A large-scale MRP parallel scheduling method and system based on material-dependent LLC topology hierarchy
By using a material-dependent LLC topology hierarchical method, the BOM is decomposed into a flat array for parallel computation, which solves the problems of long computation time and stability in large-scale MRP systems, and achieves efficient parallel computing and improved system stability.
Patent Information
- Application Number
- CN202611285369.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-24
- Publication Date
- 2026-11-10
Smart Images

Figure CN123070592A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of Material Requirements Planning (MRP) and Advanced Planning and Scheduling (APS). Specifically, it is a large-scale MRP parallel scheduling method and system based on material-dependent LLC topology hierarchy. Background Technology
[0002] In discrete manufacturing industries (such as electronics assembly, automobile manufacturing, and heavy equipment), Material Requirements Planning (MRP) is the core engine of supply chain operations. Its basic function is to deduct inventory and supplies in transit layer by layer from top to bottom along the Bill of Materials (BOM) structure based on customer orders and forecasted demand, and generate suggested procurement plans or production plans (Planned Orders) at material gaps.
[0003] In large-scale discrete manufacturing plants, MRP calculations face extremely high computational pressure and time-sensitive challenges:
[0004] 1. The bottleneck of "computation serialization" caused by the depth of material structure dependence (BOM cascading).
[0005] There are strong parent-child cascading dependencies between materials (e.g., finished product FG consumes semi-finished product Sub, and semi-finished product Sub consumes raw material Raw). After calculating the suggested replenishment quantity of the parent item, the resulting demand is automatically passed on to the gross demand of the child item. This means that MRP must be calculated strictly from top to bottom according to the material structure, and blind global parallelism is not possible. Traditional MRP systems can only use a single thread to sequentially traverse hundreds of thousands of SKU nodes, resulting in large-scale planning calculations taking up to several hours, making it impossible to integrate into What-If sandbox decision-making processes that require only seconds.
[0006] 2. Multi-level BOM loop deadlock causes computation to be suspended.
[0007] When Enterprise Engineering Changes (ECNs) are frequent or master data maintenance is chaotic, circular dependencies between materials are easily generated (e.g., A's sub-components include B, B's sub-components include C, and C's sub-components refer back to A). If the scheduling engine blindly performs multi-threaded calculations without static loop detection, the calculation will fall into infinite recursion (stack overflow) or deadlock between threads, causing the calculation engine to crash and hang completely. In complex network graph structures, dynamic loop detection is extremely expensive, and traditional systems can only throw errors or limit the calculation level during calculation.
[0008] 3. Multi-threaded concurrent lock contention and cache miss.
[0009] If a traditional system attempts to use simple object-oriented (OOP) multithreading techniques, that is, assigning an independent computing thread to each material node, the frequent demand passing and state updates between parent and child nodes require threads to frequently lock the shared material inventory and allocation ledger. High thread contention and complex pointer jumps cause frequent invalidations of the CPU's L1 / L2 cache, and the computing performance of multi-core processors may actually degrade or even fall below that of single-threaded systems.
[0010] Therefore, there is an urgent need for a large-scale MRP parallel scheduling method and system based on material-dependent LLC topology hierarchy to solve the technical problems of traditional MRP in large-scale computing, such as extremely long serial time, BOM loop deadlock and collapse, and severe multi-threaded concurrent lock contention, so as to achieve second-level static topology loop prevention, absolute parallelism with zero conflict within the layer, and extremely fast planning calculation close to the physical limit of the CPU. Summary of the Invention
[0011] Therefore, the technical problem to be solved by the present invention is to provide a large-scale MRP parallel elimination scheduling method and system based on material-dependent LLC topology hierarchy. This method solves the technical problems of traditional MRP in large-scale computing, such as extremely long serial time, BOM loop deadlock and collapse, and severe multi-threaded concurrent lock contention. It achieves second-level static topology loop prevention, absolute parallelism with zero conflict within the layer, and extremely fast planning calculation close to the physical limit of the CPU.
[0012] To solve the above-mentioned technical problems, the present invention provides the following technical solution:
[0013] A large-scale MRP parallel elimination scheduling method based on material-dependent LLC topology hierarchy includes the following steps:
[0014] Step P1: Load the material and BOM basic data, and generate the low-level code corresponding to each material through depth-first topology traversal with loop identification.
[0015] Step P2: Compile the product structure BOM topology into a flat array arranged contiguously in physical memory;
[0016] Step P3: After the topology compilation is completed, the multi-layer parallel computing grid is divided according to the low-level code, and the material supply and demand calculation is completed concurrently in multiple threads within the grid. The MRP parallel consumption scheduling is completed through the lock-free cross-layer demand transmission and hierarchical synchronization mechanism.
[0017] The above-mentioned large-scale MRP parallel elimination scheduling method based on material-dependent LLC topology hierarchy, in step P1, the specific process of depth-first topology traversal with loop identification includes:
[0018] Step P1-1: Traverse the directed hypergraph of materials starting from the finished product material, and use a three-color state marking method to distinguish the traversal state of each material node.
[0019] Step P1-2: During the traversal, calculate the maximum depth of all dependent paths for each material and generate the corresponding low-level code;
[0020] Steps P1-3: If a circular dependency is identified during the traversal, the static loop detection interceptor is triggered.
[0021] In the above-mentioned large-scale MRP parallel elimination scheduling method based on material-dependent LLC topology hierarchy, the three-color status marking method in step P1-1 is as follows: white mark indicates unvisited, gray mark indicates being visited, and black mark indicates visit completed.
[0022] In the above-mentioned large-scale MRP parallel elimination scheduling method based on material-dependent LLC topology hierarchy, in steps P1-2, the low-level code of the finished product is 0, and the low-level code of the bottom raw material is the largest.
[0023] In the above-mentioned large-scale MRP parallel elimination scheduling method based on material-dependent LLC topology hierarchy, in steps P1-3, the cyclic dependency relationship is identified as: encountering gray-marked nodes during the search process;
[0024] Static loop detection interceptor execution:
[0025] A. Abort topology calculation;
[0026] B. Extract the material dependency chain in reverse through the recursive call stack, and persistently store the abnormal information containing the chain nodes and paths into the audit database.
[0027] The above-mentioned large-scale MRP parallel elimination scheduling method based on material-dependent LLC topology hierarchy includes the following specific steps in step P2:
[0028] Step P2-1: Construct a material code mapping dictionary to uniformly convert string-format material codes into fixed-length numeric identifiers;
[0029] Step P2-2: Decompose the tree-shaped BOM dependency relationship into standardized flat association records and store them uniformly as an array structure;
[0030] Steps P2-3: Sort all flat, related records according to their parent material identifiers, ensuring that all child material records corresponding to the same parent material are stored contiguously in physical memory, and construct a head-tail offset index vector to achieve... Constant-time memory location and CPU cache prefetching.
[0031] The above-mentioned large-scale MRP parallel scheduling method based on material-dependent LLC topology hierarchy, in step P3, the specific process of dividing multiple parallel computing grids according to the low-level code and completing the material supply and demand calculation concurrently in multiple threads within the grid includes the following steps:
[0032] Step P3-1: Group all materials according to the maximum low-level code corresponding to each material to form a multi-layer parallel computing grid;
[0033] Step P3-2: Set up the outer timing loop, perform calculations layer by layer in ascending order of the lower-level codes, distribute multi-core concurrent calculations within a single-layer grid using the OpenMP thread pool, and perform hierarchical barrier synchronization after each layer of calculation is completed before proceeding to the next layer of grid calculation;
[0034] Step P3-3: Based on the characteristic that materials in the same lower layer code have no mutual dependence, the supply and demand deduction calculation is performed concurrently by multiple threads for all materials in a single layer grid.
[0035] In the above-mentioned large-scale MRP parallel consumption scheduling method based on material-dependent LLC topology hierarchy, in step P3, when a gap is generated in the material supply and demand calculation, a planned order is generated, and the pull amount transmitted to downstream sub-materials is calculated in combination with the single-item consumption coefficient.
[0036] In the above-mentioned large-scale MRP parallel absorption scheduling method based on material-dependent LLC topology hierarchy, step P3 utilizes the hierarchical isolation characteristic of the sub-material's lower-level code being higher than the parent material's code, and uses hardware-level atomic instructions to directly accumulate physical memory, superimposing the pull amount into the corresponding sub-material's demand data, eliminating thread mutex lock overhead and ensuring the absolute correctness of the accumulation operation.
[0037] A large-scale MRP parallel elimination scheduling system based on material-dependent LLC topology hierarchy, used to implement the above-mentioned large-scale MRP parallel elimination scheduling method, includes:
[0038] The low-level code generation module is used to load material and BOM basic data and generate the low-level code corresponding to each material through depth-first topology traversal with loop recognition.
[0039] The topology compilation module is used to compile the product structure BOM into a flat array that is arranged contiguously in physical memory;
[0040] The supply and demand calculation module is used to divide the low-level code into multiple parallel computing grids after the topology is compiled, and to complete the supply and demand deduction in multiple threads within the grid through the OpenMP thread pool. It also realizes lock-free cross-layer demand transmission and hierarchical barrier synchronization based on hardware-level atomic instructions, and completes MRP parallel absorption scheduling.
[0041] The technical solution of the present invention achieves the following beneficial technical effects:
[0042] This invention achieves lock-free concurrent solving within a hierarchy based on the dependency-free characteristics of materials at the same level in lower-level codes. This eliminates the lock waiting overhead caused by multi-threaded data contention, thereby eliminating conflicts arising from parent-child dependencies in same-level calculations. Computational efficiency increases linearly with the number of CPU cores. For large-scale MRP cascaded deployment scenarios involving 50,000 types of materials and 3 million orders, existing serial solutions typically take several hours. This method can compress the overall solution time to less than 800 milliseconds, achieving a significant leap in efficiency for large-scale supply and demand calculations.
[0043] This invention performs a full material dependency graph loop detection using a three-color-marked depth-first traversal (DFS) before the MRP solution begins, identifying and intercepting BOM circular dependency anomalies in advance, and simultaneously retaining complete audit records of the anomaly links. This fundamentally avoids program crashes such as recursive stack overflows and multi-threaded deadlocks during the solution process, achieving 100% pre-interception of material topology anomalies and significantly reducing the probability of system malfunctions.
[0044] This invention adopts a flat, contiguous memory storage architecture oriented towards data DOD, coupled with a hierarchical dynamic batch scheduling mechanism (Dynamic Chunking), to reconstruct the memory fragmentation problem caused by traditional object-oriented discrete storage, improve the CPU cache line hit rate to over 95%, significantly reduce the number of cache misses, completely avoid the frequent cache misses caused by traditional OOP structures, achieve high-speed data reading with lock-free concurrency and no redundant memory copying, and further reduce computing power consumption. Attached Figure Description
[0045] Figure 1 This is a flowchart of a large-scale MRP parallel absorption scheduling method based on material-dependent LLC topology hierarchy according to the present invention. Detailed Implementation
[0046] Example 1 discloses a large-scale MRP parallel scheduling method based on material-dependent LLC topology hierarchy, such as... Figure 1 As shown, it includes the following steps:
[0047] Step P1: Load material and BOM basic data. During the planned ignition phase, the solver traverses the flat directed hypergraph network, runs an efficient depth-first search (DFS) topology sort, and compiles absolute low-level code (LLC) for tens of thousands of material nodes across the network. The specific process includes the following steps:
[0048] Step P1-1: Starting with the finished product, traverse the directed hypergraph of materials, using a three-color status marker to distinguish the traversal status of each material node. The three-color status marker is: white indicates unvisited, gray indicates being visited, and black indicates visited. Through a depth-first search using three-color markers, the entire material topology traversal, low-level code calculation, and loop pre-identification are completed simultaneously. This allows for the global material hierarchy partitioning to be completed in one go before the formal MRP calculation, accurately obtaining the unique hierarchy identifier for each material and providing a reliable hierarchical basis for subsequent lock-free parallel computation.
[0049] Steps P1-2: During the traversal, calculate the maximum depth of all dependent paths for each material and generate the corresponding low-level code. The low-level code is defined as the maximum depth of the directed path of a material in the supply chain network structure. The low-level code of finished products is 0, and the low-level code of raw materials at the bottom layer is the largest.
[0050] Steps P1-3: If a circular dependency is identified during the traversal, the static loop detection interceptor is triggered. A circular dependency is identified when a gray-marked node is encountered during the search.
[0051] Static loop detection interceptor execution:
[0052] A. Abort topology calculation;
[0053] B. Extracting circular material dependency chains by reversing the recursive call stack (e.g.) This method persistently stores anomaly information, including process nodes and paths, in the audit database. During the traversal process, it captures material circular dependencies in real time, proactively blocking abnormal calculation flows and fundamentally preventing recursive stack overflows and multi-threaded deadlocks during the MRP solution phase. This significantly improves the stability of large-scale material network solutions and the efficiency of pre-processing.
[0054] Step P2: Compile the product structure BOM topology into a flat array arranged contiguously in physical memory; the specific process includes:
[0055] Step P2-1: Construct a material code mapping dictionary to uniformly convert string-format material codes into fixed-length numeric identifiers;
[0056] Step P2-2: Decompose the tree-shaped BOM dependency relationship into standardized flat association records and store them uniformly as an array structure;
[0057] Steps P2-3: Sort all flat, related records by parent material identifier, ensuring that all child material records corresponding to the same parent material are stored contiguously in physical memory, and construct a head-to-tail offset index vector (parent_to_bom_indices) to achieve this. Constant-time memory location and CPU cache prefetching.
[0058] In this example, to completely eliminate cache misses caused by multi-level pointer jumps, the system uses Data-Oriented Design (DOD) to compile the product structure BOM topology into a flat, contiguous SoA (Structure of Arrays) array in physical memory.
[0059] Material codes (e.g., FG_Mate70) are mapped to consecutive uint32_t type numeric IDs in one go during reading via the PartVocab dictionary. The parent-child relationships of the BOM and the material consumption ratio (perqty) are stored in a flat structure array called FlatBomItem.
[0060]
[0061] In physical memory, all FlatBomItems are sorted in ascending order by the value of parent_id, ensuring that all child items of the same parent item are recorded contiguously in memory, providing a perfect hardware-friendly channel for CPU instruction prefetching.
[0062] This step digitizes the text code by establishing a material code mapping and decomposes the tree-shaped BOM into a flat array of contiguous memory. The child component data is arranged centrally according to the parent material identifier, eliminating the cache failure problem caused by random access of multi-level pointers in the traditional tree structure. This significantly improves the CPU cache prefetch hit rate and reduces memory access overhead. The digitized identifier also simplifies the comparison and accumulation calculation of material association data, providing a regular and efficient data storage foundation for the subsequent multi-threaded lock-free parallel demand transmission, and improving the overall processing speed of large-scale BOM data traversal and calculation.
[0063] Step P3: After the topology compilation is completed, the multi-layer parallel computing grid is divided according to the low-level code, and the material supply and demand calculation is completed concurrently in multiple threads within the grid. The MRP parallel consumption scheduling is completed through the lock-free cross-layer demand transmission and hierarchical synchronization mechanism.
[0064] The specific process includes the following steps:
[0065] Step P3-1: Group all materials according to the maximum low-level code corresponding to each material to form a multi-layer parallel computing grid;
[0066] Step P3-2: Set up the outer timing loop, perform calculations layer by layer in ascending order of the lower-level code, distribute multi-core concurrent calculations within a single-layer grid using the OpenMP thread pool, and perform hierarchical barrier synchronization after each layer of calculation is completed before proceeding to the next layer of grid calculation (Barrier Synchronization).
[0067] Step P3-3: Based on the characteristic that materials in the same lower layer code have no mutual dependence, the supply and demand deduction calculation is performed concurrently by multiple threads for all materials in a single layer grid.
[0068] When a gap is found in the material supply and demand calculation, a planned order is generated. The pull amount transmitted to downstream sub-materials is calculated based on the unit consumption coefficient. Taking advantage of the hierarchical isolation characteristic that the lower level code of the sub-material is higher than that of the parent material, a hardware-level atomic instruction (#pragma omp atomic) is used to directly accumulate the pull amount in physical memory. This pull amount is then added to the demand data of the corresponding sub-material, eliminating the overhead of thread mutex locks and ensuring the absolute correctness of the accumulation operation.
[0069] In this example, after the topology compilation is completed, the system divides the entire network of SKUs into a series of gridded parallel netting grid ranks based on the maximum LLC value of the material.
[0070] The computational core contains an outer LLC timing loop, strictly following the ascending order of LLC values (0, 1, 2, ..., LLC_{max}). Since any material is only consumed by its parent material whose lower-level key is less than its own (i.e., structurally higher), therefore:
[0071] Core theorem: All material nodes within the same LLC level are structurally independent and undependent (there is no parent-child consumption relationship or logical transfer between any materials of the same level).
[0072] Based on the above theorem, this invention completely decouples thread locking within each LLC mesh layer. For all materials within the current LLC layer, OpenMP multithreading is directly invoked, distributing the deduction calculations to each CPU core for fully concurrent deduction calculations without data conflicts.
[0073]
[0074] This embodiment uses a simplified material structure of an automotive parts assembly plant as an example to demonstrate the topology compilation, static loop detection, and LLC hierarchical parallel computing full lifecycle of the present invention.
[0075] 1. Supply chain topology initialization
[0076] The system parses the master data and reads the material and BOM dependencies:
[0077] Material nodes:
[0078] PART_A (Complete vehicle, lower layer code to be determined), PART_B (Engine, lower layer code to be determined), PART_C (Piston, lower layer code to be determined), PART_D (Piston ring, lower layer code to be determined), PART_E (Sealing rubber, lower layer code to be determined).
[0079] BOM cascading relationships:
[0080] PART_A Consume 1 PART_B
[0081] PART_B Consumes 4 PART_C
[0082] PART_C Consume 3 PART_D
[0083] PART_D Consume 2 PART_E
[0084] PART_B Consumes 1 PART_E (Note: There are cross-level consumption relationships in the BOM structure)
[0085] 2. DFS topology compilation and static anti-loop verification
[0086] The system starts compiling the three-color DFS topology from the root node PART_A:
[0087] Enter PART_A (gray), traverse its child components, enter PART_B (gray), traverse its child components, enter PART_C (gray), traverse its child components, enter PART_D (gray), traverse its child components, and enter PART_E (gray).
[0088] PART_E has no child components. Backtracking, PART_E is marked as black, resulting in a depth of 4.
[0089] Returning to PART_D, the traversal is complete. Mark PART_D as black, and the depth is 3.
[0090] Returning to PART_C, the traversal is complete. Mark PART_C as black, and the depth is 2.
[0091] Returning to PART_B, continue traversing the other branch PART_E (black), without conflict. After completing the traversal of PART_B, mark PART_B as black; its path to the deepest sub-component is A. B C D E (depth is 4, therefore the maximum depth of PART_B, i.e., LLC code, is 1).
[0092] Returning to PART_A, mark it as black, and set LLC to 0.
[0093] If an abnormal loop occurs (e.g., a child component of PART_E is misconfigured as PART_A): When the DFS traverses to PART_E and attempts to access its child component PART_A, the system detects that the current state of PART_A is "grayed out" and immediately triggers the static loop detection interceptor, reporting the error "BOM Loop Detected: PART_A". PART_B PART_C PART_D PART_E PART_A” and exit ignition to ensure system safety.
[0094] 3. LLC Parallel Mesh Generation and Computation Execution
[0095] In a loop-free state, the system assigns LLC level materials to all materials:
[0096] LLC0:PART_A
[0097] LLC1:PART_B
[0098] LLC2:PART_C
[0099] LLC3:PART_D
[0100] LLC4:PART_E
[0101] The execution details of the parallel grid elimination scheduling are shown in Table 1:
[0102] Table 1 Parallel Grid Hierarchical Consumption Reconciliation Calculation Table
[0103]
[0104] The entire consumption process follows the LLC time-series unidirectional downward rolling pattern. Within each LLC grid layer, since there are no prerequisite dependencies between materials, the calculation threads are fully concurrent, without the need to set any thread locks on inventory or BOM nodes. After rapidly executing barrier synchronization, the process switches to the next layer, completing high-throughput multi-core computation.
[0105] Example 2: A large-scale MRP parallel elimination scheduling system based on material-dependent LLC topology hierarchy, used to implement the above-mentioned large-scale MRP parallel elimination scheduling method, including:
[0106] The low-level code generation module is used to load material and BOM basic data and generate the low-level code corresponding to each material through depth-first topology traversal with loop recognition.
[0107] The topology compilation module is used to compile the product structure BOM into a flat array that is arranged contiguously in physical memory;
[0108] The supply and demand calculation module is used to divide the low-level code into multiple parallel computing grids after the topology is compiled, and to complete the supply and demand deduction in multiple threads within the grid through the OpenMP thread pool. It also realizes lock-free cross-layer demand transmission and hierarchical barrier synchronization based on hardware-level atomic instructions, and completes MRP parallel absorption scheduling.
[0109] The above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of the claims of this patent application.
Claims
1. A large-scale MRP parallel scheduling method based on material-dependent LLC topology hierarchy, characterized in that, Includes the following steps: Step P1: Load the material and BOM basic data, and generate the low-level code corresponding to each material through depth-first topology traversal with loop identification. Step P2: Compile the product structure BOM topology into a flat array arranged contiguously in physical memory; Step P3: After the topology compilation is completed, the multi-layer parallel computing grid is divided according to the low-level code, and the material supply and demand calculation is completed concurrently in multiple threads within the grid. The MRP parallel consumption scheduling is completed through the lock-free cross-layer demand transmission and hierarchical synchronization mechanism.
2. The large-scale MRP parallel scheduling method based on material-dependent LLC topology hierarchy according to claim 1, characterized in that, In step P1, the specific process of depth-first topology traversal with loop identification includes: Step P1-1: Traverse the directed hypergraph of materials starting from the finished product material, and use a three-color state marking method to distinguish the traversal state of each material node. Step P1-2: During the traversal, calculate the maximum depth of all dependent paths for each material and generate the corresponding low-level code; Steps P1-3: If a circular dependency is identified during the traversal, the static loop detection interceptor is triggered.
3. The large-scale MRP parallel scheduling method based on material-dependent LLC topology hierarchy according to claim 2, characterized in that, In step P1-1, the three-color status marking method is as follows: white mark indicates not visited, gray mark indicates visiting, and black mark indicates visiting completed.
4. The large-scale MRP parallel elimination scheduling method based on material-dependent LLC topology hierarchy according to claim 3, characterized in that, In steps P1-2, the bottom layer code of the finished product is 0, while the bottom layer code of the raw material is the largest.
5. A large-scale MRP parallel elimination scheduling method based on material-dependent LLC topology hierarchy according to claim 3, characterized in that, In steps P1-3, the circular dependency is identified as: encountering gray-marked nodes during the search process; Static loop detection interceptor execution: A. Abort topology calculation; B. Extract the material dependency chain in reverse through the recursive call stack, and persistently store the abnormal information containing the chain nodes and paths into the audit database.
6. The large-scale MRP parallel scheduling method based on material-dependent LLC topology hierarchy according to claim 1, characterized in that, The specific process of step P2 includes: Step P2-1: Construct a material code mapping dictionary to uniformly convert string-format material codes into fixed-length numeric identifiers; Step P2-2: Decompose the tree-shaped BOM dependency relationship into standardized flat association records and store them uniformly as an array structure; Steps P2-3: Sort all flat, related records according to their parent material identifiers, ensuring that all child material records corresponding to the same parent material are stored contiguously in physical memory, and construct a head-tail offset index vector to achieve... Constant-time memory location and CPU cache prefetching.
7. A large-scale MRP parallel scheduling method based on material-dependent LLC topology hierarchy according to claim 1, characterized in that, In step P3, the specific process of dividing the material supply and demand calculation into multiple parallel computing grids based on the low-level code and performing the calculation concurrently in multiple threads within the grid includes the following steps: Step P3-1: Group all materials according to the maximum low-level code corresponding to each material to form a multi-layer parallel computing grid; Step P3-2: Set up the outer timing loop, perform calculations layer by layer in ascending order of the lower-level codes, distribute multi-core concurrent calculations within a single-layer grid using the OpenMP thread pool, and perform hierarchical barrier synchronization after each layer of calculation is completed before proceeding to the next layer of grid calculation; Step P3-3: Based on the characteristic that materials in the same lower layer code have no mutual dependence, the supply and demand deduction calculation is performed concurrently by multiple threads for all materials in a single layer grid.
8. A large-scale MRP parallel scheduling method based on material-dependent LLC topology hierarchy according to claim 7, characterized in that, In step P3, when a gap is generated in the material supply and demand calculation, a planned order is generated, and the pull amount transmitted to downstream sub-materials is calculated in combination with the unit consumption coefficient.
9. A large-scale MRP parallel scheduling method based on material-dependent LLC topology hierarchy according to claim 8, characterized in that, In step P3, taking advantage of the hierarchical isolation characteristic that the lower-level code of the sub-material is higher than that of the parent material, hardware-level atomic instructions are used to directly accumulate the data in physical memory, and the pull amount is superimposed on the demand data of the corresponding sub-material, eliminating the overhead of thread mutex locks and ensuring the absolute correctness of the accumulation operation.
10. A large-scale MRP parallel elimination scheduling system based on material-dependent LLC topology hierarchy, used to implement the large-scale MRP parallel elimination scheduling method according to any one of claims 1-9, characterized in that, include: The low-level code generation module is used to load material and BOM basic data and generate the low-level code corresponding to each material through depth-first topology traversal with loop recognition. The topology compilation module is used to compile the product structure BOM into a flat array that is arranged contiguously in physical memory; The supply and demand calculation module is used to divide the low-level code into multiple parallel computing grids after the topology is compiled, and to complete the supply and demand deduction in multiple threads within the grid through the OpenMP thread pool. It also realizes lock-free cross-layer demand transmission and hierarchical barrier synchronization based on hardware-level atomic instructions, and completes MRP parallel absorption scheduling.
