Computing power-power cross-time and space collaborative scheduling method, system and storage medium

Abstract

The application discloses a kind of computing power-power cross-time and space collaborative scheduling method, system and storage medium, wherein, based on the computing power-power cross-time and space collaborative scheduling method of the embodiment of the application, by constructing computing power-power heterogeneous mapping model, to solve the semantic gap problem between logical computing power resources and physical power load, realize the digitization of adjustment capacity, and by introducing user intent perception and double-level buffer queue mechanism, to solve the multi-objective game problem between computing power cost and service quality under the premise of guaranteeing service level agreement (SLA), and by cross-domain mirror preheating and cost perception routing, to solve the delay bottleneck of computing power task in wide-area space "green by green", realize the space-time dynamic matching of computing power flow and power flow. Thus, a plurality of heterogeneous computing power clusters are uniformly managed downward, and power trading and dispatching systems are connected upward, to realize active and fine adjustment of computing power load.

Description

Technical Field

[0001] This invention relates to the field of power system operation and control technology, and in particular to a computing power-power cross-temporal and spatial collaborative scheduling method, a computer-readable storage medium, and a computing power-power cross-temporal and spatial collaborative scheduling system. Background Technology

[0002] With the in-depth implementation of the national "East Data West Computing" project and the explosive growth of AI big data models, computing centers have evolved into key load nodes in the power system. Since the computing power and power systems have long been in a state of separation in terms of operation mechanism, related technologies mainly realize resource scheduling between the two through scheduling systems.

[0003] However, the problems with the related technologies are: 1) The scheduling system only focuses on the logical computing metrics (such as "requesting 1 GPU card") and cannot perceive the physical power consumption characteristics of the underlying heterogeneous hardware (different models of GPU / NPU); 2) The scheduling system usually simplifies jobs as homogeneous instant tasks and ignores the differences in latency sensitivity of different business scenarios. For example, large-scale offline training tasks usually have a high latency tolerance and should be executed as "valley filling" resources during periods of low electricity prices; 3) The traditional scheduling strategies adopted by the scheduling system are limited to the resource fragmentation within a single computing center (time dimension) or only based on static strategies for off-site disaster recovery (spatial dimension). Summary of the Invention

[0004] This invention aims to at least partially solve one of the technical problems in related technologies. Therefore, the first objective of this invention is to propose a cross-temporal and spatial collaborative scheduling method for computing power and electricity, which can uniformly manage multiple heterogeneous computing power clusters at the lower level and connect to the power trading and scheduling system at the upper level, realizing proactive and refined adjustment of computing power load.

[0005] A second objective of this invention is to provide a computer-readable storage medium.

[0006] The third objective of this invention is to propose a computing power-electricity cross-temporal and spatial collaborative scheduling system.

[0007] To achieve the above objectives, the computing power-power cross-temporal and spatial collaborative scheduling method proposed in the first aspect of the present invention includes: performing multi-dimensional perception of the source-load environment and user operation intentions to obtain energy status data from the power grid side, resource operation data from the computing power side, and operation micro-architecture feature vectors and service quality intention information from the user side; generating an estimated runtime and average power profile of the user operation under a specific hardware environment based on the operation micro-architecture feature vectors, hardware parameters, and a pre-set energy consumption mapping model, and obtaining the time elasticity index and spatial elasticity index of the user operation, wherein the time elasticity index is the time window that can be postponed without default, and the spatial elasticity index is... The elasticity index represents the feasibility of cross-domain migration. Based on the energy status data, resource operation data, the expected runtime and average power profile of the user job under a specific hardware environment, the service quality intent information, and the time elasticity index and spatial elasticity index of the user job, a spatiotemporal collaborative decision-making process based on two-level buffering and cost routing is performed on the user job. The spatiotemporal collaborative decision-making process includes time-domain load shaping and spatial traffic redirection. The power resources on the grid side and the computing resources on the computing power side are dynamically matched according to the time-domain adjustment instructions of the time-domain load shaping and the spatial routing instructions of the spatial traffic redirection to execute the user job.

[0008] According to embodiments of the present invention, the computing power-power cross-temporal and spatial collaborative scheduling method solves the semantic gap between logical computing power resources and physical power load by constructing a computing power-power heterogeneous mapping model, thereby realizing the digitalization of regulation capabilities. Furthermore, by introducing user intent awareness and a two-level buffer queue mechanism, it solves the multi-objective game problem between computing power cost and service quality while ensuring the Service Level Agreement (SLA). Moreover, by using cross-domain inference service mirror preheating and cost-aware routing, it solves the latency bottleneck of computing power tasks "following the green" in a wide area, thereby realizing the spatiotemporal dynamic matching of computing power flow and power flow.

[0009] Thus, multiple heterogeneous computing power clusters are uniformly managed downwards and connected to the power trading and dispatching system upwards, enabling proactive and refined adjustment of computing power load.

[0010] In addition, the computing power-electricity cross-temporal and spatial collaborative scheduling method according to the above embodiments of the present invention may also have the following additional technical features: According to one embodiment of the present invention, the energy status data includes the predicted value of spot electricity price, the predicted value of renewable energy output and carbon emission intensity factor; the resource operation data includes the load level of each heterogeneous node, the latency of network links and the backlog length of job queues; the job microarchitecture feature vector includes the deep learning framework, batch size, number of iterations and dataset size; and the service quality intent information includes a Boolean flag indicating whether the user should enable economic mode and the maximum acceptable tolerable waiting time for the user.

[0011] According to one embodiment of the present invention, generating the expected runtime and average power profile of a user job under a specific hardware environment, and obtaining the time elasticity index and spatial elasticity index of the user job, includes: inputting the job microarchitecture feature vector and the hardware parameters into the preset energy consumption mapping model to obtain the expected runtime and average power profile of the user job under a specific hardware environment; obtaining the time elasticity index of the user job based on the user's maximum acceptable tolerable waiting time and the expected runtime; and obtaining the spatial elasticity index of the user job based on the cross-domain transmission bandwidth and the job data volume.

[0012] According to one embodiment of the present invention, the time-domain load shaping includes: constructing a two-level buffer mechanism of a real-time queue and a delayed queue; if the predicted spot electricity price is greater than a first preset electricity price threshold and the Boolean flag indicates that the user should activate the economic mode, then the user job is injected into the delayed queue; otherwise, it is injected into the real-time queue; if the predicted spot electricity price is less than a second preset electricity price threshold or the predicted renewable energy output is greater than a first output threshold, then each user job in the delayed queue is dynamically released to the real-time queue according to its priority; or, if the current waiting time of the user job is greater than or equal to a preset proportion of the user's maximum acceptable tolerable waiting time, then the user job is forcibly promoted to the real-time queue.

[0013] According to one embodiment of the present invention, the method further includes obtaining the priority of each user job based on the waiting urgency and unit energy consumption economy of each user job.

[0014] According to one embodiment of the present invention, the airspace traffic guidance includes: obtaining the comprehensive cost potential of each computing node, wherein the comprehensive cost potential is a weighted sum of electricity cost, carbon emission intensity, network latency and migration penalty; if the comprehensive cost potential of the current node is less than a preset potential threshold, generating a cross-domain scheduling instruction to drive the current node to pull the inference service image in advance and start the shadow instance; and after the shadow instance is ready, dynamically adjusting the routing weight configuration of the global gateway to guide the service traffic to migrate smoothly to the current node.

[0015] According to an embodiment of the present invention, the step of dynamically matching the power resources on the grid side and the computing resources on the computing power side based on the time-domain load shaping time-domain adjustment command and the spatial traffic-oriented spatial routing command includes: responding to the time-domain adjustment command by dynamically adjusting resource quotas or modifying the priority class of user jobs, and responding to the spatial routing command by calling the network gateway interface to update the distribution weight table of the load balancer.

[0016] According to an embodiment of the present invention, the method further includes: after executing the user job, obtaining the actual energy efficiency and SLA achievement rate; if the difference between the actual energy efficiency and the SLA achievement rate is greater than or equal to the difference between the actual energy efficiency and the SLA achievement rate, updating the preset energy consumption mapping model.

[0017] To achieve the above objectives, a computer-readable storage medium is provided in the second aspect of the present invention, which stores a computing power-power intertemporal collaborative scheduling program thereon. When the computing power-power intertemporal collaborative scheduling program is executed by a processor, it implements the computing power-power intertemporal collaborative scheduling method of the present invention described above.

[0018] According to embodiments of the present invention, a computer-readable storage medium can execute a computing power-electricity cross-temporal and spatial collaborative scheduling program stored thereon to uniformly manage multiple heterogeneous computing power clusters downwards and connect to the power trading and dispatching system upwards, thereby realizing proactive and refined adjustment of computing power load.

[0019] To achieve the above objectives, the computing power-power cross-temporal collaborative scheduling system proposed in the third aspect of the present invention includes: a multi-dimensional intent and state perception module, used to perform multi-dimensional perception of the source-load environment and user operation intent, and acquire energy status data from the grid side, resource operation data from the computing power side, and operation micro-architecture feature vectors and service quality intent information from the user side; and a heterogeneous resource mapping and quantification module, used to generate the expected runtime and average power profile of the user operation under a specific hardware environment based on the operation micro-architecture feature vectors, hardware parameters, and a pre-set energy consumption mapping model, and to acquire the time elasticity index and spatial elasticity index of the user operation, wherein the time elasticity index is the time window that can be postponed without default. The spatial elasticity index represents the feasibility of cross-domain migration. The spatiotemporal collaborative decision-making module is used to perform spatiotemporal collaborative decision-making based on two-level buffering and cost routing on the user job, according to the energy status data, resource operation data, the expected runtime and average power profile of the user job under a specific hardware environment, the service quality intent information, and the time elasticity index and spatial elasticity index of the user job. The spatiotemporal collaborative decision-making includes time-domain load shaping and spatial traffic guidance. The execution interaction module is used to dynamically match the power resources on the grid side and the computing resources on the computing power side according to the time-domain adjustment instructions for time-domain load shaping and the spatial routing instructions for spatial traffic guidance, in order to execute the user job.

[0020] According to embodiments of the present invention, the computing power-power cross-temporal and spatial collaborative scheduling system solves the semantic gap between logical computing power resources and physical power load by constructing a computing power-power heterogeneous mapping model, thereby realizing the digitalization of regulation capabilities. Furthermore, by introducing user intent awareness and a two-level buffer queue mechanism, it solves the multi-objective game problem between computing power cost and service quality while ensuring the Service Level Agreement (SLA). Moreover, by using cross-domain mirroring preheating and cost-aware routing, it solves the latency bottleneck of computing power tasks "following the green" in a wide area, thereby realizing the spatiotemporal dynamic matching of computing power flow and power flow.

[0021] Thus, multiple heterogeneous computing power clusters are uniformly managed downwards and connected to the power trading and dispatching system upwards, enabling proactive and refined adjustment of computing power load.

[0022] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the computing power-electricity cross-temporal collaborative scheduling method according to an embodiment of the present invention; Figure 2This is a flowchart illustrating a computing power-electricity cross-temporal collaborative scheduling method according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating a computing power-electricity cross-temporal collaborative scheduling method according to an embodiment of the present invention; Figure 4 This is a flowchart illustrating a computing power-electricity cross-temporal collaborative scheduling method according to a specific embodiment of the present invention; Figure 5 This is a flowchart illustrating a computing power-electricity cross-temporal collaborative scheduling method according to an embodiment of the present invention; Figure 6 This is a flowchart illustrating a computing power-electricity cross-temporal collaborative scheduling method according to a specific embodiment of the present invention; Figure 7 This is a flowchart illustrating a computing power-electricity cross-temporal collaborative scheduling method according to an embodiment of the present invention; Figure 8 This is a block diagram of a computing power-electricity cross-temporal collaborative scheduling system according to an embodiment of the present invention; Figure 9 This is a schematic diagram of the connection topology of a computing power-electricity cross-temporal collaborative scheduling system according to an embodiment of the present invention. Detailed Implementation

[0024] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0025] The following describes, with reference to the accompanying drawings, a computing power-power cross-temporal and spatial collaborative scheduling method, a computer-readable storage medium, and a computing power-power cross-temporal and spatial collaborative scheduling system according to embodiments of the present invention.

[0026] Figure 1 This is a flowchart illustrating the computing power-electricity cross-temporal collaborative scheduling method according to an embodiment of the present invention.

[0027] Specifically, in some embodiments of the present invention, such as Figure 1 As shown, the computing power-power cross-temporal collaborative scheduling method includes: S101 performs multi-dimensional perception of the source-load environment and user operation intentions, and obtains energy status data from the power grid side, resource operation data from the computing power side, and operation micro-architecture feature vectors and service quality intention information from the user side.

[0028] It is understood that in this embodiment of the present invention, by performing multi-dimensional perception of the source-load environment and user operation intentions, real-time collection of energy status data from the grid side, resource operation data from the computing power side, and operation micro-architecture feature vectors and service quality intention information from the user side, a panoramic view is established for subsequent decision-making.

[0029] Specifically, in some embodiments of the present invention, the energy status data includes the spot electricity price forecast, the new energy output forecast, and the carbon emission intensity factor; the resource operation data includes the load level of each heterogeneous node, the latency of the network link, and the job queue backlog length; the job microarchitecture feature vector includes the deep learning framework, batch size, number of iterations, and dataset size; and the service quality intent information includes a Boolean flag indicating whether the user should enable the economic mode and the maximum tolerable waiting time that the user can accept.

[0030] Specifically, in this embodiment of the present invention, the source load environment and the user's operational intent are perceived in multiple dimensions through the following methods: Source load environmental perception: Through the data management interface, real-time energy status data from the grid side is retrieved, including spot electricity price forecasts. Forecast values ​​of new energy power output With carbon emission intensity factor .

[0031] Through the cluster monitoring interface, resource operation data on the computing power side is collected, including the load level of each heterogeneous node and the latency of network links. And the length of the job queue backlog.

[0032] User task intent awareness: Provide interactive options at the job submission portal (UI / API) to collect the user's job micro-architecture feature vector. This includes deep learning frameworks, batch size, number of iterations (epochs), and dataset size.

[0033] Collect user service quality (SLA) intent information, including Boolean flags used to indicate whether the user should activate economy mode. and the maximum acceptable tolerable waiting time for users .

[0034] S102 generates the expected runtime and average power profile of the user job under a specific hardware environment based on the job microarchitecture feature vector, hardware parameters and preset energy consumption mapping model, and obtains the time elasticity index and spatial elasticity index of the user job.

[0035] It is understood that, in this embodiment of the present invention, a non-intrusive quantization model is established to address the problem of opaque power consumption characteristics of heterogeneous hardware (different models of GPU / NPU) in order to realize job profile generation and spatiotemporal elasticity calculation, thereby realizing the assessment of adjustment capability based on heterogeneous resource mapping.

[0036] Therefore, in the above embodiments of the present invention, by constructing an "information-energy coupling" mapping model, the physical power consumption profile of user operations is accurately deduced based on logical operation characteristics (e.g., batch size, epochs, etc.), thereby breaking through the "semantic gap" between heterogeneous computing power and power load, realizing the digitalization of regulation capabilities, enabling computing centers to transform complex computing tasks into standard, predictable power load curves, and possessing the ability to participate in power market demand response as accurately as industrial loads.

[0037] Specifically, in this embodiment of the invention, such as Figure 2 As shown, the system generates a profile of the expected runtime and average power consumption of user jobs under a specific hardware environment, and obtains the time elasticity index and spatial elasticity index of user jobs, including: S201 inputs the job microarchitecture feature vector and hardware parameters into the preset energy consumption mapping model to obtain the expected runtime and average power profile of the user job under a specific hardware environment.

[0038] It is understood that, in this embodiment of the present invention, the job microarchitecture feature vector collected in the aforementioned step S101 is... The data is input into a pre-defined energy consumption mapping model (e.g., preferably a regression model based on Transformer or GBDT). This model is trained jointly using historical job logs and real-time power consumption monitoring data. The inputs to this model also include environmental parameters such as hardware model and cooling efficiency. The output is the estimated runtime of the user's job under a specific hardware environment. and average power profile .

[0039] Therefore, in the above embodiments of the present invention, the estimated runtime of user jobs under a specific hardware environment is output through a preset energy consumption mapping model. and average power profile Thus, logical computational tasks are transformed into physical electrical loads, for example, ,in, To create a profile of the average power, This is a mapping function between the microarchitecture feature vectors and hardware parameters and the average power profile. For the microarchitecture feature vector of the task, These are hardware parameters.

[0040] It should be noted that the pre-built energy consumption mapping model supports online learning and can dynamically update the weights based on the actual power consumption data from the closed-loop feedback in subsequent steps.

[0041] S202, based on the user's maximum acceptable tolerable waiting time and expected runtime, obtain the user job's time elasticity index, and based on the cross-domain transmission bandwidth and job data volume, obtain the user job's spatial elasticity index. The time elasticity index is the time window that can be postponed under the premise of no default, and the spatial elasticity index is the feasibility of cross-domain migration.

[0042] It is understood that, in this embodiment of the present invention, the maximum tolerable waiting time for the user is also considered. and expected runtime The system obtains the time elasticity index of user jobs (i.e., the time window during which execution can be postponed without defaulting) and the spatial elasticity index of user jobs (i.e., the feasibility of cross-domain migration) based on cross-domain transmission bandwidth and job data volume.

[0043] S103, based on energy status data, resource operation data, the expected runtime and average power profile of user jobs under specific hardware environments, service quality intent information, and the time elasticity index and spatial elasticity index of user jobs, performs spatiotemporal collaborative decision-making on user jobs based on two-level buffering and cost routing. The spatiotemporal collaborative decision-making includes time-domain load shaping and spatial traffic redirection.

[0044] It is understood that in this embodiment of the present invention, by adopting a "spatiotemporal decoupling" strategy, time-domain load shaping and spatial traffic guidance are executed in parallel, and spatiotemporal collaborative decision-making based on two-level buffering and cost routing is performed on user jobs.

[0045] The specific processes of time-domain load shaping and spatial-domain flow guidance are described below with reference to specific embodiments of the present invention: Furthermore, in some embodiments of the present invention, such as Figure 3 As shown, time-domain load shaping includes: S301 establishes a two-level buffering mechanism consisting of a real-time queue and a delayed queue.

[0046] It is understood that in this embodiment of the present invention, a two-level buffering mechanism of real-time queue and delayed queue is constructed to achieve dynamic scheduling based on the two-level buffered queue.

[0047] S302, if the predicted spot electricity price is greater than the first preset electricity price threshold and the Boolean flag indicates that the user should enable the economic mode, then the user's job will be injected into the delayed queue; otherwise, it will be injected into the real-time queue.

[0048] Understandably, when new user orders arrive, combined with the spot electricity price forecast... With Boolean tags Select the appropriate queue for the new user job; specifically, if the spot electricity price forecast value... Greater than the first preset electricity price threshold P1 and Boolean flag Instructing users to activate the economy mode (i.e.) If the user job is not found to be in the queue, it will be injected into the delayed queue for "water storage" and the expected savings will be returned to the user. Otherwise, the user job will be directly injected into the real-time queue for execution.

[0049] S303, if the predicted spot electricity price is less than the second preset electricity price threshold or the predicted new energy output is greater than the first output threshold, then according to the priority of each user's job in the delay queue, the user's job in the delay queue will be dynamically released to the real-time queue; or, if the current waiting time of a user's job is greater than or equal to a preset proportion of the user's maximum acceptable tolerable waiting time, then the user's job will be forcibly promoted to the real-time queue.

[0050] It is understood that, in this embodiment of the present invention, by monitoring the power grid status in real time, a spot electricity price forecast is detected. Less than the second preset electricity price threshold P2 or the predicted value of renewable energy output When the output time exceeds the first output threshold E1, a batch release signal is triggered. Based on the priority of each user job in the delay queue, these jobs are dynamically released to the real-time queue. Simultaneously, the delay queue is continuously scanned; once the current waiting time of a user job is detected to be greater than the threshold, a batch release signal is triggered. Or equal to the user's maximum acceptable tolerable waiting time. The preset ratio (e.g., reaching 90%) will forcefully trigger the escape mechanism, regardless of the electricity price, and immediately force the user's work to be promoted to the real-time queue to ensure that the business does not default.

[0051] Therefore, in the above embodiments of the present invention, by introducing a "user intent perception" and a "two-level buffer queue" mechanism, the user's "cost-saving intention" is identified. The delayed queue is used to "store" user operations during peak electricity price periods and "release" user operations during off-peak electricity price periods. At the same time, in conjunction with the SLA timeout circuit breaker mechanism, the lowest cost execution window is automatically found while strictly adhering to the user's maximum waiting time commitment. This satisfies the user's low-cost needs while also meeting the grid's peak shaving and valley filling needs, thereby achieving effective cost reduction and SLA guarantee, and realizing refined flexible scheduling.

[0052] Furthermore, in some embodiments of the present invention, the priority of each user job is obtained based on the urgency of waiting for each user job and the economic efficiency of unit energy consumption.

[0053] It is understood that, in this embodiment of the present invention, the priority score of each user job in the delay queue is calculated based on the waiting urgency of each user job and the economic efficiency per unit energy consumption. Jobs with higher priority scores are placed in the real-time queue. The priority score is proportional to the urgency of the user job's wait (e.g., whether it's about to time out) and the energy efficiency of the user job (e.g., how much money can be saved). Specifically, the priority score of a user job... ,in, , These are configurable weighting coefficients. Waiting time for the assignment This represents the maximum tolerable waiting time for the task. To estimate cost savings, Profiling the average power of the operation. The estimated runtime for the operation.

[0054] The following is in conjunction with the appendix Figure 4 The time-domain adjustment process of the embodiments of the present invention will continue to be described, for example, as follows: Figure 4 As shown, in some embodiments of the present invention, after the user submits the job, step S10 is executed: S10 performs user job intent perception on user jobs, and collects user job microarchitecture feature vectors and service quality (SLA) intent information.

[0055] S11. Determine whether the user has enabled the economy mode. If yes, proceed to step S12; otherwise, proceed to step S18.

[0056] S12, determine whether the predicted spot electricity price is higher than the first preset electricity price threshold. If yes, proceed to step S13; otherwise, proceed to step S18.

[0057] S13, inject the user job into the delay queue.

[0058] S14 monitors each user job in the delay queue in parallel.

[0059] S15, determine whether each user job in the delay queue has triggered the release condition. If the electricity price drops to a trough or green power output surges, proceed to step S16. If the waiting time of a user job is close to the user's maximum acceptable tolerable waiting time... If the SLA circuit breaker fails, proceed to step S17.

[0060] S16, calculate the release priority of each user job in the delay queue, and execute step S18.

[0061] S17, forcibly trigger the escape mechanism and execute step S18.

[0062] S18 injects the user job into the real-time queue.

[0063] Furthermore, in some embodiments of the present invention, such as Figure 5 As shown, airspace traffic guidance includes: S401 obtains the comprehensive cost potential energy of each computing node. The comprehensive cost potential energy is a weighted sum of electricity cost, carbon emission intensity, network latency and migration penalty.

[0064] It is understood that, in this embodiment of the present invention, each computing node is calculated in real time. Comprehensive cost potential energy To achieve a comprehensive cost potential energy Cross-domain migration, in which the comprehensive cost potential energy For electricity costs Carbon emission intensity Network latency and relocation penalties The weighted sum (i.e. ),in, This is a weighting factor for electricity costs. This is a weighting factor for carbon emission intensity. This is a weighting factor for network latency. This is the weighting coefficient for migration penalties.

[0065] S402 If the overall cost potential energy of the current node is less than the preset potential energy threshold, a cross-domain scheduling instruction is generated to drive the current node to pull the inference service image in advance and start the shadow instance. After the shadow instance is ready, the routing weight configuration of the global gateway is dynamically adjusted to guide the service traffic to migrate smoothly to the current node.

[0066] It is understood that, in this embodiment of the present invention, when the comprehensive cost potential of the current node (e.g., the western node) is detected... Less than the preset potential threshold (For example, the potential energy resulting from the abundance of green electricity) When the traffic is significantly reduced, a cross-domain scheduling instruction is generated. This instruction is used to implement the warm-up phase (i.e., the aforementioned driving the current node to pull the inference service image in advance (including the environment and code that the task depends on) and start the shadow instance, at which time no traffic is connected) and the switching phase (i.e., after the shadow instance is ready, the routing weight configuration of the global gateway is dynamically adjusted to guide the service traffic to migrate smoothly to the current node).

[0067] Therefore, in the above embodiments of the present invention, by using "cost potential energy assessment" and "cooperative preheating mechanism", the potential energy advantage of areas with abundant green electricity can be perceived in advance, and the operating environment (shadow instance) can be prepared in advance in the area. Then, when green electricity surges or electricity prices are inverted, millisecond-level seamless switching of business traffic can be achieved, truly implementing the concept of "computing power following green development", significantly improving the proportion of renewable energy consumption, thereby overcoming the latency bottleneck of cross-domain scheduling and improving the level of green electricity consumption.

[0068] The following is in conjunction with the appendix Figure 6 The airspace routing process of this embodiment of the invention will be further described. For example, as follows: Figure 6 As shown, the airspace migration process is divided into four phases: Phase 1: Monitoring and Assessment; Phase 2: Collaborative Warm-up; Phase 3: Seamless Switching; and Phase 4: Resource Release. In Phase 1, the combined cost potential difference ΔJ between the source and target regions is calculated. If ΔJ exceeds the migration threshold, Phase 2 begins. In Phase 2, the target region (e.g., the western region) is controlled by pulling the mirror image and starting a shadow instance. The source region (e.g., the eastern region) is controlled by maintaining service and unchanged traffic, proceeding to Phase 3. In Phase 3, the global gateway weight is modified, and traffic is smoothly diverted to the target region, proceeding to Phase 4. In Phase 4, the instance in the source region is shut down. This achieves cross-domain migration.

[0069] S104 dynamically matches the power resources on the grid side and the computing resources on the computing side according to the time-domain adjustment instructions for time-domain load shaping and the spatial routing instructions for spatial traffic guidance, in order to execute user jobs.

[0070] Furthermore, in some embodiments of the present invention, dynamically matching the power resources on the grid side and the computing resources on the computing power side according to the time-domain load shaping time-domain adjustment instructions and the spatial routing instructions for spatial traffic guidance includes: dynamically adjusting resource quotas or modifying the priority class of user jobs in response to the time-domain adjustment instructions, and calling the network gateway interface to update the distribution weight table of the load balancer in response to the spatial routing instructions.

[0071] It is understood that, in this embodiment of the present invention, in response to the time-domain adjustment command, the resource quota or job priority class is dynamically adjusted by calling the interface of the underlying container orchestration engine (such as Kubernetes), thereby controlling the operation of the physical server and task execution, and in response to the airspace routing command, the distribution weight table of the load balancer is updated by calling the network gateway interface.

[0072] Furthermore, in some embodiments of the present invention, such as Figure 7 As shown, the method also includes: S105 obtains the actual energy efficiency and SLA achievement rate after executing user tasks.

[0073] It is understood that, in this embodiment of the present invention, the actual energy efficiency (unit computing power, such as the electricity cost per PFlops) and SLA achievement rate after executing user jobs are collected in order to adaptively learn the preset energy consumption mapping model.

[0074] S106, If there is a difference between the actual energy efficiency and the SLA achievement rate, update the preset energy consumption mapping model.

[0075] It is understood that in this embodiment of the present invention, when the difference between the actual energy efficiency and the SLA achievement rate is greater than the deviation threshold, the model parameters of the preset energy consumption mapping model in the aforementioned step S201 are corrected using an online learning algorithm, so as to further optimize the accuracy of real-time control of computing power resources and power resources.

[0076] In summary, the computing power-power cross-temporal collaborative scheduling method according to embodiments of the present invention addresses the semantic gap between logical computing power resources and physical power load by constructing a heterogeneous mapping model of computing power and power, thereby digitizing the regulation capability. Furthermore, by introducing user intent awareness and a two-level buffer queue mechanism, it resolves the multi-objective game between computing power cost and service quality while ensuring the Service Level Agreement (SLA). Finally, through cross-domain mirroring preheating and cost-aware routing, it addresses the latency bottleneck of computing tasks "following the green" in a wide-area space, achieving dynamic spatiotemporal matching of computing power flow and power flow. Thus, it unifies the management of multiple heterogeneous computing power clusters downwards and connects upwards to the power trading and scheduling system, enabling proactive and refined adjustment of computing load.

[0077] Based on the computing power-power cross-temporal and spatial collaborative scheduling method of the above-mentioned embodiments of the present invention, the present invention also proposes a computer-readable storage medium storing a computing power-power cross-temporal and spatial collaborative scheduling program thereon. When the computing power-power cross-temporal and spatial collaborative scheduling program is executed by a processor, it implements the computing power-power cross-temporal and spatial collaborative scheduling method of the above-mentioned embodiments of the present invention.

[0078] It should be understood that the specific implementation of the computer-readable storage medium in the embodiments of the present invention can be found in the specific implementation of the computing power-electricity cross-temporal and spatial collaborative scheduling method in the foregoing embodiments of the present invention. To reduce redundancy, it will not be repeated here.

[0079] In summary, the computer-readable storage medium according to embodiments of the present invention, by executing the computing power-power cross-temporal and spatial collaborative scheduling program stored thereon, can uniformly manage multiple heterogeneous computing power clusters downwards and connect to the power trading and scheduling system upwards, thereby realizing proactive and refined adjustment of computing power load.

[0080] Figure 8 This is a block diagram of a computing power-electricity cross-temporal collaborative scheduling system according to an embodiment of the present invention.

[0081] Specifically, in some embodiments of the present invention, such as Figure 8 As shown, the computing power-electricity cross-temporal collaborative scheduling system 100 includes: an intent and state multi-dimensional perception module 10, a heterogeneous resource mapping and quantification module 20, a spatiotemporal collaborative decision-making module 30, and an execution interaction module 40.

[0082] The intent and state multidimensional perception module 10 is used to perform multidimensional perception of the source-load environment and user operation intent, and to obtain energy status data from the grid side, resource operation data from the computing power side, and operation microarchitecture feature vectors and service quality intent information from the user side. The heterogeneous resource mapping and quantification module 20 is used to generate the expected runtime and average power profile of user operations under a specific hardware environment based on the operation microarchitecture feature vector, hardware parameters, and a preset energy consumption mapping model, and to obtain the time elasticity index and spatial elasticity index of user operations. The time elasticity index is the time window that can be postponed under the premise of no default, and the spatial elasticity index is the cross-regional elasticity index. The feasibility of domain migration; the spatiotemporal collaborative decision-making module 30 is used to perform spatiotemporal collaborative decision-making on user jobs based on two-level buffering and cost routing, according to energy status data, resource operation data, the expected runtime and average power profile of user jobs in a specific hardware environment, service quality intent information, and the time elasticity index and spatial elasticity index of user jobs. The spatiotemporal collaborative decision-making includes time-domain load shaping and spatial traffic guidance; the execution interaction module 40 is used to dynamically match the power resources on the grid side and the computing resources on the computing power side according to the time-domain adjustment instructions of time-domain load shaping and the spatial routing instructions of spatial traffic guidance, so as to execute user jobs.

[0083] Specifically, such as Figure 9 As shown, the computing power-power cross-temporal and spatial collaborative scheduling system 100 serves as the core processing unit. Based on the dual-source drive of the computing power user terminal (job characteristics / SLA intent) and the power grid trading / scheduling system (electricity price / green electricity / carbon potential) at the input end, and combined with the intent and state multi-dimensional perception module 10, heterogeneous resource mapping and quantification module 20, spatiotemporal collaborative decision module 30 (including time domain: two-level buffer queue decision engine and spatial domain: cost-aware routing decision engine) and execution interaction module 40, it performs corresponding control on the underlying heterogeneous computing power clusters (e.g., K8s / Volcano, etc., used to execute quota / priority instructions and closed-loop feedback of actual electricity cost / SLA achievement rate) and network gateway devices (e.g., Gateway / DNS, etc., used to execute routing weight instructions) in the heterogeneous infrastructure at the execution end.

[0084] It should be understood that the specific implementation of the computing power-power cross-temporal and spatial collaborative scheduling system 100 in the embodiments of the present invention can refer to the specific implementation of the computing power-power cross-temporal and spatial collaborative scheduling method in the foregoing embodiments of the present invention. To reduce redundancy, it will not be repeated here.

[0085] In summary, the computing power-power cross-temporal collaborative scheduling system according to embodiments of the present invention addresses the semantic gap between logical computing resources and physical power load by constructing a heterogeneous mapping model of computing power and power, thereby digitizing the regulation capability. Furthermore, by introducing user intent awareness and a two-level buffer queue mechanism, it resolves the multi-objective game between computing power cost and service quality while ensuring the Service Level Agreement (SLA). Finally, through cross-domain mirroring preheating and cost-aware routing, it overcomes the latency bottleneck of computing tasks "following the green" in a wide-area space, achieving dynamic spatiotemporal matching of computing power flow and power flow. Thus, it unifies the management of multiple heterogeneous computing power clusters downwards and connects upwards to the power trading and scheduling system, enabling proactive and refined adjustment of computing load.

[0086] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0087] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0088] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. 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 suitable manner in one or more embodiments or examples.

[0089] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0090] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0091] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0092] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "over," and "on top" of the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0093] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for cross-temporal and spatial collaborative scheduling of computing power and power, characterized in that, The method includes: Multidimensional perception of the source-load environment and user operation intentions is achieved to obtain energy status data from the power grid side, resource operation data from the computing power side, and operation micro-architecture feature vectors and service quality intention information from the user side. Based on the job microarchitecture feature vector, hardware parameters, and preset energy consumption mapping model, the expected runtime and average power profile of the user job in a specific hardware environment are generated, and the time elasticity index and spatial elasticity index of the user job are obtained. The time elasticity index is the time window that can be postponed without default, and the spatial elasticity index is the feasibility of cross-domain migration. Based on the energy status data, the resource operation data, the expected runtime and average power profile of the user job under a specific hardware environment, the service quality intent information, and the time elasticity index and spatial elasticity index of the user job, a spatiotemporal collaborative decision based on two-level buffering and cost routing is performed on the user job, wherein the spatiotemporal collaborative decision includes time-domain load shaping and spatial traffic redirection. The power resources on the grid side and the computing resources on the computing side are dynamically matched according to the time-domain load shaping time-domain adjustment command and the spatial routing command for spatial flow guidance in order to execute the user job. The energy status data includes the predicted spot electricity price, the predicted output of new energy sources, and the carbon emission intensity factor. The resource operation data includes the load level of each heterogeneous node, the latency of the network link, and the backlog length of the job queue. The job microarchitecture feature vector includes the deep learning framework, batch size, number of iterations, and dataset size. The service quality intent information includes a Boolean flag indicating whether the user should enable the economic mode and the maximum tolerable waiting time that the user can accept. The time-domain load shaping includes: Construct a two-level buffering mechanism consisting of a real-time queue and a delayed queue; If the predicted spot electricity price is greater than the first preset electricity price threshold and the Boolean flag indicates that the user should enable the economic mode, then the user's job will be injected into the delayed queue; otherwise, it will be injected into the real-time queue. If the predicted spot electricity price is less than the second preset electricity price threshold or the predicted new energy output is greater than the first output threshold, then each user job in the delay queue is dynamically released to the real-time queue according to the priority of each user job in the delay queue; or, if the current waiting time of the user job is greater than or equal to a preset proportion of the user's maximum acceptable tolerable waiting time, then the user job is forcibly promoted to the real-time queue.

2. The computing power-power cross-temporal collaborative scheduling method according to claim 1, characterized in that, The process of generating the expected runtime and average power profile of the user job under a specific hardware environment, and obtaining the time elasticity index and spatial elasticity index of the user job, includes: The job microarchitecture feature vector and the hardware parameters are input into the preset energy consumption mapping model to obtain the expected runtime and average power profile of the user job under a specific hardware environment. Based on the user's maximum acceptable tolerable waiting time and the expected runtime, the time elasticity index of the user job is obtained, and based on the cross-domain transmission bandwidth and job data volume, the spatial elasticity index of the user job is obtained.

3. The computing power-power cross-temporal collaborative scheduling method according to claim 1, characterized in that, The priority of each user job is obtained based on the urgency of waiting and the economic efficiency of unit energy consumption.

4. The computing power-power cross-temporal collaborative scheduling method according to claim 1, characterized in that, The airspace traffic guidance includes: Obtain the comprehensive cost potential energy of each computing node, wherein the comprehensive cost potential energy is a weighted sum of electricity cost, carbon emission intensity, network latency and migration penalty; If the overall cost potential energy of the current node is less than the preset potential energy threshold, a cross-domain scheduling instruction is generated to drive the current node to pull the inference service image in advance and start the shadow instance. After the shadow instance is ready, the routing weight configuration of the global gateway is dynamically adjusted to guide the service traffic to migrate smoothly to the current node.

5. The computing power-power cross-temporal collaborative scheduling method according to any one of claims 1-4, characterized in that, The dynamic matching of the power resources on the grid side and the computing resources on the computing side based on the time-domain load shaping time-domain adjustment command and the spatial traffic-oriented spatial routing command includes: In response to the time-domain adjustment command, the resource quota is dynamically adjusted or the priority class of user jobs is modified. In response to the spatial routing command, the network gateway interface is invoked to update the load balancer's distribution weight table.

6. The computing power-power cross-temporal collaborative scheduling method according to claim 1, characterized in that, The method further includes: After the user task is executed, the actual energy efficiency and SLA achievement rate are obtained; If the difference between the actual energy efficiency and the SLA achievement rate is greater than or equal to the actual energy efficiency, the preset energy consumption mapping model is updated.

7. A computer-readable storage medium, characterized in that, It stores a computing power-power cross-temporal and spatial collaborative scheduling program, which, when executed by the processor, implements the computing power-power cross-temporal and spatial collaborative scheduling method as described in any one of claims 1-6.

8. A computing power-power cross-temporal collaborative scheduling system, characterized in that, The system includes: The intent and state multidimensional perception module is used to perform multidimensional perception of the source load environment and user operation intent, and to obtain energy status data from the grid side, resource operation data from the computing power side, and operation micro-architecture feature vectors and service quality intent information from the user side. The energy status data includes the predicted value of spot electricity price, the predicted value of renewable energy output, and the carbon emission intensity factor. The resource operation data includes the load level of each heterogeneous node, the latency of network links, and the backlog length of the job queue. The operation micro-architecture feature vector includes the deep learning framework, batch size, number of iterations, and dataset size. The service quality intent information includes a Boolean flag indicating whether the user should activate the economic mode and the maximum acceptable tolerable waiting time for the user. The heterogeneous resource mapping and quantification module is used to generate the expected runtime and average power profile of the user job in a specific hardware environment based on the job microarchitecture feature vector, hardware parameters and preset energy consumption mapping model, and to obtain the time elasticity index and spatial elasticity index of the user job. The time elasticity index is the time window that can be postponed without default, and the spatial elasticity index is the feasibility of cross-domain migration. The spatiotemporal collaborative decision-making module is used to perform spatiotemporal collaborative decision-making based on two-level buffering and cost routing on user jobs according to the energy status data, resource operation data, the expected runtime and average power profile of user jobs under specific hardware environments, service quality intent information, and the time elasticity index and spatial elasticity index of user jobs. The spatiotemporal collaborative decision-making includes time-domain load shaping and spatial traffic guidance. The time-domain load shaping includes: constructing a two-level buffer mechanism of a real-time queue and a delay queue; if the predicted spot electricity price is greater than a first preset electricity price threshold and the Boolean flag indicates that the user is in economic mode, then the user job is injected into the delay queue; otherwise, it is injected into the real-time queue; if the predicted spot electricity price is less than a second preset electricity price threshold or the predicted new energy output is greater than a first output threshold, then each user job in the delay queue is dynamically released to the real-time queue according to its priority; or, if the current waiting time of the user job is greater than or equal to a preset proportion of the user's maximum acceptable tolerable waiting time, then the user job is forcibly promoted to the real-time queue. The execution interaction module is used to dynamically match the power resources on the grid side and the computing resources on the computing power side according to the time-domain load shaping time-domain adjustment instructions and the spatial traffic-oriented spatial routing instructions, so as to execute the user job.

Images