing first-principles calculations based on plane-wave pseudopotential density functional theory, characterized in that... include: The communication carrier construction module is used to obtain a cross-computing node aggregate communication plane based on the topology partitioning of the same computing node, and to create shared memory windows within each computing node. The data partitioning module is used to divide the data used in each round of self-consistent field iteration or molecular dynamics iteration into streaming data and recalcible data based on whether the data used in each round of iteration is updated with changes in atomic positions, whether it can be independently and redundantly recalculated within a single computing node, and whether it is necessary to maintain a globally consistent view across computing nodes. A streaming data transmission module is used to transmit the streaming data between computing nodes through the cross-computing node aggregate communication plane; The recalcible data processing module is used to independently and redundantly recalculate the complete recalcible data on each computing node, and to share the recalcible data with multiple processes within the same computing node using the shared memory window within each computing node. In each round of self-consistent field iteration or molecular dynamics iteration, the transmission process of the streaming data and the sharing process of the recalculated data are carried out in concert.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.
