Global cross-timezone computing power scheduling method

Abstract

To address the problem of how to more effectively utilize computing resources across different time zones, this invention provides a global cross-time zone computing power scheduling method. The computing platform executes the following steps: Step 1: Receiving computing power demand information from at least one time zone and storing it in a storage module; Step 2: Calculating a computing power scheduling scheme that maximizes the profit margin and utilization rate of the computing platform based on a machine learning module using one of the following algorithms: a target-oriented algorithm, a constraint-oriented algorithm, a heuristic algorithm, a machine learning algorithm, or a deep learning algorithm; Step 3: Scheduling GPU resources from at least one time zone according to the computing power scheduling scheme.

Claims

1. A global cross-time zone computing power scheduling method, wherein the computing power platform executes the following: Step 1: Receive computing power demand information from at least one time zone, collect all parameter data and conditions, store them in a storage module, and transmit them to the machine learning module for filtering and calculation based on these conditions. This storage module also has an active monitoring and communication mechanism, enabling real-time communication with edge nodes, user applications, or cloud platforms in various locations. It automatically updates the conditions based on user behavior, power changes, and task load patterns. Simultaneously, the storage module continuously adjusts parameters with the machine learning module, instantly labeling sudden demand and abnormal trends such as soaring electricity prices, and notifying the machine learning module to redistribute or relearn the weights of the input parameter data. Step 2: Based on the objective and constraint criteria, the machine learning module, using one of the heuristic, machine learning, or deep learning algorithms, calculates a computing power scheduling scheme that maximizes the profit margin and utilization rate of the computing platform. The machine learning module generates a GPU scheduling and pricing scheme to optimize TP based on the objective and constraint criteria. In conjunction with UR, the machine learning module generates output parameters, and does not output results all at once. Instead, it performs rolling optimization based on feedback from the monitoring module. After each output, the monitoring module sends back the GPU simulated load and expected TP / UR effect. The machine learning module compares the GPU simulated load and expected TP / UR effect returned by the monitoring module with the actual situation, adjusts its strategy, and outputs the next solution. The machine learning module autonomously adjusts the α and β weight combination to seek a better solution based on GPU resource bottlenecks and historical performance in different time zones, without relying on fixed parameter data. Step 3: According to the computing power scheduling scheme, regional convection scheduling and complementary scheduling of GPU resources in at least one time zone are performed from the same and / or different time zones, and rental and pricing are carried out. The computing power platform not only performs resource flow and rental operations, but the monitoring module also automatically generates KPIs based on the actual GPU resource utilization rate, the difference between expected and actual revenue, and other factors after execution. The report is sent back to the machine learning module, triggering the module to perform short-term internal parameter re-optimization, mid-term fine-tuning of weight coefficients α and β, and long-term overall model retraining. Based on changes in regional conditions, the scheduling strategy automatically evolves to achieve self-consistent optimization of GPU resources. The objective function is: max = (), and the constraint functions include: UR = ; , , , ; TP = -, = min (+), ; The objective variable is TP: total profit within the global time zone time period [1,24]; UR: GPU utilization rate within the global time zone time period [1,24]; The decision variable is: GPU rental price for each region in each time period; The parameter data includes: the number of GPUs available for scheduling in each region, where i represents the region; The parameter data includes: the power demand for each region in each time period; : the computing power demand for each region in each time period; =: the number of GPUs scheduled from the region to be used at time t; : the number of idle GPUs in each region in each time period; : the number of GPUs used in each region in each time period;The variables are: ,,…,: n regions globally; ,,…,: each time period of the day; : electricity price for each time zone during each time period; : electricity cost for each region during each time period when the GPU is idle; : electricity cost for each region during each time period when the GPU is in use.

2. The global cross-time zone computing power scheduling method as described in request item 1, wherein, The target variables include: total profit within the specified period and GPU utilization rate.

3. The global cross-time zone computing power scheduling method as described in request item 1, wherein, This constraint can calculate profit margin, utilization rate, and scheduling constraints.

4. The global cross-time zone computing power scheduling method as described in claim 1, wherein, The machine learning module, based on one of the heuristic algorithms, machine learning algorithms, or deep learning algorithms, dynamically adjusts the computing power scheduling scheme according to real-time changes in computing power demand information, available GPU resources, and electricity prices.

5. The global cross-time zone computing power scheduling method as described in claim 1, wherein, This machine learning module, based on one of the heuristic algorithms, machine learning algorithms, or deep learning algorithms, predicts future resource needs by analyzing trends in historical and real-time data, thereby adjusting GPU resource allocation in advance.

6. The global cross-time zone computing power scheduling method as described in claim 5, wherein, This real-time data includes electricity prices for any combination of different time zones, different time periods, GPU resource usage, and GPU resource idleness.

7. The global cross-time zone computing power scheduling method as described in claim 6, wherein, This machine learning module, based on one of the heuristic algorithms, machine learning algorithms, or deep learning algorithms, can extract and summarize user or system behavior patterns from historical data, thereby adjusting GPU resource allocation in advance.

Images