MLBM-DEM pipeline scheduling optimization method oriented to heterogeneous multi-domain processor

Abstract

The invention provides an MLBM-DEM pipeline scheduling optimization method oriented to a heterogeneous multi-domain processor. The method comprises the following steps: S1, executing data division based on application awareness and a heterogeneous multi-domain processor, and respectively calculating the calculation amount of MLBM and DEM and the calculation capability proportion of the MLBM and the DEM; s2, computing resources of a CPU end and an acceleration domain cluster are allocated according to the computing power proportion of the MLBM and the DEM, and a pipeline design module oriented to a heterogeneous multi-domain processor is constructed; and S3, obtaining residual computing resources except the CPU end and the acceleration domain cluster, and executing a Manager-Worker real-time scheduling mechanism. The MLBM-DEM pipeline scheduling optimization calculation is realized under the condition that the calculation precision is ensured, and the overall calculation performance is further improved.

Description

[Technical field]

[0001] The present invention relates to the fields of fluid mechanics and high performance computing, and in particular to an MLBM-DEM pipeline scheduling optimization method for heterogeneous multi-domain processors. [Background technology]

[0002] In today's era of rapid technological development, high-performance computing plays an important role in many fields. Fluid mechanics simulation, as the basis of scientific research and engineering design, requires large-scale computing resources to achieve accurate simulation. The Discrete Element Multi-Level Lattice Boltzmann Method (MLBM-DEM) is a numerical simulation technology that combines the Discrete Element Method (DEM) and the Multi-Level Lattice Boltzmann Method (MLBM), which aims to simulate and analyze complex systems composed of fluids and granular materials. This method combines the advantages of DEM and MLBM, and is used to study the dynamic behavior of particle-fluid interactions, with broad application prospects. Due to its complex scenarios, MLBM-DEM has a larger amount of calculations, so using the heterogeneous multi-domain processor acceleration of the domestic E-class supercomputer system can effectively reduce the amount of calculations and improve computing performance and simulation accuracy. However, since the core calculation of MLBM-DEM lies in the collision migration calculation in MLBM and the interaction force between particles in DEM, the calculation of these two parts is carried out in parallel on a heterogeneous multi-domain processor, and the rest of the boundary processing and other work are placed on the CPU side. This will produce a pipeline with strong data dependencies, leading to challenges in data partitioning and load balancing, especially when applied to domestic E-class supercomputer systems.

[0003] Nowadays, the scale and complexity of computational fluid dynamics research are increasing. A single computing node can no longer meet the computing needs and must rely on high-performance computers. my country's supercomputers are developing from P-level to E-level, and heterogeneous multi-processors and accelerators have become a development trend. With the bottleneck of single-core performance improvement after Dennard expansion, it has become necessary to build multi-core high-performance processors.

[0004] How to implement the pipeline scheduling optimization method of MLBM-DEM on heterogeneous multi-domain processors to fully utilize the performance of heterogeneous multi-domain processors and improve the computational efficiency of MLBM-DEM, as well as how to design and optimize the pipeline scheduling algorithm with strong data dependence have become urgent issues to be solved. [Summary of the invention]

[0005] In view of the shortcomings of the prior art, the present application provides an MLBM-DEM pipeline scheduling optimization method for heterogeneous multi-domain processors, realizes data partitioning based on application perception and heterogeneous multi-domain processors, constructs a pipeline design module for heterogeneous multi-domain processors, and designs a Manager-Worker real-time scheduling mechanism to optimize pipeline scheduling, thereby further improving the overall computing performance.

[0006] The present application provides an MLBM-DEM pipeline scheduling optimization method for heterogeneous multi-domain processors, comprising the following steps:

[0007] S1, perform data partitioning based on application awareness and heterogeneous multi-domain processors, and calculate the computational amount of MLBM and DEM and the computing capacity ratio of MLBM and DEM respectively;

[0008] S2. Allocate computing resources of the CPU end and the acceleration domain cluster according to the computing power ratio of the MLBM and the DEM, and construct a pipeline design module for heterogeneous multi-domain processors;

[0009] S3: Obtain the remaining computing resources except the CPU and acceleration domain cluster, and execute the Manager-Worker real-time scheduling mechanism.

[0010] Preferably, the heterogeneous multi-domain processor includes a general area and an acceleration area, including the following features:

[0011] The heterogeneous multi-domain processor contains 16 CPUs, 96 control cores (Ctrl) and 1536 acceleration cores (Acc), achieving 11.6TFLOPS of double-precision performance and 45.4Gfops / w of power efficiency when running at 1.2GHz. The heterogeneous multi-domain processor has the following features:

[0012] (1) Heterogeneous multi-domain microarchitecture, where 16 CPUs form a general area, 96 control cores (Ctrl) and 1536 acceleration cores (Acc) form an acceleration area, and the acceleration area is evenly divided into four acceleration domain clusters;

[0013] (2) Combining very long instruction word (VLIW) and acceleration array micro-architecture, where each acceleration core works in VLIW mode, and every 16 acceleration cores (Acc) and 1 control core (Ctrl) are organized into an acceleration array;

[0014] (3) Ultra-high bandwidth and high reliability array memory, which is a vector memory that supports up to two vector load / stores. The data types loaded / stored by the array memory include half-word (32 bits), word (64 bits), and double-word (128 bits). Therefore, the array memory can provide up to 512 bytes (16 × 2 × 128 bits) to 16 accelerator cores simultaneously.

[0015] Each acceleration domain cluster has 24 control cores (Ctrl), 384 acceleration cores (Acc), on-chip shared memory (GSM), high-bandwidth shared memory (HBSM), and off-chip DDR memory. Each acceleration domain cluster executes independently of each other. The CPU in the general area can access all HBSM and DDR spaces of different acceleration domain clusters, while the control core and accelerator core can only access the GSM, HBSM, and DDR in their own acceleration domain cluster.

[0016] Data sharing between different acceleration domain clusters is achieved through the CPU side. The 16 CPUs in the general area have their own L1 and L2 caches and are connected through a mesh NoC that supports cache consistency. The general area is capable of overall task control, operating system startup, and general processing, and the acceleration area is designed for compute-intensive tasks. In each acceleration domain cluster, 1 control core (Ctrl) and 16 acceleration cores (Acc) form an acceleration array, and 24 control cores (Ctrl) and 384 acceleration cores (Acc) form 24 acceleration arrays.

[0017] Preferably, the data partitioning based on application awareness and heterogeneous multi-domain processors comprises the following steps:

[0018] In heterogeneous computing, the relationship between task size and computing power is crucial. Reasonable task division can significantly improve computing efficiency and achieve load balancing. Based on this, a dynamic data division method based on application awareness is proposed. This method dynamically divides and allocates computing grids according to the characteristics of computing tasks and the size of data. This method can dynamically adapt and optimize the use of computing resources according to the needs of different computing tasks, thereby improving computing efficiency and reducing resource waste. It includes the following steps:

[0019] S11. Calculate the adjusted parallel processing capacity

[0020]

[0021] Among them, P adj is the adjusted parallel processing capability, C i is the computing power of the ith core, f i is the proportion of this type of core in the total calculation, and n is the number of cores of this type; for heterogeneous multi-domain processors, it includes CPU and acceleration domain clusters;

[0022] S12. Obtain the required ratio of computing volume to data volume

[0023]

[0024] Among them, R is the required ratio of computing amount to data amount, C and D are the computing amount and data amount required to complete a grid block respectively;

[0025] S13. Calculating the ideal size of pipeline grid blocks on heterogeneous multi-domain processors

[0026]

[0027] Among them, B is the memory bandwidth, L is the data transmission delay;

[0028] S14. Calculate the size of pipeline grid blocks on heterogeneous multi-domain processors

[0029]

[0030] Among them, α is the adaptive adjustment factor;

[0031] S15. Calculation of MLBM and DEM and the ratio of computing power between them

[0032] Assume that the computational effort of MLBM is C MLBM , represents the amount of computation required to perform the MLBM operation, including the collision and migration steps; the amount of computation required for DEM is C DEM , represents the amount of calculation required to perform DEM operations, including the calculation of interactions between particles; the initial data volume of MLBM is D MLBM , represents the initial data volume of MLBM in the user input data scale; the initial data volume of DEM is D DEM , represents the initial data volume of DEM in the user input data scale; assuming that the computing power ratio between MLBM and DEM is R comp ,but:

[0033] The computational cost of MLBM C MLBM It is related to the number of grid points and the number of operations performed on each grid point. Assuming that an average of O MLBM operations, then:

[0034]

[0035] The amount of DEM calculation C DEM It depends on the number of particles and the computational complexity of the interaction between each pair of particles; assuming that each pair of particles performs an average of O DEM operations. Since the interaction between particles usually involves all pairs of particles, the amount of calculation is approximately:

[0036]

[0037] Therefore, the computing power ratio between MLBM and DEM is R comp , expressed as:

[0038]

[0039] Preferably, the pipeline design module for heterogeneous multi-domain processors includes, in sequence:

[0040] MLBM information update and boundary processing module: Based on the relevant information fed back by the DEM information update and feedback module, the CPU side updates the MLBM information and processes the boundary conditions;

[0041] MLBM collision migration calculation module: Based on MLBM information update and boundary processing, it is used to perform MLBM collision migration calculation in the acceleration domain cluster;

[0042] Meshing model: Based on MLBM collision migration calculation, CPU-side adaptive meshing;

[0043] Information transmission module: The CPU transmits the relevant information of MLBM to DEM;

[0044] DEM core calculation module: used to perform DEM core calculation in the acceleration domain cluster;

[0045] DEM information update and feedback module: used by the CPU to update DEM related information and feed the information back to the MLBM information update and boundary processing module.

[0046] In order to avoid pipeline blocking, the following strategies are designed:

[0047] (1) Introducing a buffer: A buffer is set between each pipeline module to store temporary calculation results and ensure smooth data transmission between processors. For example, when the MLBM collision migration calculation module completes the MLBM collision migration calculation, the result will be stored in the buffer and wait for the CPU to perform subsequent processing.

[0048] (2) Efficient inter-process communication: MPI is used for efficient inter-process communication to transmit Halo layer data in a timely manner, ensuring that the next stage of calculation is not affected by data transmission delays.

[0049] (3) Asynchronous computing and communication overlap: The use of asynchronous computing and communication overlap technology enables the processor to simultaneously receive and send data while performing calculations, further improving the overall efficiency of the pipeline.

[0050] Preferably, the Manager-Worker real-time scheduling mechanism comprises the following steps:

[0051] S31, dividing the remaining computing resources into a Manager node and multiple Worker nodes;

[0052] S32. Based on the real-time polling and dynamic scheduling mechanism, the multiple Worker nodes execute task scheduling and regularly report task progress to the Manager;

[0053] S33, the Manager receives the task progress of the Worker node, dynamically schedules the Worker node according to the real-time data, and sends a termination message to the Worker node after all tasks are completed;

[0054] S34, the Worker node receives the termination message, performs cleanup work and exits. Experiments have shown that when the method is used to optimize the MLBM-DEM pipeline scheduling for heterogeneous multi-domain processors, the computing performance of 16 CPU cores and 4 acceleration domain clusters is significantly improved compared with 16 CPU cores, especially in large-scale data sets, and even reaches more than 3 times the original computing performance.

[0055] While ensuring the calculation accuracy, the present invention realizes data partitioning based on application perception and heterogeneous multi-domain processors, designs a Manager-Worker real-time scheduling mechanism to optimize pipeline scheduling, and further improves the overall computing performance.

Brief Description of the Drawings

[0056] Figure 1 It is a flow chart of an MLBM-DEM pipeline scheduling optimization method for heterogeneous multi-domain processors of the present invention;

[0057] Figure 2 It is an acceleration domain framework diagram of a heterogeneous multi-domain processor of an MLBM-DEM pipeline scheduling optimization method for a heterogeneous multi-domain processor of the present invention;

[0058] Figure 3 This is a Manager-Worker real-time scheduling mechanism diagram of the MLBM-DEM pipeline scheduling optimization method for heterogeneous multi-domain processors of the present invention. [Specific implementation method]

[0059] The present invention discloses an MLBM-DEM pipeline scheduling optimization method for heterogeneous multi-domain processors. The present application has the characteristics of high computing performance while ensuring accuracy, and is widely used in multiple fields, including geotechnical engineering (such as soil liquefaction research and slope stability analysis), chemical engineering (such as multiphase flow reactor design and particle flow), biofluid mechanics (such as drug delivery and cell movement), environmental engineering (such as pollutant diffusion and soil erosion), energy engineering (such as particle combustion and gas-solid two-phase flow), etc.

[0060] The present application can significantly improve the computing performance while ensuring the accuracy. The specific implementation methods of the present application will be described in detail below in conjunction with the accompanying drawings. It is worth noting that what is described in this implementation method is only a common use method of the present application. The technology used in the present application should be protected until practitioners in related fields can develop better solutions.

[0061] In one embodiment, Figure 1 The figure shows the overall flow chart of an MLBM-DEM pipeline scheduling optimization method for heterogeneous multi-domain processors. It can be seen that

[0062] Step 101, performing data partitioning based on application awareness and heterogeneous multi-domain processors, and calculating the computational load of MLBM and DEM and the computational capacity ratio of MLBM and DEM respectively;

[0063] Step 102, allocating computing resources of the CPU end and the acceleration domain cluster according to the computing power ratio of the MLBM and the DEM, and constructing a pipeline design module for heterogeneous multi-domain processors;

[0064] Step 103: Obtain the remaining computing resources except the CPU end and the acceleration domain cluster, and execute the Manager-Worker real-time scheduling mechanism.

[0065] In step 101, data partitioning based on application awareness and heterogeneous multi-domain processors includes the following steps:

[0066] (1) Calculate the adjusted parallel processing capacity

[0067]

[0068] Among them, P adj is the adjusted parallel processing capability, C i is the computing power of the ith core, f i is the proportion of this type of core in the total calculation, and n is the number of cores of this type; for heterogeneous multi-domain processors, it includes CPU and acceleration domain clusters;

[0069] (2) Obtaining the required ratio of computing volume to data volume

[0070]

[0071] Among them, R is the required ratio of computing amount to data amount, C and D are the computing amount and data amount required to complete a grid block respectively;

[0072] (3) Calculating the ideal size of pipeline grid blocks on heterogeneous multi-domain processors

[0073]

[0074] Among them, B is the memory bandwidth, L is the data transmission delay;

[0075] (4) Calculating the size of pipeline grid blocks on heterogeneous multi-domain processors

[0076]

[0077] Among them, α is the adaptive adjustment factor;

[0078] (5) Calculation of the computational effort of MLBM and DEM and the computing power ratio between the two

[0079] Assume that the computational effort of MLBM is C MLBM , represents the amount of computation required to perform the MLBM operation, including the collision and migration steps; the amount of computation required for DEM is C DEM , represents the amount of calculation required to perform DEM operations, including the calculation of interactions between particles; the initial data volume of MLBM is D MLBM , represents the initial data volume of MLBM in the user input data scale; the initial data volume of DEM is D DEM , represents the initial data volume of DEM in the user input data scale; assuming that the computing power ratio between MLBM and DEM is R comp ,but:

[0080] The computational cost of MLBM C MLBM It is related to the number of grid points and the number of operations performed on each grid point. Assuming that an average of O MLBM operations, then:

[0081]

[0082] The amount of DEM calculation C DEM It depends on the number of particles and the computational complexity of the interaction between each pair of particles; assuming that each pair of particles performs an average of O DEM operations. Since the interaction between particles usually involves all pairs of particles, the amount of calculation is approximately:

[0083]

[0084] Therefore, the computing power ratio between MLBM and DEM is R comp , expressed as:

[0085]

[0086] In step 102, the pipeline design module for the heterogeneous multi-domain processor includes:

[0087] MLBM information update and boundary processing module: Based on the relevant information fed back by the DEM information update and feedback module, the CPU side updates the MLBM information and processes the boundary conditions;

[0088] MLBM collision migration calculation module: Based on MLBM information update and boundary processing, it is used to perform MLBM collision migration calculation in the acceleration domain cluster;

[0089] Meshing model: Based on MLBM collision migration calculation, CPU-side adaptive meshing;

[0090] Information transmission module: The CPU transmits the relevant information of MLBM to DEM;

[0091] DEM core calculation module: used to perform DEM core calculation in the acceleration domain cluster;

[0092] DEM information update and feedback module: used by the CPU to update DEM related information and feed the information back to the MLBM information update and boundary processing module.

[0093] In order to avoid pipeline blocking, the following strategies are designed:

[0094] (1) Introducing a buffer: A buffer is set between each pipeline module to store temporary calculation results and ensure smooth data transmission between processors. For example, when the acceleration domain cluster completes the collision migration calculation of MLBM, the result will be stored in the buffer and wait for the CPU to perform subsequent processing.

[0095] (2) Efficient inter-process communication: MPI is used for efficient inter-process communication to transmit Halo layer data in a timely manner, ensuring that the next stage of calculation is not affected by data transmission delays.

[0096] (3) Asynchronous computing and communication overlap: The use of asynchronous computing and communication overlap technology enables the processor to simultaneously receive and send data while performing calculations, further improving the overall efficiency of the pipeline.

[0097] In step 103, the Manager-Worker real-time scheduling mechanism includes the following steps:

[0098] S31, dividing the remaining computing resources into a Manager node and multiple Worker nodes;

[0099] S32. Based on the real-time polling and dynamic scheduling mechanism, the multiple Worker nodes execute task scheduling and regularly report task progress to the Manager;

[0100] S33, the Manager receives the task progress of the Worker node, dynamically schedules the Worker node according to the real-time data, and sends a termination message to the Worker node after all tasks are completed;

[0101] S34. The Worker node receives the termination message, performs cleanup work and exits.

[0102] In one embodiment, Figure 2 The figure shows the overall framework of a heterogeneous multi-domain processor using the MLBM-DEM pipeline scheduling optimization method. It can be seen that

[0103] The heterogeneous multi-domain processor contains 16 CPUs, 96 control cores (Ctrl) and 1536 acceleration cores (Acc), achieving 11.6TFLOPS of double-precision performance and 45.4Gfops / w of power efficiency when running at 1.2GHz. The heterogeneous multi-domain processor has the following features:

[0104] (1) Heterogeneous multi-domain microarchitecture, where 16 CPU cores form a general domain, 96 control cores (Ctrl) and 1536 acceleration cores (Acc) form an acceleration domain, and the acceleration domain is evenly divided into four acceleration domain clusters;

[0105] (2) Combining very long instruction word (VLIW) and acceleration array microarchitecture, where each acceleration core works in VLIW mode and every 16 acceleration cores and 1 control core are organized into an acceleration array;

[0106] (3) Ultra-high bandwidth and high reliability array memory, which is a vector memory that supports up to two vector load / stores. The data types loaded / stored by the array memory include half-word (32 bits), word (64 bits), and double-word (128 bits). Therefore, the array memory can provide up to 512 bytes (16 × 2 × 128 bits) to 16 accelerator cores simultaneously.

[0107] Each acceleration domain cluster has 24 control cores (Ctrl), 384 acceleration cores (Acc), on-chip shared memory (GSM), high-bandwidth shared memory (HBSM) and off-chip DDR memory. Each acceleration domain cluster executes independently of each other. The CPU in the general area can access all HBSM and DDR spaces of different acceleration domain clusters, while the control core and accelerator core can only access the GSM, HBSM and DDR in their respective acceleration domain clusters. Data sharing between different acceleration domain clusters is achieved through the CPU. The 16 CPUs in the general area have their own L1 and L2 caches and are connected through a mesh NoC that supports cache consistency. The general area is capable of overall task control, operating system startup and general processing, and the acceleration area is designed for compute-intensive tasks. In each acceleration domain cluster, 1 control core (Ctrl) and 16 acceleration cores (Acc) form an acceleration array, and 24 control cores (Ctrl) and 384 acceleration cores (Acc) form 24 acceleration arrays.

[0108] In one embodiment, Figure 3 The figure shows an MLBM-DEM pipeline scheduling optimization method for heterogeneous multi-domain processors. It can be seen that

[0109] In step 301, the system will first preliminarily allocate tasks on the CPU and the acceleration domain cluster according to their computing requirements and the computing capabilities of their respective platforms. The entire computing cluster is divided into a Manager node and multiple Worker nodes, where Worker nodes include both CPU cores and computing nodes in the acceleration domain cluster. These nodes are not responsible for the initial fixed task allocation, but are dynamically managed by the Manager node.

[0110] In step 302, the Manager monitors the execution of all tasks in real time, paying particular attention to the data dependencies between tasks and the remaining computing amount of each task. Based on real-time data such as idle and busy status, the Manager can flexibly schedule Worker nodes to process computing tasks that are about to be completed and have important data dependencies on subsequent tasks, so as to avoid any task delays or interruptions. In addition, when the Worker node is close to completing the current task, the Manager continues to monitor the computing load of the entire system to ensure that the Worker node can be immediately rescheduled to process other tasks that need to be accelerated after completing a task. This real-time polling and dynamic task scheduling mechanism is designed to maintain the efficient operation of all computing nodes, avoid the idle state of any node, and ensure that tasks that are closely connected due to data dependencies can be executed smoothly and continuously.

[0111] By adopting this real-time monitoring and dynamic scheduling method, we can not only ensure the smoothness and efficiency of task execution, but also maximize the use of domestic E-class supercomputer system resources to achieve efficient and balanced allocation of computing tasks. This method is particularly suitable for processing complex computing tasks with strong data dependencies, and can effectively improve overall computing performance and resource utilization.

[0112] Those skilled in the art should understand that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0113] Finally, it should be noted that although the present invention is mainly based on specific cases in terms of example display and design, and the software involved is in the form of a system. However, the various components involved in this system should still be protected, including but not limited to data partitioning, Manager-Worker real-time scheduling mechanism and its code, so any modification, improvement, use, etc. of the implementation of the present invention should be covered in the claims and description of the present invention.

Claims

1. A MLBM-DEM pipeline scheduling optimization method for heterogeneous multi-domain processors, characterized in that: The following steps are involved: S1, perform data partitioning based on application awareness and heterogeneous multi-domain processors to obtain the computing power ratio of MLBM and DEM; S2. Allocate computing resources of the CPU and the acceleration domain cluster according to the computing power ratio of the MLBM and the DEM, and construct a pipeline design module for heterogeneous multi-domain processors; S3: Obtain the remaining computing resources except the CPU and acceleration domain cluster, and execute the Manager-Worker real-time scheduling mechanism.

2. The method according to claim 1, characterized in that The heterogeneous multi-domain processor includes a general area and an acceleration area, the general area includes 16 CPU cores; the acceleration area is evenly divided into four acceleration domain clusters, the acceleration domain clusters include 24 control cores Ctrl, 384 acceleration cores Acc, on-chip shared memory GSM, high-bandwidth shared memory HBSM and off-chip memory DDR.

3. The method according to claim 1, characterized in that The method of obtaining the computing power ratio of MLBM and DEM based on application awareness and data partitioning of heterogeneous multi-domain processors includes the following steps: S11. Calculate the adjusted parallel processing capacity Among them, P adj is the adjusted parallel processing capability, C i is the computing power of the ith core, f i is the proportion of this type of core in the total calculation, and n is the number of cores of this type; for heterogeneous multi-domain processors, it includes CPU and acceleration domain clusters; S12. Obtain the required ratio of computing volume to data volume Among them, R is the required ratio of computing amount to data amount, C and D are the computing amount and data amount required to complete a grid block respectively; S13. Calculating the ideal size of pipeline grid blocks on heterogeneous multi-domain processors Among them, B is the memory bandwidth, L is the data transmission delay; S14. Calculate the size of pipeline grid blocks on heterogeneous multi-domain processors Among them, α is the adaptive adjustment factor; S15. Calculation of MLBM and DEM and the ratio of computing power between them Assume that the computational effort of MLBM is C MLBM , represents the amount of computation required to perform the MLBM operation, including the collision and migration steps; the amount of computation required for DEM is C DEM , represents the amount of calculation required to perform DEM operations, including the calculation of interactions between particles; the initial data volume of MLBM is D MIBM , represents the initial data volume of MLBM in the user input data scale; the initial data volume of DEM is D DEM , represents the initial data volume of DEM in the user input data scale; assuming that the computing power ratio between MLBM and DEM is R comp ,but: The computational cost of MLBM C MLBM It is related to the number of grid points and the number of operations performed on each grid point. Assuming that an average of O MLBM operations, then: The amount of DEM calculation C DEM It depends on the number of particles and the computational complexity of the interaction between each pair of particles; assuming that each pair of particles performs an average of O DEM operations. Since the interaction between particles usually involves all pairs of particles, the amount of calculation is approximately: Therefore, the computing power ratio between MLBM and DEM is R comp , expressed as:

4. The method according to claim 3, characterized in that: The pipeline design module for heterogeneous multi-domain processors includes: MLBM information update and boundary processing module: Based on the relevant information fed back by the DEM information update and feedback module, the CPU side updates the MLBM information and processes the boundary conditions; MLBM collision migration calculation module: Based on MLBM information update and boundary processing, it is used to perform MLBM collision migration calculation in the acceleration domain cluster; Meshing model: Based on MLBM collision migration calculation, CPU-side adaptive meshing; Information transmission module: The CPU transmits the relevant information of MLBM to DEM; DEM core calculation module: used to perform DEM core calculation in the acceleration domain cluster; DEM information update and feedback module: used by the CPU to update DEM related information and feed the information back to the MLBM information update and boundary processing module.

5. The method according to claim 4, characterized in that A buffer is set between each pipeline design module to store temporary calculation results.

6. The method according to claim 1, characterized in that The Manager-Worker real-time scheduling mechanism includes the following steps: S31, dividing the remaining computing resources into a Manager node and multiple Worker nodes; S32. Based on the real-time polling and dynamic scheduling mechanism, the multiple Worker nodes execute task scheduling and regularly report task progress to the Manager; S33, the Manager receives the task progress of the Worker node, dynamically schedules the Worker node according to the real-time data, and sends a termination message to the Worker node after all tasks are completed; S34. The Worker node receives the termination message, performs cleanup work and exits.

Images