An automatic aggregation scheduling method for improving GPU computing power utilization rate

Abstract

The present application relates to the technical field of data processing, and more particularly to an automatic aggregation scheduling method for improving GPU computing power utilization rate: GPU running data and a computing task set are collected, resource information library is constructed by processing the running data, standard task features are obtained by extracting the computing task set features, and a task queue is generated by classification; the local computing power pool is formed by aggregating the GPU fragments meeting the conditions based on the resource information library, and the scheduling mode is determined according to the comparison result of the total computing power of the computing power pool and the task demand; the matching accuracy of the task queue and the local computing power pool is obtained, and the effectiveness of the improvement of the GPU computing power utilization rate is determined. If it is invalid, it is determined whether to increase the GPU fragment elastic redundancy threshold; if not, the running data loss rate is obtained to determine the data collection integrity; if it is not complete, it is determined whether to reduce the node resource load adaptation factor; if not, the GPU fragment architecture fluctuation suppression threshold is determined based on the computing power fluctuation amplitude. The present application improves the effectiveness of the improvement of the GPU computing power utilization rate.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an automatic aggregation scheduling method for improving GPU computing power utilization. Background Technology

[0002] In existing technologies, heterogeneous GPU clusters have become the core infrastructure supporting large-scale computing power demands. However, existing GPU computing power aggregation and scheduling technologies still have limitations: on the one hand, GPU load has the characteristics of rapid dynamic changes, cross-node data synchronization is prone to delays and deviations, and network jitter and hardware interruptions at edge nodes can easily lead to data loss or corruption. The original monitoring data can also be affected by noise introduced by host scheduling and instance interference, resulting in scheduling decisions based on low-quality data and causing misjudgments in resource allocation; on the other hand, existing scheduling strategies lack quantitative control over data quality and dynamic parameter optimization mechanisms, resulting in insufficient effectiveness in improving GPU computing power utilization.

[0003] Chinese Patent Publication No. CN119917224A discloses a cloud platform AI computing power scheduling method, apparatus, and device. The method includes: after receiving a virtual machine migration instruction, extracting the target model, target slot, and target quantity from the VGPU information of the source virtual machine. The target model is the model of the GPU that provides VGPUs to the source virtual machine, the target slot is the slot where the GPU that provides VGPUs to the source virtual machine is located, and the target quantity is the number of VGPUs provided by the GPU in the target slot to the source virtual machine. The target slot and the target quantity correspond one-to-one.

[0004] The method involves identifying candidate computing nodes from other computing nodes, where each candidate node's GPU model in each target slot is the target model. It then checks if a first computing node exists among the candidate nodes, where the first computing node's idle VGPU count in each target slot is greater than or equal to the corresponding target count. If a first computing node exists, the target computing node is determined from among these first computing nodes. However, the aforementioned cloud platform AI computing power scheduling method, device, and equipment suffer from several problems. Firstly, due to the lack of preprocessing of GPU monitoring data, direct node selection is susceptible to short-term noise interference. Secondly, the absence of a resource fragment aggregation mechanism prevents the integration of discrete GPU fragments, leading to resource waste and biased optimal node selection, ultimately resulting in insufficient effectiveness in improving GPU computing power utilization. Summary of the Invention

[0005] To address this, the present invention provides an automatic aggregation scheduling method to improve GPU computing power utilization. This method overcomes the problems in the prior art where, due to the lack of preprocessing of GPU monitoring data, direct node selection is susceptible to short-term noise interference, the absence of a resource fragment aggregation mechanism, and the inability to integrate discrete GPU fragments, resulting in resource waste and bias in the optimal node judgment, thus leading to insufficient effectiveness in improving GPU computing power utilization.

[0006] To achieve the above objectives, the present invention provides an automatic aggregation scheduling method for improving GPU computing power utilization, comprising:

[0007] GPU running data and computing task sets are collected separately. The running data is processed to obtain a resource information database. The computing task sets are feature extracted to obtain standard task features. The computing task sets are classified based on the standard task features to obtain task queues.

[0008] Based on the resource information database, GPU fragments that meet the preset conditions in each node are aggregated to obtain a local computing power pool. The scheduling method of the task queue is determined based on the comparison result between the total computing power of the local computing power pool and the computing power requirement of the task.

[0009] Obtain the matching accuracy between the task queue and the local computing power pool, and determine whether the effectiveness of improving GPU computing power utilization meets the requirements based on the matching accuracy between the task queue and the local computing power pool.

[0010] If the effectiveness of improving GPU computing power utilization does not meet the requirements, then it is determined whether it is necessary to increase the elastic redundancy threshold of GPU fragments.

[0011] If it is not necessary to increase the elastic redundancy threshold of GPU fragments, then obtain the missing rate of running data per unit time to determine whether the completeness of the running data collection meets the requirements.

[0012] If the completeness of the collected operational data does not meet the requirements, then determine whether it is necessary to reduce the node resource load adaptation factor.

[0013] If it is not necessary to reduce the node resource load adaptation factor, the architecture fluctuation suppression threshold of GPU fragments is determined based on the fluctuation range of the local computing power pool per unit time.

[0014] Furthermore, the effectiveness of improving GPU computing power utilization is determined based on the matching accuracy between the task queue and the local computing power pool, including:

[0015] The matching accuracy of the task queue with the local computing power pool is compared with the preset second accuracy.

[0016] If the matching accuracy between the task queue and the local computing power pool is greater than or equal to the preset second accuracy, then the effectiveness of improving GPU computing power utilization is determined to meet the requirements.

[0017] If the matching accuracy between the task queue and the local computing power pool is less than the preset second accuracy, then the effectiveness of improving GPU computing power utilization is determined to be unsatisfactory.

[0018] Further, determine whether it is necessary to increase the elastic redundancy threshold for GPU fragmentation, including:

[0019] The matching accuracy of the task queue with the local computing power pool is compared with the preset first accuracy and the preset second accuracy, respectively.

[0020] If the matching accuracy between the task queue and the local computing power pool is less than or equal to the preset first accuracy, then it is determined that the elastic redundancy threshold of GPU fragments needs to be increased.

[0021] If the matching accuracy between the task queue and the local computing power pool is greater than a preset first accuracy and less than a preset second accuracy, then it is determined that there is no need to increase the elastic redundancy threshold of GPU fragments.

[0022] Furthermore, the increase in the elastic redundancy threshold of the GPU fragments is determined by the difference between the preset first accuracy and the matching accuracy of the task queue and the local computing power pool.

[0023] Furthermore, the completeness of the collected operational data is determined based on the missing rate of operational data per unit time, including:

[0024] Compare the missing rate of the data processed per unit time with the preset first missing rate;

[0025] If the missing rate of the running data per unit time is less than or equal to the preset first missing rate, then it is determined that the completeness of the running data collection meets the requirements, and it is determined whether the elastic redundancy threshold of the GPU fragment meets the requirements.

[0026] If the missing rate of the running data per unit time is greater than the preset first missing rate, then it is determined that the completeness of the collected running data does not meet the requirements.

[0027] Further, determine whether it is necessary to reduce the node resource load adaptation factor, including:

[0028] The missing rate of the running data within the unit time period is compared with the preset first missing rate and the preset second missing rate, respectively;

[0029] If the missing rate of the running data per unit time is greater than the preset first missing rate and less than the preset second missing rate, then it is determined that the node resource load adaptation factor needs to be reduced.

[0030] If the missing rate of the running data per unit time is greater than or equal to the preset second missing rate, then it is determined that there is no need to reduce the node resource load adaptation factor.

[0031] Furthermore, the reduction rate of the node resource load adaptation factor is determined by the difference between the missing rate of running data per unit time and the preset first missing rate.

[0032] Furthermore, the architecture fluctuation suppression threshold for GPU fragments is determined based on the fluctuation range of the local computing power pool per unit time, including:

[0033] Compare the fluctuation range of the local computing power pool within a unit of time with the preset fluctuation range;

[0034] If the computing power fluctuation of the local computing power pool within the unit time is less than or equal to the preset fluctuation range, then it is determined that the validity of the running data meets the requirements, there is no need to reduce the architecture fluctuation suppression threshold of the GPU fragment, and it is determined whether the node resource load adaptation factor meets the requirements.

[0035] If the fluctuation range of the local computing power pool within the unit time is greater than the preset fluctuation range, it is determined that the validity of the running data does not meet the requirements, and the architecture fluctuation suppression threshold of GPU fragments needs to be reduced.

[0036] Furthermore, the fluctuation range of the local computing power pool per unit time is the difference between the maximum and minimum computing power of the local computing power pool per unit time.

[0037] Furthermore, the reduction in the architecture fluctuation suppression threshold of the GPU fragment is determined by the difference between the fluctuation amplitude of the local computing power pool per unit time and the preset fluctuation amplitude.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows: The method of the present invention adjusts the elastic redundancy threshold of GPU fragments according to the matching accuracy of the task queue and the local computing power pool. Because there is a static deviation between the computing power representation dimension of GPU fragments of different architectures and the matching rules of standard task features in heterogeneous GPU clusters, the GPU fragments aggregated in the local computing power pool cannot accurately meet the task computing power requirements. By increasing the elastic redundancy threshold of GPU fragments, a buffer space can be reserved to adapt to the dynamic computing power fluctuations of tasks during computing power matching, compensating for the computing power efficiency deviation between heterogeneous GPU architectures, correcting the mismatch problem of matching computing power values ​​but insufficient actual execution efficiency, and improving the matching accuracy of standard task features and the local computing power pool. Furthermore, the method adjusts the node resource load adaptation factor according to the missing data rate of running data per unit time. In heterogeneous GPU clusters, the communication link between edge nodes and core scheduling nodes often spans multiple network devices, making them susceptible to network jitter caused by factors such as fiber optic link signal attenuation. Additionally, the GPU devices of edge nodes are in a high-load computing state for extended periods, which can easily lead to short-term frequency reduction when the efficiency of the cooling system decreases, resulting in data packet loss during resource status data transmission. Loss or damage can cause incomplete data in the resource information database. By reducing the node resource load adaptation factor, the scheduling system's reliance on missing and complete data can be reduced. Complete data collected locally in real time can be prioritized for load adaptation calculations, while triggering a retransmission mechanism for missing data. This ensures the integrity of GPU runtime data in the resource information database, avoids GPU fragment aggregation deviations caused by data loss, and ensures accurate calculation of the total computing power of the local computing pool. The architecture fluctuation suppression threshold for GPU fragments is adjusted based on the computing power fluctuation amplitude of the local computing pool per unit time. In heterogeneous GPU clusters, different architectures of GPUs have inherent computing power scheduling noise. Combined with resource contention interference from multiple virtual instances within the same physical GPU, this leads to instantaneous computing power fluctuation noise mixed into the GPU runtime data, misjudging normal interference as real computing power anomalies. By reducing the architecture fluctuation suppression threshold for GPU fragments, the fluctuation judgment criteria for different architectures of GPUs can be tightened, specifically filtering instantaneous noise interference that is incompatible with the architecture characteristics, retaining real computing power fluctuations strongly correlated with task load, improving the effectiveness of GPU runtime data, and enhancing the effectiveness of GPU computing power utilization.

[0039] Furthermore, the method of the present invention adjusts the elastic redundancy threshold of GPU fragments by setting a preset first accuracy rate and a preset second accuracy rate. Since there is a static deviation between the computing power representation dimension of GPU fragments of different architectures and the matching rules of standard task features in heterogeneous GPU clusters, the GPU fragments aggregated in the local computing power pool cannot accurately meet the computing power requirements of the task. By increasing the elastic redundancy threshold of GPU fragments, a buffer space can be reserved to adapt to the dynamic computing power fluctuations of the task during computing power matching, to compensate for the computing power performance deviation between heterogeneous GPU architectures, to correct the mismatch problem of matching computing power values ​​but insufficient actual execution efficiency, to improve the matching accuracy of standard task features and the local computing power pool, and to further improve the effectiveness of improving GPU computing power utilization.

[0040] Furthermore, the method of the present invention adjusts the node resource load adaptation factor by setting a preset first missing rate and a preset second missing rate. In heterogeneous GPU clusters, the communication link between edge nodes and core scheduling nodes often spans multiple network devices, which is susceptible to network jitter caused by factors such as fiber optic link signal attenuation. Moreover, the GPU devices of edge nodes are in a high-load computing state for a long time, which can easily cause short-term frequency reduction when the efficiency of the heat dissipation system decreases. This leads to data packet loss or damage during the transmission of resource status data, resulting in incomplete data in the resource information database. By reducing the node resource load adaptation factor, the scheduling system's dependence on missing data can be reduced, and complete data collected locally in real time can be used first for load adaptation calculation. At the same time, the retransmission mechanism of missing data is triggered to ensure the integrity of GPU running data in the resource information database, avoid GPU fragment aggregation deviation caused by data missing, ensure the accuracy of the total computing power calculation of the local computing power pool, and further improve the effectiveness of improving GPU computing power utilization.

[0041] Furthermore, the method described in this invention adjusts the architecture fluctuation suppression threshold of GPU fragments by setting a preset fluctuation amplitude. Since different architecture GPUs in a heterogeneous GPU cluster have inherent computing power scheduling noise, coupled with resource contention interference from multiple virtual instances within the same physical GPU, instantaneous computing power fluctuation noise is mixed into the GPU running data, misjudging normal interference as real computing power anomalies. By reducing the architecture fluctuation suppression threshold of GPU fragments, the fluctuation judgment criteria of different architecture GPUs can be tightened, and instantaneous noise interference that does not match the architecture characteristics can be filtered in a targeted manner, while retaining the real computing power fluctuation that is strongly correlated with the task load, thereby improving the effectiveness of GPU running data and further improving the effectiveness of improving GPU computing power utilization. Attached Figure Description

[0042] Figure 1 This is an overall flowchart of the automatic aggregation scheduling method for improving GPU computing power utilization according to an embodiment of the present invention;

[0043] Figure 2 The flowchart illustrates the process of determining the elastic redundancy threshold of GPU fragments in the automatic aggregation scheduling method for improving GPU computing power utilization according to an embodiment of the present invention.

[0044] Figure 3 A flowchart illustrating the process of determining the node resource load adaptation factor in the automatic aggregation scheduling method for improving GPU computing power utilization according to an embodiment of the present invention.

[0045] Figure 4 The flowchart illustrates the process of determining the architecture fluctuation suppression threshold for GPU fragments in the automatic aggregation scheduling method for improving GPU computing power utilization, as described in this embodiment of the invention. Detailed Implementation

[0046] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0047] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0048] Please see Figure 1 As shown, it is an overall flowchart of the automatic aggregation scheduling method for improving GPU computing power utilization according to an embodiment of the present invention.

[0049] This invention provides an automatic aggregation scheduling method for improving GPU computing power utilization, comprising:

[0050] Step S1: Collect GPU running data and computing task set respectively, process the running data to obtain resource information database, extract features from the computing task set to obtain standard task features, and classify the computing task set based on the standard task features to obtain task queue.

[0051] Step S2: Aggregate GPU fragments that meet preset conditions in each node based on the resource information database to obtain a local computing power pool, and determine the scheduling method of the task queue based on the comparison result of the total computing power of the local computing power pool and the computing power requirement of the task.

[0052] Step S3: Obtain the matching accuracy between the task queue and the local computing power pool, and determine whether the effectiveness of improving GPU computing power utilization meets the requirements based on the matching accuracy between the task queue and the local computing power pool.

[0053] Step S4: If the effectiveness of improving GPU computing power utilization does not meet the requirements, determine whether it is necessary to increase the elastic redundancy threshold of GPU fragments.

[0054] Step S5: If it is not necessary to increase the elastic redundancy threshold of GPU fragments, then obtain the missing rate of running data per unit time to determine whether the completeness of the running data collection meets the requirements.

[0055] Step S6: If the completeness of the collected running data does not meet the requirements, determine whether it is necessary to reduce the node resource load adaptation factor.

[0056] Step S7: If it is not necessary to reduce the node resource load adaptation factor, then determine the architecture fluctuation suppression threshold of GPU fragments based on the computing power fluctuation amplitude of the local computing power pool per unit time.

[0057] Specifically, the operational data includes the GPU hardware's rated computing power, video memory usage, and GPU core temperature.

[0058] Specifically, the computing task set includes task ID, peak computing power requirement, and peak video memory usage.

[0059] Specifically, the process of processing the runtime data to obtain the resource information database involves dividing resources into three levels: node, graphics card, and computing power slice, marking dedicated GPUs and shared GPUs, and forming the resource information database.

[0060] Specifically, the resource information database is a multi-dimensional collection of resource status data that is structured and stored in a three-level architecture of node-GPU-fragment after the collected GPU running data has been cleaned, denoised, and standardized.

[0061] Specifically, standard task characteristics include the task's computing power fluctuation coefficient, the type of GPU architecture supported by the task, and the minimum video memory capacity required by the task.

[0062] Specifically, the task queues include high-priority short task queues, ordinary batch processing queues, and long-term training queues.

[0063] Specifically, the high-priority short task queue includes autonomous driving perception reasoning, emergency event data statistical analysis, and financial risk control real-time decision reasoning.

[0064] Specifically, the general batch processing queue includes e-commerce product image classification and reasoning, business data ETL processing, and segmented scenario model training.

[0065] Specifically, the long-term training queue includes pre-training of enterprise-level general-purpose large models, training of autonomous driving full-scenario models, and training of smart city decision-making models.

[0066] Specifically, the preset conditions are that the remaining video memory on each node is greater than the minimum video memory requirement of the task, and the computing power utilization rate is less than 50%.

[0067] Specifically, GPU fragments are the smallest units of computing power that are independently schedulable and have resource isolation characteristics, formed by splitting a single physical GPU according to the dimensions of computing power and video memory resources through technologies such as hardware virtualization and software logical partitioning.

[0068] Specifically, the process of aggregating GPU fragments based on the resource information database to obtain a local computing power pool is executed step by step according to the logic of task demand-driven, precise fragment selection, computing power redundancy verification, collaborative aggregation construction, and dynamic adaptation optimization to obtain the local computing power pool.

[0069] Specifically, GPU fragments within the same node are preferentially aggregated to form a local computing power pool. When local fragments are insufficient, GPU fragments from multiple nodes within the same network segment are aggregated based on network bandwidth thresholds to form a cross-node computing power pool.

[0070] Specifically, the process of determining the task queue scheduling method based on the comparison between the total computing power of the local computing power pool and the computing power requirements of the tasks is as follows:

[0071] If the total computing power of the resource-adaptive computing power pool is greater than or equal to the computing power requirement of the task and the total video memory is greater than or equal to the maximum video memory requirement of the task, and the communication latency is less than 1ms, the local computing power pool will be matched first; if the communication latency is less than 10ms, the cross-node computing power pool in the same network segment will be matched second.

[0072] If the resource adaptability requirements are not met (total computing power of the computing power pool ≥ task computing power requirement and total video memory ≥ task maximum video memory requirement), then task queue scheduling will not be performed.

[0073] Specifically, scheduling methods include local scheduling, cross-node computing power aggregation scheduling, and task splitting scheduling.

[0074] Specifically, local scheduling methods include real-time target detection inference for autonomous vehicles, real-time risk control decision inference for financial transactions, and real-time target tracking inference for intelligent security cameras.

[0075] Specifically, cross-node computing power aggregation and scheduling methods include batch inference for product image classification on e-commerce platforms, batch inference for content review on short video platforms, and batch calculation for short-term forecasting of meteorological data.

[0076] Specifically, the task splitting and scheduling method includes LLaMA-13B large model pre-training, urban traffic flow prediction model training, and global ocean circulation simulation calculation.

[0077] Specifically, the matching accuracy rate between the task queue and the local computing power pool is the ratio of the number of times the task queue and the local computing power pool are accurately matched to the total number of matches.

[0078] Specifically, accurate matching means that a single task in the task queue meets the requirements of resource adaptation, stable operation, and efficiency in relation to the allocated local computing power pool.

[0079] Specifically, the elastic redundancy threshold of GPU fragments is the critical value of the proportion of pre-reserved redundant computing power to its rated computing power when GPU fragments participate in the aggregation of local computing power pool.

[0080] Specifically, the missing data rate per unit time is the ratio of the amount of missing data to the total amount of data in operation per unit time.

[0081] Specifically, the node resource load adaptation factor is a weighted adjustment coefficient used to dynamically balance the resource load status of edge nodes with the reliability of operational data.

[0082] Specifically, the GPU fragmentation architecture fluctuation suppression threshold is a critical parameter used to control memory fragmentation in the GPU architecture. It suppresses performance fluctuations caused by frequent allocation / release of video memory by optimizing the memory allocation algorithm.

[0083] In practice, the beneficial effects of this invention are as follows: The method of this invention adjusts the elastic redundancy threshold of GPU fragments based on the matching accuracy between the task queue and the local computing power pool. Because there is a static deviation between the computing power representation dimensions of GPU fragments of different architectures and the matching rules of standard task features in heterogeneous GPU clusters, the GPU fragments aggregated in the local computing power pool cannot accurately meet the task's computing power requirements. By increasing the elastic redundancy threshold of GPU fragments, a buffer space can be reserved during computing power matching to accommodate dynamic computing power fluctuations in tasks, compensating for computing power performance deviations between heterogeneous GPU architectures, correcting the mismatch problem of matching computing power values ​​but insufficient actual execution efficiency, and improving the matching accuracy between standard task features and the local computing power pool. Furthermore, the method adjusts the node resource load adaptation factor based on the missing data rate of running data per unit time. In heterogeneous GPU clusters, the communication links between edge nodes and core scheduling nodes often span multiple network devices, making them susceptible to network jitter caused by factors such as fiber optic link signal attenuation. Additionally, the GPU devices of edge nodes are under high load for extended periods, making them prone to short-term frequency reduction when the cooling system efficiency decreases, leading to packet loss during resource status data transmission. Loss or damage can cause incomplete data in the resource information database. By reducing the node resource load adaptation factor, the scheduling system's reliance on missing and complete data can be reduced. Complete data collected locally in real time can be prioritized for load adaptation calculations, while triggering a retransmission mechanism for missing data. This ensures the integrity of GPU runtime data in the resource information database, avoids GPU fragment aggregation deviations caused by missing data, and ensures accurate calculation of the total computing power of the local computing pool. The architecture fluctuation suppression threshold for GPU fragments is adjusted based on the computing power fluctuation amplitude of the local computing pool per unit time. In heterogeneous GPU clusters, different architectures of GPUs have inherent computing power scheduling noise. Combined with resource contention interference from multiple virtual instances within the same physical GPU, this leads to instantaneous computing power fluctuation noise mixed into the GPU runtime data, misjudging normal interference as real computing power anomalies. By reducing the architecture fluctuation suppression threshold for GPU fragments, the fluctuation judgment criteria for different architectures of GPUs can be tightened, specifically filtering instantaneous noise interference that is incompatible with the architecture characteristics, retaining real computing power fluctuations strongly correlated with task load, improving the effectiveness of GPU runtime data, and enhancing the effectiveness of GPU computing power utilization.

[0084] Please continue reading. Figure 2 The diagram shown is a logical flowchart illustrating the process of determining the elastic redundancy threshold of GPU fragments in the automatic aggregation scheduling method for improving GPU computing power utilization according to an embodiment of the present invention.

[0085] Specifically, determining whether the improvement in GPU computing power utilization meets the requirements based on the matching accuracy between the task queue and the local computing power pool includes:

[0086] The matching accuracy of the task queue with the local computing power pool is compared with the preset second accuracy.

[0087] If the matching accuracy between the task queue and the local computing power pool is greater than or equal to the preset second accuracy, then the effectiveness of improving GPU computing power utilization is determined to meet the requirements.

[0088] If the matching accuracy between the task queue and the local computing power pool is less than the preset second accuracy, then the effectiveness of improving GPU computing power utilization is determined to be unsatisfactory.

[0089] One reason why the improvement in GPU computing power utilization might not be effective enough could be that the integrity of the collected runtime data is not up to standard, or that the elastic redundancy threshold for GPU fragments is not sufficient. The next step is to determine which specific cause it is, which is essentially the process of deciding whether to increase the elastic redundancy threshold for GPU fragments.

[0090] Specifically, determining whether the elastic redundancy threshold for GPU fragmentation needs to be increased includes:

[0091] The matching accuracy of the task queue with the local computing power pool is compared with the preset first accuracy and the preset second accuracy, respectively.

[0092] If the matching accuracy between the task queue and the local computing power pool is less than or equal to the preset first accuracy, then it is determined that the elastic redundancy threshold of GPU fragments needs to be increased.

[0093] If the matching accuracy between the task queue and the local computing power pool is greater than a preset first accuracy and less than a preset second accuracy, then it is determined that there is no need to increase the elastic redundancy threshold of GPU fragments.

[0094] Specifically, if the matching accuracy between the task queue and the local computing power pool is less than or equal to a preset first accuracy rate, it indicates that the reason for the failure to improve GPU computing power utilization is that the elastic redundancy threshold of GPU fragments does not meet the requirements. Therefore, the elastic redundancy threshold of GPU fragments needs to be increased. If the matching accuracy between the task queue and the local computing power pool is greater than a preset first accuracy rate but less than a preset second accuracy rate, it can be preliminarily determined that the completeness of the collected running data does not meet the requirements. Next, it is necessary to make a final determination on whether the completeness of the collected running data meets the requirements based on the missing data rate of the running data per unit time, that is, to determine whether the reason for the failure to improve GPU computing power utilization is the incompleteness of the collected running data.

[0095] Understandably, the preset first accuracy rate is lower than the preset second accuracy rate. The three intervals divided by the preset first accuracy rate and the preset second accuracy rate correspond to three different scenarios:

[0096] The first interval is when the matching accuracy between the task queue and the local computing power pool is less than or equal to the preset first accuracy. The corresponding situation is: due to the static deviation between the computing power representation dimension of GPU fragments of different architectures and the matching rules of standard task characteristics in heterogeneous GPU clusters, the GPU fragments aggregated in the local computing power pool cannot accurately meet the computing power requirements of the task. At this time, it is necessary to adjust the elastic redundancy threshold of GPU fragments.

[0097] The second interval is when the matching accuracy between the task queue and the local computing power pool is greater than the preset first accuracy but less than the preset second accuracy. The corresponding situation is as follows: In heterogeneous GPU clusters, the communication link between edge nodes and core scheduling nodes often spans multiple network devices, which is susceptible to network jitter caused by factors such as fiber optic link signal attenuation. In addition, the GPU devices of edge nodes are in a high-load computing state for a long time, which is prone to short-term frequency reduction when the efficiency of the heat dissipation system decreases. This leads to data packet loss or damage during the transmission of resource status data, resulting in incomplete data in the resource information database. At this time, it is necessary to further determine whether the integrity of the collected running data meets the requirements.

[0098] The third interval is when the matching accuracy between the task queue and the local computing power pool is greater than or equal to the preset second accuracy. The corresponding situation is: the effectiveness of improving GPU computing power utilization is confirmed to meet the requirements, and no adjustment is needed.

[0099] Understandably, in the automatic aggregation and scheduling of GPU computing power utilization, the effectiveness of improving computing power utilization is characterized by preset first accuracy and preset second accuracy. The core logic is to transform the abstract effect of utilization improvement into a quantifiable judgment of matching accuracy ranges. The preset first accuracy serves as the boundary between adjusting the GPU fragment elastic redundancy threshold and judging the integrity of running data. The preset second accuracy serves as the critical point for determining whether the effectiveness of utilization improvement meets the standard, providing a clear basis for targeted optimization. The preset first accuracy and preset second accuracy can be set according to actual working conditions. The setting of the preset first accuracy and preset second accuracy aims to ensure the effectiveness and practicality of improving GPU computing power utilization. Optionally, the preset first accuracy and preset second accuracy are determined through a limited number of experiments by evaluating the effect of different accuracies on GPU computing power utilization improvement. The determined preset first accuracy and preset second accuracy should satisfy the condition that they are neither too small nor cause excessive interference to the process of improving GPU computing power utilization. For example, the preset first accuracy rate is generally selected in the range of [80%, 85%], and the preset second accuracy rate is generally selected in the range of [86%, 90%].

[0100] Preferably, the first accuracy rate is 82% in the preferred embodiment, and the second accuracy rate is 88% in the preferred embodiment.

[0101] Specifically, the increase in the elastic redundancy threshold of the GPU fragments is determined by the difference between a preset first accuracy rate and the matching accuracy rate between the task queue and the local computing power pool.

[0102] Specifically, when the difference between the preset first accuracy rate and the matching accuracy rate between the task queue and the local computing power pool is within 5%, the elastic redundancy threshold of the GPU fragment is increased to 1.2 times the original value. When the difference between the preset first accuracy rate and the matching accuracy rate between the task queue and the local computing power pool exceeds 5%, in addition to increasing to 1.2 times the original value, for every 1% exceeding 5%, the elastic redundancy threshold of the GPU fragment increases by 2%. For example, if the difference between the preset first accuracy rate and the matching accuracy rate between the task queue and the local computing power pool is 7%, and the current elastic redundancy threshold of the GPU fragment is 8%, the increased elastic redundancy threshold of the GPU fragment is 8%×1.2+2%×2=13.6%.

[0103] In practice, the method of this invention adjusts the elastic redundancy threshold of GPU fragments by setting a preset first accuracy rate and a preset second accuracy rate. Since there is a static deviation between the computing power representation dimension of GPU fragments of different architectures and the matching rules of standard task features in heterogeneous GPU clusters, the GPU fragments aggregated in the local computing power pool cannot accurately meet the computing power requirements of the task. By increasing the elastic redundancy threshold of GPU fragments, a buffer space can be reserved to adapt to the dynamic computing power fluctuations of the task during computing power matching, to compensate for the computing power performance deviation between heterogeneous GPU architectures, to correct the mismatch problem of matching computing power values ​​but insufficient actual execution efficiency, to improve the matching accuracy of standard task features and the local computing power pool, and to further improve the effectiveness of improving GPU computing power utilization.

[0104] Please continue reading. Figure 3 The diagram shown is a logical flowchart illustrating the process of determining the node resource load adaptation factor in the automatic aggregation scheduling method for improving GPU computing power utilization according to an embodiment of the present invention.

[0105] Specifically, the completeness of the collected operational data is determined based on the missing rate of operational data per unit time, including:

[0106] Compare the missing rate of the data processed per unit time with the preset first missing rate;

[0107] If the missing rate of the running data per unit time is less than or equal to the preset first missing rate, then it is determined that the completeness of the running data collection meets the requirements, and it is determined whether the elastic redundancy threshold of the GPU fragment meets the requirements.

[0108] If the missing rate of the running data per unit time is greater than the preset first missing rate, then it is determined that the completeness of the collected running data does not meet the requirements.

[0109] Specifically, when the missing rate of running data per unit time is less than or equal to the preset first missing rate, it is determined that the completeness of the running data collection meets the requirements. However, if the effectiveness of improving GPU computing power utilization has not been determined to meet the requirements, it is necessary to further determine whether the elastic redundancy threshold of GPU fragments meets the requirements.

[0110] In implementation, the elastic redundancy threshold of the actual GPU fragment is compared with the predetermined redundancy threshold benchmark value to determine whether the elastic redundancy threshold of the GPU fragment meets the requirements. If the elastic redundancy threshold of the actual GPU fragment is less than the predetermined redundancy threshold benchmark value, the elastic redundancy threshold of the GPU fragment is determined to be unacceptable. The predetermined redundancy threshold benchmark value is the average value of the elastic redundancy threshold of the GPU fragment monitored in the previous three months of the historical period.

[0111] If the elastic redundancy threshold of GPU fragments does not meet the requirements, then increase the elastic redundancy threshold of GPU fragments; if the elastic redundancy threshold of GPU fragments meets the requirements, then re-collect the matching accuracy of the task queue and the local computing power pool, and re-determine whether the effectiveness of improving GPU computing power utilization meets the requirements.

[0112] If the missing rate of the running data per unit time exceeds a preset first missing rate, it can be determined that the reason for the failure to improve GPU computing power utilization is that the completeness of the running data collection is not up to standard. The reasons for this may be that the node resource load adaptation factor is not up to standard, or that the validity of the running data is not up to standard. The next step is to determine which specific cause it is, which is essentially the process of deciding whether to reduce the node resource load adaptation factor.

[0113] Specifically, determining whether to reduce the node resource load adaptation factor includes:

[0114] The missing rate of the running data within the unit time period is compared with the preset first missing rate and the preset second missing rate, respectively;

[0115] If the missing rate of the running data per unit time is greater than the preset first missing rate and less than the preset second missing rate, then it is determined that the node resource load adaptation factor needs to be reduced.

[0116] If the missing rate of the running data per unit time is greater than or equal to the preset second missing rate, then it is determined that there is no need to reduce the node resource load adaptation factor.

[0117] Specifically, if the missing rate of the running data within a unit time is greater than a preset first missing rate but less than a preset second missing rate, the reason for the non-compliance of the running data collection integrity is determined to be that the node resource load adaptation factor is not up to standard, and therefore the node resource load adaptation factor needs to be reduced. If the missing rate of the running data within a unit time is greater than or equal to the preset second missing rate, it can be preliminarily determined that the running data validity is not up to standard. Next, it is necessary to make a final determination on whether the running data validity meets the requirements based on the fluctuation range of the local computing power pool within a unit time, that is, to determine whether the reason for the non-compliance of the running data collection integrity is the non-compliance of the running data validity.

[0118] It is understandable that the preset first missing rate is less than the preset second missing rate, and the three intervals divided by the preset first missing rate and the preset second missing rate correspond to three different scenarios:

[0119] The first interval is when the missing rate of the running data per unit time is less than or equal to the preset first missing rate. The corresponding situation is: the completeness of the running data collection meets the requirements. At this time, it is necessary to further determine whether the elastic redundancy threshold of the GPU fragments meets the requirements.

[0120] The second interval is defined as the missing rate of running data within a unit of time being greater than the preset first missing rate and less than the preset second missing rate. The corresponding situation is as follows: In heterogeneous GPU clusters, the communication link between edge nodes and core scheduling nodes often spans multiple network devices, which is susceptible to network jitter caused by factors such as fiber optic link signal attenuation. In addition, the GPU devices of edge nodes are in a high-load computing state for a long time, which is prone to short-term frequency reduction when the efficiency of the heat dissipation system decreases. This leads to data packet loss or damage during the transmission of resource status data, resulting in incomplete data in the resource information database.

[0121] The third interval is when the missing rate of running data per unit time is greater than or equal to the preset second missing rate. The corresponding situation is: due to the inherent computing power scheduling noise of different architecture GPUs in heterogeneous GPU clusters, coupled with the resource contention interference of multiple virtual instances within the same physical GPU, instantaneous computing power fluctuation noise is mixed into the GPU running data, misjudging normal interference as real computing power anomalies. At this time, it is necessary to further determine whether the validity of the running data meets the requirements.

[0122] Understandably, in the automatic aggregation and scheduling process of GPU computing power utilization, the introduction of preset first missing rate and preset second missing rate to characterize the completeness of the collected running data is crucial. The core logic is to transform data completeness into a quantifiable missing rate range judgment. The preset first missing rate is the boundary between the threshold for confirming elastic redundancy and the node resource load adaptation factor that needs adjustment, while the preset second missing rate is the critical point between determining the node resource load adaptation factor that needs adjustment and determining the data collection frequency, providing a quantitative basis for targeted optimization. The preset first missing rate and preset second missing rate can be set according to actual working conditions. The setting of the preset first missing rate and preset second missing rate aims to ensure the effectiveness and practicality of improving GPU computing power utilization. Optionally, the preset first missing rate and preset second missing rate are determined through a limited number of experiments by evaluating the effect of different missing rates on improving GPU computing power utilization. The determined preset first missing rate and preset second missing rate should satisfy the condition that they are neither too small nor cause excessive interference to the process of improving GPU computing power utilization. For example, the preset first missing rate is generally selected in the range of [3%, 7%], and the preset second missing rate is generally selected in the range of [8%, 10%].

[0123] Preferably, the first missing rate is 5% in the preferred embodiment, and the second missing rate is 9% in the preferred embodiment.

[0124] Specifically, the reduction in the node resource load adaptation factor is determined by the difference between the missing rate of running data per unit time and a preset first missing rate.

[0125] Specifically, when the difference between the missing data rate and the preset first missing rate within a unit time is within 2%, the node resource load adaptation factor is reduced to 0.9 times the original value. When the difference between the missing data rate and the preset first missing rate within a unit time exceeds 2%, in addition to reducing it to 0.9 times the original value, the node resource load adaptation factor is further reduced by 0.03 for every 0.8% exceeding the original value. For example, when the difference between the missing data rate and the preset first missing rate within a unit time is 3.6%, the current node resource load adaptation factor is 2.0, and the reduced node resource load adaptation factor is 2.0×0.9-0.03×2=1.74.

[0126] In implementation, the method of this invention adjusts the node resource load adaptation factor by setting a preset first missing rate and a preset second missing rate. In heterogeneous GPU clusters, the communication links between edge nodes and core scheduling nodes often span multiple network devices, making them susceptible to network jitter caused by factors such as fiber optic link signal attenuation. Furthermore, the GPU devices on edge nodes are constantly under high load, making them prone to short-term frequency reduction when the cooling system efficiency decreases. This leads to data packet loss or corruption during resource status data transmission, resulting in incomplete data in the resource information database. By reducing the node resource load adaptation factor, the scheduling system's reliance on missing data can be reduced, prioritizing the use of complete data collected locally in real time for load adaptation calculations. Simultaneously, a retransmission mechanism for missing data is triggered, ensuring the integrity of GPU running data in the resource information database, avoiding GPU fragment aggregation deviations caused by data loss, ensuring accurate calculation of the total computing power of the local computing pool, and further improving the effectiveness of GPU computing power utilization.

[0127] Please continue reading. Figure 4 The diagram shown is a logical flowchart illustrating the process of determining the architecture fluctuation suppression threshold of GPU fragments in the automatic aggregation scheduling method for improving GPU computing power utilization according to an embodiment of the present invention.

[0128] Specifically, the architecture fluctuation suppression threshold for GPU fragments is determined based on the fluctuation range of the local computing power pool per unit time, including:

[0129] Compare the fluctuation range of the local computing power pool within a unit of time with the preset fluctuation range;

[0130] If the computing power fluctuation of the local computing power pool within the unit time is less than or equal to the preset fluctuation range, then it is determined that the validity of the running data meets the requirements, there is no need to reduce the architecture fluctuation suppression threshold of the GPU fragment, and it is determined whether the node resource load adaptation factor meets the requirements.

[0131] If the fluctuation range of the local computing power pool within the unit time is greater than the preset fluctuation range, it is determined that the validity of the running data does not meet the requirements, and the architecture fluctuation suppression threshold of GPU fragments needs to be reduced.

[0132] Specifically, the fluctuation range of the local computing power pool per unit time is the difference between the maximum and minimum computing power of the local computing power pool per unit time.

[0133] Specifically, if the fluctuation range of the local computing power pool within a time period is less than or equal to the preset fluctuation range, it is determined that the validity of the running data meets the requirements. However, if the completeness of the collected running data has not been determined to meet the requirements, it is necessary to further determine whether the node resource load adaptation factor meets the requirements.

[0134] In implementation, the node resource load adaptation factor is determined to meet the requirements based on the comparison between the actual node resource load adaptation factor and the predetermined adaptation factor threshold. If the actual node resource load adaptation factor is greater than the predetermined adaptation factor threshold, the node resource load adaptation factor is determined to not meet the requirements. The predetermined adaptation factor threshold is the average value of the node resource load adaptation factor monitored in the previous three months of the historical period.

[0135] If the node resource load adaptation factor does not meet the requirements, the node resource load adaptation factor is reduced; if the node resource load adaptation factor meets the requirements, the missing rate of running data per unit time is re-collected, and the completeness of the collected running data is re-evaluated.

[0136] When the fluctuation range of the local computing power pool exceeds the preset fluctuation range within a unit of time, it can be determined that the reason for the failure of the validity of the running data is that the architecture fluctuation suppression threshold of the GPU fragment does not meet the requirements. Therefore, it is necessary to reduce the architecture fluctuation suppression threshold of the GPU fragment.

[0137] It is understandable that the two preset fluctuation ranges correspond to two different scenarios:

[0138] The first interval is when the local computing power pool's computing power fluctuation is less than or equal to the preset fluctuation range per unit time. The corresponding situation is: the validity of the running data is confirmed to meet the requirements. At this time, it is necessary to further determine whether the node resource load adaptation factor meets the requirements.

[0139] The second interval is when the local computing power pool's computing power fluctuation amplitude is greater than the preset fluctuation amplitude within a unit of time. The corresponding situation is as follows: In heterogeneous GPU clusters, different architecture GPUs have inherent computing power scheduling noise. In addition, within the same physical GPU, there is resource contention interference from multiple virtual instances, which causes instantaneous computing power fluctuation noise to be mixed into the GPU running data. Normal interference is misjudged as real computing power anomalies. At this time, it is necessary to reduce the architecture fluctuation suppression threshold of GPU fragments.

[0140] Understandably, in the automatic aggregation and scheduling of GPU computing power utilization, the preset fluctuation range is used to characterize the validity of the running data. The core logic is to transform data validity into a quantifiable judgment of computing power fluctuation range. The preset fluctuation range serves as the critical point to distinguish between the node resource load adaptation factor that needs to be confirmed and the architecture fluctuation suppression threshold that needs to be adjusted, clarifying the targeted processing direction under different fluctuation scenarios and providing a quantitative basis for ensuring the accuracy of computing power scheduling decisions. The preset fluctuation range can be set according to actual working conditions. The setting of the preset fluctuation range aims to ensure the effectiveness and practicality of improving GPU computing power utilization. Optionally, the preset fluctuation range is determined through a limited number of experiments by evaluating the effect of different fluctuation ranges on improving GPU computing power utilization. The determined preset fluctuation range should meet the requirement of being neither too small nor causing excessive interference to the improvement effect of GPU computing power utilization. For example, the preset fluctuation range is generally selected in the range of [6%, 10%].

[0141] Preferably, the preset fluctuation range is 8% in this preferred embodiment.

[0142] Specifically, the reduction in the architecture fluctuation suppression threshold of the GPU fragment is determined by the difference between the fluctuation amplitude of the local computing power pool per unit time and the preset fluctuation amplitude.

[0143] Specifically, when the difference between the local computing power fluctuation amplitude and the preset fluctuation amplitude within a unit time is within 1%, the architecture fluctuation suppression threshold of the GPU fragment is reduced to 0.85 times the original value. When the difference between the local computing power fluctuation amplitude and the preset fluctuation amplitude within a unit time exceeds 1%, the architecture fluctuation suppression threshold of the GPU fragment is reduced by 0.04 for every 0.5% increase beyond the original value of 0.85 times. For example, when the difference between the local computing power fluctuation amplitude and the preset fluctuation amplitude within a unit time is 2.0%, the current architecture fluctuation suppression threshold of the GPU fragment is 0.3. The reduced architecture fluctuation suppression threshold of the GPU fragment is 0.3×0.85-0.04×2=0.175.

[0144] In practice, the method described in this invention adjusts the architecture fluctuation suppression threshold of GPU fragments by setting a preset fluctuation amplitude. Since there is inherent computing power scheduling noise in heterogeneous GPU clusters with different architectures, coupled with resource contention interference from multiple virtual instances within the same physical GPU, instantaneous computing power fluctuation noise is mixed into the GPU running data. Normal interference is misjudged as abnormal real computing power passing through. Reducing the architecture fluctuation suppression threshold of GPU fragments can tighten the fluctuation judgment criteria for GPUs with different architectures, specifically filter instantaneous noise interference that does not match the architecture characteristics, retain the real computing power fluctuation that is strongly correlated with the task load, improve the effectiveness of GPU running data, and further improve the effectiveness of improving GPU computing power utilization.

[0145] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. An automatic aggregation scheduling method for improving GPU computing power utilization, characterized in that, include: GPU running data and computing task sets are collected separately. The running data is processed to obtain a resource information database. The computing task sets are feature extracted to obtain standard task features. The computing task sets are classified based on the standard task features to obtain task queues. Based on the resource information database, GPU fragments that meet the preset conditions in each node are aggregated to obtain a local computing power pool. The scheduling method of the task queue is determined based on the comparison result between the total computing power of the local computing power pool and the computing power requirement of the task. Obtain the matching accuracy between the task queue and the local computing power pool, and determine whether the effectiveness of improving GPU computing power utilization meets the requirements based on the matching accuracy between the task queue and the local computing power pool. If the effectiveness of improving GPU computing power utilization does not meet the requirements, then it is determined whether it is necessary to increase the elastic redundancy threshold of GPU fragments. If it is not necessary to increase the elastic redundancy threshold of GPU fragments, then obtain the missing rate of running data per unit time to determine whether the completeness of the running data collection meets the requirements. If the completeness of the collected operational data does not meet the requirements, then determine whether it is necessary to reduce the node resource load adaptation factor. If it is not necessary to reduce the node resource load adaptation factor, the architecture fluctuation suppression threshold of GPU fragments is determined based on the computing power fluctuation amplitude of the local computing power pool per unit time. Determining whether the improvement in GPU computing power utilization meets the requirements based on the matching accuracy between the task queue and the local computing power pool includes: The matching accuracy of the task queue with the local computing power pool is compared with the preset second accuracy. If the matching accuracy between the task queue and the local computing power pool is greater than or equal to the preset second accuracy, then the effectiveness of improving GPU computing power utilization is determined to meet the requirements. If the matching accuracy between the task queue and the local computing power pool is less than the preset second accuracy, then the effectiveness of improving GPU computing power utilization is determined to be unsatisfactory. Determine whether the elastic redundancy threshold for GPU fragmentation needs to be increased, including: The matching accuracy of the task queue with the local computing power pool is compared with the preset first accuracy and the preset second accuracy, respectively. If the matching accuracy between the task queue and the local computing power pool is less than or equal to the preset first accuracy, then it is determined that the elastic redundancy threshold of GPU fragments needs to be increased. If the matching accuracy between the task queue and the local computing power pool is greater than a preset first accuracy and less than a preset second accuracy, then it is determined that there is no need to increase the elastic redundancy threshold of GPU fragments. The increase in the elastic redundancy threshold of the GPU fragments is determined by the difference between the preset first accuracy and the matching accuracy of the task queue and the local computing power pool.

2. The automatic aggregation scheduling method for improving GPU computing power utilization according to claim 1, characterized in that, The completeness of the collected operational data is determined based on the missing rate of operational data per unit time, including: Compare the missing rate of the data processed per unit time with the preset first missing rate; If the missing rate of the running data per unit time is less than or equal to the preset first missing rate, then it is determined that the completeness of the running data collection meets the requirements, and it is determined whether the elastic redundancy threshold of the GPU fragment meets the requirements. If the missing rate of the running data per unit time is greater than the preset first missing rate, then it is determined that the completeness of the collected running data does not meet the requirements.

3. The automatic aggregation scheduling method for improving GPU computing power utilization according to claim 2, characterized in that, Determine whether it is necessary to reduce the node resource load adaptation factor, including: The missing rate of the running data within the unit time period is compared with the preset first missing rate and the preset second missing rate, respectively; If the missing rate of the running data per unit time is greater than the preset first missing rate and less than the preset second missing rate, then it is determined that the node resource load adaptation factor needs to be reduced. If the missing rate of the running data per unit time is greater than or equal to the preset second missing rate, then it is determined that there is no need to reduce the node resource load adaptation factor.

4. The automatic aggregation scheduling method for improving GPU computing power utilization according to claim 3, characterized in that, The reduction rate of the node resource load adaptation factor is determined by the difference between the missing rate of running data per unit time and the preset first missing rate.

5. The automatic aggregation scheduling method for improving GPU computing power utilization according to claim 4, characterized in that, The architecture fluctuation suppression threshold for GPU fragments is determined based on the fluctuation range of the local computing power pool per unit time, including: Compare the fluctuation range of the local computing power pool within a unit of time with the preset fluctuation range; If the computing power fluctuation of the local computing power pool within the unit time is less than or equal to the preset fluctuation range, then it is determined that the validity of the running data meets the requirements, there is no need to reduce the architecture fluctuation suppression threshold of the GPU fragment, and it is determined whether the node resource load adaptation factor meets the requirements. If the fluctuation range of the local computing power pool within the unit time is greater than the preset fluctuation range, it is determined that the validity of the running data does not meet the requirements, and the architecture fluctuation suppression threshold of GPU fragments needs to be reduced.

6. The automatic aggregation scheduling method for improving GPU computing power utilization according to claim 5, characterized in that, The fluctuation range of the local computing power pool per unit time is the difference between the maximum and minimum computing power of the local computing power pool per unit time.

7. The automatic aggregation scheduling method for improving GPU computing power utilization according to claim 6, characterized in that, The reduction in the architecture fluctuation suppression threshold of the GPU fragment is determined by the difference between the fluctuation amplitude of the local computing power pool per unit time and the preset fluctuation amplitude.

Images