Multi-core parallel particle tracking acceleration method for large-scale ocean drift targets

Abstract

The application belongs to the technical field of offshore target drift tracking, and specifically discloses a multi-core parallel particle tracking acceleration method for large-range marine drifting targets. First, an offshore target drift prediction model is constructed; a forecast area satellite remote sensing image is acquired to obtain an initial distribution position of the offshore target; model calculation area, particle number and prediction time length model parameters are set; model driving fields such as wind field and flow field are acquired; parallel setting is performed on a multi-core processor, including the number of calculation CPUs and the number of physical cores, the number of running threads, task division, one-time thread pool creation and scheduling strategy setting; particle state is initialized, a time advancement and drift numerical simulation process is performed, large-scale drift particles are tracked and calculated in parallel, and finally a particle trajectory set is predicted. The method realizes high-fidelity and fine drift prediction within a minute-level time limit.

Description

Technical Field

[0001] This invention belongs to the field of marine target drift tracking technology, and specifically relates to a multi-core parallel particle tracking acceleration method for large-scale ocean drifting targets, which is suitable for scenarios involving parallel tracking and accelerated computation of large-scale drifting particles. Background Technology

[0002] Both marine oil spills and seaweed blooms are characterized by their wide-ranging origins, highly sensitive drift paths to wind-wave-current coupling, and the irreversible economic and ecological damage they cause once they make landfall. In actual emergency response, it is crucial to promptly ascertain the target's drift location and affected area over the next 3 to 7 days to formulate timely response plans.

[0003] Existing target drift prediction technologies generally employ single-threaded particle tracking models based on the Lagrange framework to predict the trajectory of drifting targets at sea (such as oil spills, floating hazardous materials, and seaweed blooms).

[0004] These models are driven by three-dimensional ocean currents, wind fields, and wave fields, and solve the particle motion equations through numerical integration.

[0005] Because such models need to simulate the trajectories of tens of thousands of particles over a timescale of 168 hours (i.e., 7 days) or even longer, traditional single-threaded computation is extremely time-consuming and cannot meet the timeliness requirements for emergency response to marine green tide disasters, pollution, etc.

[0006] In order to output results within the "golden emergency window", existing target drift prediction systems usually sacrifice spatial resolution or particle size, resulting in large errors in predicting the drift positions of scattered particle points.

[0007] In summary, the current method of using a single thread to predict target drift has the following drawbacks:

[0008] 1. Limited computational efficiency: Existing methods typically use a single thread to serially advance particle trajectories, resulting in insufficient utilization of multi-core CPU resources (typically less than 10%), making it difficult to support large-scale, long-duration simulations and severely limiting scalability.

[0009] 2. Inefficient parallel synchronization: To ensure data consistency, single-threaded architectures often rely on global synchronization mechanisms, which can cause idle waiting during computation and significant delays in particle cross-region migration, thus affecting overall timeliness.

[0010] 3. High memory access overhead: Due to the unreasonable layout of the particle data structure, the access pattern is non-continuous, resulting in a cache miss rate of over 30%, which significantly reduces memory bandwidth utilization efficiency and CPU pipeline performance.

[0011] 4. Insufficient load balancing: Under large-scale complex flow fields, the single-threaded approach cannot dynamically adapt to the uneven particle distribution, and some processing cores remain idle for a long time, resulting in a serious imbalance in the overall computing load. Summary of the Invention

[0012] The purpose of this invention is to propose a multi-core parallel particle tracking acceleration method for large-scale ocean drift targets, which utilizes a multi-core computing platform to perform parallel tracking calculations on a large number of drifting particles, thereby completing drift prediction within a limited time limit.

[0013] To achieve the above objectives, the present invention adopts the following technical solution:

[0014] A multi-kernel parallel particle tracking acceleration method for large-scale ocean drift targets includes the following steps:

[0015] Step 1. Based on the Lagrange particle tracking method, considering the effects of wind, waves and currents on marine targets, construct a marine target drift prediction model;

[0016] Step 2. Acquire satellite remote sensing images of the forecast area and obtain the initial distribution locations of maritime targets based on the satellite remote sensing images;

[0017] Step 3. Set the model parameters, including the model computation region, number of particles, prediction duration, and time step;

[0018] Step 4. Obtain the model driving field, including wind field, flow field, and Stokes drift caused by wave-current interaction;

[0019] Step 5. Configure parallel processing on multi-core processors, including calculating the number of CPUs and their physical cores, setting the number of running threads, task partitioning, creating a thread pool all at once, and setting the scheduling policy;

[0020] Step 6. Initialize the particle state, then execute the time-progression and drift numerical simulation process to perform parallel tracking calculations on large-scale drifting particles, and finally predict the set of all particle trajectories.

[0021] The present invention has the following advantages:

[0022] As described above, this invention discloses a multi-core parallel particle tracking acceleration method for large-scale ocean drift targets. This method can achieve highly scalable and highly parallel particle tracking acceleration on multi-core processors, thereby enabling the completion of calculations within minutes (i.e., less than 10 minutes) and within a high-resolution (kilometer-scale) model space computation range. The 7-day and long-term (within 15 days) drift prediction or source tracing simulation of particles of this magnitude provides real-time decision-making basis for marine oil spill emergency command or marine target salvage and control such as seaweed green tides, directly supporting emergency decision-making for oil spill containment and seaweed interception. Attached Figure Description

[0023] Figure 1 This is a flowchart of the multi-core parallel particle tracking acceleration method for large-scale ocean drifting targets in this invention;

[0024] Figure 2 This is the system framework for the multi-core parallel particle tracking acceleration method for large-scale ocean drifting targets in this invention. Detailed Implementation

[0025] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0026] Currently, methods for predicting target drift using a single thread are available in 10... 5 km² level sea area, 10 6 At the particle scale, it would take 3 to 6 hours to forecast the drift of a maritime target over a period of 72 hours, which is clearly insufficient to meet the timeliness requirements of emergency response.

[0027] To address the aforementioned problems, the method of this invention can implement a highly scalable and highly parallel efficient particle tracking acceleration method on multi-core processors, thereby enabling high-fidelity and refined drift prediction within a minute-level timeframe.

[0028] like Figure 1 As shown, the multi-kernel parallel particle tracking acceleration method for large-scale ocean drift targets includes the following steps:

[0029] Step 1. Based on the Lagrange particle tracking method, consider the effects of wind, waves and currents on marine target patches, and construct a marine target drift (or source tracing) prediction model.

[0030] The movement of a target object in seawater can be viewed as the physical movement of a point mass following ocean currents. Therefore, this invention employs the Lagrange particle tracking method to construct a drift prediction model for targets at sea.

[0031] The particle tracking method uses a random particle movement pattern to simulate particle motion. The displacement variable of each particle is determined by the Lagrange equation. A drift prediction model for marine targets is constructed, as shown in formula (1):

[0032] (1)

[0033] in, To determine the displacement of a single particle, after determining the displacement of a single particle per unit time, the unit displacement is integrated over time to achieve dynamic integration of the particle. It is a surface ocean current; The wind is at an altitude of 10 meters above the sea. The wind effect coefficient; B is the Stokes drift caused by wave-current interaction; B is the diffusion coefficient of a single particle; Z is an independent random number.

[0034] Wind action coefficient This represents the direct drag effect of wind on a single particle, and its value depends on the type of object. For example, the wind effect coefficient for *Ulva prolifera* patches. The wind effect coefficient of oil pollution ranges from 0.011 to 0.013. It is around 0.025.

[0035] Step 2. Acquire satellite remote sensing images of the forecast area and obtain the initial distribution locations of maritime targets based on the satellite remote sensing images.

[0036] After acquiring satellite remote sensing images, we will conduct scattered point extraction of maritime targets.

[0037] First, the satellite remote sensing images are preprocessed with radiometric calibration and atmospheric correction to obtain the radiometric products of the lower atmosphere.

[0038] The preprocessed images are then analyzed using remote sensing methods for marine targets. Common methods for remote sensing of marine targets include the Normalized Difference Vegetation Index (NDVI).

[0039] Then, the calculated image is visually interpreted, and the boundary value of the pixel that can be identified as a marine target is selected. The detection index is used as the threshold to extract the marine target correlation, thereby obtaining the initial distribution position of the marine target.

[0040] It should be noted that the process of obtaining the initial distribution location of maritime targets in this embodiment is relatively conventional and will not be described in detail here. Additionally, the maritime targets in this embodiment include oil spills, floating hazardous materials, or green algae blooms, etc.

[0041] Step 3. Set the model parameters, including the model computation region, number of particles, prediction duration, and time step.

[0042] I. Model Calculation Area: The spatial extent of the model calculation area is greater than the area where the marine target is monitored and the sea area where all marine targets may drift during the drift numerical simulation.

[0043] II. Forecast Duration: The forecast duration can be determined according to the forecast requirements, and is generally greater than 72 hours. In this embodiment, the forecast duration can be set to, for example, 168 hours (7 days), or 15 days.

[0044] III. Number of target particles: Based on the initial distribution of marine targets detected by remote sensing, the number of particle points is counted to obtain the total number of particles. This part is fairly standard and will not be elaborated upon here.

[0045] IV. Time step: The value range is determined based on the model's accuracy requirements; for example, it can be set as follows: .

[0046] Step 4. Obtain the model driving field, including wind field, flow field, and Stokes drift caused by wave-current interaction.

[0047] Acquire hydrometeorological data on surface currents, sea surface winds, and the Stokes drift.

[0048] Surface flow Wind at a height of 10 meters above the sea The data originates from measured data of the sea area where the target is located or from numerical simulation results based on ocean and meteorological models. (Stokes drifting) Calculated using an ocean numerical model.

[0049] The data duration for all model-driven fields is no less than the duration of the drift numerical simulation, i.e., the prediction duration.

[0050] Step 5. Configure parallel processing on multi-core processors, including calculating the number of CPUs and their physical cores, setting the number of running threads, task partitioning, creating a thread pool all at once, and setting the scheduling policy.

[0051] To achieve multi-core parallel particle tracking acceleration for large-scale ocean drift targets, this invention constructs a multi-core parallel particle tracking acceleration system architecture for large-scale ocean drift targets, such as... Figure 2 As shown, the system comprises the following five-layer architecture:

[0052] The data management layer, parallel runtime layer, physical update layer, I / O and checkpoint layer, and configuration and monitoring layer.

[0053] The data management layer is responsible for loading, slicing, and caching model-driven field and grid data, and provides a unified interface for accessing data by time and space.

[0054] The parallel runtime layer is used for thread pools, task queues, work stealers, NUMA-aware memory management, thread affinity, and scheduling strategies.

[0055] The physics update layer is used for particle propulsion (advection + wind and wave effects), boundary and shoreline treatment, cross-subdomain migration, random diffusion term, and anomaly / missing field compensation.

[0056] The I / O and checkpoint layer is used for asynchronous batch writes, double buffering, and fault recovery checkpoints.

[0057] The configuration and monitoring layer is used for parameter configuration, performance counters (throughput, queue length, cache hit rate approximations), logs, and reproducibility tags (random seed, version number).

[0058] Based on the above system architecture, the parallel configuration steps in this embodiment are as follows:

[0059] Step 5.1. First, detect the number of CPUs running on the node and record the number of physical cores of the CPU.

[0060] It should be noted here that if the running node contains more than two CPUs, a NUMA-bound CPU core strategy is adopted to avoid performance loss caused by the migration of threads between different CPU cores.

[0061] Step 5.2. If the user has set the number of threads participating in the calculation, the number of running threads will be set to the number specified by the user; otherwise, the number of running threads will be set to the number of physical cores of the running node.

[0062] Step 5.3. If the user specifies the particle block size, then the user-specified setting will be used; otherwise, the system default even distribution setting will be used.

[0063] Step 5.4. Create a thread pool once for use by computational tasks throughout the entire drift numerical simulation.

[0064] A thread pool is a data structure used to store persistent threads used in parallel simulation example tracing. Here, persistence means that the thread's lifecycle covers the entire program execution cycle; that is, it can be created once and reused multiple times.

[0065] Step 5.5. If a user specifies a scheduling policy, the policy will be determined according to the user's settings; otherwise, the static scheduling policy will be used by default.

[0066] User scheduling strategies include static scheduling strategies and dynamic scheduling strategies. Both have task queues, but the difference is that dynamic scheduling strategies need to retrieve tasks from the task queue multiple times, while static scheduling strategies retrieve them all at once.

[0067] like Figure 2 As shown, thread pool initialization settings are performed at the parallel runtime layer, i.e., setting... (Hardware concurrency, user-specified upper limit), create a fixed-size thread pool.

[0068] CPU affinity and NUMA are performed at the parallel runtime layer, binding threads to physical cores and cross-NUMA clustering; using The strategy initializes the data slices responsible for each thread, reducing remote memory access.

[0069] like Figure 2 As shown, memory and alignment are handled at the data management layer: particle attributes are stored using SoA (Structure of Arrays) and aligned to 32 / 64 bytes for SIMD vectorization; the driving field slices are cached in blocks. Utilizing structured storage (SoA) and NUMA-aware layout significantly improves the locality of memory access, thereby reducing the cache miss rate.

[0070] like Figure 2 As shown, time steps and interpolators are initialized in the configuration and monitoring layer, and grid indexes and interpolation coefficient tables are pre-built, which helps to avoid redundant calculations.

[0071] Randomization and reproducibility are performed at the configuration and monitoring layers. An independent counter-type PRNG (such as Philox / PCG) is assigned to each thread, and child streams are derived using (global seed, thread ID) to ensure that the results of multi-threaded reruns are reproducible.

[0072] By introducing multi-threaded task partitioning and a work-stealing scheduling mechanism, CPU utilization was improved and good scalability was achieved.

[0073] like Figure 2 As shown, hybrid tasks are partitioned at the parallel runtime layer. The partitioning process is as follows:

[0074] I. Task division based on particle chunking:

[0075] When the initial particle distribution is uniform, the particle array is divided into segments of size . The continuous blocks ensure consistent task granularity, which is beneficial for SIMD. In this embodiment, the particle array is a data structure that stores and represents the particles to be tracked in computer representation, representing the entire particle swarm.

[0076] SIMD is an abbreviation for Single Instruction, Multiple Data. It is a parallel computing model in which the same instruction performs the same operation on multiple pieces of data in parallel at the same time.

[0077] II. Divide tasks according to spatial tiling.

[0078] When particle clusters / hot spots appear, the model computation region is divided into M×N tiles based on a grid, and each tile maintains a local particle list. The criteria for determining clusters / hot spots are as follows: First, the number of particles in each tile can be obtained statistically. The criterion for determining clusters / hot spots is that the number of particles in a tile is greater than 4 times the average of all tiles.

[0079] III. Adaptive task partitioning:

[0080] At runtime, based on queue congestion and the variance of execution time per thread, the system automatically switches between task partitioning by particle blocks or by spatial subdomain to ensure load balancing.

[0081] like Figure 2 As shown, in the parallel runtime layer, each thread maintains a local lock-free circular queue, and a lightweight task library is available globally.

[0082] In the parallel runtime layer, when the local queue runs out of work, a thread steals work from the tail of another thread's queue.

[0083] At the parallel runtime layer, contention is reduced through hierarchical scheduling (local priority, accessing the global queue when necessary).

[0084] In the parallel runtime layer, batch submission is performed: statistics and migration are performed after processing K particles or completing a tile, reducing atomic operations and lock overhead.

[0085] This invention achieves load balancing by leveraging the aforementioned hybrid task partitioning and adaptive scheduling (if the user does not specify a scheduling method, the system will decide which scheduling method to use), thereby fundamentally improving the system's stability and computational efficiency.

[0086] Step 6. Initialize the particle state, then execute the time-progression and drift numerical simulation process to perform parallel tracking calculations on large-scale drifting particles, and finally predict the set of all particle trajectories.

[0087] In this embodiment, the particle state is initialized, that is, given a particle... exist The latitude and longitude information at each moment is obtained through remote sensing interpretation. Given the initial time value of the model driving field, where... .

[0088] Specifically, the numerical simulation process of time progression and drift is as follows:

[0089] Initialize particle state ; Obtain the model driving field, including wind field, flow field, and Stokes drift caused by wave-current interaction; Set the time step ;in This indicates the start time of the drift numerical simulation duration, i.e., the prediction duration.

[0090] Step 6.1. Main loop.

[0091] right According to time steps To proceed; This refers to the end time of the drift numerical simulation.

[0092] Step 6.2. Field interpolation to calculate the physical driving field.

[0093] For each particle position Bilinear / trilinear interpolation was used to obtain the flow velocity, wind vector, and wave-current interaction term; where the flow velocity was... The wind vector is Wave-current interaction is .

[0094] If the particle is in a hot spot tile, the most recent interpolation coefficient table is reused to reduce the amount of computation.

[0095] Specifically, based on the particle's position, the driving field data such as wind, flow, and stoke drift from the neighboring grid are interpolated to the particle's location to drive the particle's motion.

[0096] Step 6.3. Perform multi-threaded parallel particle updates.

[0097] I. Advection:

[0098] .

[0099] II. Wind and Wave Drift Additional Items:

[0100] ;

[0101] in, The coefficient of wind effect.

[0102] III. Random diffusion:

[0103] Based on isotropic or anisotropic diffusion tensors, from normal distribution arrays Medium sampling spread noise ;

[0104] .

[0105] IV. Coastline / Boundary:

[0106] like If the target hits the shoreline / boundary, it will execute a stay, absorb, or dissipate strategy.

[0107] Step 6.4. Cross-subdomain migration.

[0108] Particles are added to the corresponding tile's migration buffer vector based on their position. After local aggregation within a thread, they are moved to the target tile's task queue all at once. In this embodiment, each tile maintains a migration buffer vector. Local tracking is performed within the thread to accelerate the process, and then batch aggregation is performed before moving the particles to the target tile's task queue all at once, thereby avoiding fine-grained lock contention.

[0109] Step 6.5. Handling missing measurement fields.

[0110] If there are missing measurements in the model's driving field, perform time extrapolation or neighborhood spatial interpolation on the relevant variables. The relevant variables refer to the Stokes drift caused by wind field, flow field, or wave-current interaction.

[0111] Step 6.6. If the write time point is reached, perform an asynchronous write operation.

[0112] Write the trajectory data of each thread at the current time step to the output file (CSV / NetCDF / binary custom) in batches (or asynchronously). Compress the data in batches by particle block or tile to reduce write jitter.

[0113] This invention employs an asynchronous cross-domain migration and batch submission strategy, which effectively eliminates synchronous idle time and reduces migration latency.

[0114] Step 6.7. If the checkpoint interval is reached, periodically save data such as particle state.

[0115] In this embodiment, the time point is written out as the output interval. The particle trajectory can be output once per hour or once every half hour, which is set by the user according to their own needs. Similarly, the checkpoint is also set by the user.

[0116] Step 6.8. When the drift numerical simulation is completed, i.e. when At that time, the set of particle trajectories and checkpoint data are output.

[0117] The multi-core parallel particle tracking method proposed in this invention effectively solves the problem of tracking large-scale ocean drifting targets within a limited time (less than 10 minutes). The problem of long-term (within 15 days) drift / source tracing calculation of particles of the order of magnitude.

[0118] Of course, the above description is only a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. A multi-kernel parallel particle tracking acceleration method for large-scale ocean drifting targets, characterized in that, Includes the following steps: Step 1. Based on the Lagrange particle tracking method, considering the effects of wind, waves and currents on marine targets, construct a marine target drift prediction model; Step 2. Acquire satellite remote sensing images of the forecast area and obtain the initial distribution locations of maritime targets based on the satellite remote sensing images; Step 3. Set the model parameters, including the model computation region, number of particles, prediction duration, and time step; Step 4. Obtain the model driving field, including wind field, flow field, and Stokes drift caused by wave-current interaction; Step 5. Configure parallel processing on multi-core processors, including calculating the number of CPUs and their physical cores, setting the number of running threads, task partitioning, creating a thread pool all at once, and setting the scheduling policy; Step 6. Initialize the particle state, then execute the time-progression and drift numerical simulation process, perform parallel tracking calculations on large-scale drifting particles, and finally predict the set of all particle trajectories; In step 6, the numerical simulation process of time progression and drift is as follows: Initialize particle state Obtain the model-driven field and set the time step. ; Step 6.

1. Main loop; right According to time steps To proceed; among them This indicates the start time of the drift numerical simulation duration, i.e., the prediction duration; This refers to the end time of the drift numerical simulation. Step 6.

2. Field interpolation to calculate the physical driving field; For each particle position Bilinear / trilinear interpolation was used to obtain the flow velocity, wind vector, and wave-current interaction term; where the flow velocity was... The wind vector is Wave-current interaction is ; Step 6.

3. Perform multi-threaded parallel particle updates; I. Advection: ； II. Wind and Wave Drift Additional Items: ； in, The wind effect coefficient; III. Random diffusion: Based on isotropic or anisotropic diffusion tensors, from normal distribution arrays Medium sampling spread noise ; ； IV. Coastline / Boundary: If particles If the target hits the shoreline / boundary, execute a stay, attract, or absorb strategy. Step 6.

4. Cross-subdomain migration: Particles are added to the migration buffer of the corresponding tile according to their position, and after local aggregation within the thread, they are moved to the task queue of the target tile all at once. Step 6.

5. Handling missing measurement fields; If there are missing measurements in the model's driving field, perform time extrapolation or neighborhood spatial interpolation on the relevant variables; where the relevant variables refer to the Stokes drift caused by wind field, flow field, or wave-current interaction. Step 6.

6. If the write-out time point is reached, write the particle trajectory data predicted by each thread at the current time step to the output file in batches, compressing them by particle block or tile batch to reduce write-out jitter. Step 6.

7. If the checkpoint interval is reached, save the particle state periodically; Step 6.

8. When the drift numerical simulation is completed, i.e. when At that time, the set of particle trajectories and checkpoint data are output.

2. The multi-core parallel particle tracking acceleration method for large-scale ocean drifting targets according to claim 1, characterized in that, In step 1, the particle tracking method uses a random particle movement pattern to simulate particle motion. The displacement variable of each particle is determined by the Lagrange equation, and a drift prediction model for marine targets is constructed, as shown in formula (1): （1） in To determine the displacement of a single particle, after determining the displacement of a single particle per unit time, the unit displacement is integrated over time to achieve dynamic integration of the particle. The wind effect coefficient; It is a surface ocean current; The wind is at an altitude of 10 meters above the sea. Stokes drift caused by wave-current interaction; The diffusion coefficient of a single particle; It is an independent random number.

3. The multi-core parallel particle tracking acceleration method for large-scale ocean drifting targets according to claim 1, characterized in that, In step 2, the process of obtaining the initial distribution location of the maritime target is as follows: First, after acquiring satellite remote sensing images, radiometric calibration and atmospheric correction preprocessing are performed on the satellite remote sensing images to obtain atmospheric low-level emissivity products; then, the marine target remote sensing detection algorithm is used to calculate the preprocessed images. Then, the calculated image is visually interpreted, the boundary value of the pixel that can be identified as a marine target is selected, and the detection index is used as the threshold to extract the marine target correlation and obtain the initial distribution position of the marine target.

4. The multi-core parallel particle tracking acceleration method for large-scale ocean drifting targets according to claim 1, characterized in that, In step 3, the specific settings for the model parameters are as follows: I. Model Calculation Area: The spatial extent of the model calculation area is greater than the area where the marine target is monitored and the sea area where all marine targets may drift during the drift numerical simulation; II. Prediction Duration: Set the prediction duration to be greater than 72 hours; III. Particle Count: Based on the initial distribution of targets at sea, count the number of particle points to obtain the total particle count. ; IV. Time Step: Settings Its value range is set to .

5. The multi-core parallel particle tracking acceleration method for large-scale ocean drifting targets according to claim 1, characterized in that, In step 4, the process of obtaining the model driving field is as follows: Acquire hydrometeorological data on surface currents, sea surface winds, and the Stokes drift; Surface flow Wind at a height of 10 meters above the sea The data originates from measured data of the sea area where the target is located or from numerical simulation results based on ocean and meteorological models. (Stokes drifting) Calculated using an ocean numerical model; The data duration for all model-driven fields is no less than the duration of the drift numerical simulation, i.e., the prediction duration.

6. The multi-core parallel particle tracking acceleration method for large-scale ocean drifting targets according to claim 1, characterized in that, In step 5, the parallel setup steps are as follows: Step 5.

1. First, detect the number of CPUs running on the node and record the number of physical cores of the CPU; Step 5.

2. If the user has set the number of threads participating in the calculation, the number of running threads will be set to the number specified by the user; otherwise, the number of running threads will be set to the number of physical cores of the running node. Step 5.

3. If the user specifies the particle block size, then the user-specified setting will be used; otherwise, the default even distribution setting will be used. Step 5.

4. Create a thread pool all at once for computational tasks to use throughout the entire drift numerical simulation; Step 5.

5. If a user specifies a scheduling policy, the policy will be determined according to the user's settings; otherwise, the static scheduling policy will be used by default.

7. The multi-core parallel particle tracking acceleration method for large-scale ocean drifting targets according to claim 6, characterized in that, In step 5.5, the user scheduling strategy includes a static scheduling strategy and a dynamic scheduling strategy.

8. The multi-core parallel particle tracking acceleration method for large-scale ocean drifting targets according to claim 1, characterized in that, In step 6, the particle state is initialized, i.e., given the particle... exist The latitude and longitude information at the moment is provided, and the model driving field at the initial moment of the drift numerical simulation is also given, including the wind field, flow field and the Stokes drift caused by wave-current interaction; in, , The total number of particles, This indicates the start time of the drift numerical simulation duration.

9. The multi-core parallel particle tracking acceleration method for large-scale ocean drifting targets according to claim 1, characterized in that, The marine targets include oil spills, floating hazardous materials, or green algae blooms.

Images