A tidal computing power scheduling system and method for a heterogeneous computing power resource pool

Abstract

This application relates to the technical field of artificial intelligence computing power, and in particular to a tidal computing power scheduling system and method for heterogeneous computing power resource pools. It includes: a computing power request parsing module, used to receive and parse computing power service requests to determine the computing power requirements corresponding to the tasks; a resource pool management module, used to manage the computing power resource pool; a priority scheduling module, used to obtain request information corresponding to the computing power service requests to determine task priorities, and configure corresponding resource allocation strategies based on task priorities; a node filtering module, used to filter computing nodes in the computing power resource pool according to computing power requirements; and a tidal dynamic scaling scheduling module, used to calculate the tidal period determination results and the load data of the computing power resource pool, and dynamically generate tidal dynamic actions in conjunction with a preset multi-level load threshold pipeline for tidal decision adjustment. This application achieves efficient utilization and intelligent elastic scaling of computing power resources.

Description

Technical Field

[0001] This application relates to the technical field of artificial intelligence computing power, and in particular to a tidal computing power scheduling system and scheduling method for heterogeneous computing power resource pools. Background Technology

[0002] Against the backdrop of the continuous advancement of AI-powered intelligent computing and large-scale modeling to accelerate enterprise digital transformation, computing power has become a core production factor and crucial infrastructure driving the development of the digital economy. Enterprises and institutions are experiencing exponential growth in their demand for computing power during their digitalization and intelligentization processes. To support diverse and complex computing tasks, the current computing power resource pool is evolving towards a highly heterogeneous structure, encompassing various computing architectures such as general-purpose computing (CPU), graphics processing unit (GPU), neural network processing unit (NPU), programmable gate array (FPGA), and application-specific integrated circuit (ASIC). These heterogeneous computing resources are widely deployed in ultra-large-scale cloud data centers, edge computing nodes, and local private deployment environments, forming a cross-regional, cross-platform, and multi-level computing power supply system.

[0003] However, traditional computing power scheduling systems primarily focus on homogeneous resource management, lacking the ability to deeply perceive and efficiently schedule heterogeneous computing power resource pools. They are limited to optimizing scheduling for static or short-cycle tasks and cannot achieve dynamic, elastic supply of computing power across clusters and regions based on load fluctuations. Faced with significant differences in performance characteristics, power consumption, task adaptability, and cost structure among heterogeneous computing resources, existing scheduling mechanisms struggle to dynamically optimize resource allocation based on task characteristics, resulting in low computing power resource utilization, high energy consumption, and complex operation and maintenance. Summary of the Invention

[0004] To achieve efficient utilization and intelligent elastic scaling of computing resources, this application provides a tidal computing power scheduling system and method for heterogeneous computing power resource pools.

[0005] Firstly, this application provides a tidal computing power scheduling system for heterogeneous computing power resource pools, employing the following technical solution:

[0006] A tidal computing power scheduling system for heterogeneous computing power resource pools includes:

[0007] The computing power request parsing module is used to receive computing power service requests and parse them to determine the computing power requirements corresponding to the task;

[0008] The resource pool management module is used to manage a predefined computing resource pool, wherein the computing resource pool includes several types of computing nodes, and each computing node is configured with a corresponding tidal attribute label, current load information, hardware resource rules and tidal time period configuration;

[0009] The priority scheduling module obtains the request information corresponding to the computing power service request to determine the task priority, and configures the corresponding resource allocation strategy for the task based on the task priority. The task priority includes high priority and low priority.

[0010] The node filtering module is used to filter computing nodes that meet the conditions in the computing resource pool according to the computing power requirements. The conditions to meet the requirements include at least meeting the resource requirements, meeting the resource allocation strategy requirements, and meeting the tidal time period requirements.

[0011] The tidal dynamic scaling scheduling module is used to calculate the tidal period determination result and the load data of the computing power resource pool, and dynamically generate tidal dynamic actions in combination with the preset multi-level load threshold pipeline to make tidal decision adjustments to the corresponding computing nodes in the computing power resource pool. The tidal dynamic actions include expansion, reduction, pause, and full-speed operation.

[0012] In some embodiments, a computing power integration module is also included, specifically for:

[0013] Determine whether there are two or more fragmented computing powers in each of the aforementioned computing nodes;

[0014] If it exists, then determine whether integrating the fragmented computing power will release the complete computing power of at least one of the computing nodes;

[0015] If possible, all tasks in the computing nodes of the same type will be migrated to any one of the computing nodes.

[0016] In some embodiments, the tidal attribute label corresponding to the computing node includes tidal nodes and non-tidal nodes, and the priority scheduling module is specifically used for:

[0017] If the task priority is high, it is assigned to the non-tidal node to obtain computing power;

[0018] If the task priority is low, then determine whether the current load status of the computing power resource library is high.

[0019] If yes, the low-priority task is suspended; otherwise, the task is assigned to the tidal node in the corresponding tidal period to obtain computing power.

[0020] In some embodiments, the tidal dynamic scaling scheduling module calculates the tidal period determination result, specifically including:

[0021] Obtain the start and end times of the tidal task to generate tidal periods;

[0022] Determine whether the current time is within the tidal period;

[0023] If so, the result of the tidal period determination is 1;

[0024] If not, the result of the tidal period determination is 0.

[0025] In some embodiments, the tidal dynamic scaling scheduling module calculates the load data of the computing resource pool, specifically including:

[0026] The load data is calculated based on the following formula:

[0027] ;

[0028] in, The average load of the computing resource pool at time t is represented by the average load of the computing resource pool at time t. The resource utilization rate of the i-th node at time t is represented by N, and the total number of computing nodes is represented by N.

[0029] In some embodiments, the tidal dynamic scaling scheduling module is specifically used for:

[0030] Obtain several threshold nodes of the multi-level load threshold pipeline, wherein the threshold nodes include protection thresholds. Stretch trigger threshold Recovery threshold ,in, < < ;

[0031] Determine the numerical relationship between the load data and each of the threshold nodes;

[0032] like < < Then the tidal dynamic action is called shrinkage, which is characterized by reducing the computing power occupancy ratio of tidal nodes and reducing concurrency to shrink the computing power resources of tidal tasks.

[0033] like < < Then the tidal dynamic action is expansion, which is characterized by increasing the computing power occupancy ratio of tidal nodes and increasing concurrency to increase the computing power resources of tidal tasks.

[0034] like ≥ If the tidal dynamic action is paused, the pause is characterized by stopping the tasks in the tidal node or migrating the tasks in several of the tidal nodes to other low-load non-tidal nodes.

[0035] like ≤ Then the tidal dynamic action is to run at full speed, which means that the computing power occupancy ratio and concurrency of all the tidal nodes are adjusted to the highest.

[0036] In some embodiments, the tidal dynamic scaling scheduling module is further configured to:

[0037] When the tidal period determination result is 1 and ≤ When the tidal task is activated, computing power can be scheduled to the tidal node;

[0038] When the tidal period determination result is 0, the tidal task cannot be activated by default.

[0039] In some of these embodiments, when the task is paused or migrated, it is determined whether the task belongs to a task type that supports resuming from breakpoint.

[0040] If so, the current status data corresponding to the task is obtained and mounted in the distributed storage. When the task is restored or migrated, the current status data is extracted from the stored distributed storage and loaded so that the task can continue to execute from the interruption point.

[0041] In some embodiments, a peak / slow peak adjustment module is also included, specifically for:

[0042] Based on the current load data of the computing power resource pool, determine whether it is in a peak or off-peak period of computing power usage;

[0043] If the computing power usage is at its peak, then the tasks in the tidal node will be restricted or suspended.

[0044] If the computing power usage is at a low point, then the tasks in the tidal node are resumed.

[0045] Secondly, this application provides a tidal computing power scheduling method for heterogeneous computing power resource pools, adopting the following technical solution:

[0046] A tidal computing power scheduling method for heterogeneous computing power resource pools is proposed, based on the above-mentioned system implementation.

[0047] The technical solutions provided by the embodiments of this application have the following technical effects:

[0048] This system enables unified perception, intelligent scheduling, and dynamic orchestration of various types of computing resources, as well as tidal resource reuse and intelligent task priority scheduling, achieving efficient utilization and intelligent elastic scaling of computing resources. It not only supports a variety of heterogeneous computing resources (including CPUs, GPUs, NPUs, FPGAs, etc.), but also dynamically senses load changes, balancing peak-hour real-time tasks with off-peak-hour tidal task demands, reducing energy consumption and improving computing power utilization.

[0049] By combining the load function L(t) and the tidal time window function C(t) for judgment, and combining the tidal scheduling decision function S(t) defined by the load water level model (including the recovery threshold Tl, the scaling trigger threshold Tm, and the protection threshold Tu), this mechanism can perform intelligent, flexible, and cost-optimized dynamic scheduling management of tidal computing power scheduling tasks in real time and efficiently under the environment of fluctuating computing power demand and complex heterogeneous resource pools. This significantly improves the utilization rate of computing power resources, while ensuring the performance requirements of high-priority real-time tasks and the continuous and stable operation of the system. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the module connection of a tidal computing power scheduling system for heterogeneous computing power resource pools provided in this embodiment.

[0051] Figure 2 This is a dynamic tidal scheduling framework diagram provided in the embodiments of this application.

[0052] Figure 3 This is a schematic diagram of the computing power scheduling logic provided in the embodiments of this application. Detailed Implementation

[0053] To better understand the purpose, technical solutions, and advantages of this application, it has been described and illustrated below with reference to the accompanying drawings and embodiments. However, those skilled in the art should understand that this application can be implemented without these details. In some cases, to avoid obscuring various aspects of this application due to unnecessary description, well-known methods, processes, systems, components, and / or circuits already described at a higher level will not be elaborated upon. It will be apparent to those skilled in the art that various modifications can be made to the embodiments disclosed in this application, and the general principles defined in this application can be applied to other embodiments and application scenarios without departing from the principles and scope of this application. Therefore, this application is not limited to the illustrated embodiments, but conforms to the broadest scope consistent with the scope of protection claimed in this application.

[0054] It should be noted that the descriptions of these embodiments are for the purpose of aiding understanding the present invention, but do not constitute a limitation thereof. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0055] In the description of this application, "several" means one or more, "more than" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0056] In the description of this application, the terms "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples.

[0057] like Figure 1 and Figure 2 As shown in the figure, this application discloses a tidal computing power scheduling system for heterogeneous computing power resource pools, including:

[0058] The computing power request parsing module is used to receive computing power service requests and parse them to determine the computing power requirements corresponding to the task.

[0059] First, the scheduling platform responds to the received computing power service requests, which come from the user or the upper-layer service platform.

[0060] When a computing power service request is received and the system has corresponding idle computing power, a computing power task will be generated. The system will parse the computing power service request to extract the basic parameters of the task to determine the computing power requirement.

[0061] The computing power requirement includes at least the required computing power value (number of CPU cores, GPU model and memory capacity, memory, storage I / O, etc.), resource specifications, task priority, task type (real-time inference, online service, offline training, batch processing, etc.), and estimated duration (minutes, hours, days). After parsing, the task information corresponding to the computing power request enters the scheduling queue, triggering the subsequent resource selection process.

[0062] The resource pool management module is used to manage predefined computing resource pools. The computing resource pool includes several types of computing nodes, and each computing node is configured with corresponding tidal attribute tags, current load information, hardware resource rules, and tidal time period configurations.

[0063] The computing resource pool is a pre-set database on the platform, which stores detailed characteristic information of all computing nodes, including but not limited to: unique node identifier, physical location, GPU type, computing node type (GPU, CPU, NPU, FPGA, etc.), number of GPU cores, memory capacity, GPU video memory capacity, storage capacity and type, amount of idle computing resources, node load balancing status, task queue length and number of tasks, bandwidth, network latency, etc.

[0064] It also includes real-time load resource utilization, node health status (fault, maintenance, normal), tidal attribute tags corresponding to nodes (tidal computing power nodes or non-tidal computing power nodes), tidal time period configuration (22:00 at night to 6:00 the next day), tidal task priority, historical utilization rate of tidal time periods, utilization strategy of tidal nodes, historical fluctuation data, etc.

[0065] Tidal computing nodes are characterized by allocating computing power for computation during specified tidal periods, while not allocating computing power during non-tidal periods. They are suitable for continuous tasks with low time requirements, such as offline tasks and AI training.

[0066] Non-tidal nodes can perform computing tasks at any time of day without any downtime. As soon as a computing task is assigned, computing power can be allocated immediately, making them suitable for real-time tasks with high timeliness requirements.

[0067] It can also include node energy consumption information (real-time power, energy efficiency ratio), operating cost parameters, as well as security level and access control status.

[0068] In this embodiment, the key data for subsequent node selection and computing resource scheduling are tidal attribute tags, current load information, and tidal time period configuration. Subsequent appropriate node selection requires a comprehensive evaluation based on the resource occupancy, available computing power, and tidal conditions of each computing node in the computing resource library.

[0069] The priority scheduling module obtains the request information corresponding to the computing power service request to determine the task priority, and configures the corresponding resource allocation strategy for the task based on the task priority.

[0070] The priority of a task is dynamically set based on its urgency and importance. Task priorities include high priority and low priority. Different priority tasks correspond to different types of computing nodes and computing resource allocation strategies to avoid affecting the normal operation of important and urgent tasks.

[0071] In the embodiments of this application, real-time tasks are assigned to high priority. These tasks have a short timeframe and need to be completed in real time. Offline training, big data analysis, and other tasks have a longer duration and do not have an immediate timeframe requirement, so they are assigned to low priority.

[0072] The tasks implemented include real-time inference and online services.

[0073] Different resource allocation strategies are dynamically configured based on the priority information of the task type corresponding to a computing power request.

[0074] The node filtering module is used to filter computing nodes that meet certain conditions from the computing resource pool based on computing power requirements. The conditions for meeting these conditions include at least meeting resource requirements, meeting resource allocation strategy requirements, and meeting tidal time period requirements.

[0075] Node filtering is used to select computing nodes to match various computing power request tasks. After filtering out computing nodes that meet the conditions, the computing power tasks are assigned to the corresponding computing nodes for computing power allocation.

[0076] Specifically, in the embodiments of this application, the requirements include three main aspects: first, the computing power resources of the computing node need to meet the resource requirements of the task; second, the priority strategy requirements corresponding to tasks with different priorities need to be met; and third, for tasks that need to be computed on the tidal node, the working time requirements of the tidal node need to be matched.

[0077] Specifically, based on the above main requirements, in this embodiment of the application, when filtering nodes, nodes that meet at least the following conditions are prioritized for filtering:

[0078] (1) The real-time resource utilization rate of the computing node is lower than the first threshold;

[0079] (2) The available computing power of the computing node to be selected is greater than or equal to the computing power requirement of the task;

[0080] (3) If the task is a tidal task, prioritize matching tidal nodes that are in the tidal priority period;

[0081] (4) If the task is a high-priority real-time task, non-tidal nodes should be selected first to ensure task performance and stability.

[0082] The tidal dynamic scaling scheduling module is used to calculate the tidal period determination results and the load data of the computing resource pool, and dynamically generate tidal dynamic actions in combination with the preset multi-level load threshold pipeline to make tidal decision adjustments to the corresponding computing nodes in the computing resource pool. The tidal dynamic actions include expansion, reduction, pause, and full-speed operation.

[0083] The scheduling platform dynamically manages the start, pause, resume, expansion, and contraction of tidal tasks based on the tidal time period settings and real-time load changes, combined with the threshold management method of the load pipeline, to achieve smooth load adjustment and maximize resource utilization.

[0084] The load pipeline has three water level nodes. Based on the real-time load of each computing node in the current computing resource pool and the numerical management between each water level node, different tidal actions are dynamically selected to adjust the computing power allocation in the computing nodes. The purpose of the adjustment is to adjust the computing power allocation ratio between tidal nodes and non-tidal nodes, so as to dynamically optimize the computing power allocation strategy based on the number of computing power tasks of different importance and the load during peak task periods.

[0085] like Figure 3 As shown, the above scheme enables unified perception, intelligent scheduling, and dynamic orchestration of various types of computing resources, as well as tidal resource reuse and intelligent task priority scheduling, achieving efficient utilization and intelligent elastic scaling of computing resources. This system not only supports various heterogeneous computing resources (including CPUs, GPUs, NPUs, FPGAs, etc.), but also dynamically senses load changes, balancing peak-hour real-time tasks with off-peak tidal task demands, reducing energy consumption, and improving computing power utilization.

[0086] like Figure 3 As shown, in some other embodiments, a computing power integration module is also included, specifically for:

[0087] Determine whether there are two or more fragmented computing powers in each computing node.

[0088] If it exists, then determine whether integrating several fragmented computing powers will release the complete computing power of at least one computing node.

[0089] If possible, all tasks on the same type of compute nodes will be migrated to any compute node.

[0090] When the idle resources of a single computing node are insufficient to match the task requirements, the system initiates fragmented computing power integration. It determines whether there is fragmented computing power in the target node that meets the tidal type, and at the same time determines whether there are two or more computing nodes with fragmented computing power. If both exist, the fragmented computing power is broken down into smaller units through task migration, and the computing power fragments of several computing nodes are migrated to the same computing node, releasing the usable complete computing power unit.

[0091] Specifically, the system calculates the consolidation space of multiple low-utilization nodes based on the current load of each node. Using task migration actions, low-priority tasks that support interrupt recovery or state persistence are migrated to the same node. The criterion is that after the consolidation operation is completed, at least one complete GPU, NPU, or CPU computing unit is released, and this released computing power unit is used for the current computing power request.

[0092] When migrating tasks during fragment integration, the principle of least impact is followed, and the lowest priority tasks among several low-priority tasks are selected for migration to avoid affecting the normal operation of other tasks.

[0093] If the current load utilization of all four tidal nodes is 25%, then by migrating the fragmented computing power of any three tidal nodes to one tidal node, the current computing power of that tidal node will become 100%, while the current load utilization of the other three tidal nodes will change to 0%.

[0094] In other embodiments, the tidal attribute label corresponding to the computing node includes tidal nodes and non-tidal nodes, and the priority scheduling module is specifically used for:

[0095] If the task has a high priority, it will be assigned to a non-tidal node to obtain computing power.

[0096] If the task priority is low, then determine whether the load status in the current computing resource pool is high.

[0097] If yes, the low-priority task will be suspended; otherwise, the task will be assigned to a tidal node in the corresponding tidal period to obtain computing power.

[0098] High-priority tasks are given priority in obtaining computing power when resources are scarce. They need to be guaranteed to obtain the corresponding computing power immediately. Therefore, high-priority tasks need to be matched to the best-performing non-tidal nodes.

[0099] Low-priority tasks will be dynamically allocated to tidal nodes based on the current load of the system's computing resources pool.

[0100] If the current computing power resource pool is under high load, it means that the tidal nodes and / or non-tidal nodes are running at high speed and there are not many computing power resources left. At this time, low-priority tasks can be temporarily suspended and allocated to tidal nodes when the computing power resource pool is no longer under high load.

[0101] At the same time, if the tidal node that can currently allocate computing power is in the tidal period, it will be allocated to the tidal node first.

[0102] It is important to note that the system does not restrict low-priority tasks to be assigned to tidal nodes and high-priority tasks to be assigned to non-tidal nodes. The priority allocation decisions mentioned above all correspond to priority allocation. If the system is currently under very low load, low-priority tasks can also be assigned to non-tidal nodes to quickly allocate computing power for execution. If the current tidal nodes are all saturated but new high-priority computing tasks appear, in order to ensure a fast computing power response for high-priority tasks, if there are available tidal nodes during the tidal period, high-priority tasks can be assigned to tidal nodes first to allocate computing power for computation as quickly as possible.

[0103] In other embodiments, when high-priority tasks suddenly increase, the system can perform actions such as pausing, scaling down, or migrating low-priority tasks to release necessary computing resources.

[0104] In other embodiments, the tidal dynamic scaling scheduling module calculates the tidal period determination result, specifically including:

[0105] Obtain the start and end times of the tidal task to generate tidal periods.

[0106] Users or system administrators can configure and pre-set the tidal periods for each computing node corresponding to each object (GPU, CPU, etc.) in the scheduling platform. The tidal periods can be set daily, weekly, or dynamically adjusted according to the business adaptive strategy.

[0107] The configuration of tidal periods includes the start time, end time, and corresponding resource usage eligibility rules.

[0108] Determine whether the current time is within the tidal period. If yes, the tidal period determination result is 1; otherwise, the tidal period determination result is 0.

[0109] Specifically, the determination of tidal periods is based on the following formula:

[0110] .

[0111] in, Characterized by the start time of the tides, The end time of the tide. It is characterized by tidal periods.

[0112] The tidal period assessment result generates a unique result based on whether the current time belongs to the tidal period. The tidal period assessment result is used to combine the load data with the threshold comparison results in the multi-level load pipeline to assess whether the tidal task can be started.

[0113] In other embodiments, the tidal dynamic scaling scheduling module calculates the load data of the computing resource pool, specifically including:

[0114] The load data is calculated based on the following formula:

[0115] ;

[0116] in, Characterized by the average load of the computing resource pool at time t. The resource utilization rate of the i-th node at time t is represented by N, and the total number of computing nodes is represented by N.

[0117] The calculated load data is used to characterize the overall load situation in the current resource pool, and it can reflect the remaining available computing power resources in the current resource pool.

[0118] like Figure 3 As shown, in some other embodiments, the tidal dynamic scaling scheduling module is specifically used for:

[0119] Obtain several threshold nodes for a multi-level load threshold pipeline, including protection thresholds. Stretch trigger threshold Recovery threshold ,in, < < .

[0120] Different thresholds represent reference values ​​for subsequent dynamic adjustments, and threshold segments formed by different thresholds represent different levels of computing power load utilization.

[0121] Determine the numerical relationship between the load data and each threshold node.

[0122] Specifically, the scheduling decisions for tidal tasks are generated based on the following function formula:

[0123] .

[0124] in, To pause To reduce volume, To expand capacity, To run at full speed.

[0125] like < < The tidal dynamic action is shrinkage, which is characterized by reducing the computing power occupancy ratio of tidal nodes and reducing concurrency to shrink the computing power resources of tidal tasks.

[0126] like < < The dynamic action of tidal expansion is to increase the computing power occupancy ratio of tidal nodes and increase the concurrency to increase the computing power resources of tidal tasks.

[0127] like ≥ If the tidal dynamic action is paused, the pause is characterized by stopping the tasks in the tidal node or migrating the tasks in several tidal nodes to other low-load non-tidal nodes.

[0128] like ≤ If the tidal dynamic action is to run at full speed, then the full speed operation is characterized by adjusting the computing power occupancy ratio and concurrency of all tidal nodes to the highest level.

[0129] During the tidal task operation, the system continuously monitors the overall load changes in real time. When the load is detected to further decrease to near the recovery threshold, the system will take action. The system will then automatically expand the capacity of tidal tasks based on the available remaining computing resources, dynamically increasing the concurrency of tidal tasks to improve task efficiency and fully utilize the idle computing power in the resource pool during off-peak periods.

[0130] When the load rises to the scaling trigger threshold and protection threshold During this period, the system enters a resource contraction phase, gradually reducing the computing power occupancy ratio of tidal tasks and decreasing concurrency to prioritize the computing power needs of real-time high-priority tasks.

[0131] If the load continues to rise and exceeds the protection threshold The system immediately suspends tidal tasks or migrates them to other low-load nodes according to the load distribution to free up critical computing resources and ensure the performance of real-time tasks and the overall stability of the system.

[0132] By combining the load function L(t) and the tidal time window function C(t) for judgment, and combining the tidal scheduling decision function S(t) defined by the load water level model (including the recovery threshold Tl, the scaling trigger threshold Tm, and the protection threshold Tu), this mechanism can perform intelligent, flexible, and cost-optimized dynamic scheduling management of tidal computing power scheduling tasks in real time and efficiently under the environment of fluctuating computing power demand and complex heterogeneous resource pools. This significantly improves the utilization rate of computing power resources, while ensuring the performance requirements of high-priority real-time tasks and the continuous and stable operation of the system.

[0133] In other embodiments, the system also introduces a load trend prediction mechanism, enabling the scheduling of tidal tasks to not only be based on a joint judgment of the time window function and the current real-time load, but also to support comprehensive decision-making by combining multi-dimensional intelligent scheduling decisions, and to predict load change trends in advance for resource scaling up and down, specifically including:

[0134] Based on cost optimization strategies:

[0135] Under the peak-valley electricity pricing model, tidal tasks are prioritized to run during periods of low electricity price, while the execution of unnecessary tidal tasks is reduced during periods of high electricity price, thereby reducing the overall computing power cost.

[0136] Based on an energy-first strategy:

[0137] Dynamically prioritize nodes with higher energy efficiency and lower PUE to allocate tidal tasks, thereby improving resource utilization efficiency and reducing energy consumption;

[0138] Task type-aware strategy:

[0139] For task types that support breakpoint resuming, when a tidal task is paused due to load changes or needs to be migrated, the system automatically executes the interruption suspension mechanism. After the task is resumed or migrated, it can quickly and seamlessly resume from the interruption point, ensuring the integrity and continuity of the task.

[0140] Furthermore, the scheduling platform continuously monitors the energy consumption, load, and operating costs of nodes. For nodes that are idle for extended periods or under low load, the platform automatically enters hibernation, offline, or shutdown mode to reduce energy consumption. Based on changes in task load, when an upward trend in load occurs, the system supports quickly waking up low-energy nodes and adding them to the resource pool.

[0141] In addition to the cost-first and energy-first modes mentioned above, scheduling strategies can also include performance-first and adaptive balancing modes, which can be dynamically switched according to different business scenarios.

[0142] In other embodiments, the tidal dynamic scaling scheduling module is further used for:

[0143] When the tidal period determination result is 1 and ≤ When the tidal task is activated, computing power can be scheduled to the tidal node.

[0144] When the tidal period determination result is 0, the tidal task cannot be activated by default.

[0145] When the system detects that the current time has entered the preset tidal time window (i.e., the tidal period determination function C(t)=1), and the overall load is below the scaling trigger threshold... When the system automatically initiates tidal task scheduling, it prioritizes the use of idle computing resources in tidal nodes to execute low-priority tidal tasks, such as AI training, offline inference, batch processing, and big data analysis.

[0146] It should also be noted that if the available resources of tidal nodes are insufficient, the system can use scheduling strategies to supplement the low-priority available computing resources of some non-tidal nodes to carry out tidal tasks.

[0147] If the current time is not within the tidal time window (i.e., the tidal time period determination function C(t)=0), the tidal task is not schedulable by default. It can only actively participate in scheduling or be manually allocated by the administrator when there are additional idle resources.

[0148] In other embodiments, when a task is paused or migrated, it is determined whether the task belongs to a task type that supports resuming from breakpoint.

[0149] If so, the current state data corresponding to the task is obtained and mounted in the distributed storage. When the task is restored or migrated, the current state data is extracted from the saved distributed storage and loaded so that the task can continue to execute from the interruption point.

[0150] When task interruptions occur due to computing power consolidation, task migration, or changes in tidal strategies, the system supports task state persistence.

[0151] For task types that support resumeable computation (such as image analysis, video stream processing, batch data computation, etc.), the system saves the current computation state, data cache, and progress before the task is paused. This state information is persisted to distributed storage. When the task resumes, the scheduling platform extracts the necessary information from the saved state and continues execution directly from the point of interruption, ensuring the continuity and correctness of task computation.

[0152] In other embodiments, a peak / trough adjustment module is also included, specifically for:

[0153] Determine whether the computing power resource pool is in a peak or off-peak period based on its current load data.

[0154] If it is during peak computing power usage, then the tasks in the tidal node will be restricted or suspended.

[0155] If it is a period of low computing power usage, then resume the tasks in the tidal node.

[0156] The scheduling platform prioritizes the resource needs of high-priority real-time tasks during peak computing power usage periods, and makes full use of idle computing power to schedule tidal tasks during off-peak periods, thereby balancing the load and improving overall resource utilization.

[0157] Based on real-time load monitoring and prediction, the system automatically limits or suspends the operation of low-priority tidal tasks when it detects a peak period, freeing up computing resources to ensure the operation of real-time tasks. When entering an off-peak period, if L(t) is detected to be lower than Tm, the system automatically resumes the operation of tidal tasks and automatically expands their parallelism based on idle resources to improve the processing efficiency of tidal tasks.

[0158] like Figure 3 As shown, this application also discloses a tidal computing power scheduling method for heterogeneous computing power resource pools, based on the system implementation described above.

[0159] The implementation principle is as follows:

[0160] This system enables unified perception, intelligent scheduling, and dynamic orchestration of various types of computing resources, as well as tidal resource reuse and intelligent task priority scheduling, achieving efficient utilization and intelligent elastic scaling of computing resources. It not only supports a variety of heterogeneous computing resources (including CPUs, GPUs, NPUs, FPGAs, etc.), but also dynamically senses load changes, balancing peak-hour real-time tasks with off-peak-hour tidal task demands, reducing energy consumption and improving computing power utilization.

[0161] By combining the load function L(t) and the tidal time window function C(t) for judgment, and combining the tidal scheduling decision function S(t) defined by the load water level model (including the recovery threshold Tl, the scaling trigger threshold Tm, and the protection threshold Tu), this mechanism can perform intelligent, flexible, and cost-optimized dynamic scheduling management of tidal computing power scheduling tasks in real time and efficiently under the environment of fluctuating computing power demand and complex heterogeneous resource pools. This significantly improves the utilization rate of computing power resources, while ensuring the performance requirements of high-priority real-time tasks and the continuous and stable operation of the system.

[0162] It should be understood that although the steps in the flowcharts in the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily performed in the order indicated by the arrows. Unless otherwise expressly stated herein, there is no strict order in which these steps are performed, and they may be performed in other orders.

[0163] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A tidal computing power scheduling system for heterogeneous computing power resource pools, characterized in that, include: The computing power request parsing module is used to receive computing power service requests and parse them to determine the computing power requirements corresponding to the task; The resource pool management module is used to manage a predefined computing resource pool, wherein the computing resource pool includes several types of computing nodes, and each computing node is configured with a corresponding tidal attribute label, current load information, hardware resource rules and tidal time period configuration; The priority scheduling module obtains the request information corresponding to the computing power service request to determine the task priority, and configures the corresponding resource allocation strategy for the task based on the task priority. The task priority includes high priority and low priority. The node filtering module is used to filter computing nodes that meet the conditions in the computing resource pool according to the computing power requirements. The conditions to meet the requirements include at least meeting the resource requirements, meeting the resource allocation strategy requirements, and meeting the tidal time period requirements. The tidal dynamic scaling scheduling module is used to calculate the tidal period determination result and the load data of the computing power resource pool, and dynamically generate tidal dynamic actions in combination with the preset multi-level load threshold pipeline to make tidal decision adjustments to the corresponding computing nodes in the computing power resource pool. The tidal dynamic actions include expansion, reduction, pause, and full-speed operation. The computing power integration module determines whether there are two or more fragmented computing powers in each computing node. If it exists, then determine whether integrating the fragmented computing power will release the complete computing power of at least one of the computing nodes; If possible, all tasks in the computing nodes of the same type will be migrated to any one of the computing nodes. The tidal dynamic scaling scheduling module calculates the tidal period determination result, specifically including: Obtain the start and end times of the tidal task to generate tidal periods; Determine whether the current time is within the tidal period; If so, the result of the tidal period determination is 1; If not, the result of the tidal period determination is 0; The tidal dynamic scaling scheduling module calculates the load data of the computing resource pool, specifically including: The load data is calculated based on the following formula: ； in, The average load of the computing resource pool at time t is represented by the average load of the computing resource pool at time t. The resource utilization rate of the i-th node at time t is represented by N, and the total number of computing nodes is represented by N. The tidal dynamic scaling scheduling module is specifically used for: Obtain several threshold nodes of the multi-level load threshold pipeline, wherein the threshold nodes include protection thresholds. Stretch trigger threshold Recovery threshold ,in, < < ; Determine the numerical relationship between the load data and each of the threshold nodes; like < < Then the tidal dynamic action is called shrinkage, which is characterized by reducing the computing power occupancy ratio of tidal nodes and reducing concurrency to shrink the computing power resources of tidal tasks. like < < Then the tidal dynamic action is expansion, which is characterized by increasing the computing power occupancy ratio of tidal nodes and increasing concurrency to increase the computing power resources of tidal tasks. like If ≥, then the tidal dynamic action is paused, and pause is characterized by stopping the tasks in the tidal node or migrating the tasks in several of the tidal nodes to other low-load non-tidal nodes. If ≤ Then the tidal dynamic action is to run at full speed, which means that the computing power occupancy ratio and concurrency of all the tidal nodes are adjusted to the highest.

2. The tidal computing power scheduling system for heterogeneous computing power resource pools according to claim 1, characterized in that, The tidal attribute labels corresponding to the computing nodes include tidal nodes and non-tidal nodes, and the priority scheduling module is specifically used for: If the task priority is high, it is assigned to the non-tidal node to obtain computing power; If the task priority is low, then determine whether the current load status of the computing power resource library is high. If yes, the low-priority task is suspended; otherwise, the task is assigned to the tidal node in the corresponding tidal period to obtain computing power.

3. The tidal computing power scheduling system for heterogeneous computing power resource pools according to claim 1, characterized in that, The tidal dynamic scaling scheduling module is further used for: When the tidal period determination result is 1 and ≤ When the tidal task is activated, computing power can be scheduled to the tidal node; When the tidal period determination result is 0, the tidal task cannot be activated by default.

4. The tidal computing power scheduling system for heterogeneous computing power resource pools according to claim 3, characterized in that, When the task is paused or migrated, determine whether the task belongs to the type of task that supports resuming from breakpoint. If so, the current status data corresponding to the task is obtained and mounted in the distributed storage. When the task is restored or migrated, the current status data is extracted from the stored distributed storage and loaded so that the task can continue to execute from the interruption point.

5. The tidal computing power scheduling system for heterogeneous computing power resource pools according to claim 1, characterized in that, It also includes a peak / off-peak adjustment module, specifically used for: Based on the current load data of the computing power resource pool, determine whether it is in a peak or off-peak period of computing power usage; If the computing power usage is at its peak, then the tasks in the tidal node will be restricted or suspended. If the computing power usage is at a low point, then the tasks in the tidal node are resumed.

6. A tidal computing power scheduling method for heterogeneous computing power resource pools, characterized in that, Based on the system implementation as described in any one of claims 1-5.

Images