GPU computing resource scheduling method and device, and storage medium

Abstract

The application relates to the technical field of computers and discloses a GPU computing resource scheduling method and device and a storage medium, and adaptive regulation and control of time slice intervals are realized by constructing a multi-dimensional evaluation model chain; first, task computing density and display memory requirements are used to generate task computing characteristic coefficients; performance efficiency coefficients are output in combination with task average response time and preemption overhead proportion; then, ready queue length and context switching frequency are fused to generate a load synergy factor under the constraints of task characteristics and performance coefficients; meanwhile, hardware health coefficients are evaluated based on hardware temperature and cache hit rate changes; and finally, the optimal time slice interval is dynamically calculated through an interval decision model according to the load synergy factor and the hardware health coefficient. The application solves the problem that fixed time slice scheduling cannot adapt to dynamic changes of mixed workloads, and realizes the collaborative optimization of resource utilization, task response speed and system stability.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, and in particular relates to a method, apparatus and storage medium for scheduling GPU computing resources. Background Technology

[0002] With the rapid development of artificial intelligence and high-performance computing technologies, graphics processing units (GPUs), as core computing infrastructure, have a decisive impact on overall system performance due to their resource scheduling efficiency. In multi-user, multi-task shared cluster environments, existing scheduling strategies face significant challenges. Exclusive allocation, which fixes the entire GPU resource to a single task, is simple to implement but results in high resource idle rates, making it difficult to adapt to dynamically fluctuating workloads. Static hardware partitioning provides resource isolation through physical partitioning but sacrifices elastic scalability and easily leads to resource fragmentation. Queue-based scheduling, while ensuring the sequential execution of basic tasks, cannot effectively differentiate task priorities, often resulting in long-running tasks blocking urgent tasks and causing a decrease in system responsiveness. Time-slice round-robin scheduling, a common method to improve resource utilization, allows multiple tasks to share GPU resources through time-sharing, but its traditional implementation relies on fixed-length time slice parameters. This fixed configuration has inherent flaws: when the time slice is too long, the response latency of interactive tasks increases significantly, affecting user experience; when the time slice is too short, frequent context switching introduces additional performance overhead, reducing computational throughput. This contradiction is particularly pronounced in mixed workload scenarios, such as when long-duration training tasks coexist with short-duration inference tasks, where fixed time slices cannot adaptively balance response speed and execution efficiency. Although some studies have attempted to introduce task prioritization or preemption mechanisms, existing solutions generally lack the ability to dynamically perceive multi-dimensional runtime states. Scheduling decisions are often based solely on a single metric, such as throughput or response time, failing to comprehensively consider factors such as task computational characteristics, system load fluctuations, and hardware health status. This makes it difficult to coordinate and optimize resource utilization, task response speed, and system stability.

[0003] Therefore, there is an urgent need for an intelligent management mechanism that can perceive multi-dimensional operating status in real time and dynamically adjust scheduling parameters. Summary of the Invention

[0004] The purpose of this invention is to provide a GPU computing resource scheduling method, apparatus, and storage medium to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A GPU computing resource scheduling method, the method comprising: Based on task computation density and task memory requirements, the task computation feature coefficients are output by a computation feature profiling model. Based on the average response time and preemption overhead of the task, the performance efficiency coefficient is output through the performance evaluation model. Under the current task's calculated characteristic coefficients and performance efficiency coefficients, the load coordination factor is output through the load coordination model based on the ready queue length and context switching frequency. Based on changes in hardware temperature and cache hit rate, a hardware health coefficient is output through a hardware diagnostic model. Based on the length-frequency matching factor and hardware health coefficient, the time slice interval is adjusted in real time through an interval decision model.

[0006] As a further aspect of the present invention: the specific steps for real-time adjustment of the time slice interval based on the length-frequency matching factor and hardware health coefficient through the interval decision model are as follows: The load coordination factor and the hardware health coefficient are multiplied and fused to obtain a comprehensive control factor. Based on the comprehensive control factor, the final time slice interval is dynamically calculated and output between the minimum and maximum time slice intervals allowed by the system using a linear interpolation function.

[0007] As a further aspect of the present invention: the specific steps for outputting a hardware health coefficient based on hardware temperature status and cache hit rate changes through a hardware diagnostic model are as follows: The current hardware temperature is compared with the maximum safe temperature provided by the GPU manufacturer, and then the temperature index is obtained by using the min function to limit the upper limit to 1. The difference between the current cache hit rate and the baseline cache hit rate is compared with the baseline cache hit rate, and then the upper limit is limited to 1 using the min function to obtain the hit rate index. The temperature index and hit rate index are weighted and combined to output the hardware health coefficient. A decrease in the hardware health coefficient indicates that the hardware is in poor condition. The temperature index is inversely proportional to the hardware health coefficient, while the hit rate index is directly proportional to the hardware health coefficient.

[0008] As a further aspect of the present invention: the specific steps for outputting the load coordination factor through the load coordination model based on the ready queue length and context switching frequency, under the current task's calculated characteristic coefficients and performance efficiency coefficients, are as follows: The ratio of the current ready queue length to the system peak queue length is processed, and then the upper limit is limited to 1 using the min function to obtain the ready queue length exponent. The current context switching frequency is compared with the system peak switching frequency, and then the upper limit is limited to 1 using the min function to obtain the context switching frequency index. Based on the task-calculated characteristic coefficients and performance efficiency coefficients, dynamic adjustment coefficients are obtained. The task computation characteristic coefficient is directly proportional to the dynamic adjustment coefficient, while the performance efficiency coefficient is inversely proportional to the dynamic adjustment coefficient. Import the ready queue length index, context switching frequency index, and dynamic adjustment coefficient into the formula. Obtain the load coordination factor ,in, The index of the ready queue length. This is a context switching frequency index. This is the dynamic adjustment coefficient.

[0009] As a further aspect of the present invention: the specific steps for outputting task computation feature coefficients by calculating the feature profiling model based on task computation density and task memory requirements are as follows: The task computation density index is obtained by comparing the difference between the current task computation density and the minimum task computation density with the difference between the maximum task computation density and the minimum task computation density. The task memory requirement index is obtained by comparing the difference between the current task memory requirement and the minimum task memory requirement with the difference between the maximum task memory requirement and the minimum task memory requirement. The task computation density index and the task memory requirement index are weighted and fused to obtain the task computation characteristic coefficient. Both the task computation density index and the task memory requirement index are proportional to the task computation characteristic coefficient.

[0010] As a further aspect of the present invention: the specific steps for outputting the performance efficiency coefficient based on the average task response time and preemption overhead ratio through the performance evaluation model are as follows: The average task response time is compared with the system baseline response time, and then the upper limit of the min function is set to 1 to obtain the exponent of the average task response time. The preemption overhead ratio is the ratio of context switching time to task execution time; After weighting and fusing the task average response time index and the preemption cost ratio, the difference between 1 and the weighted fusion value is taken to obtain the performance efficiency coefficient. Both the task average response time index and the preemption cost ratio are inversely proportional to the task's calculated characteristic coefficient.

[0011] The present invention also provides a GPU computing resource scheduling device, the device comprising: The computational feature profiling module outputs task computational feature coefficients based on task computation density and task memory requirements through the computational feature profiling model. The performance evaluation module outputs a performance efficiency coefficient based on the average task response time and preemption overhead ratio through a performance evaluation model. The load coordination module, based on the current task's calculated characteristic coefficients and performance efficiency coefficients, outputs a load coordination factor through the load coordination model, taking into account the ready queue length and context switching frequency. The hardware health assessment module outputs a hardware health coefficient based on hardware temperature status and cache hit rate changes through a hardware diagnostic model. The interval decision module adjusts the time slice interval in real time based on the length-frequency matching factor and the hardware health coefficient through the interval decision model.

[0012] The present invention also provides a storage medium storing a computer program. The computer-readable storage medium stores a computer program, and the computer program is executed by a processor to perform the steps of the GPU computing resource scheduling method.

[0013] Compared with the prior art, the beneficial effects of the present invention are: This invention achieves adaptive fine-grained scheduling: by comprehensively sensing task computation characteristics, system performance status, load pressure and hardware health status, the time slice interval is dynamically adjusted, overcoming the insufficient adaptability of fixed time slice parameters under mixed workloads, and effectively balancing the contradiction between task response latency and context switching overhead.

[0014] Improve overall system performance: The dynamic adjustment scheduling strategy based on load coordination factor can effectively prevent ready queue backlog and excessive context switching, and significantly improve GPU resource utilization and system throughput while ensuring task responsiveness.

[0015] Enhance system stability: By monitoring GPU temperature and cache performance in real time through hardware health coefficients, scheduling decisions are ensured to stay within hardware safety boundaries, effectively preventing system failures caused by overheating or performance degradation and extending hardware lifespan.

[0016] Optimize energy efficiency: The adaptive time slice adjustment mechanism avoids unnecessary frequent switching and resource idleness, reducing the overall energy consumption of the system while meeting performance goals, and is particularly suitable for long-term operation scenarios of large-scale GPU clusters. Attached Figure Description

[0017] Figure 1 The overall flowchart of the GPU computing resource scheduling method is shown.

[0018] Figure 2 A structural diagram of the GPU computing resource scheduling device is shown. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0021] like Figure 1 As shown, an embodiment of the present invention provides a GPU computing resource scheduling method, the method comprising: Step S100: Based on the task's computational density and memory requirements, output the task's computational feature coefficients by calculating the feature profile model; Task computation characteristic coefficients can be understood as indicators used to quantify the computational intensity and memory usage characteristics of a task, and they can be implemented in various ways. For example, a linear regression model can be built based on the number of instructions executed and the proportion of memory allocation for the task, or cluster analysis can be performed using historical task execution data to generate characteristic coefficients. These coefficients are mainly used to provide differentiated criteria at the task level, enabling the system to identify tasks with high computational density or high memory requirements.

[0022] Step S200: Based on the average response time and preemption overhead ratio of the task, output the performance efficiency coefficient through the performance evaluation model; The performance efficiency coefficient can be understood as an indicator used to comprehensively evaluate the efficiency of task execution, and it can be implemented in various ways. For example, an evaluation function can be constructed based on a weighted combination of task throughput and latency, or the efficiency coefficient can be generated by normalizing the ratio of task completion time to resource occupancy time. This coefficient is mainly used to reflect the actual impact of scheduling strategies on system throughput and latency, providing a performance benchmark for dynamic adjustments.

[0023] Step S300: With the characteristic coefficients and performance efficiency coefficients of the current task calculated, output the load coordination factor based on the ready queue length and context switching frequency through the load coordination model; The load balancing factor can be understood as an indicator reflecting the degree of load balancing in a system, and it can be implemented in various ways. For example, a load assessment model can be constructed based on a weighted sum of task priorities and queue waiting times, or a graph structure analysis can be performed on the resource competition relationships between tasks to generate the balancing factor. This factor is mainly used to coordinate the competition relationships between multiple tasks and prevent queue backlog or excessive switching.

[0024] Step S400: Based on the hardware temperature status and cache hit rate changes, output the hardware health coefficient through the hardware diagnostic model; Hardware health coefficients can be understood as indicators used to monitor the physical state of a GPU, and they can be implemented in various ways. For example, a health assessment model can be built based on the correlation analysis between hardware power consumption and thermal efficiency, or a health coefficient can be generated by predicting the statistical relationship between hardware runtime and failure rate. This coefficient is mainly used to ensure that scheduling decisions do not exceed hardware safety boundaries and maintain long-term operational stability.

[0025] Step S500: Based on the length-frequency matching factor and hardware health coefficient, the time slice interval is adjusted in real time through the interval decision model; By employing a multi-dimensional real-time perception and dynamic adjustment mechanism, this paper addresses the inherent shortcomings of fixed-time-slice scheduling. Specifically, traditional methods typically rely on single-dimensional indicators for scheduling decisions, making it difficult to achieve an effective balance between resource utilization, task response speed, and system stability. This application, however, constructs an interconnected chain of evaluation models, integrating task characteristics, system performance, load status, and hardware health into a unified decision-making basis. This avoids the limitations of single-dimensional scheduling, thereby enabling adaptive adjustment of the time-slice interval. This mechanism can adaptively balance response speed, resource utilization, and system stability under mixed workloads, meeting the complex and ever-changing needs of practical applications.

[0026] This approach achieves dynamic adjustment of time slice intervals through multiple interrelated evaluation models, thereby resolving the contradiction between task response latency and context switching overhead caused by fixed-time-slice round-robin methods. Specifically, firstly, based on task computation density and memory requirements, a computational feature profiling model outputs task computational feature coefficients. These coefficients quantify the computational intensity and memory usage characteristics of tasks, providing a differentiated basis for subsequent scheduling and avoiding mismatches between resource allocation and task requirements. Furthermore, based on average task response time and preemption overhead ratios, a performance evaluation model outputs performance efficiency coefficients. These coefficients comprehensively reflect task execution efficiency and, through the trade-off between response time and switching overhead, demonstrate the actual impact of scheduling strategies on system throughput and latency, providing a performance benchmark for dynamic adjustments.

[0027] Building upon this foundation, and considering the current task's characteristic coefficients and performance efficiency coefficients, a load coordination factor is output through a load coordination model, taking into account the ready queue length and context switching frequency. This factor integrates queue status and switching frequency based on task characteristics, outputting an indicator reflecting the system's load balancing level, coordinating multi-task competition, and preventing queue backlog or excessive switching. Simultaneously, based on hardware temperature status and cache hit rate changes, a hardware diagnostic model outputs a hardware health coefficient. This coefficient monitors the GPU's physical state, and through correlation analysis between temperature and cache performance, ensures that scheduling decisions do not exceed hardware safety boundaries, maintaining long-term operational stability.

[0028] Finally, based on the length-frequency matching factor and hardware health coefficient, the time slice interval is adjusted in real time through an interval decision model. This process utilizes the synergy between the load coordination factor and the hardware health coefficient to dynamically calculate the optimal time slice length, avoiding response delays caused by excessively long time slices or switching overhead caused by excessively short time slices, thereby achieving precise matching between scheduling parameters and runtime state. Thus, the above technical features form a closed loop: task computation characteristics and performance efficiency coefficients provide context for load coordination, and the load coordination factor and hardware health coefficient jointly input the interval decision model, enabling time slice adjustment to simultaneously consider task requirements, system load, and hardware reliability, thereby adaptively balancing response speed, resource utilization, and system stability under mixed workloads.

[0029] As a preferred embodiment of the technical solution of the present invention, the specific steps of real-time adjustment of the time slice interval based on the length-frequency matching factor and hardware health coefficient through the interval decision model are as follows: The load coordination factor and the hardware health coefficient are multiplied and fused to obtain a comprehensive control factor. This factor comprehensively reflects the instantaneous load pressure and long-term hardware reliability of the system, and its value range is 0-1; The comprehensive control factor refers to the value obtained by combining the load coordination factor, which reflects the system load status, and the hardware health coefficient, which reflects the hardware health status, through mathematical operations. It can be implemented by multiplication operations. The purpose is to use the coupling characteristics of multiplication operations to amplify negative effects and ensure that the system prioritizes stability under the dual pressure of load and hardware.

[0030] Based on the comprehensive control factor, the final time slice interval is dynamically calculated and output between the minimum and maximum time slice intervals allowed by the system using a linear interpolation function. In practical applications, the linear interpolation function can be understood as a mathematical mapping tool that can scale proportionally between the preset minimum time slice interval and the maximum time slice interval according to the comprehensive control factor. The purpose is to make the time slice length transition smoothly with the system state and adapt to the instantaneous changes in sudden load and hardware temperature rise in mixed workloads.

[0031] The linear interpolation function is: ,in, For the final time slice interval, This is the minimum time slice interval allowed by the system. This is the maximum time slice interval allowed by the system. As a comprehensive regulatory factor; Used to ensure basic system responsiveness and prevent task starvation. Used to limit the maximum continuous execution time of a single task, controlling the upper limit of context switching latency.

[0032] This paper addresses the lack of a refined mechanism for integrating load pressure and hardware health status by constructing an adaptive time-slice control mechanism. First, a comprehensive control factor is obtained by multiplying and fusing the load coordination factor and the hardware health coefficient. When the load coordination factor reflects high load pressure or the hardware health coefficient indicates deterioration of the device status, the multiplication amplifies the negative effects, causing the comprehensive control factor to decay rapidly, avoiding the indicator cancellation problem that may occur with simple weighted averaging. Second, the time-slice interval is dynamically calculated based on the comprehensive control factor using a linear interpolation function. This design abandons the coarse strategy of fixed threshold switching, allowing the time slice length to smoothly transition with the system state. The minimum time-slice interval is set based on the rigid requirement of avoiding task starvation, ensuring timely response of interactive tasks by guaranteeing a minimum execution window. The maximum time-slice interval addresses the inherent overhead of context switching, preventing long tasks from monopolizing resources by limiting the longest execution time of a single task. Together, they constitute the safety boundary of the control, preventing performance loss caused by excessively short time slices and response latency caused by excessively long time slices. Based on this, the above scheme, combined with the aforementioned GPU computing resource scheduling method, can better achieve synergistic optimization of resource utilization, task response speed, and system stability in complex and ever-changing real-world application environments.

[0033] Regarding step S400, the specific steps for outputting the hardware health coefficient based on hardware temperature status and cache hit rate changes through the hardware diagnostic model are as follows: The current hardware temperature is compared with the maximum safe temperature provided by the GPU manufacturer, and then the temperature index is obtained by using the min function to limit the upper limit to 1. The temperature index is a normalized indicator generated by comparing the current hardware temperature with the maximum safe temperature. It can be implemented using hardware devices such as thermistors and digital temperature sensors in conjunction with software algorithms. Its purpose is to quantify temperature risk in a standardized way, avoiding assessment biases caused by differences in temperature units.

[0034] The difference between the current cache hit rate and the baseline cache hit rate is compared with the baseline cache hit rate, and then the upper limit is limited to 1 using the min function to obtain the hit rate index. The baseline cache hit rate is dynamically calculated by the exponentially weighted moving average of historical cache hit rates. To avoid the denominator being 0 when the baseline cache hit rate is 0, the max function is used to limit the denominator to a very small constant when the baseline cache hit rate is 0. The hit rate index is a metric generated based on the ratio of changes in cache hit rate to a baseline hit rate. It can be implemented using a Performance Monitoring Unit (PMU) or dedicated cache analysis tools. Its purpose is to accurately capture real-time fluctuations in cache performance while ensuring stable output of valid data even under extreme conditions.

[0035] The temperature index and hit rate index are weighted and combined to output a hardware health coefficient ranging from 0 to 1. A decrease in the hardware health coefficient indicates poor hardware condition. The temperature index is inversely proportional to the hardware health coefficient, while the hit rate index is directly proportional to the hardware health coefficient. The hardware health coefficient is a comprehensive indicator used to quantify the operating status of GPU hardware. It can be implemented in various ways, such as by combining data collected by temperature sensors with cache performance statistics for multi-dimensional evaluation. The purpose of introducing this coefficient is to provide a reliable basis for reflecting the hardware health status in real time, thereby laying the foundation for dynamic adjustment of time slice intervals.

[0036] The weighted fusion formula for the temperature index and hit rate index is as follows: ,in, For hardware health coefficient, Temperature index The hit rate index and All are weighting coefficients. and All are greater than or equal to 0 and less than or equal to 1. One way to obtain the weighting coefficient is to sum the temperature index and the hit rate index, and then take the ratio of the temperature index to the sum to obtain the weighting coefficient. The hit rate index is then compared with the sum of the above values ​​to obtain the weighting coefficient. .

[0037] First, the current hardware temperature is collected using a temperature sensor and compared to the maximum safe temperature provided by the GPU manufacturer to obtain a preliminary temperature assessment value. Considering that extreme conditions may lead to abnormal values, a min function is used to limit the result, ensuring the temperature index is strictly controlled within the 0-1 range. For cache hit rate evaluation, the difference between the current cache hit rate and the baseline cache hit rate is calculated and compared to the baseline value to obtain a preliminary hit rate assessment value. To avoid the denominator being zero, a max function is used to handle the baseline cache hit rate at the boundary; when it is zero, it is automatically replaced with a very small constant. The baseline cache hit rate is dynamically calculated using an exponentially weighted moving average of historical data, giving higher weight to recent data, allowing the evaluation results to quickly respond to workload changes.

[0038] Building upon this, the temperature index is weighted and fused with the hit rate index in a 1-tn format to construct the final hardware health coefficient. This design ensures that the temperature index and hardware health coefficient are inversely proportional, directly related to the risk of hardware overheating; while the hit rate index is directly proportional to the hardware health coefficient, reflecting performance health. The weighting coefficients are dynamically adjusted based on the relative magnitude of the two. When there is an abnormal surge in temperature, the weight of the temperature index is automatically increased to prioritize avoiding thermal risks; when cache performance drops sharply, the weight of the hit rate index is increased to quickly respond to performance degradation. The final output is a hardware health coefficient in the range of 0-1, providing a continuous and reliable quantitative basis for time-slice control.

[0039] The above technical solution enables multi-dimensional dynamic evaluation of GPU hardware status, effectively solving the problems of incomplete hardware status monitoring and computational instability. Especially in mixed workload scenarios, it can promptly detect subtle changes in hardware status and adjust resource allocation strategies accordingly, avoiding hardware failures or performance crashes caused by delayed evaluation.

[0040] As a preferred embodiment of the present invention, the specific steps of outputting the load coordination factor through the load coordination model based on the ready queue length and context switching frequency, under the current task's calculated characteristic coefficients and performance efficiency coefficients, are as follows: The ratio of the current ready queue length to the system peak queue length is processed, and then the upper limit is limited to 1 using the min function to obtain the ready queue length exponent. The ready queue length index is a normalized ratio of the current ready queue length to the system peak queue length. It can be achieved by dividing the current ready queue length by the system peak queue length and using the min function to limit the upper limit to 1. The purpose is to eliminate the influence of absolute values ​​on the evaluation and make the queue lengths comparable under different hardware scales.

[0041] The current context switching frequency is compared with the system peak switching frequency, and then the upper limit is limited to 1 using the min function to obtain the context switching frequency index. The context switching frequency index can be understood as the normalized ratio of the current context switching frequency to the system peak switching frequency. It can be achieved by dividing the current context switching frequency by the system peak switching frequency and using the min function to limit the upper limit to 1. The purpose is to provide a consistent benchmark for evaluating switching behavior.

[0042] The system peak queue length and switching frequency are dynamically updated using the maximum value of a sliding window of historical observations. The dynamic adjustment coefficient is obtained by comparing the task calculation characteristic coefficient with the performance efficiency coefficient. The task computation characteristic coefficient is directly proportional to the dynamic adjustment coefficient, while the performance efficiency coefficient is inversely proportional to the dynamic adjustment coefficient. The dynamic adjustment coefficient is the ratio of the task calculation characteristic coefficient to the performance efficiency coefficient. It can be achieved through simple mathematical ratio calculations and aims to dynamically adjust the focus of load assessment based on task characteristics and system health. The load coordination factor is a comprehensive indicator calculated using a specific formula based on the ready queue length index, context switching frequency index, and dynamic adjustment coefficient. Its purpose is to quantify the degree of coordination of system load, thereby providing a high-fidelity load indication for time-slice interval control.

[0043] Import the ready queue length index, context switching frequency index, and dynamic adjustment coefficient into the formula. Obtain the load coordination factor with an output range of 0-1. A higher load coordination factor indicates a better system load condition. The index of the ready queue length. This is a context switching frequency index. As a dynamic adjustment coefficient, in the above ratio processing, to avoid the denominator being 0, the denominator term is added to a very small positive number.

[0044] First, normalization is used to convert the ready queue length and context switching frequency into proportional values ​​relative to system capacity, making observational data comparable across different hardware scales and load intensities. Based on this, the system's peak queue length and switching frequency are dynamically updated using a sliding window maximum value, ensuring that the evaluation benchmark can be adjusted in real time according to the system's operating status, avoiding the lag problem caused by fixed thresholds. The introduction of a dynamic adjustment coefficient establishes a coupling mechanism between task characteristics and system performance. When the task's computational characteristic coefficient is high, the dynamic adjustment coefficient increases, making the load assessment focus more on queue backlog; conversely, when the performance efficiency coefficient is high, the dynamic adjustment coefficient decreases, reducing the weight of switching overhead. Finally, the above indices and coefficients are integrated into a load coordination factor using a specific formula. This formula constructs a non-linear relationship using the ratio of the absolute difference to the sum. When the queue length index matches the adjusted switching frequency index, the factor approaches 1, indicating optimal load coordination; a mismatch reflects an imbalance. Furthermore, the existence of the dynamic adjustment coefficient allows the matching point to automatically shift with task characteristics, thus adapting to the load characteristics of different types of tasks. Overall, these features work together to solve the problem of inaccurate evaluation caused by static weight settings in mixed workload scenarios, and provide a reliable load indication basis for time slice interval adjustment.

[0045] Based on the task calculation characteristic coefficients and performance efficiency coefficients, and combined with the ready queue length and context switching frequency, a dynamic adjustment mechanism was constructed to achieve adaptive generation of load coordination factors. This effectively solved the problem of inaccurate evaluation under mixed workloads and provided accurate load indication for time slice interval control.

[0046] As a preferred embodiment of the present invention, the specific steps for outputting task computation feature coefficients by calculating the feature profiling model based on task computation density and task memory requirements are as follows: The task computation density index is obtained by comparing the difference between the current task computation density and the minimum task computation density with the difference between the maximum task computation density and the minimum task computation density. Task computation density is the product of the number of floating-point operations performed by the task per unit time and the GPU computing unit utilization. FLOPs are obtained through GPU performance counters, and GPU utilization is monitored in real time through the driver interface. A sliding window average is used for long-running tasks, while instantaneous values ​​are used for short-running tasks. The maximum and minimum task computation density are the historically observed maximum and minimum values. The task computation density index refers to the normalized result of the current task computation density relative to historical observation extremes. It can be calculated by taking the ratio of the difference between the current task computation density and the minimum value to the difference between the maximum and minimum values. The purpose of introducing this index is to avoid the influence of absolute value bias on the representation of task characteristics and to ensure that the historical proportion of task computation intensity can be dynamically reflected under different workload scales.

[0047] The task memory requirement index is obtained by comparing the difference between the current task memory requirement and the minimum task memory requirement with the difference between the maximum task memory requirement and the minimum task memory requirement. The video memory requirement is the peak video memory usage during task execution, including: video memory usage for model parameters, video memory for intermediate activation values, and video memory for the working buffer. The maximum and minimum video memory requirements for the task are the historically observed maximum and minimum values. The task memory requirement index can be understood as a normalized result of the current task memory requirement relative to historical extreme values, and it can be achieved through a normalization method similar to that of the task computation density index. The introduction of this index compensates for the limitations of a single computation density index, enabling task characteristics to simultaneously reflect the differences between computationally intensive and memory-intensive tasks.

[0048] The task computation density index and the task memory requirement index are weighted and fused to obtain the task computation feature coefficient. The output range of the task computation feature coefficient is 0-1. Both the task computation density index and the task memory requirement index are proportional to the task computation feature coefficient. The weighted fusion formula is as follows: ,in, Calculate the characteristic coefficients for the task. Calculate the density index for the task. This represents the memory requirement index for the task. and All are weighting coefficients. and All are greater than or equal to 0 and less than or equal to 1. .

[0049] One method to obtain the weighting coefficient is to sum the task computation density index and the task memory requirement index, and then take the ratio of the task computation density index to the sum to obtain the weighting coefficient. The weighting coefficient is obtained by comparing the task's memory requirement index with the sum of the above values. The weighting coefficient is a fusion parameter dynamically generated based on the real-time proportional relationship between the task's computational density index and its memory requirement index. It can be calculated by comparing each index with the sum. The purpose of this mechanism is to enable the weighted fusion process to automatically respond to real-time changes in task characteristics, ensuring that the fusion result better reflects the actual resource bottlenecks of the task.

[0050] A dual-dimensional feature fusion mechanism was employed to dynamically construct task computation feature coefficients. First, based on historical data-driven normalization, a task computation density index and a task memory requirement index were generated separately. This not only resolved the absolute value bias issue but also improved the timeliness of feature extraction by continuously updating historical extreme values. Second, a dynamic weight generation mechanism was used to weight and fuse the two indices according to their real-time proportional relationship. This adaptive fusion method accurately reflects the comprehensive resource requirements of the task. Based on this, the output standardized task computation feature coefficients provided a unified input for the subsequent load coordination model, effectively avoiding interference from features of different magnitudes on scheduling decisions.

[0051] Considering that task computational characteristic coefficients need to serve as important inputs to the load coordination model, this scheme constructs a more comprehensive task resource requirement representation system by integrating the dual-dimensional features of computational density and memory requirements. This multi-dimensional feature collaborative expression not only solves the problem of one-sided task feature representation but also significantly improves the balance between resource utilization and response speed in the scheduling system. Especially in mixed workload scenarios, it can accurately capture the resource consumption characteristics of different types of tasks, providing a reliable basis for intelligent scheduling decisions.

[0052] As a preferred embodiment of the present invention, the specific steps for outputting the performance efficiency coefficient based on the average task response time and preemption overhead ratio through the performance evaluation model are as follows: The average task response time is compared with the system baseline response time, and then the upper limit of the min function is set to 1 to obtain the exponent of the average task response time. Context switching time includes GPU register state saving / restoration time, memory context switching overhead, and command buffer flushing time. System baseline response time is dynamically calculated using an exponentially weighted moving average of historical average response times, giving higher weight to recent data. The task average response time index is the ratio of the task's average response time to the system baseline response time. It can be achieved through a simple division operation and aims to quantify the deviation of the current task response time from the system baseline. The system baseline response time can be understood as a dynamically updated reference value based on an exponentially weighted moving average of historical data. Its sensitivity to recent data can be adjusted by changing the weighting factors, so that the baseline value can quickly adapt to changes in system load trends.

[0053] The preemption overhead ratio is the ratio of context switching time to task execution time; Preemption overhead can be defined as the ratio of context switching time to task execution time. It can be calculated by monitoring context switching time and task execution time in real time, with the aim of measuring the impact of task switching on overall performance.

[0054] After weighted fusion of the task average response time index and the preemption cost ratio, the difference between 1 and the weighted fusion value is taken to obtain the performance efficiency coefficient. The output range of the performance efficiency coefficient is 0-1. Both the task average response time index and the preemption cost ratio are inversely proportional to the task's calculated characteristic coefficient. The performance efficiency coefficient is the result of subtracting 1 from the weighted average response time index of the task and the preemption overhead ratio. It can be achieved by a linear weighting formula, and its purpose is to comprehensively reflect the balance between task response speed and context switching overhead.

[0055] The formula for calculating the performance efficiency coefficient is: ,in, The performance efficiency coefficient. This is the index of average task response time. In order to seize the proportion of expenses, and All are weighting coefficients. and All are greater than or equal to 0 and less than or equal to 1. .

[0056] One method to obtain the weighting coefficient is to sum the average task response time index and the preemption cost ratio, and then compare the average task response time index with this sum to obtain the weighting coefficient. The weighting coefficient is obtained by comparing the percentage of preemptive expenditure with the sum of the above values. .

[0057] First, the average task response time index is generated by processing the ratio of the average task response time to the system baseline response time. A min function is then used to limit the index value, ensuring it remains within a reasonable range and preventing outliers from interfering with subsequent calculations. The system baseline response time employs an exponentially weighted moving average dynamic update mechanism, which can sensitively capture system performance change trends, thus providing an accurate reference benchmark for the response time index. The preemption overhead ratio is calculated by the ratio of context switching time to task execution time, achieving task-related normalization of overhead metrics and effectively distinguishing between the high proportion of normal short tasks and true performance bottlenecks. Subsequently, the average task response time index and the preemption overhead ratio are weighted and fused. A dynamic weight allocation mechanism ensures that the two remain relatively balanced under mixed workloads, preventing a single indicator from dominating the evaluation results. Finally, the performance efficiency coefficient is obtained by taking the difference between 1 and the weighted fusion value. Its standardized 0-1 output range facilitates direct integration into subsequent models and intuitively reflects the positive effects of reduced response time and lower overhead.

[0058] This solution addresses the issue of inaccurate performance efficiency quantification during scheduling, achieving a dynamic balance assessment between task response speed and context switching overhead, thus improving the accuracy of resource scheduling decisions and the overall system performance optimization capability. Building upon this, the solution, combined with the aforementioned task calculation characteristic coefficients and load coordination factors, forms a complete system for multi-dimensional operational status awareness and adaptive scheduling parameter adjustment, effectively addressing resource management challenges in complex and ever-changing real-world application environments.

[0059] As a preferred embodiment of the technical solution of the present invention, the present invention also discloses a GPU computing power resource scheduling device, the device 10 comprising: The computational feature profiling module 11 outputs task computational feature coefficients based on task computation density and task memory requirements through the computational feature profiling model. Performance evaluation module 12 outputs performance efficiency coefficients based on the average response time and preemption overhead ratio of tasks through the performance evaluation model; Load coordination module 13, under the current task's calculated characteristic coefficients and performance efficiency coefficients, outputs load coordination factors based on the ready queue length and context switching frequency through the load coordination model; Hardware health assessment module 14 outputs hardware health coefficients based on hardware temperature status and cache hit rate changes through a hardware diagnostic model. The interval decision module 15 adjusts the time slice interval in real time based on the length-frequency matching factor and the hardware health coefficient through the interval decision model.

[0060] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A GPU computing resource scheduling method, characterized in that, The method includes: Based on task computation density and task memory requirements, the task computation feature coefficients are output by a computation feature profiling model. Based on the average response time and preemption overhead of the task, the performance efficiency coefficient is output through the performance evaluation model. Under the current task's calculated characteristic coefficients and performance efficiency coefficients, the load coordination factor is output through the load coordination model based on the ready queue length and context switching frequency. Based on changes in hardware temperature and cache hit rate, a hardware health coefficient is output through a hardware diagnostic model. Based on the length-frequency matching factor and hardware health coefficient, the time slice interval is adjusted in real time through an interval decision model.

2. The GPU computing resource scheduling method according to claim 1, characterized in that, The specific steps for real-time adjustment of the time slice interval based on the length-frequency matching factor and hardware health coefficient using the interval decision model are as follows: The load coordination factor and the hardware health coefficient are multiplied and fused to obtain a comprehensive control factor. Based on the comprehensive control factor, the final time slice interval is dynamically calculated and output between the minimum and maximum time slice intervals allowed by the system using a linear interpolation function.

3. The GPU computing resource scheduling method according to claim 2, characterized in that, The specific steps for outputting the hardware health coefficient based on hardware temperature status and cache hit rate changes through a hardware diagnostic model are as follows: The current hardware temperature is compared with the maximum safe temperature provided by the GPU manufacturer, and then the temperature index is obtained by using the min function to limit the upper limit to 1. The difference between the current cache hit rate and the baseline cache hit rate is compared with the baseline cache hit rate, and then the upper limit is limited to 1 using the min function to obtain the hit rate index. The temperature index and hit rate index are weighted and combined to output the hardware health coefficient. A decrease in the hardware health coefficient indicates that the hardware is in poor condition. The temperature index is inversely proportional to the hardware health coefficient, while the hit rate index is directly proportional to the hardware health coefficient.

4. The GPU computing resource scheduling method according to claim 3, characterized in that, The specific steps for outputting task computation feature coefficients by calculating the feature profiling model based on task computation density and task memory requirements are as follows: The task computation density index is obtained by comparing the difference between the current task computation density and the minimum task computation density with the difference between the maximum task computation density and the minimum task computation density. The task memory requirement index is obtained by comparing the difference between the current task memory requirement and the minimum task memory requirement with the difference between the maximum task memory requirement and the minimum task memory requirement. The task computation density index and the task memory requirement index are weighted and fused to obtain the task computation characteristic coefficient. Both the task computation density index and the task memory requirement index are proportional to the task computation characteristic coefficient.

5. The GPU computing resource scheduling method according to claim 4, characterized in that, The specific steps for outputting the performance efficiency coefficient based on the average task response time and preemption overhead ratio through the performance evaluation model are as follows: The average task response time is compared with the system baseline response time, and then the upper limit of the min function is set to 1 to obtain the exponent of the average task response time. The preemption overhead ratio is the ratio of context switching time to task execution time; After weighting and fusing the task average response time index and the preemption cost ratio, the difference between 1 and the weighted fusion value is taken to obtain the performance efficiency coefficient. Both the task average response time index and the preemption cost ratio are inversely proportional to the task's calculated characteristic coefficient.

6. A GPU computing resource scheduling device, characterized in that, The device includes: The computational feature profiling module outputs task computational feature coefficients based on task computation density and task memory requirements through the computational feature profiling model. The performance evaluation module outputs a performance efficiency coefficient based on the average task response time and preemption overhead ratio through a performance evaluation model. The load coordination module, based on the current task's calculated characteristic coefficients and performance efficiency coefficients, outputs a load coordination factor through the load coordination model, taking into account the ready queue length and context switching frequency. The hardware health assessment module outputs a hardware health coefficient based on hardware temperature status and cache hit rate changes through a hardware diagnostic model. The interval decision module adjusts the time slice interval in real time based on the length-frequency matching factor and the hardware health coefficient through the interval decision model.

7. A storage medium, characterized in that, The storage medium stores a computer program, and the computer program stored on the computer-readable storage medium executes the steps of the GPU computing resource scheduling method as described in any one of claims 1 to 5 when the processor runs the computer program.

Images