Computing power adjustment method and computing device

Abstract

The application relates to the technical field of computing, in particular to a computing power adjustment method and a computing device. The method comprises the following steps: in response to a computing power request triggered by a user, determining a first quota of computing power resources available to the user; acquiring a long-term utilization deviation generated by the user in using the computing power resources, the long-term utilization deviation being used to represent a deviation degree of the quota and the load of the computing power resources of the user in the time dimension; determining an adaptive computing power adjustment mode from a long-term adjustment mode and an instantaneous adjustment mode according to the long-term utilization deviation, the long-term adjustment mode being used to control real-time computing power to converge to a preset computing power quota, and the instantaneous adjustment mode being used to adjust the computing power for burst traffic; adjusting the first quota to obtain a second quota according to the computing power adjustment mode; and allocating new computing power resources to the user according to the second quota, the new computing power resources being used to execute a computing task corresponding to the computing power request.

Description

Technical Field

[0001] This application relates to the field of computing technology, and in particular to a computing power adjustment method and computing device. Background Technology

[0002] With the rapid development of artificial intelligence technology, the efficient utilization and refined management of computing resources are key technologies in commercial applications. Currently, users typically apply for computing resources in a flexible manner, rather than monopolizing the entire computing equipment, in order to improve resource utilization and reduce user costs.

[0003] However, existing computing power adjustment technologies generally adjust the corresponding computing power resources according to the quota applied by the user. This method of computing power adjustment is prone to instantaneous overload or underload of computing power resources. Summary of the Invention

[0004] This application provides a computing power adjustment method and computing device, which improves the utilization rate of computing power resources by adaptively adjusting computing power using different computing power modes.

[0005] According to a first aspect of the embodiments of this application, a computing power adjustment method is provided, applied to a computing device, the method comprising: In response to a user-triggered computing power request, determine the first quota of computing power resources available to the user.

[0006] The long-term utilization deviation of computing resources by users is obtained. The long-term utilization deviation is used to characterize the degree of deviation of the user's computing resource quota and load over time.

[0007] Based on the long-term utilization deviation, a suitable computing power adjustment mode is determined from the long-term adjustment mode and the instantaneous adjustment mode. The long-term adjustment mode is used to control the real-time computing power to converge to the preset computing power quota, and the instantaneous adjustment mode is used to adjust the computing power for sudden traffic.

[0008] Based on the computing power adjustment mode, the first quota is adjusted to obtain the second quota.

[0009] According to the second quota, new computing resources are allocated to users, and these new computing resources are used to execute the computing tasks corresponding to the computing requests.

[0010] In this embodiment, in response to a user-triggered computing power request, a first quota corresponding to the computing power resources available to the user can be determined first. Then, the long-term utilization deviation, which characterizes the degree of deviation between the quota and the actual load over a long period, is introduced as the core judgment criterion. This allows for intelligent matching of either a long-term adjustment mode or an instantaneous adjustment mode. The long-term adjustment mode can smoothly converge real-time computing power to a preset baseline quota, correcting deviations from long-term resource idleness or over-allocation. The instantaneous adjustment mode can quickly respond to sudden business traffic and complete temporary computing power adaptation. Furthermore, the first quota is dynamically optimized according to the selected computing power adjustment mode to obtain a precisely adapted second quota. Finally, new computing power resources are allocated to the user according to the second quota to carry the corresponding computing tasks. This achieves an upgrade of computing power quota to a refined dynamic control that combines long-term and short-term effects and load linkage. It can eliminate long-term computing power waste and the risk of continuous overload, while efficiently covering sudden traffic impacts, improving the overall computing power resource utilization, task scheduling stability, and intelligent adaptation capability of quota allocation, and ensuring balanced operation of cluster computing power load.

[0011] According to a first aspect of the embodiments of this application, in some embodiments of the first aspect, obtaining the long-term utilization deviation caused by the user's use of computing resources includes: Multiple historical utilization rates are obtained from the first-in-first-out (FIFO) queue. These historical utilization rates are used to characterize the efficiency of a user's use of allocated computing resources.

[0012] Calculate the average utilization rate based on multiple historical utilization rates.

[0013] The long-term utilization deviation is determined based on the difference between the average utilization rate and the first quota.

[0014] In this embodiment, when obtaining long-term utilization deviation, multiple historical utilization data can be retrieved in an orderly manner from the FIFO queue. This allows for the systematic aggregation of the actual efficiency of user computing resources usage according to time sequence. The average utilization rate is then calculated based on the average of multiple periodic samples. The difference between this average utilization rate and the preset first quota determines the long-term utilization deviation. Using multiple historical utilization rates as key parameters for subsequent long-term utilization deviation can smooth out instantaneous traffic fluctuations and filter out sudden load interference. This ensures that the long-term utilization deviation can objectively and stably characterize the true deviation between the user's actual computing load and the quota configuration over a long period. This provides accurate and reliable quantitative calculation basis for subsequently differentiating between long-term and short-term adjustment modes and dynamically optimizing computing quotas, improving the stability and data accuracy of computing deviation assessment.

[0015] According to a first aspect of the embodiments of this application, in some embodiments of the first aspect, determining an appropriate computing power adjustment mode from a long-term adjustment mode and a momentary adjustment mode based on long-term utilization deviation includes: If the long-term utilization rate deviation exceeds the preset first quota threshold, the long-term adjustment mode will be determined as the computing power adjustment mode.

[0016] If the long-term utilization deviation is less than or equal to the first quota threshold, the instantaneous adjustment mode will be determined as the computing power adjustment mode.

[0017] In this embodiment, a preset first quota threshold is used as the judgment criterion. When the long-term utilization deviation is greater than the first quota threshold, the long-term adjustment mode is precisely activated to continuously correct the long-standing problem of excessive computing power occupation, and promote the actual computing power load to steadily converge to the compliant quota range. When the long-term utilization deviation is less than or equal to the first quota threshold, the instantaneous adjustment mode is switched to only perform lightweight computing power fine-tuning for temporary burst traffic to handle the impact of burst traffic. The intelligent switching and precise adaptation of the two adjustment modes are achieved through threshold linkage. This can not only specifically address the hidden dangers of long-term resource over-provisioning and overload, and continuous crowding of cluster computing power, but also avoid the problems of frequent quota fluctuations and unstable business operation caused by indiscriminate long-term parameter tuning. It takes into account the regularity of long-term computing power configuration and the flexibility of adapting to burst scenarios, so that the computing power control strategy fits the user's real load characteristics, and further improves the accuracy of dynamic quota adjustment and the stability of the overall cluster computing power scheduling.

[0018] According to a first aspect of the embodiments of this application, in some embodiments of the first aspect, adjusting the first quota according to a computing power adjustment mode to obtain a second quota includes: Determine the quota adjustment coefficient based on the coefficient adjustment strategy corresponding to the computing power adjustment mode.

[0019] Based on the quota adjustment coefficient and the dynamic scaling factor, the first quota is adjusted to obtain the second quota.

[0020] In this embodiment, a dedicated coefficient adjustment strategy is invoked based on the matched computing power adjustment mode to accurately generate quota adjustment coefficients that adapt to the current load characteristics. Then, the initial first quota is adjusted in conjunction with a dynamic scaling factor to obtain the second quota. The mode-specific strategy ensures that the coefficient selection meets the control requirements of long-term steady-state correction or instantaneous burst adaptation. The introduction of a dynamic scaling factor can further adapt to real-time computing power fluctuations, node load conditions, and resource margins to achieve flexible scaling. This allows quota adjustment to have both strategy-specificity and real-time dynamic adaptability, significantly improving the precision and rationality of iterative optimization from the first quota to the second quota. This achieves precise adjustment of computing power resources and effectively avoids resource waste or adjustment imbalance caused by rigid quota adjustments, excessive amplitude, or adaptation deviations.

[0021] According to a first aspect of the embodiments of this application, in some embodiments of the first aspect, it further includes: When the average utilization rate of computing resources by users is less than the preset quota threshold, the dynamic scaling factor is reduced.

[0022] If a user's instantaneous utilization of computing resources exceeds a preset threshold, the dynamic scaling factor will be increased.

[0023] In this embodiment, the dynamic scaling factor is adaptively adjusted based on a combination of the user's average and instantaneous utilization of computing resources. When the average utilization is less than a preset quota threshold, the dynamic scaling factor is proactively reduced to suppress blind quota expansion, converge idle and redundant computing power, and avoid long-term resource waste caused by inflated quotas. When the instantaneous utilization exceeds a preset limit threshold, the dynamic scaling factor can be increased accordingly, enabling rapid expansion of quota adjustments to adapt to sudden high-load traffic, timely filling instantaneous computing power gaps, and alleviating business computing power pressure. By adjusting the dynamic scaling factor, long-term resource utilization and short-term business stability are effectively balanced, making quota adjustments more closely match the user's actual computing power consumption patterns and improving the overall flexibility and accuracy of computing power allocation.

[0024] According to a first aspect of the embodiments of this application, in some embodiments of the first aspect, when the computing power adjustment mode is a long-term adjustment mode, the quota adjustment coefficient is determined according to the coefficient adjustment strategy corresponding to the computing power adjustment mode, including: Using the number of SMs per GPU, the maximum number of threads per SM, the dynamic scaling factor, and the long-term utilization deviation as input data, and inputting the coefficient adjustment strategy corresponding to the long-term adjustment mode, the quota adjustment coefficient is obtained.

[0025] In this embodiment, hardware fundamental parameters such as the number of SMs per GPU and the maximum number of threads per SM, as well as trend parameters such as dynamic scaling factors and long-term utilization deviation, are used as input data. This comprehensively and meticulously covers the key factors affecting GPU computing power allocation and utilization. By inputting this data into the coefficient adjustment strategy corresponding to the long-term adjustment mode, the algorithm and logic within the strategy can perform in-depth analysis and calculation based on actual and complex operating conditions, ultimately obtaining a scientifically reasonable quota adjustment coefficient. This achieves refined management and dynamic optimization of GPU resources, enabling the adjustment of computing power quotas to tend towards controlling real-time computing power convergence to the preset computing power quota, effectively improving resource utilization, avoiding resource waste or insufficiency, and ensuring stable and efficient system operation under high load and dynamic changing environments.

[0026] According to a first aspect of the embodiments of this application, in some embodiments of the first aspect, when the computing power adjustment mode is an instantaneous adjustment mode, the quota adjustment coefficient is determined according to the coefficient adjustment strategy corresponding to the computing power adjustment mode, including: The instantaneous utilization rate deviation is obtained based on the instantaneous utilization rate of computing resources used by the user and the first quota.

[0027] Using the number of SMs per GPU, the maximum number of threads per SM, the dynamic scaling factor, and the instantaneous utilization deviation as input data, and inputting the coefficient adjustment strategy corresponding to the instantaneous adjustment mode, the quota adjustment coefficient is obtained.

[0028] In this embodiment, the instantaneous utilization deviation is obtained based on the instantaneous utilization rate and the first quota. Then, hardware fundamental parameters such as the number of SMs per GPU and the maximum number of threads per SM, as well as parameters reflecting the GPU's operating status and resource usage fluctuations in real time, such as the dynamic scaling factor and the instantaneous utilization deviation, are used as input data. This comprehensively and meticulously covers the key factors affecting GPU computing power allocation and utilization. Inputting these parameters into the coefficient adjustment strategy corresponding to the instantaneous adjustment mode allows for rapid response to instantaneous GPU load changes. Based on real-time data and preset strategies, instantaneous computing power is adjusted to ensure that quota adjustment coefficients are provided to match the rapidly changing instantaneous computing power demand. This achieves rapid adjustment of computing power resources and avoids performance bottlenecks or resource idleness.

[0029] According to a first aspect of the embodiments of this application, in some embodiments of the first aspect, after allocating new computing resources to users according to a second quota, the method further includes: If the new computing resources meet the computing power requirements of the computing task corresponding to the computing power request, the target computing resources for executing the computing task corresponding to the computing power request are allocated from the new computing resources so as to utilize the target computing resources to execute the computing task.

[0030] If the new computing resources do not meet the computing power requirements of the computing task corresponding to the computing power request, the computing power request will be blocked.

[0031] In this embodiment, after receiving a user's computing power request, the user's initial quota is first obtained. Then, long-term utilization deviation is introduced, and the initial quota is optimized and corrected by intelligently matching a long-term adjustment mode or an instantaneous adjustment mode to obtain a second quota. Adapted new computing power resources are generated based on the optimized second quota configuration. Simultaneously, a computing power demand verification step is added. When the new computing power resources can fully support the task's computing power consumption, the target computing power resources are accurately allocated to ensure the normal execution of the computing task. If the resource volume cannot match the demand, the computing power request is blocked and intercepted in a timely manner. This ensures that the entire computing power adjustment process can achieve refined quota iteration by combining historical load deviations, taking into account both normal computing power convergence and the ability to adapt to sudden traffic, and can also prevent operational anomalies caused by insufficient resources through a final verification mechanism. This improves the fit of single-user computing power allocation and resource utilization while strengthening the overall stability, security, and risk control capabilities of computing power scheduling.

[0032] According to a first aspect of the embodiments of this application, in some embodiments of the first aspect, it further includes: Determine the first number of tokens in the token pool of the new computing resource mapping.

[0033] Determine the number of second tokens required for the computing task corresponding to the computing power request.

[0034] If the number of second tokens is less than or equal to the number of first tokens, new computing resources are determined to meet the computing power requirements of the computing task corresponding to the computing power request.

[0035] If the number of second tokens is greater than the number of first tokens, it is determined that the new computing resources do not meet the computing resources required for the computing task corresponding to the computing request.

[0036] In this embodiment, the number of tokens is used to represent computing power resources. By determining the first token quantity corresponding to at least one token in the token pool for a new computing power resource, the new computing power resource is measured using the number of tokens. Furthermore, the second token quantity required by the computing task corresponding to the computing power request is clearly defined, quantifying the token quantity from the demand side. By comparing the first and second token quantities, it is possible to quickly and accurately determine whether the new computing power resource meets the computing power requirements of the computing task. That is, when the second token quantity is less than or equal to the first token quantity, it is directly determined that the requirement is met, avoiding over-allocation of resources. When the second token quantity is greater than the first token quantity, it is determined that the requirement is not met, preventing task execution failure or performance degradation due to insufficient resources. This token quantity comparison method improves the efficiency and accuracy of computing power resource allocation, effectively ensuring that computing tasks can utilize computing power resources reasonably and efficiently.

[0037] According to a second aspect of the embodiments of this application, a computing power adjustment device is provided, which may include: The request response unit is used to respond to a computing power request triggered by a user and determine the first quota of computing power resources available to the user.

[0038] The long-term analysis unit is used to obtain the long-term utilization deviation caused by the user's use of computing resources. The long-term utilization deviation is used to characterize the degree of deviation of the user's computing resource quota and load over time.

[0039] The mode selection unit is used to determine the appropriate computing power adjustment mode from the long-term adjustment mode and the instantaneous adjustment mode based on the long-term utilization deviation. The long-term adjustment mode is used to control the real-time computing power to converge to the preset computing power quota, and the instantaneous adjustment mode is used to adjust the computing power for sudden traffic.

[0040] The quota adjustment unit is used to adjust the first quota and obtain the second quota according to the computing power adjustment mode.

[0041] The resource allocation unit is used to allocate new computing resources to users according to the second quota. The new computing resources are used to execute the computing tasks corresponding to the computing power requests.

[0042] According to a third aspect of the embodiments of this application, a computing device is provided, which includes at least a memory and a processor, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to implement the computing power adjustment method of any of the above embodiments.

[0043] According to a fourth aspect of the embodiments of this application, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, it implements any computing power adjustment method.

[0044] According to a fifth aspect of the embodiments of this application, a computer product is provided, comprising: the step of implementing any computing power adjustment method when the computer product is executed by a processor.

[0045] It should be understood that both the foregoing general description and the following detailed description are exemplary and intended to provide further illustration of the claimed technology. Attached Figure Description

[0046] The above and other objects, features, and advantages of the embodiments of this application will become more apparent from the more detailed description of the embodiments in conjunction with the accompanying drawings. The accompanying drawings are used to provide a further understanding of the embodiments of this application and constitute a part of the specification. They are used together with the embodiments of this application to explain the embodiments of this application and do not constitute a limitation thereof. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0047] Figure 1 The figure shows a schematic diagram of a computing power adjustment system according to an embodiment of this application; Figure 2 The figure shows a flowchart of a computing power adjustment method according to an embodiment of this application; Figure 3 The figure shows an example diagram of a physical GPU according to an embodiment of this application; Figures 4a-4b The figure shows an example diagram of a GPU computing power quota according to an embodiment of this application; Figure 5 The figure shows another flowchart of a computing power adjustment method according to an embodiment of this application; Figure 6 The figure shows an example diagram of a computing power adjustment method according to an embodiment of this application; Figure 7 The figure shows an example diagram of a computing power adjustment framework according to an embodiment of this application; Figure 8The figure shows a signaling diagram of a computing power allocation method according to an embodiment of this application; Figure 9 The figure shows a schematic diagram of a computing power adjustment device according to an embodiment of this application; Figure 10 The figure shows a hardware block diagram of a computing device according to an embodiment of this application. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the embodiments of this application more apparent, exemplary embodiments according to the embodiments of this application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the embodiments of this application, and not all embodiments of the embodiments of this application. It should be understood that the embodiments of this application are not limited to the exemplary embodiments described herein.

[0049] The technical solution of this application embodiment can be applied to computing power allocation scenarios such as cloud computing. By setting two computing power adjustment mechanisms with two time scales, the GPU computing power allocation can meet both long-term stable control and instantaneous computing power demand, achieving a balance between instantaneous load and long-term convergence, making the computing power supply highly consistent with the user quota, and improving the user experience and service quality.

[0050] The technical solutions of the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0051] like Figure 1 The diagram shown is a structural schematic of a computing power adjustment system provided in an embodiment of this application. The computing power adjustment system may include: user equipment 10 and server 20.

[0052] In this system, user equipment 10 can collect users' computing power needs at the application layer, generate computing power requests, and send the users' computing power requests to server 20. Server 20 can receive computing power requests sent by user equipment 10.

[0053] Server 20 can execute the computing power adjustment method provided in this application embodiment, adjust the user's computing power resource quota, and then execute the computing task corresponding to the computing power request through the new computing power resources.

[0054] like Figure 2 The diagram shown is a flowchart of a computing power adjustment method provided in an embodiment of this application. The computing power adjustment method may include: S201. In response to a user-triggered computing power request, determine the first quota of computing power resources available to the user.

[0055] Optionally, a computing power request may include a computing task, specifically a request for computing resources initiated by the user to the server through the application layer. This request can inform the task scheduler of the computing power requirements of the current task, which can be represented by the number of tokens or GPU (Graphics Processing Unit) quotas. For example, the computing power requirement of the computing task corresponding to the computing power request could be 200 tokens or 0.1 GPU quotas.

[0056] Optionally, computing resources can refer to resources such as physical GPUs, hard drives, and network cards. The first quota of computing resources can refer to the share / proportion of usable physical GPUs allocated to the user. For example, if a user is initially allocated 0.3 GPU quotas, the first quota is 0.3. The share / proportion can refer to the proportion of physical GPUs usable by the user out of the total physical GPUs. In the formulas of the following embodiments, "quote" represents the first quota.

[0057] GPU quotas can refer to the share allocated to a user from a physical GPU, such as a share of 0.1, 0.25, or 0.5 physical GPU devices, rather than monopolizing the entire expensive physical GPU device. This allows for on-demand allocation, significantly reducing user costs while maximizing resource utilization.

[0058] Figure 3 An example diagram of a physical GPU is shown. For instance, a physical GPU 300 with 40GB of memory can be divided into four GPU quotas: 10%, 2%, 20%, and 68%. This translates to four GPU quotas: 301 (8GB quota), 302 (2GB quota), 303 (10GB quota), and 304 (20GB quota). Each GPU can be housed in a container, and users can configure corresponding containers to utilize their allocated GPU quotas.

[0059] A token is a unit of computing power resource, with each token corresponding to a fixed amount of computing power. A token pool can consist of a preset number of tokens. A quota threshold can be set, which can refer to the upper limit of tokens a single user can claim, or the maximum number of tokens. A complete physical GPU can be divided into several tokens.

[0060] For example, a complete physical GPU can be divided into 1000 tokens. If user A is allocated 0.3 physical GPUs, then user A's token pool includes 300 tokens. That is to say, the first quota allocated to user A corresponds to 300 tokens, and these 300 tokens can form user A's token pool.

[0061] It is also understandable that users can register first to obtain an initial quota of computing power resources. The first quota can be the initial quota or the second quota obtained after the last computing power adjustment.

[0062] Computing power requests can also include user information. Based on the user information, the initial quota of computing power resources allocated to the user can be retrieved, and the token pool corresponding to the initial quota can be obtained.

[0063] S202. Obtain the long-term utilization deviation of computing resources used by users. The long-term utilization deviation is used to characterize the degree of deviation of the user's computing resource quota and load in the time dimension.

[0064] Long-term utilization deviation is used to characterize the degree of deviation between a user's computing resource quota and load within the first time period in the time dimension. The start time of the first time period is the current time, and the duration of the first time period is a preset duration.

[0065] The larger the absolute value of the long-term utilization rate, the higher the degree of deviation; the smaller the absolute value of the long-term utilization rate, the lower the degree of deviation.

[0066] S203. Based on the long-term utilization deviation, determine the appropriate computing power adjustment mode from the long-term adjustment mode and the instantaneous adjustment mode. The long-term adjustment mode is used to control the real-time computing power to converge to the preset computing power quota, and the instantaneous adjustment mode is used to adjust the computing power for sudden traffic.

[0067] The computing power adjustment mode can be either a long-term adjustment mode or an instantaneous adjustment mode.

[0068] like Figure 4a The example graph showing GPU computing power quotas uses curve 401 to represent real-time GPU computing power. Without using long-term or instantaneous adjustment modes, curve 401 exhibits significant instantaneous and mean deviations. Curve 401 shows several peak computing power levels. The average computing power corresponding to curve 401 is 75%, which is greater than the 50% computing power quota.

[0069] like Figure 4b The example graph showing GPU computing power quotas uses curve 402 to represent real-time GPU computing power. After adjusting computing power using either long-term or instantaneous adjustment modes, the mean deviation of curve 402 gradually converges, meaning that the average computing power of 1 and the computing power quota of 50% are gradually approaching each other. In other words, the average computing power of 1 is approaching the computing power quota of 50%. Furthermore, computing power adjustments can be made promptly for sudden fluctuations in computing power.

[0070] S204. Adjust the first quota according to the computing power adjustment mode to obtain the second quota.

[0071] Optionally, the second quota can be obtained by adjusting the computing power of the first quota through a computing power adjustment mode.

[0072] S205. According to the second quota, allocate new computing resources to the user. The new computing resources are used to execute the computing tasks corresponding to the computing request.

[0073] Optionally, if the computing power resource is a physical GPU, a new token pool can be allocated to the user based on the second quota corresponding to the GPU, and the new token pool can be used to execute the computing task corresponding to the computing power request.

[0074] Furthermore, according to the token bucket algorithm, tokens can be allocated from physical GPUs to the token pool according to the second quota to achieve precise control of GPU computing power.

[0075] In this embodiment, in response to a user-triggered computing power request, a first quota corresponding to the computing power resources available to the user can be determined first. Then, the long-term utilization deviation, which characterizes the degree of deviation between the quota and the actual load over a long period, is introduced as the core judgment criterion. This allows for intelligent matching of either a long-term adjustment mode or an instantaneous adjustment mode. The long-term adjustment mode can smoothly converge real-time computing power to a preset baseline quota, correcting deviations from long-term resource idleness or over-allocation. The instantaneous adjustment mode can quickly respond to sudden business traffic and complete temporary computing power adaptation. Furthermore, the first quota is dynamically optimized according to the selected computing power adjustment mode to obtain a precisely adapted second quota. Finally, new computing power resources are allocated to the user according to the second quota to carry the corresponding computing tasks. This achieves an upgrade of computing power quota to a refined dynamic control that combines long-term and short-term effects and load linkage. It can eliminate long-term computing power waste and the risk of continuous overload, while efficiently covering sudden traffic impacts, improving the overall computing power resource utilization, task scheduling stability, and intelligent adaptation capability of quota allocation, and ensuring balanced operation of cluster computing power load.

[0076] As one example, obtaining the long-term utilization deviation caused by the user's use of computing resources includes: Multiple historical utilization rates are obtained from a FIFO (First In First Out) queue. These historical utilization rates characterize the efficiency with which users utilize allocated computing resources. An average utilization rate is calculated based on these historical utilization rates. The long-term utilization deviation is determined by the difference between the average utilization rate and the first quota.

[0077] Optionally, historical utilization can refer to the utilization of the GPU's SM (Streaming Multiprocessor). The SM is the basic independent computing unit ("big core") of the GPU, which contains CUDA (Compute Unified Device Architecture) cores, Tensor cores, schedulers, registers, shared memory, and other modules.

[0078] SM utilization can refer to the percentage of time (or the percentage of active SMs) within a sampling period, specifically used to characterize the actual load intensity and resource utilization efficiency of the GPU computing engine.

[0079] Average utilization rate The calculation formula is as follows:

[0080] in, This represents the utilization rate of the i-th historical period.

[0081] The formula for calculating long-term utilization deviation is:

[0082] in, This refers to the average utilization rate. Refers to the first share.

[0083] Optionally, a low-level interface can be provided to detect the user's utilization rate of the corresponding computing resources and send the utilization rate to a FIFO queue. The length of the FIFO queue can be fixed or dynamically adjusted, but the possibility of the queue growing indefinitely is low. If the FIFO queue is full, the oldest historical utilization rate stored in the FIFO can be deleted, and a new utilization rate can be stored in the queue.

[0084] In this embodiment, when obtaining long-term utilization deviation, multiple historical utilization data can be retrieved in an orderly manner from the FIFO queue. This allows for the systematic aggregation of the actual efficiency of user computing resources usage according to time sequence. The average utilization rate is then calculated based on the average of multiple periodic samples. The difference between this average utilization rate and the preset first quota determines the long-term utilization deviation. Using multiple historical utilization rates as key parameters for subsequent long-term utilization deviation can smooth out instantaneous traffic fluctuations and filter out sudden load interference. This ensures that the long-term utilization deviation can objectively and stably characterize the true deviation between the user's actual computing load and the quota configuration over a long period. This provides accurate and reliable quantitative calculation basis for subsequently differentiating between long-term and short-term adjustment modes and dynamically optimizing computing quotas, improving the stability and data accuracy of computing deviation assessment.

[0085] As an example, based on the long-term utilization deviation, an appropriate computing power adjustment mode is determined from both long-term and instantaneous adjustment modes, including: If the long-term utilization rate deviation exceeds the preset first quota threshold, the long-term adjustment mode will be determined as the computing power adjustment mode.

[0086] If the long-term utilization deviation is less than or equal to the first quota threshold, the instantaneous adjustment mode will be determined as the computing power adjustment mode.

[0087] If the deviation is significant, it indicates that the computing power resource quota and load are not well-matched, frequently resulting in overload or resource shortages. In this case, the long-term adjustment mode can be selected as the computing power adjustment mode. If the deviation is minor, it indicates that the computing power resource quota and load are well-matched, with fewer instances of overload or resource shortages, and the quota and load are relatively balanced. In this case, the instantaneous adjustment mode can be selected.

[0088] The first quota threshold can be determined based on the pre-configured quota threshold and threshold adjustment coefficient.

[0089] If (|utilizationDiffAvg|>α × quota_threshold), the long-term adjustment mode is determined as the computing power adjustment mode. Otherwise, the instantaneous adjustment mode is determined as the computing power adjustment mode.

[0090] Where |utilizationDiffAvg| represents the long-term utilization deviation, α×quota_threshold represents the first quota threshold, quota_threshold represents the pre-configured quota threshold, and α represents the preset threshold adjustment coefficient.

[0091] In this embodiment, a preset first quota threshold is used as the judgment criterion. When the long-term utilization deviation is greater than the first quota threshold, the long-term adjustment mode is precisely activated to continuously correct the long-standing problem of excessive computing power occupation, and promote the actual computing power load to steadily converge to the compliant quota range. When the long-term utilization deviation is less than or equal to the first quota threshold, the instantaneous adjustment mode is switched to only perform lightweight computing power fine-tuning for temporary burst traffic to handle the impact of burst traffic. The intelligent switching and precise adaptation of the two adjustment modes are achieved through threshold linkage. This can not only specifically address the hidden dangers of long-term resource over-provisioning and overload, and continuous crowding of cluster computing power, but also avoid the problems of frequent quota fluctuations and unstable business operation caused by indiscriminate long-term parameter tuning. It takes into account the regularity of long-term computing power configuration and the flexibility of adapting to burst scenarios, so that the computing power control strategy fits the user's real load characteristics, and further improves the accuracy of dynamic quota adjustment and the stability of the overall cluster computing power scheduling.

[0092] like Figure 5The diagram shown is an example of a computing power adjustment method provided in an embodiment of this application. The computing power adjustment method includes: S501, Receive computing power request.

[0093] S502. In response to a user-triggered computing power request, determine the first quota of computing power resources available to the user.

[0094] S503. Obtain the long-term utilization deviation caused by the user's use of computing resources.

[0095] S504. Based on the long-term utilization deviation, determine the appropriate computing power adjustment mode from the long-term adjustment mode and the instantaneous adjustment mode.

[0096] S505. Adjust the first quota according to the computing power adjustment mode to obtain the second quota.

[0097] S506. According to the second quota, allocate new computing resources to the user. The new computing resources are used to execute the computing tasks corresponding to the computing request.

[0098] S507. Determine whether the new computing resources meet the computing power requirements of the computing task corresponding to the computing power request. If yes, proceed to S508; otherwise, proceed to S509.

[0099] S508. Allocate target computing resources from the new computing resources to execute the computing task corresponding to the computing request, so as to utilize the target computing resources to execute the computing task.

[0100] In other words, if the new computing resources meet the computing power requirements of the computing task corresponding to the computing power request, the target computing resources corresponding to the computing task corresponding to the computing power request are allocated from the new computing resources so as to utilize the target computing resources to execute the computing task.

[0101] S509, Blocking computing power requests.

[0102] In other words, if the new computing resources do not meet the computing power requirements of the computing task corresponding to the computing power request, the computing power request will be blocked.

[0103] Optionally, after blocking the computing power request, S509 can update the first quota of computing power resources available to the user to the second quota, so that the computing power resources available to the user are the first quota. Then, S503 can continue to be executed.

[0104] In this embodiment, after receiving a user's computing power request, the user's initial quota is first obtained. Then, long-term utilization deviation is introduced, and the initial quota is optimized and corrected by intelligently matching a long-term adjustment mode or an instantaneous adjustment mode to obtain a second quota. Adapted new computing power resources are generated based on the optimized second quota configuration. Simultaneously, a computing power demand verification step is added. When the new computing power resources can fully support the task's computing power consumption, the target computing power resources are accurately allocated to ensure the normal execution of the computing task. If the resource volume cannot match the demand, the computing power request is blocked and intercepted in a timely manner. This ensures that the entire computing power adjustment process can achieve refined quota iteration by combining historical load deviations, taking into account both normal computing power convergence and the ability to adapt to sudden traffic, and can also prevent operational anomalies caused by insufficient resources through a final verification mechanism. This improves the fit of single-user computing power allocation and resource utilization while strengthening the overall stability, security, and risk control capabilities of computing power scheduling.

[0105] For ease of understanding, such as Figure 6 The diagram shown is an example of a computing power adjustment method provided in an embodiment of this application. The computing power adjustment algorithm may include: 601. Receive computing power request. It can also respond to the computing power request by obtaining the first quota of computing power resources corresponding to the user.

[0106] In addition, multiple historical utilization rates can be obtained from the FIFO queue 602. These historical utilization rates are transmitted to a background thread 603, which performs a series of calculations, such as calculating long-term utilization deviation and quota adjustment coefficients, and selects an appropriate computing power adjustment mode from long-term adjustment mode 604 and instantaneous adjustment mode 605. Based on the computing power adjustment mode, 606 dynamically adjusts the first quota to obtain a second quota. The second quota is then used to update the user's computing power resources, resulting in a new token pool 607 corresponding to the new computing power resources.

[0107] Furthermore, a dynamic scaling factor 608 can be set and adjusted. This dynamic scaling factor can participate in the calculation of the second quota; details can be found in other embodiments and will not be repeated here.

[0108] After the token pool is updated in step 607, step 609 can be executed to determine if the number of the first tokens in the token pool is greater than the number of the second tokens corresponding to the computing power request. If so, step 610 to allow the request. That is, step 611 determines that the number of tokens to be deducted is the second number of tokens, and then steps 612 to deduct the tokens, i.e., deduct the second number of tokens from the token pool in step 607. If not, step 613 to block the computing power request. Then, the process can return to step 601 to continue execution.

[0109] In one possible design, after obtaining new computing resources, the following is also included: Determine the first number of tokens in the token pool of the new computing resource mapping.

[0110] Determine the number of second tokens required for the computing task corresponding to the computing power request.

[0111] If the number of second tokens is less than or equal to the number of first tokens, new computing resources are determined to meet the computing power requirements of the computing task corresponding to the computing power request.

[0112] If the number of second tokens is greater than the number of first tokens, it is determined that the new computing resources do not meet the computing resources required for the computing task corresponding to the computing request.

[0113] In this embodiment, the number of tokens is used to represent computing power resources. By determining the first token quantity corresponding to at least one token in the token pool for a new computing power resource, the new computing power resource is measured using the number of tokens. Furthermore, the second token quantity required by the computing task corresponding to the computing power request is clearly defined, quantifying the token quantity from the demand side. By comparing the first and second token quantities, it is possible to quickly and accurately determine whether the new computing power resource meets the computing power requirements of the computing task. That is, when the second token quantity is less than or equal to the first token quantity, it is directly determined that the requirement is met, avoiding over-allocation of resources. When the second token quantity is greater than the first token quantity, it is determined that the requirement is not met, preventing task execution failure or performance degradation due to insufficient resources. This token quantity comparison method improves the efficiency and accuracy of computing power resource allocation, effectively ensuring that computing tasks can utilize computing power resources reasonably and efficiently.

[0114] As one example, adjusting the first quota according to the computing power adjustment mode to obtain the second quota includes: Determine the quota adjustment coefficient based on the coefficient adjustment strategy corresponding to the computing power adjustment mode.

[0115] Based on the quota adjustment coefficient and the dynamic scaling factor, the first quota is adjusted to obtain the second quota.

[0116] Alternatively, the second quota can be calculated using the following formula: Second quota = Quota adjustment coefficient (First Quota / Dynamic Scaling Factor) In the embodiments of the present application, according to the matching obtained computing power adjustment mode, the exclusive coefficient adjustment strategy is called to accurately generate a quota adjustment coefficient adapted to the current load characteristics, and then the initial first quota is adjusted in combination with the dynamic scaling factor to obtain the second quota. By means of the mode-exclusive policy guarantee coefficient selection, it conforms to the regulation requirements of long-term steady-state deviation correction or instantaneous burst adaptation. Introducing the dynamic scaling factor can further conform to real-time computing power fluctuations, node load conditions and resource margins to achieve flexible scaling, making the quota adjustment have both policy pertinence and real-time dynamic adaptability, greatly improving the fineness and rationality of the iteration optimization from the first quota to the second quota, realizing the precise adjustment of computing power resources, and effectively avoiding problems such as resource waste or adjustment imbalance caused by rigid quota adjustment, excessive amplitude or adaptation deviation.

[0117] It can be understood that the dynamic scaling factor can be adjusted in real time. Specifically, it further includes: When the average utilization rate of the computing power resources by the user is less than the preset quota threshold, the dynamic scaling factor is reduced.

[0118] When the instantaneous utilization rate of the computing power resources by the user is greater than the preset limit threshold, the dynamic scaling factor is increased.

[0119] Among them, the situation where the average utilization rate of the computing power resources by the user is less than the preset quota threshold can be expressed by the following formula: avg < quota - ε, where avg is the average utilization rate, quota - ε is the preset quota threshold, Quota is the first share, and ε is the preset constant threshold.

[0120] The situation where the instantaneous utilization rate of the computing power resources by the user is greater than the preset limit threshold can be expressed by the following formula: cur_spike > limit, where cur_spike is the instantaneous utilization rate and limit is the preset limit threshold.

[0121] In the embodiments of the present application, the dynamic scaling factor is adaptively regulated according to the combination of the average utilization rate and the instantaneous utilization rate of the computing power resources by the user. When the average utilization rate is less than the preset quota threshold, the dynamic scaling factor is actively reduced, which can inhibit the blind expansion of the quota, converge idle redundant computing power, and avoid long-term resource waste caused by inflated quotas. When the instantaneous utilization rate is greater than the preset limit threshold, the dynamic scaling factor can be correspondingly increased, which can quickly amplify the quota adjustment amplitude to adapt to sudden high-load traffic, timely fill the instantaneous computing power gap, and relieve the computing power pressure of the service. By adjusting the dynamic scaling factor, the long-term resource utilization rate and short-term service stability are effectively balanced, making the quota adjustment more conform to the real computing power consumption law of the user, and improving the flexibility and adaptation accuracy of the overall computing power allocation.

[0122] The following combines Figure 7The computing power adjustment architecture that may be applied in the embodiments of this application is described. For example... Figure 7 The diagram illustrates an example of a computing power adjustment architecture. It may include an upper-layer application 701, which can interact with the user, collect the user's computing power requirements, generate a computing power request, and execute 702 to issue the computing power request. In this embodiment, the computing power adjustment method can be implemented through a request scheduling module 703 and a dual-timescale dynamic control module 704.

[0123] The dual-timescale dynamic control module 704 may include a token pool control module 7041 and a scaling factor adjustment module 7042.

[0124] The token pool control module 7041 can adjust the user's computing power resource quota, such as adjusting it from the first quota to the second quota.

[0125] The scaling factor adjustment module 7042 can perform dynamic scaling factor adjustment to apply the adjusted dynamic scaling factor to the token pool control module 7041.

[0126] The request scheduling module 703 may include a resource verification module 7031 and a resource request module 7032.

[0127] The resource verification module 7031 can be used to verify whether the tokens in the token pool can meet the computing power requirements corresponding to the computing power request. If they can, the computing power request is allowed; if they cannot, the computing power request is blocked.

[0128] The resource request module 7032 can allocate resources to the computing task corresponding to the computing power request if the resource verification passes. If the number of tokens to be deducted is determined, the token deduction is executed.

[0129] Furthermore, the various modules can interact and transmit information. After resource verification is passed and computing resources are allocated, computing power requests can be sent to the corresponding upper-layer application 705. The upper-layer application 705 can also monitor GPU utilization and store the utilization in a FIFO queue.

[0130] As an example, when the computing power adjustment mode is a long-term adjustment mode, the quota adjustment coefficient is determined according to the coefficient adjustment strategy corresponding to the computing power adjustment mode, including: Using the number of SMs per GPU, the maximum number of threads per SM, the dynamic scaling factor, and the long-term utilization deviation as input data, and inputting the coefficient adjustment strategy corresponding to the long-term adjustment mode, the quota adjustment coefficient is obtained.

[0131] Optionally, the coefficient adjustment strategy corresponding to the long-term adjustment mode can be expressed using the following formula: incrementAvg=(g_sm_num²×g_max_thread_per_sm) / g_increment_factor×utilizationDiffAvg / 10 Where g_sm_num is the number of SMs per GPU obtained through the CUDA API.

[0132] g_max_thread_per_sm is the maximum number of threads per SM.

[0133] g_increment_factor is the dynamic scaling factor.

[0134] utilizationDiffAvg represents the long-term utilization bias.

[0135] incrementAvg is the calculated quota adjustment coefficient.

[0136] In this embodiment, hardware fundamental parameters such as the number of SMs per GPU and the maximum number of threads per SM, as well as trend parameters such as dynamic scaling factors and long-term utilization deviation, are used as input data. This comprehensively and meticulously covers the key factors affecting GPU computing power allocation and utilization. By inputting this data into the coefficient adjustment strategy corresponding to the long-term adjustment mode, the algorithm and logic within the strategy can perform in-depth analysis and calculation based on actual and complex operating conditions, ultimately obtaining a scientifically reasonable quota adjustment coefficient. This achieves refined management and dynamic optimization of GPU resources, enabling the adjustment of computing power quotas to tend towards controlling real-time computing power convergence to the preset computing power quota, effectively improving resource utilization, avoiding resource waste or insufficiency, and ensuring stable and efficient system operation under high load and dynamic changing environments.

[0137] As another embodiment, when the computing power adjustment mode is instantaneous adjustment mode, the quota adjustment coefficient is determined according to the coefficient adjustment strategy corresponding to the computing power adjustment mode, including: The instantaneous utilization rate deviation is obtained based on the instantaneous utilization rate of computing resources used by the user and the first quota.

[0138] Using the number of SMs per GPU, the maximum number of threads per SM, the dynamic scaling factor, and the instantaneous utilization deviation as input data, and inputting the coefficient adjustment strategy corresponding to the instantaneous adjustment mode, the quota adjustment coefficient is obtained.

[0139] Instantaneous utilization deviation can be expressed by the following formula:

[0140] in, This is used to calculate the instantaneous utilization deviation.

[0141] This refers to instantaneous utilization rate. For quotas.

[0142] The coefficient adjustment strategy corresponding to the instantaneous adjustment mode can be expressed by the following formula: increment_Cur=(g_sm_num²×g_max_thread_per_sm / g_increment_factor)× utilizationDiffCur / 10 Where g_sm_num is the number of SMs per GPU obtained through the CUDA API.

[0143] g_max_thread_per_sm is the maximum number of threads per SM.

[0144] g_increment_factor is the dynamic scaling factor.

[0145] utilizationDiffCur represents the instantaneous utilization rate deviation.

[0146] increment_cur is the calculated quota adjustment coefficient.

[0147] In this embodiment, the instantaneous utilization deviation is obtained based on the instantaneous utilization rate and the first quota. Then, hardware fundamental parameters such as the number of SMs per GPU and the maximum number of threads per SM, as well as parameters reflecting the GPU's operating status and resource usage fluctuations in real time, such as the dynamic scaling factor and the instantaneous utilization deviation, are used as input data. This comprehensively and meticulously covers the key factors affecting GPU computing power allocation and utilization. Inputting these parameters into the coefficient adjustment strategy corresponding to the instantaneous adjustment mode allows for rapid response to instantaneous GPU load changes. Based on real-time data and preset strategies, instantaneous computing power is adjusted to ensure that quota adjustment coefficients are provided to match the rapidly changing instantaneous computing power demand. This achieves rapid adjustment of computing power resources and avoids performance bottlenecks or resource idleness.

[0148] like Figure 8 The diagram shown is a signaling diagram of a computing power allocation method provided in an embodiment of this application. The computing power allocation method may include: 800. Continuous cyclic FIFP sequence with time T0-T1 (e.g., each time slice T=10ms).

[0149] 801. When the GPU task finishes execution, send the current SM utilization to the host sampling module.

[0150] 802. The host sampling module submits the current SM utilization value to the FIFO queue.

[0151] 803. The central scheduling engine calculates long-term utilization deviation.

[0152] 804. The central scheduling engine calculates the instantaneous utilization deviation.

[0153] 805. The central scheduling engine selects the long-term adjustment mode and sends it to the dynamic token controller so that the dynamic token controller can determine the quota adjustment coefficient according to the coefficient adjustment strategy corresponding to the long-term adjustment mode.

[0154] 806. The central scheduling engine selects the instantaneous adjustment mode and sends it to the dynamic token controller, so that the dynamic token controller can determine the quota adjustment coefficient according to the coefficient adjustment strategy corresponding to the instantaneous adjustment mode.

[0155] 807. The dynamic token controller returns the adjusted second quota to the central scheduling engine.

[0156] 808. The central scheduling engine adjusts the dynamic scaling factor.

[0157] 809. The central scheduling engine uses dynamic scaling factors and quota adjustment coefficients to adjust the first quota and obtain the second quota.

[0158] 810. The central scheduling engine issues a second quota to the GPU task so that the GPU task can be allocated the computing resources corresponding to the second quota.

[0159] 811. GPU tasks execute computing power requests until the task is completed or the computing power resources are preempted by higher priority tasks.

[0160] It is understandable that the aforementioned GPU tasks, host sampling modules, central scheduling engines, and dynamic token controllers are all computer program modules configured in the computing device to perform corresponding operation controls.

[0161] like Figure 9 As shown, a schematic diagram of a computing power adjustment device 900 is provided, which may include: The request response unit 901 is used to respond to a computing power request triggered by a user and determine the first quota of computing power resources available to the user.

[0162] The long-term analysis unit 902 is used to obtain the long-term utilization deviation caused by the user's use of computing resources. The long-term utilization deviation is used to characterize the degree of deviation of the user's computing resource quota and load in the time dimension.

[0163] The mode selection unit 903 is used to determine the appropriate computing power adjustment mode from the long-term adjustment mode and the instantaneous adjustment mode based on the long-term utilization deviation. The long-term adjustment mode is used to control the real-time computing power to converge to the preset computing power quota, and the instantaneous adjustment mode is used to adjust the computing power for burst traffic.

[0164] The quota adjustment unit 904 is used to adjust the first quota according to the computing power adjustment mode to obtain the second quota.

[0165] Resource allocation unit 905 is used to allocate new computing resources to users according to the second quota. The new computing resources are used to execute the computing tasks corresponding to the computing request.

[0166] As one embodiment, the long-term analysis unit 902 includes: The historical acquisition module is used to obtain multiple historical utilization rates from the FIFO queue. The historical utilization rate is used to characterize the efficiency of the user in utilizing the allocated computing resources.

[0167] The average calculation module is used to calculate the average utilization rate based on multiple historical utilization rates.

[0168] The deviation determination module is used to determine the long-term utilization deviation based on the difference between the average utilization rate and the first quota.

[0169] As another embodiment, the mode selection module is specifically used to: determine the long-term adjustment mode as the computing power adjustment mode when the long-term utilization deviation is greater than a preset first quota threshold; and determine the instantaneous adjustment mode as the computing power adjustment mode when the long-term utilization deviation is less than or equal to the first quota threshold.

[0170] As another embodiment, the quota adjustment unit 904 includes: The coefficient determination module is used to determine the quota adjustment coefficient based on the coefficient adjustment strategy corresponding to the computing power adjustment mode.

[0171] The quota determination module is used to adjust the first quota based on the quota adjustment coefficient and the dynamic scaling factor to obtain the second quota.

[0172] As yet another embodiment, it also includes: The dynamic scaling adjustment unit is used to decrease the dynamic scaling factor when the average utilization rate of computing resources by a user is less than a preset quota threshold, and to increase the dynamic scaling factor when the instantaneous utilization rate of computing resources by a user is greater than a preset limit threshold.

[0173] As another embodiment, when the computing power adjustment mode is the long-term adjustment mode, the coefficient determination module is specifically used to: take the number of SMs per GPU, the maximum number of threads per SM, the dynamic scaling factor, and the long-term utilization deviation as input data, input the coefficient adjustment strategy corresponding to the long-term adjustment mode, and obtain the quota adjustment coefficient.

[0174] As another embodiment, when the computing power adjustment mode is instantaneous adjustment mode, the coefficient determination module is specifically used to: obtain the instantaneous utilization rate deviation based on the instantaneous utilization rate of computing power resources used by the user and the first quota. Using the number of SMs per GPU, the maximum number of threads per SM, the dynamic scaling factor, and the instantaneous utilization rate deviation as input data, the coefficient adjustment strategy corresponding to the instantaneous adjustment mode is input to obtain the quota adjustment coefficient.

[0175] As yet another embodiment, it also includes: The first allocation unit is used to allocate target computing resources from the new computing resources to execute the computing task corresponding to the computing request, provided that the new computing resources meet the computing power requirements of the computing task corresponding to the computing request.

[0176] The second allocation unit is used to block computing requests when new computing resources do not meet the computing power requirements of the computing task corresponding to the computing request.

[0177] As yet another embodiment, it also includes: The token deduction unit is used to determine the first number of tokens in the token pool of the new computing power resource mapping. It also determines the second number of tokens required for the computing task corresponding to the computing power request. If the second number of tokens is less than or equal to the first number of tokens, it is determined that the new computing power resources meet the computing power requirements of the computing task corresponding to the computing power request. If the second number of tokens is greater than the first number of tokens, it is determined that the new computing power resources do not meet the computing power requirements of the computing task corresponding to the computing power request.

[0178] In the embodiments of this application, Figure 9 The device shown can also be a chip or a chip system, such as a system on chip (SoC) or a baseboard management controller (BMC).

[0179] Figure 10 This is a hardware block diagram of a computing device provided in an embodiment of this application. The computing device 1000 according to an embodiment of this application includes at least a memory 1001 and a processor 1002. The memory 1001 is used to store computer programs. The processor 1002 is used to execute the computer programs to implement the computing power adjustment method of any of the above embodiments.

[0180] In addition, both the memory 1001 and the processor 1002 are electrically connected to the bus 1003.

[0181] Furthermore, embodiments of this application also provide a computer-readable storage medium for storing a computer program. When executed by a processor, the computer program implements the computing power adjustment method of any of the preceding embodiments of this application.

[0182] Furthermore, embodiments of this application also provide a computer-readable storage medium for storing a computer program. When executed by a processor, the computer program implements the computing power adjustment method of any of the preceding embodiments of this application.

[0183] Computer-readable storage media include, but are not limited to, volatile storage media and / or non-volatile storage media. Volatile storage media may include, for example, random access storage media (RAM) and / or cache storage media. Non-volatile storage media may include, for example, read-only storage media (ROM), hard disks, flash memory, optical disks, magnetic disks, etc.

[0184] This application also provides a computer program product, including a BIOS program, which, when executed by a processor, implements the computing power adjustment method of any of the preceding embodiments of this application.

[0185] The basic principles of the embodiments of this application have been described above with reference to specific examples. However, it should be noted that the advantages, benefits, and effects mentioned in the embodiments of this application are merely examples and not limitations, and should not be considered as essential features of each embodiment of this application. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the embodiments of this application from necessarily employing the aforementioned specific details.

[0186] The block diagrams of devices, apparatuses, devices, and systems involved in the embodiments of this application are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context explicitly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0187] Additionally, as used herein, the "or" used in a list of items beginning with "at least one" indicates a separate list, such that a list of, for example, "at least one of A, B, or C" means A or B or C, or AB or AC or BC, or ABC (i.e., A and B and C). Furthermore, the word "exemplary" does not imply that the described example is preferred or better than other examples.

[0188] It should also be noted that in the systems and methods of this application embodiment, each component or step can be decomposed and / or recombined. These decompositions and / or recombinations should be considered as equivalent solutions of the embodiments of this application.

[0189] Various changes, substitutions, and modifications can be made to the technology herein without departing from the teachings defined by the appended claims. Furthermore, the scope of the claims of the embodiments of this application is not limited to the specific aspects of the processes, machines, manufactures, events, means, methods, and actions described above. Currently existing or later-developed processes, machines, manufactures, events, means, methods, or actions that perform substantially the same function or achieve substantially the same result as the corresponding aspects herein can be utilized. Therefore, the appended claims include such processes, machines, manufactures, events, means, methods, or actions within their scope.

[0190] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use embodiments of this application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of embodiments of this application. Therefore, embodiments of this application are not intended to be limited to the aspects shown herein, but rather to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0191] The above description has been given for illustrative and descriptive purposes. Furthermore, this description is not intended to limit the embodiments of this application to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.

Claims

1. A method for adjusting computing power, characterized in that, include: In response to a user-triggered computing power request, a first quota of computing power resources available to the user is determined; The long-term utilization deviation of the computing resources used by the user is obtained, and the long-term utilization deviation is used to characterize the degree of deviation of the user's computing resource quota and load in the time dimension. Based on the long-term utilization deviation, a suitable computing power adjustment mode is determined from the long-term adjustment mode and the instantaneous adjustment mode. The long-term adjustment mode is used to control the real-time computing power to converge to the preset computing power quota, and the instantaneous adjustment mode is used to adjust the computing power for burst traffic. According to the computing power adjustment mode, the first quota is adjusted to obtain the second quota; According to the second quota, new computing resources are allocated to the user, and the new computing resources are used to execute the computing tasks corresponding to the computing request.

2. The method according to claim 1, characterized in that, The acquisition of the long-term utilization deviation caused by the user's use of the computing resources includes: Multiple historical utilization rates are obtained from the first-in-first-out (FIFO) queue, and these historical utilization rates are used to characterize the user's efficiency in utilizing the allocated computing resources. Calculate the average utilization rate based on the aforementioned historical utilization rates; The long-term utilization deviation is determined based on the difference between the average utilization rate and the first quota.

3. The method according to claim 1, characterized in that, The step of determining the appropriate computing power adjustment mode from the long-term adjustment mode and the instantaneous adjustment mode based on the long-term utilization deviation includes: If the long-term utilization deviation is greater than the preset first quota threshold, the long-term adjustment mode will be determined as the computing power adjustment mode. If the long-term utilization deviation is less than or equal to the first quota threshold, the instantaneous adjustment mode is determined as the computing power adjustment mode.

4. The method according to claim 1, characterized in that, The step of adjusting the first quota according to the computing power adjustment mode to obtain the second quota includes: The quota adjustment coefficient is determined based on the coefficient adjustment strategy corresponding to the computing power adjustment mode. Based on the quota adjustment coefficient and the dynamic scaling factor, the first quota is adjusted to obtain the second quota.

5. The method according to claim 4, characterized in that, Also includes: If the average utilization rate of the computing resources by the user is less than the preset quota threshold, the dynamic scaling factor is reduced. If the instantaneous utilization rate of the computing resources by the user exceeds a preset limit threshold, the dynamic scaling factor is increased.

6. The method according to claim 4, characterized in that, When the computing power adjustment mode is a long-term adjustment mode, determining the quota adjustment coefficient according to the coefficient adjustment strategy corresponding to the computing power adjustment mode includes: Using the number of streaming multiprocessors (SMs) per GPU, the maximum number of threads per SM, the dynamic scaling factor, and the long-term utilization deviation as input data, the coefficient adjustment strategy corresponding to the long-term adjustment mode is input to obtain the quota adjustment coefficient.

7. The method according to claim 4, characterized in that, When the computing power adjustment mode is an instantaneous adjustment mode, determining the quota adjustment coefficient according to the coefficient adjustment strategy corresponding to the computing power adjustment mode includes: The instantaneous utilization rate deviation is obtained based on the instantaneous utilization rate of the computing power resources used by the user and the first quota; Using the number of SMs per GPU, the maximum number of threads per SM, the dynamic scaling factor, and the instantaneous utilization deviation as input data, the coefficient adjustment strategy corresponding to the instantaneous adjustment mode is input to obtain the quota adjustment coefficient.

8. The method according to any one of claims 1-7, characterized in that, After allocating new computing resources to the user according to the second quota, the process further includes: If the new computing power resources meet the computing power requirements of the computing task corresponding to the computing power request, a target computing power resource for executing the computing task corresponding to the computing power request is allocated from the new computing power resources, so as to use the target computing power resource to execute the computing task. If the new computing resources do not meet the computing power requirements of the computing task corresponding to the computing power request, the computing power request will be blocked.

9. The method according to claim 8, characterized in that, Also includes: Determine the first number of tokens in the token pool of the new computing power resource mapping; Determine the number of second tokens required for the computing task corresponding to the computing power request; If the number of the second tokens is less than or equal to the number of the first tokens, it is determined that the new computing power resources meet the computing power requirements of the computing task corresponding to the computing power request. If the number of the second tokens is greater than the number of the first tokens, it is determined that the new computing power resources do not meet the computing power resources required for the computing task corresponding to the computing power request.

10. A computing device, characterized in that, include: Memory and processor; The memory is used to store computer programs, which are invoked by the processor to execute any of the above-described computing power adjustment methods.

Images