Methods and apparatus for assessing computing power requirements in intelligent computing power business scenarios
By analyzing the concurrent requests and data processing characteristics of intelligent computing services, the storage and computing power requirements are dynamically assessed, and a precise resource allocation plan is generated. This solves the problem of resource mismatch in existing technologies and improves the utilization efficiency and scheduling accuracy of computing clusters.
Patent Information
- Application Number
- CN202611125873.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-28
- Publication Date
- 2026-11-17
AI Technical Summary
Existing methods for assessing computing power demand rely on static, fixed configurations based on historical extreme points. These methods fail to detect load fluctuations and data processing changes in intelligent computing services, leading to resource idleness or mismatch and reducing the utilization efficiency of computing power clusters.
By analyzing the storage state evolution based on the timing characteristics of concurrent requests and the distribution of data processing length, and combining the available resource boundaries of the computing cluster with the business processing load, computing load fusion and hardware resource combination are performed to generate a precise computing resource configuration scheme, ensuring dynamic matching of the coupling relationship between storage and computing power.
It enables dynamic adaptation of computing resources, avoiding resource idleness or mismatch caused by static peak configuration, and improving the overall utilization efficiency and resource scheduling accuracy of the computing cluster.
Smart Images

Figure CN123114646A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and apparatus for assessing computing power requirements for intelligent computing power business scenarios. Background Technology
[0002] With the convergence of cloud computing and edge computing, intelligent computing services (such as large-scale model inference and real-time video analytics) place extremely high demands on the resource scheduling of underlying infrastructure. Existing methods for assessing computing power requirements typically employ a static peak allocation method. This method uses resource consumption data recorded during the highest load moments of a business system's historical operation, such as maximum CPU utilization and maximum memory usage, as the fixed computing power allocation standard during system deployment or expansion. For example, if the highest concurrent requests for a certain intelligent video analytics service in the past month resulted in GPU memory usage reaching 80%, the existing method would directly lock and allocate 80% of the GPU memory resources to that service instance.
[0003] However, the load of intelligent computing services exhibits significant tidal effects and bursts, and the duration of storage resource occupation and release rhythm vary significantly depending on the length of input data. Existing methods, relying solely on historical extreme points for static configuration, fail to perceive the dynamic evolution of resource occupation caused by fluctuations in concurrent request timing and changes in data processing length during intelligent computing service operation. This easily overlooks the dynamic coupling relationship between storage status and computing load during service operation, resulting in a large amount of allocated computing resources being idle or operating inefficiently due to a lack of corresponding dynamic storage support during off-peak periods or short data processing scenarios. Conversely, in long data burst scenarios, the mismatch between storage resource allocation and actual data flow may lead to computing power waiting, resulting in serious waste of computing resources and reducing the overall utilization efficiency of the computing cluster. Summary of the Invention
[0004] This invention provides a method and apparatus for evaluating computing power requirements in intelligent computing power business scenarios. It enables precise adaptation based on the dynamic coupling relationship between storage and computing power during business operation and the actual collaborative efficiency of hardware, avoiding resource idleness or mismatch caused by static peak configuration, and improving the overall utilization efficiency and resource scheduling accuracy of computing power clusters in intelligent computing power business scenarios.
[0005] In a first aspect, the present invention provides a method for assessing computing power requirements for intelligent computing power service scenarios, including: Based on the concurrent request timing characteristics and data processing length distribution of intelligent computing services, runtime storage state evolution analysis is performed to obtain the storage resource occupancy status of future time windows. Based on the storage resource occupancy status, the available storage resource boundary of the computing power cluster, and the basic load of business processing, the computing power load is fused to obtain the scenario-based total computing power load including storage loss. Based on the scenario-based total computing power load and the heterogeneous hardware topology of the computing power cluster, security resource boundaries are defined to obtain candidate hardware resource combinations that meet business continuity requirements. Based on the network interconnection bandwidth characteristics of the candidate hardware resource combinations and computing power clusters, an evolution analysis of communication collaboration efficiency is performed to obtain the effective computing power output index under different hardware resource combinations. Based on the effective computing power output index and the response time constraints of intelligent computing power services, a target computing power resource allocation scheme is generated.
[0006] Secondly, the present invention also provides a computing power demand assessment device for intelligent computing power service scenarios, applied to the computing power demand assessment method for intelligent computing power service scenarios as described in the first aspect; the computing power demand assessment device for intelligent computing power service scenarios includes: The storage status prediction module is used to perform runtime storage status evolution analysis based on the concurrent request timing characteristics and data processing length distribution of intelligent computing power services, and to obtain the storage resource occupancy status of future time windows. The scenario-based load fusion module is used to fuse computing loads based on the storage resource occupancy status, the available storage resource boundary of the computing power cluster, and the basic load of business processing, to obtain the scenario-based total computing load including storage loss. The security boundary delineation module is used to delineate security resource boundaries based on the scenario-based total computing power load and the heterogeneous hardware topology of the computing power cluster, and obtain candidate hardware resource combinations that meet business continuity requirements. The collaborative efficiency evolution module is used to perform communication collaborative efficiency evolution analysis based on the candidate hardware resource combinations and the network interconnection bandwidth characteristics of the computing power cluster, and to obtain the effective computing power output index under different hardware resource combinations. The timeliness adaptation decision module is used to generate a target computing power resource configuration scheme based on the effective computing power output index and the response timeliness constraints of intelligent computing power services.
[0007] Thirdly, the present invention also provides an electronic device, comprising: a memory for storing computer software programs; and a processor for reading and executing the computer software programs, thereby realizing the computing power demand assessment method for intelligent computing power business scenarios as described above.
[0008] Fourthly, the present invention also provides a non-transitory computer-readable storage medium storing a computer software program, which, when executed by a processor, implements the computing power demand assessment method for intelligent computing power service scenarios as described above.
[0009] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the computing power demand assessment method for intelligent computing power business scenarios as described above.
[0010] The computing power demand assessment method for intelligent computing power service scenarios provided in this invention analyzes the runtime storage state evolution of concurrent request timing characteristics and data processing length distribution of intelligent computing power services. This yields a storage resource occupancy status for future time windows that reflects the dynamic changes in the business, overcoming the shortcomings of existing technologies that rely solely on historical static peak values and cannot perceive the impact of data length changes on resource occupancy rhythm. The method integrates computing power load based on storage resource occupancy status, the available storage resource boundaries of the computing power cluster, and the basic load of business processing. This transforms dynamically changing storage demands into a scenario-based total computing power load including storage losses, achieving a shift from a single-dimensional static indicator to a multi-dimensional dynamically coupled load, ensuring real-time matching between computing power demand assessment and actual business data flow. Finally, it defines security resource boundaries based on the scenario-based total computing power load and the heterogeneous hardware topology of the computing power cluster. The system selects candidate hardware resource combinations that meet business continuity requirements, providing a physical resource foundation for dynamic load. Based on the network interconnection bandwidth characteristics of the candidate hardware resource combinations and the computing power cluster, it conducts communication collaboration efficiency evolution analysis, quantifies the data throughput and computing collaboration capabilities of different hardware combinations in a real network environment, and obtains the effective computing power output index under different hardware resource combinations. This solves the problem of the disconnect between theoretical computing power and actual output caused by ignoring communication overhead and hardware differences. Based on the effective computing power output index and the response time constraints of intelligent computing power services, a target computing power resource configuration scheme is generated. Ultimately, it achieves precise adaptation based on the dynamic coupling relationship between storage and computing power during service operation and the actual hardware collaboration efficiency, avoiding resource idleness or mismatch caused by static peak configuration, and improving the overall utilization efficiency and resource scheduling accuracy of the computing power cluster in intelligent computing power service scenarios. Attached Figure Description
[0011] Figure 1 This is a flowchart illustrating the computing power demand assessment method for intelligent computing power service scenarios provided in this embodiment of the invention. Figure 2 This is a schematic diagram of the structure of the computing power demand assessment device for intelligent computing power business scenarios provided in an embodiment of the present invention; Figure 3 An embodiment diagram of the electronic device provided in this invention; Figure 4 An embodiment diagram of a computer-readable storage medium provided in accordance with the present invention. Detailed Implementation
[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0013] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.
[0014] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.
[0015] See Figure 1 , Figure 1 This is a flowchart illustrating the computing power demand assessment method for intelligent computing power service scenarios provided by the present invention. In this embodiment of the invention, the execution entity of the computing power demand assessment method for intelligent computing power service scenarios is an assessment device. Therefore, the computing power demand assessment method for intelligent computing power service scenarios includes: Step 10: Based on the concurrent request timing characteristics and data processing length distribution of intelligent computing services, perform runtime storage state evolution analysis to obtain the storage resource occupancy status of future time windows.
[0016] Optionally, the evaluation device captures business request stream data within a future time window in real time from the request access gateway or log collection of intelligent computing power services, and parses the concurrent request timing characteristics and data processing length distribution from this data. The concurrent request timing characteristics refer to the absolute timestamp sequence of all arriving intelligent computing power service requests within the future time window, as well as the statistical measure of the time interval between adjacent requests; the data processing length distribution refers to the set of input data scale parameters carried by each service request, which, depending on the service type, are specifically represented by at least one of the following: the number of tokens, the total number of image pixels, the number of video frames, or the number of time-series data points.
[0017] After obtaining the timing characteristics of concurrent requests and the distribution of data processing lengths, the evaluation device simulates the dynamic occupancy of storage resources within a future time window. During the simulation, it not only considers the accumulation of basic data volume but also focuses on quantifying the risk of storage access failures caused by low storage access alignment efficiency and concurrency conflicts, and deriving the additional resource residency latency caused by reduced storage efficiency. Finally, it superimposes the basic occupancy with latency inflation to generate a storage resource occupancy trend reflecting the effect of processing efficiency degradation. This is specifically illustrated in steps 101 to 105, and the generated storage resource occupancy trend refers to the storage resource occupancy trend curve that changes over time within a future time window.
[0018] Step 20: Based on the storage resource occupancy status, the available storage resource boundary of the computing power cluster, and the basic load of business processing, the computing power load is fused to obtain the scenario-based total computing power load including storage loss.
[0019] Optionally, after determining the storage resource occupancy status, the evaluation device reads the available storage resource boundaries of the computing cluster from the resource manager or hardware monitoring agent of the computing cluster, and extracts the basic workload of the business processing from the model configuration file of the intelligent computing service. Here, the available storage resource boundaries represent storage resources within the cluster that are not currently monopolized by other high-priority tasks and can be supplied to this intelligent computing service. Specifically, this includes the physical capacity limit of the storage medium, the remaining allocable bandwidth threshold, and the IOPS (input / output operations per second) quota. The basic workload of the business processing refers to the theoretical total computation (in floating-point operations) required to complete a single inference or training task of the intelligent computing service under ideal lossless conditions, as well as the total duration of the future time window.
[0020] Based on this, the evaluation device identifies high-pressure periods of storage resource scarcity by combining the obtained storage resource occupancy status with the available storage resource boundaries of the computing cluster and the basic load of business processing. Within these periods, it quantifies the resulting local storage access losses, converting these losses into equivalent computational downtime, and then derives the additional computing power compensation value required to compensate for these downtimes. Finally, it merges this compensation with the theoretical baseline computing power value to obtain a scenario-based total computing power load encompassing the impact of storage performance bottlenecks, as detailed in steps 201 to 205. This scenario-based total computing power load is a computing power demand sequence with a time dimension, reflecting the dynamic total demand for computing resources at different times within a future time window.
[0021] Step 30: Based on the scenario-based total computing load and the heterogeneous hardware topology of the computing cluster, define the security resource boundaries to obtain candidate hardware resource combinations that meet business continuity requirements.
[0022] Optionally, after receiving the scenario-based total computing power load, the evaluation device begins to acquire the heterogeneous hardware topology of the computing power cluster. This topology describes relevant information about all available hardware nodes within the computing power cluster, specifically including the physical connection relationships between hardware nodes, node type identifiers, type attribute parameters of each node, and hierarchical relationships between nodes. The node type attribute parameters include, but are not limited to, the number of CPU cores, GPU memory capacity, peak computing power of tensor processing units, memory bandwidth limit, and the precision types supported by the node.
[0023] Subsequently, the evaluation device first performs single-node computing power capacity mapping. It traverses each independent hardware node in the heterogeneous hardware topology, extracting the node type attribute parameters of that independent hardware node. Based on the model structure characteristics of the intelligent computing power service, it determines the minimum precision support and operator compatibility constraints required for the service. The peak computing power requirement in the scenario-based total computing power load is compared with the peak computing power of the tensor processing unit of the independent hardware node. If the peak computing power of the tensor processing unit of the independent hardware node is less than the peak computing power requirement in the scenario-based total computing power load, the independent hardware node is marked as an unavailable single-node node; if it is greater than or equal to, it is marked as a potentially available single-node node. For potentially available single-node nodes, the evaluation device further verifies whether its memory capacity meets the maximum corrected resource occupancy within the future time window in step 1053. If it does, the independent hardware node is added to the basic available node pool.
[0024] Next, the evaluation device generates two types of initial hardware resource combinations based on the basic available node pool: single-node combinations and multi-node combinations. A single-node combination uses each independent hardware node in the basic available node pool as a separate initial hardware resource combination; a multi-node combination selects two or more independent hardware nodes from the basic available node pool to form an initial hardware resource combination. For multi-node combinations, the evaluation device also checks whether there are physical links or indirect paths via switches between all independent hardware nodes within the combination. If any nodes are disconnected, the multi-node combination is removed; if they are connected, it is retained.
[0025] After obtaining the initial set of hardware resource combinations, the evaluation device verifies the performance of each combination. For a single-node combination, its aggregate computing power limit is the peak computing power of the tensor processing unit of that node; for a multi-node combination, the aggregate computing power limit is calculated as the sum of the peak computing power of the tensor processing unit of all independent hardware nodes in the combination multiplied by a preset parallel speedup coefficient. If the aggregate computing power limit of an initial hardware resource combination is less than the peak computing power requirement in the scenario's total computing power load, the combination is removed from the set; if it is greater than or equal to the peak computing power requirement, it is retained.
[0026] Furthermore, the evaluation device also needs to define the security resource boundary based on the retained initial hardware resource combination set. The process is as follows: First, a minimum resource redundancy threshold is set within the security resource boundary to meet business continuity requirements. This business continuity requirement stipulates that even if a single hardware node fails, the remaining hardware nodes can still maintain a minimum level of service responsiveness. Therefore, the evaluation device verifies the redundancy of the retained initial hardware resource combination set. If the combination is a single-node combination, it determines whether the single node supports hot backup or rapid migration mechanisms; if not, it is directly removed. If the combination is a multi-node combination, it simulates the removal of any independent hardware node in the combination and calculates the lower limit of the aggregate computing power of the remaining nodes. If the lower limit of the aggregate computing power of the remaining nodes is still greater than or equal to the basic load component in the scenario-based total computing power load, then the combination meets the security resource boundary requirements and is considered a compliant combination; otherwise, it is removed. Here, the lower limit of aggregate computing power refers to the minimum guaranteed computing power that the remaining nodes can provide after deducting communication overhead under the worst-case single-point failure scenario. The basic load component refers to the theoretical baseline computing power value after removing the equivalent compensation computing power value.
[0027] Finally, the evaluation device performs resource fragmentation filtering on the compliant combinations. It checks the resource idle status of individual hardware nodes within a compliant combination at the current moment. If an individual hardware node is already locked and occupied by other high-priority services, and its remaining idle computing power is insufficient to share the corresponding share of the scenario-based total computing power load, then the combination containing that individual hardware node is removed from the compliant combination. The remaining compliant combinations are then integrated to obtain candidate hardware resource combinations. That is, the candidate hardware resource combination is a list of hardware nodes that meet the constraints of computing power requirements, topology connectivity, security redundancy, and current resource availability.
[0028] For example, assuming the intelligent computing power service is intelligent video analytics, its peak total computing power load requirement is 500 trillion operations per second, with a GPU memory requirement of 160GB. The basic load component is 300 trillion operations per second. The safety redundancy requirement is: after any single node failure, the minimum aggregate computing power of the remaining nodes must be ≥300 TFLOPS. The parallel speedup coefficient is: 0.9 for a two-node combination, 0.8 for a three-node combination, and 0.75 for four nodes or more (simulated communication overhead increases with the number of nodes).
[0029] The evaluation device acquires the heterogeneous hardware topology of the computing cluster, including nodes A, B, C, D, and E. Specifically, node A has 200 TFLOPS of computing power and 80GB of GPU memory, currently idle; node B has 200 TFLOPS of computing power and 80GB of GPU memory, currently idle; node C has 150 TFLOPS of computing power and 60GB of GPU memory, currently idle; node D has 300 TFLOPS of computing power and 120GB of GPU memory, currently idle; and node E has 100 TFLOPS of computing power and 40GB of GPU memory, currently idle. The connections are as follows: A is directly connected to B, B is directly connected to C, C is directly connected to D, D is directly connected to E, and A and D are connected via a switch.
[0030] Assuming all nodes (AEs) are compatible with model accuracy, the basic available node pool is {A, B, C, D, E}. Single node combinations are: {A}, {B}, {C}, {D}, {E}, discarding any that do not meet the requirements. Multi-node combinations are: {A, B}: 80+80=160GB VRAM, connected. {A, D}: 80+120=200GB VRAM, connected. {B, C, D}: 80+60+120=260GB VRAM, connected. {A, B, D}: 80+80+120=280GB VRAM, connected. {C, D, E}: 60+120+40=220GB VRAM, connected.
[0031] In multi-node combinations: {A, B}: Maximum aggregate computing power = (200+200)×0.9 = 360 TFLOPS. 360 < 500, discard. {A, D}: Maximum aggregate computing power = (200+300)×0.9 = 450 TFLOPS. 450 < 500, discard. {C, D, E}: Maximum aggregate computing power = (150+300+100)×0.8 = 440 TFLOPS. 440 < 500, discard. {A, B, D}: Maximum aggregate computing power = (200+200+300)×0.8 = 560 TFLOPS. 560 ≥ 500, retain. {B, C, D}: Maximum aggregate computing power = (200+150+300)×0.8 = 520 TFLOPS. 520 ≥ 500, retain.
[0032] When performing security resource boundary delineation (redundancy verification), for the retained combination {A, B, D}: Removing A: Remaining computing power of {B, D} = (200 + 300) × 0.9 = 450 TFLOPS. Removing B: Remaining computing power of {A, D} = (200 + 300) × 0.9 = 450 TFLOPS. Removing D: Remaining computing power of {A, B} = (200 + 200) × 0.9 = 360 TFLOPS. The lower limit of aggregate computing power = min(450, 450, 360) = 360 TFLOPS. 360 ≥ 300 (basic load component), satisfying the security boundary. Similarly, the verification for the retained combination {B, C, D} also shows that it satisfies the security boundary.
[0033] Assume that nodes A, B, C, D, and E are currently completely idle. All nodes in combinations {A, B, D} and {B, C, D} are unlocked and have sufficient idle computing power. The candidate hardware resource combinations obtained by the evaluation device are: Combination 1: {Node A, Node B, Node D}; Combination 2: {Node B, Node C, Node D}.
[0034] Step 40: Based on the network interconnection bandwidth characteristics of candidate hardware resource combinations and computing power clusters, conduct communication collaboration efficiency evolution analysis to obtain the effective computing power output index under different hardware resource combinations.
[0035] Optionally, after determining the candidate hardware resource combination, the evaluation device obtains the network interconnection bandwidth characteristics of the computing cluster from the network controller or topology discovery service of the computing cluster. These network interconnection bandwidth characteristics include the network interconnection topology and network transmission congestion characteristics. The network interconnection topology refers to the connection method of the physical links between computing nodes in the computing cluster, the rated bandwidth of the links, and the switch hierarchy architecture. Network transmission congestion characteristics refer to parameters reflecting network transmission quality, such as the measured baseline of packet transmission delay, packet loss rate, and congestion window size, either directly measured or based on historical statistics in the current network environment.
[0036] Optionally, the evaluation device, based on the candidate hardware resource combinations and the network interconnection bandwidth characteristics of the acquired computing power cluster, first analyzes the dispersion of computing capabilities among nodes and the matching relationship between business parallel topology and network topology to identify key collaborative interaction moments. Then, it extrapolates the actual collaborative blocking duration caused by differences in computing speed and network transmission obstruction. Finally, it calculates the collaborative efficiency loss coefficient based on the blocking duration and performs nonlinear collapse conversion and standard performance normalization on the theoretical peak computing power to obtain an effective computing power output index reflecting the actual operating effect. This is illustrated in steps 401 to 405. This effective computing power output index, after standardization by resource scheduling granularity constraints, reflects the actual computing power level that the candidate hardware resource combinations can provide in a real network environment.
[0037] Step 50: Based on the effective computing power output index and the response time constraints of intelligent computing power services, generate a target computing power resource allocation scheme.
[0038] Optionally, after receiving the effective computing power output index under different candidate hardware resource combinations, the evaluation device obtains the response time constraint of the intelligent computing service, and also obtains the total computing task volume in the scenario-based total computing power load. The response time constraint refers to the maximum allowed processing time of the intelligent computing service; the computing task volume refers to the theoretical total amount of computation required to complete the intelligent computing service processing within a future time window. Based on this, a timeliness compliance screening is first performed, dividing the total computing task volume by the effective computing power output index of each candidate hardware resource combination to obtain the theoretical processing time of each combination, which is the minimum time required under ideal scheduling, considering only computation and known communication bottlenecks. If the theoretical processing time of a combination is greater than the response time constraint, the combination is determined to be unable to meet the real-time requirements and is eliminated; otherwise, the combination is retained and its timeliness margin is recorded, which is the difference between the response time constraint and the theoretical processing time.
[0039] Subsequently, the evaluation device quantifies the retained candidate hardware resource combinations from two dimensions: energy efficiency cost and stability risk. Regarding energy efficiency cost, it first obtains the energy consumption characteristic parameters of each independent hardware node in the computing cluster, specifically including full-load power consumption and idle power consumption. Then, it calculates the estimated total energy consumption of each retained candidate hardware resource combination during operation. This calculation method is: multiply the full-load power consumption of each independent hardware node within the combination by the theoretical processing time, and add the idle power consumption of other associated nodes that are not participating in the computation but are in standby mode. Simultaneously, it obtains the resource billing strategy of the computing cluster, specifically including unit pricing standards for billing by computing power unit, by memory capacity, and by network bandwidth. Based on the estimated total energy consumption and resource billing strategy, it calculates the operating economic cost of each candidate hardware resource combination.
[0040] To quantify stability risks, the assessment device analyzes the historical failure rate data of individual hardware nodes in each candidate hardware resource combination, i.e., the frequency of hardware errors or software crashes occurring in that type of node over a certain period. The overall failure probability of the combination is calculated, based on a statistical model of the failure rates of individual nodes, and increases with the number of nodes. The overall failure probability is compared to a threshold representing the maximum allowable interruption probability for the service. If the overall failure probability exceeds the threshold, the combination is marked as a high-risk combination.
[0041] The evaluation device comprehensively ranks the candidate hardware resource combinations that have passed the initial screening for timeliness compliance. The ranking is based on the following criteria: first, the margin for meeting response timeliness constraints, with a larger margin indicating stronger resistance to fluctuations; second, the economic cost of operation, with lower costs being preferred; and finally, the stability risk, with lower risks being preferred. A weighted scoring method is used to assign weight coefficients to timeliness margin, economic cost, and stability risk, calculating a comprehensive score for each combination. The candidate hardware resource combination with the highest comprehensive score is selected as the target hardware resource set.
[0042] Finally, based on the target hardware resource set, the evaluation device generates a target computing power resource configuration scheme that includes node identifiers, resource allocation ratios, network routing policies, and storage mount points. This clarifies which physical nodes will be activated, the proportion of computing power resources allocated to each node for the intelligent computing service, and the data flow path between nodes. The target computing power resource configuration scheme is then output to the resource scheduling controller for actual resource allocation.
[0043] For example, assuming candidate hardware resource combinations, combination 1: {node A, node B, node D} has an effective computing power output index of 480 TFLOPS; combination 2: {node B, node C, node D} has an effective computing power output index of 450 TFLOPS. The response time constraint for intelligent computing services is 2.0 seconds. The total computing task volume in the scenario-based total computing power load is 900 trillion floating-point operations.
[0044] The evaluation device first performs an initial screening for timeliness compliance. For combination 1, the theoretical processing time = total computational workload / effective computing power output index = 900 / 480 = 1.875 seconds. Since 1.875 s < 2.0 s, the result is: timeliness constraint is met, retain it. Its timeliness margin = 2.0 - 1.875 = 0.125 seconds. For combination 2, the theoretical processing time = 900 / 450 = 2.0 seconds. Since 2.0 s ≤ 2.0 s, the result is: timeliness constraint is met, retain it. Timeliness margin = 2.0 - 2.0 = 0 seconds.
[0045] When conducting energy efficiency cost assessments, it is assumed that the full-load power consumption of nodes A, B, and D is 1000 watts (W), and that of node C is 800 watts. Therefore, in calculating the estimated total energy consumption of combination 1, the nodes involved are A, B, and D. Its energy consumption = (1000 + 1000 + 1000) W × 1.875 s = 5625 joules (J). In calculating the estimated total energy consumption of combination 2, the nodes involved are B, C, and D. Its energy consumption = (1000 + 800 + 1000) W × 2.0 s = 5600 joules (J). That is, combination 2 has slightly lower energy consumption (5600J < 5625J), making combination 2 slightly better in terms of economic cost.
[0046] When quantifying stability risk, it is assumed that the historical failure rate of each node is 0.1%. The overall failure probability of combination 1 is: 1 - (1 - 0.001). 3 ≈0.2997%. Overall failure probability of combination 2: 1 - (1 - 0.001) 3 The risk is approximately 0.2997%. Both risks are comparable and far below the maximum permissible interruption probability threshold for the business (e.g., 1%). In this embodiment, the weights are set as follows: timeliness margin weight 0.6, economic cost weight 0.3, and stability risk weight 0.1. However, in practical applications, the evaluation device can also pre-set a weight mapping table to automatically match different weight combinations based on the type of intelligent computing service (e.g., real-time inference, offline training, critical transaction processing) to achieve more accurate scenario adaptation.
[0047] When normalizing, the timeliness margin score is calculated. Combination 1 receives a full score (1.0), while Combination 2 receives 0 points (due to a margin of 0, its resistance to fluctuations is the worst). The economic cost score is calculated using a minimum energy consumption of 5600J as a baseline of 1.0. Therefore, Combination 1's score = 5600 / 5625 ≈ 0.996, and Combination 2's score = 1.0. The stability score is the same for both, set to 1.0. Therefore, Combination 1's score = 1.0 × 0.6 + 0.996 × 0.3 + 1.0 × 0.1 = 0.6 + 0.2988 + 0.1 = 0.9988. Combination 2's score = 0 × 0.6 + 1.0 × 0.3 + 1.0 × 0.1 = 0 + 0.3 + 0.1 = 0.4. Combination 1's score is significantly higher than Combination 2's. Although combination 2 is slightly more power-efficient, its timeliness margin is zero. Any minor network fluctuation or load jitter could lead to timeouts, posing a very high risk of violating response timeliness constraints. Therefore, the evaluation device selected combination 1 as the target hardware resource set.
[0048] The evaluation device generates the following target computing power resource configuration scheme: Activated nodes: Node A, Node B, and Node D. Computing power allocation strategy: Node D (with the strongest computing power) undertakes 40% of the computing tasks. Node A and Node B each undertake 30% of the computing tasks. Communication routing strategy: Gradient synchronization data between Node A and Node B is transmitted via a direct link. Data exchange between Node A / D and Node B / D is performed via a high-speed interconnect link. Storage mounting strategy: The high-frequency access hot data blocks predicted in step 10 are preloaded into the 120GB video memory of Node D. The video memory of Node A and Node B is used as a secondary cache to store medium-frequency access data.
[0049] The embodiments of the present invention ultimately achieve precise adaptation based on the dynamic coupling relationship between storage and computing power during business operation and the actual collaborative efficiency of hardware, avoiding resource idleness or mismatch caused by static peak configuration, and improving the overall utilization efficiency and resource scheduling accuracy of computing power clusters in intelligent computing power business scenarios.
[0050] Optionally, the process of steps 101 to 105 includes: Obtain the storage resource usage status for future time windows, including: Step 101: Based on the arrival timestamp sequence of concurrent requests from the intelligent computing power service and the data scale parameter corresponding to each request, perform multi-stage storage access mode analysis to obtain the storage access step size feature sequence of each request in different processing stages.
[0051] Optionally, the evaluation device receives the concurrent request timing characteristics and data processing length distribution obtained in step 10, separates the concurrent request arrival timestamp sequence arranged in time order from the concurrent request timing characteristics, and each element in the sequence corresponds to a unique identifier and arrival time of an independent intelligent computing power service request; then, it retrieves the data scale parameter that matches the unique identifier from the data processing length distribution. This parameter quantifies the size of the input data carried by each request. For example, for large model inference services, it is represented by the number of tokens, and for video analysis services, it is represented by the number of video frames or the total number of pixels.
[0052] Based on the model structure configuration file or algorithm flow definition of the intelligent computing power service, the evaluation device divides the lifecycle of a single request into a preprocessing stage, a core computing stage, and a post-processing stage. The preprocessing stage mainly performs data loading, decoding, normalization, and format conversion operations; the core computing stage mainly performs matrix multiplication, convolution operations, or attention mechanism calculations for the model; and the post-processing stage mainly performs result encoding, serialization, and output buffering operations.
[0053] Therefore, for each of the divided processing stages, the evaluation device analyzes the read / write behavior patterns of the storage medium by the code logic or operator attributes of that stage. It determines the number of bytes of data blocks involved in a single memory or video memory read / write operation in each stage, i.e., the storage access step size. For example, if the preprocessing stage reads data from main memory in units of fixed-size image slices, the storage access step size is the size of that slice in bytes; if the core computation stage reads weight parameters from the cache in units of tensor rows, the storage access step size is the size of that tensor row in bytes. The evaluation device arranges the storage access step sizes of each request in the preprocessing, core computation, and postprocessing stages in the order of execution, forming a storage access step size feature sequence for that request. This characterizes the dynamic change trajectory of the storage medium access granularity of the request throughout the entire processing process.
[0054] Step 102: Based on the storage access step size feature sequence of each request and the preset minimum access granularity parameter of the storage medium, perform storage access alignment efficiency analysis to obtain the theoretical storage access utilization of each request.
[0055] Optionally, the minimum access granularity parameter of the storage medium refers to the smallest data unit size that the underlying storage hardware (such as DRAM, SSD or HBM) can handle in a single physical read / write operation. It is usually the cache line size or page size, such as 64 bytes, 4096 bytes, etc.
[0056] The evaluation device acquires a preset minimum access granularity parameter for the storage medium and calculates the alignment efficiency for each step value in the storage access step size feature sequence of each request. It determines whether each step value is divisible by the minimum access granularity parameter. If divisible and the starting address is aligned, the alignment efficiency of that step is 100%; otherwise, the alignment efficiency is the quotient of the step value divided by (the step value plus the padding overhead corresponding to the remainder), and the ratio of the actual effective data volume transmitted to the total physical access data volume is calculated. The alignment efficiency of the request is weighted and averaged across all processing stages, with the weight being the proportion of data access volume in each stage, to obtain the theoretical storage access utilization rate of the request. This utilization rate characterizes the effective proportion of storage bandwidth used by the request under ideal conditions without concurrency conflicts.
[0057] Step 103: Based on the arrival timestamp sequence of concurrent requests and the expected new data blocks corresponding to each request, the concurrent storage bandwidth requirements are superimposed to obtain the concurrent access conflict index within the future time window.
[0058] Optionally, the evaluation device divides the future time window into several small time slices, for example, one time slice per 1 millisecond. For each time slice, the set of concurrent requests that are active within that time slice is counted. For each active request, based on its current processing stage and storage access step size characteristic sequence, the expected new data block throughput generated by the request within that time slice is estimated. The expected new data block refers to the amount of data that the request needs to read or write from the storage medium at that moment.
[0059] The evaluation device sums the projected new data block throughput of all active requests within the same time slice to obtain the total storage bandwidth requirement for that time slice. This total storage bandwidth requirement is compared to the rated maximum bandwidth of the storage medium. If the total requirement exceeds the rated maximum bandwidth, it indicates bandwidth contention. The proportion of the excess is calculated as the instantaneous conflict intensity for that time slice. Subsequently, the instantaneous conflict intensity of all time slices throughout the entire future time window is integrated or averaged to obtain a concurrent access conflict index. This index is a value between 0 and 1, where 0 represents no conflict and 1 represents severe congestion. It quantifies the potential blocking degree caused by multiple requests simultaneously competing for storage resources.
[0060] Step 104: Based on the theoretical storage access utilization rate and the concurrent access conflict index, a comprehensive calculation of the degree of storage locality disruption is performed to obtain the storage access failure risk factor at each moment within the future time window.
[0061] Optionally, the evaluation device receives the theoretical storage access utilization and the concurrent access conflict index for each request in the future time window. Subsequently, to quantify the degree of storage locality of destruction—the risk that the storage system cannot efficiently provide data due to the combined effects of poor access patterns (low utilization) and resource contention (high conflict)—a preset conflict penalty weight is first determined. This conflict penalty weight represents the proportional coefficient of the sensitivity of concurrent conflicts to storage performance. Based on the hardware characteristic parameters of the storage media in the computing cluster (such as queue depth and bandwidth volatility) and historical performance monitoring data of the intelligent computing service, the value of this weight is determined through offline regression analysis or a lookup table method.
[0062] Next, the evaluation device calculates the basic access cost at each time step. This basic access cost is the complement of the theoretical storage access utilization (i.e., 1 minus the theoretical storage access utilization), representing the inherent performance loss due to low alignment efficiency. Then, the contention escalation cost is calculated, which is the product of the concurrent access conflict index and a preset conflict penalty weight, representing the additional risk caused by bandwidth contention. The basic access cost and the contention escalation cost are added together to obtain the original risk value. Finally, the original risk value is normalized and mapped to the range of 0 to 1, resulting in a storage access failure risk factor. The closer this factor is to 1, the higher the probability of storage access timeout, cache hit failure, or severe queuing congestion occurring at that time. This risk factor is then associated with each time slice of a future time window, forming a storage access failure risk factor sequence.
[0063] Step 105: Based on the storage access failure risk factor, the arrival timestamp sequence of concurrent requests for intelligent computing services, and the estimated basic processing time for each request, perform latency inflation correction and lifecycle reconstruction to obtain the storage resource occupancy status for future time windows.
[0064] Optionally, the evaluation device performs latency inflation correction and lifecycle reconstruction based on the storage access failure risk factor, the arrival timestamp sequence of concurrent requests for intelligent computing services, and the estimated basic processing time corresponding to each request, to obtain the storage resource occupancy status for future time windows. This is specifically illustrated in steps 1051 to 1054.
[0065] This invention, through multi-level analysis ranging from microscopic request-level storage access characteristics (step size, alignment efficiency) to macroscopic concurrency conflict situations (conflict index, risk factors), achieves refined modeling of the storage state during intelligent computing service operation. It concretizes abstract storage pressure into measurable time and capacity dimensions. This solves the problem that static configuration cannot perceive the impact of tidal effects and sudden traffic surges on storage performance, providing a precise data foundation for converting storage losses into equivalent computing power compensation in subsequent steps.
[0066] Optionally, the process of steps 1051 to 1054 includes: Step 1051: Based on the arrival timestamp sequence of concurrent requests for intelligent computing services and the estimated basic processing time for each request, construct the basic resource holding period to obtain the basic resource occupancy range for each request.
[0067] Optionally, the estimated basic processing time refers to the time required to complete the processing of a single request under ideal storage conditions (i.e., no access conflicts and 100% alignment efficiency), based on standard computing power benchmarks (such as floating-point operation capability per unit frequency) and model computational complexity (such as the number of operators and data dimensions).
[0068] The evaluation device receives the arrival timestamp sequence of concurrent requests from intelligent computing services and obtains the estimated basic processing time for each request. For each request in the arrival timestamp sequence, its arrival time is taken as the starting time point for resource occupation. This starting time point is added to the corresponding estimated basic processing time to obtain the theoretical end time point for the request. The continuous time period between the starting time point and the theoretical end time point is defined as the basic resource occupation interval for the request. This interval characterizes the time range during which the request continuously occupies computing and storage resources from start to finish under ideal conditions, ignoring storage efficiency losses. This process is performed on all concurrent requests to obtain the set of basic resource occupation intervals for all requests within a future time window.
[0069] Step 1052: Based on the storage access failure risk factor and the basic resource occupancy range, perform resource residency latency inflation to obtain the additional resource residency latency for each request caused by the reduction in storage access efficiency.
[0070] Optionally, the evaluation device receives storage access failure risk factors and the basic resource usage range for each request. It iterates through the basic resource usage range for each request, integrates or averages the storage access failure risk factors within the time period covered by that range, obtaining the average storage risk value for the request throughout the entire processing. Then, a latency inflation coefficient is preset, representing the proportion of additional time overhead corresponding to a unit risk value. The average storage risk value is multiplied by the latency inflation coefficient, and then multiplied by the expected basic processing time of the request to obtain the additional resource residency latency. This represents the additional time required beyond the basic processing time. The preset latency inflation coefficient represents the proportionality of the impact of storage access failure risk on processing time. The value of this coefficient is determined through offline regression analysis or a lookup table method based on the hardware characteristic parameters of the storage media in the computing power cluster (such as average retry latency and queue depth) and historical performance monitoring data of the intelligent computing power service.
[0071] Step 1053: Based on the basic resource occupancy range and the additional resource residency delay, reconstruct the resource lifecycle to obtain the corrected resource occupancy in each time dimension within the future time window.
[0072] Optionally, after receiving the basic resource occupancy interval and additional resource residency delay for each request, the evaluation device extends the theoretical end time of the basic resource occupancy interval backward for each request by a length equal to the additional resource residency delay of the request. The new interval formed by the extended new end time and the original start time is the corrected resource occupancy interval for the request, reflecting the actual time range of system resources occupied by the request after considering storage efficiency losses.
[0073] Subsequently, the evaluation device acquires several small time slices (e.g., 1 millisecond) defined in step 103. For each time slice, all requests within the corrected resource usage range are counted. The storage resource usage of each request is summed (i.e., the number of bytes of storage usage after converting the expected new data blocks or data size parameters for each request obtained in step 103) to obtain the corrected resource usage for that time slice. By traversing all time slices, the sequence of corrected resource usage for each time dimension within the future time window is obtained.
[0074] Step 1054: Generate a resource occupancy envelope based on the corrected resource occupancy and the time dimension sequence of the future time window to obtain the storage resource occupancy status; the storage resource occupancy status is a status curve showing the changing trend of storage resources within the future time window and the accompanying processing efficiency decay effect.
[0075] Optionally, the evaluation device receives the corrected resource occupancy and the corresponding time sequence for each time dimension within a future time window. Using time as the horizontal axis and the corrected resource occupancy as the vertical axis, it connects or fits the occupancy data points on discrete time slices to form a continuous curve, i.e., the storage resource occupancy status. This status curve not only shows the fluctuation trend of storage resource occupancy over time (such as the tidal effect) but also implicitly includes the processing efficiency degradation effect.
[0076] This invention, through its embodiments, transforms abstract storage risks into specific time and capacity-dimensional indicators, enabling dynamic reconstruction of storage resource occupancy status. This results in a storage resource occupancy status curve that not only presents the dynamic trend of storage capacity changes but also incorporates processing efficiency degradation, ensuring that subsequent computing power demand assessments can be upgraded from static resource total matching to dynamic spatiotemporal efficiency matching.
[0077] Optionally, the process of steps 201 to 205 includes: Step 201: Based on the storage resource occupancy status and the available storage resource boundary of the computing power cluster, identify the critical state of storage resources to obtain the high-pressure period of storage resources within the future time window.
[0078] Optionally, the evaluation device receives the generated storage resource occupancy status and the available storage resource boundary of the computing cluster obtained from the computing cluster resource manager. It iterates through each time point on the storage resource occupancy status curve, comparing the corrected resource occupancy at that moment with the available storage resource boundary of the computing cluster. If the corrected resource occupancy at a certain moment reaches or exceeds a preset percentage threshold (e.g., 90%) of the available storage resource boundary, then that moment is determined to be in a critical storage resource state. All discrete time points in the critical state are clustered in chronological order. If the time interval between adjacent critical state time points is less than a preset minimum discontinuity tolerance value, they are merged into a continuous time period. This preset minimum discontinuity tolerance value is a time threshold determined based on the storage controller's state switching stabilization time or the minimum execution duration of the business operator, used to filter out non-substantial high-pressure interruptions caused by instantaneous fluctuations. These merged continuous time periods are marked as storage resource high-pressure time periods. Each storage resource high-pressure time period is defined by a start timestamp and an end timestamp, representing the specific time interval within a future time window where the storage system may experience performance bottlenecks due to excessive load.
[0079] Step 202: Based on the dynamic change characteristics of storage resources during high-pressure periods, the storage medium allocation dispersion is extrapolated to obtain the local loss index of storage access during each high-pressure period.
[0080] Optionally, the evaluation device analyzes the storage access behavior of concurrent requests during each high-pressure period of storage resources, extracting the identifiers of all active intelligent computing service requests during that period. It also analyzes the physical distribution of the data blocks from these active requests within the storage medium. Since the data from concurrent requests is often scattered across non-contiguous physical addresses or different storage channels, the storage controller needs to perform frequent addressing and channel switching. Therefore, the ratio of the number of non-contiguous segments in the physical address space of these data blocks to the theoretical minimum number of segments is calculated, and combined with the alignment deviation rate of the data access step size, a storage medium allocation fragmentation index is obtained.
[0081] Simultaneously, the evaluation device acquires the parallel computing mode identifier (e.g., data parallelism, model parallelism, or pipeline parallelism) of the intelligent computing power service. If the identifier is data parallelism, the mode co-correction factor α is set to 1; if it is model parallelism or pipeline parallelism, α is set to >1 (e.g., 1.5) based on the communication frequency. The evaluation device performs nonlinear mapping calculations based on the storage medium allocation fragmentation index, the mode co-correction factor α, and a preset attenuation constant k. The formula is: Storage access local loss index = 1 - e^(-k / k) -k*离散度*αWhere k is the decay constant, obtained by fitting historical monitoring data. When historical monitoring data is lacking, k takes a default value of 2 (the theoretical fitting value based on a typical DRAM architecture). This results in a storage access locality loss index between 0 and 1, where 0 indicates no locality loss (data is completely continuous and concentrated), and 1 indicates severe locality corruption (data is extremely scattered). This quantifies the additional addressing and cache miss overhead caused by uneven data distribution.
[0082] Step 203: Estimate the additional time consumption of a single data interaction based on the local loss index of storage access and the basic access performance parameters of the storage medium to obtain the average memory access latency increment in each high-pressure period.
[0083] Optionally, the evaluation device acquires basic access performance parameters of the storage medium. These parameters include the baseline access latency (e.g., nanoseconds) and maximum sustainable bandwidth of the storage medium under ideal locality. When estimating the additional time for a single data interaction, the amplification effect of the storage access locality loss index on physical access time is first analyzed. When locality is poor, the storage controller needs to spend more time on buffer refresh, precharge, or channel arbitration. Therefore, the storage access locality loss index is multiplied by the baseline access latency of the storage medium, and then multiplied by a preset addressing overhead amplification factor to obtain the additional time for a single data interaction. This addressing overhead amplification factor is the ratio of random access to sequential access latency in the hardware architecture specification based on the storage medium, reflecting the sensitivity of different storage architectures (e.g., HBM vs DDR) to discrete access.
[0084] Subsequently, the evaluation device statistically analyzes the average number of data interactions per active request during periods of high storage resource pressure. Multiplying the additional time spent on each data interaction by the average number of interactions yields the average increase in total memory access latency per request during this high-pressure period, i.e., the average memory access latency increment. This quantifies the extra time spent on each data read compared to normal conditions due to localized storage corruption and congestion under high-pressure conditions.
[0085] Step 204: Based on the model structure characteristics and operator data dependency strength corresponding to the intelligent computing power service, perform storage latency sensitivity analysis to obtain the storage latency amplification coefficient, and based on the average memory access latency increment and the storage latency amplification coefficient, perform effective computational stagnation time conversion to obtain the equivalent computational stagnation time in each high-pressure time period.
[0086] Optionally, the evaluation device acquires the model structure characteristics and operator data dependency strength of intelligent computing services. The model structure characteristics include the number of network layers and operator types (such as convolution, fully connected, and attention mechanisms). The operator data dependency strength refers to the degree of dependence of the operator on the readiness of the input data when the operator is executed, that is, whether the computing unit must be idle and wait when the data is not ready.
[0087] Subsequently, during the storage latency sensitivity analysis, the evaluation device first parses the computation graph of the intelligent computing service, identifies the data dependencies between operators, and classifies the operators into strongly dependent operators and weakly dependent operators. Then, it calculates the proportion of the theoretical computation time of strongly dependent operators to the total computation time, obtaining the strong dependency ratio. Combined with a preset pipeline bubble impact weight (calibrated based on historical data), the storage latency amplification factor is calculated. That is, the storage latency amplification factor equals 1 plus the product of the strong dependency ratio and the pipeline bubble impact weight. This factor, greater than or equal to 1, represents the computational stall factor caused by a unit of memory access latency. Then, the average memory access latency increment obtained in step 203 is multiplied by the storage latency amplification factor to obtain the equivalent computational stall time. This represents the actual length of time the computing unit is in an idle waiting state due to the existence of storage latency.
[0088] Step 205: Based on the equivalent computing downtime and the basic load of business processing, the computing power demand is reconstructed to obtain the scenario-based total computing power load including storage loss.
[0089] Optionally, the evaluation device reconstructs the computing power demand compensation based on the equivalent computing downtime and the basic load of business processing to obtain the scenario-based total computing power load including storage loss, as in steps 2051 to 2053.
[0090] This invention establishes a dynamic coupling mapping mechanism between storage status and computing load, thereby achieving refined and scenario-based computing load assessment. The process of generating a scenario-based total computing load that includes storage losses ensures that computing load assessment is no longer limited to ideal theoretical calculations, but truly reflects the actual resource consumption of services in complex storage environments.
[0091] Optionally, the process of steps 2051 to 2053 includes: Step 2051: Based on the theoretical total computational load in the basic workload of business processing and the total duration of the future time window, calculate the total pure computing power demand under ideal conditions to obtain the theoretical baseline computing power value.
[0092] Optionally, the evaluation device receives the basic load of intelligent computing service processing and extracts the theoretical total computational load and the total duration of the future time window. The theoretical total computational load refers to the sum of floating-point operations required to complete all intelligent computing service requests within the future time window under an ideal environment with no storage latency, communication blockage, or hardware loss. The total duration of the future time window refers to the time span for resource prediction, i.e., the time from the current moment to the end of the prediction. Dividing the theoretical total computational load by the total duration of the future time window yields the average computing power required per unit time; this average computing power is the theoretical baseline computing power value. It represents the minimum constant computing power level required to maintain continuous service operation under a perfect system environment.
[0093] Step 2052: Based on the equivalent computational stagnation time and the theoretical benchmark computing power value, perform an effective computing power loss back-calculation analysis to obtain the equivalent compensation computing power value.
[0094] Optionally, the evaluation device receives the equivalent computational stagnation time within each high-pressure time period output in step 204, sums them up, and obtains the total equivalent computational stagnation time within the future time window. That is, the equivalent computational stagnation time refers to the total time that computing units are idle due to storage access latency. Subsequently, an effective computing power loss analysis is performed. In order to complete the same theoretical total computation within the total duration of the future time window, higher computing speed must be provided during non-stagnant time, or additional computing power resources must be added to compensate for the capacity loss caused by this stagnation time. Therefore, the pure computation time ratio β is first calculated, i.e., (total duration of the future time window - total equivalent computational stagnation time) / total duration of the future time window. Then, the evaluation device obtains the serial bottleneck coefficient γ (0 < γ ≤ 1) of the intelligent computing power service. This coefficient represents the proportion of the task that cannot be accelerated by increasing computing power in parallel (such as gradient synchronization, IO serialization, etc., set to 0.1 by default, which can be estimated according to Amdahl's law). Next, calculate the actual average computing power required, using the formula: Actual average computing power required = Theoretical baseline computing power / (β*(1-γ)+γ). Finally, subtract the theoretical baseline computing power from the actual average computing power required; the difference is the equivalent compensation computing power, which quantifies the amount of additional computing power resources needed to offset the time loss caused by storage latency.
[0095] Step 2053: Based on the theoretical benchmark computing power value and the equivalent compensation computing power value, the computing power demand is fused to obtain the scenario-based total computing power load.
[0096] Optionally, when the evaluation device receives the theoretical baseline computing power value and the equivalent compensated computing power value to fuse computing power requirements, it adds the theoretical baseline computing power value and the equivalent compensated computing power value to obtain the scenario-based total computing power load. This scenario-based total computing power load is a dynamic computing power requirement indicator that includes environmental losses. It not only includes the pure computing resources required by the business logic itself (theoretical baseline computing power value), but also explicitly includes the additional resource compensation requirements caused by low storage efficiency, data locality corruption, and concurrency conflicts (equivalent compensated computing power value).
[0097] This invention establishes an equivalent conversion mechanism between time loss and computing power compensation, enabling a leap from ideal theoretical values to real-world scenario values in computing power demand assessment. It not only reflects the computing density of the business but also embeds the constraints of the storage environment on computing efficiency. This allows subsequent hardware resource selection to be based on actual physical consumption rather than inflated theoretical peak values, thus avoiding resource shortages due to underestimating storage consumption or resource idleness due to overestimating peak values. This improves the resource matching accuracy and scheduling economy of computing clusters under complex dynamic loads.
[0098] Optionally, the processes of steps 401 to 405 include: Step 401: Based on the node computing performance of each node in the candidate hardware resource combination and the expected load distribution characteristics of intelligent computing power services, perform a discrete analysis of the computing capabilities between nodes to obtain the discreteness of cluster computing speed within the future time window.
[0099] Optionally, the evaluation device receives candidate hardware resource combinations, obtains the node computing performance of each node, and simultaneously acquires the expected load distribution characteristics of the intelligent computing service. Node computing performance refers to the peak floating-point arithmetic capability of each independent hardware node within the combination at a specific precision; the expected load distribution characteristics refer to the proportion or weight of data or computational tasks allocated among different nodes during parallel computing of the intelligent computing service. Subsequently, the effective processing rate of a single node is calculated. Specifically, for each independent hardware node in the candidate hardware resource combination, its node computing performance is multiplied by its corresponding expected load percentage to obtain the effective processing rate of that node when processing the current service. This rate reflects the amount of computational work that the node can complete per unit time under actual load distribution.
[0100] The evaluation device then performs a dispersion analysis of the computing capabilities among nodes. The analysis process is as follows: The arithmetic mean and standard deviation of the effective processing speed of each participating node are calculated. The standard deviation is divided by the arithmetic mean to obtain the coefficient of variation. The coefficient of variation is then normalized (e.g., using a preset linear mapping function) to obtain the dispersion of the cluster's computing speed. This dispersion is a value between 0 and 1, where 0 indicates that all nodes have completely consistent processing speeds, and 1 indicates that the speed differences between nodes are extremely large, resulting in very low collaborative efficiency.
[0101] Step 402: Based on the parallel computing topology of the intelligent computing power service and the network interconnection topology of the computing power cluster, identify the collaborative data interaction time to obtain the theoretical collaborative interaction time within the future time window.
[0102] Optionally, the evaluation device acquires the parallel computing topology and network interconnection topology of the intelligent computing power service. The parallel computing topology describes the dependencies and data exchange points (such as All-Reduce operations and parameter synchronization points) between different computing stages in the service model. Based on this, collaborative data interaction moments are identified. The specific identification process is as follows: traversing each data exchange point in the parallel computing topology, and combining the implicit processing stage execution order in the storage access step size feature sequence of each request obtained in step 101, the expected occurrence time of each exchange point on the time axis is determined. These expected occurrence times are determined as theoretical collaborative interaction moments. All theoretical collaborative interaction moments are arranged in chronological order to form a theoretical collaborative interaction moment sequence.
[0103] Step 403: Based on the cluster computing speed dispersion and the theoretical collaborative interaction time, the arrival time difference between nodes is extrapolated to obtain the theoretical collaborative delay time for each collaborative interaction time.
[0104] Optionally, after receiving the cluster computing speed dispersion and the theoretical cooperative interaction time sequence, the evaluation device performs a deduction of the arrival time difference between nodes. Due to the existence of cluster computing speed dispersion, the time it takes for different nodes to complete the current computing stage and arrive at the cooperative interaction time is inconsistent. Therefore, the specific deduction process is as follows: the evaluation device obtains the baseline computing time between two adjacent theoretical cooperative interaction times (i.e., the time required for all nodes to complete the task synchronously under ideal conditions), multiplies the baseline computing time by the cluster computing speed dispersion, and obtains the maximum completion time deviation caused by speed differences within this time period, i.e., the theoretical cooperative stall time. This time represents the forced waiting time caused only by the inconsistency of node computing speeds, without considering network transmission delay. This calculation is performed sequentially for each theoretical cooperative interaction time to obtain the theoretical cooperative stall time sequence.
[0105] Step 404: Based on the theoretical collaborative blocking duration and the network transmission blocking characteristics of the computing power cluster, analyze the communication congestion amplification effect to obtain the actual collaborative blocking duration at each collaborative interaction moment.
[0106] Optionally, the evaluation device acquires the network transmission congestion characteristics of the computing cluster and receives the theoretical collaborative congestion duration. During the analysis of communication congestion amplification effects, it is found that in a real network environment, collaborative data interaction is affected not only by node speed differences but also by network transmission conditions. When multiple nodes send synchronization data simultaneously, it may cause switch queue congestion, leading to increased transmission delay. Therefore, the evaluation device first calculates the data transmission volume at each collaborative interaction moment and estimates the basic transmission delay by combining it with the link bandwidth in the network interconnection topology. Then, it retrieves a preset congestion amplification coefficient, which is determined based on the current network load rate and historical congestion logs. Multiplying the basic transmission delay by the congestion amplification coefficient yields the network congestion delay. Finally, the theoretical collaborative congestion duration is added to the network congestion delay to obtain the actual collaborative congestion duration. This accurately reflects the non-computational time actually consumed by nodes due to collaborative synchronization in a specific network environment.
[0107] Step 405: Based on the actual collaborative blocking time and candidate hardware resource combinations, perform collaborative efficiency collapse calculation to obtain the effective computing power output index under different hardware resource combinations.
[0108] Optionally, the evaluation device performs collaborative performance collapse calculation based on the actual collaborative blocking time and candidate hardware resource combinations to obtain the effective computing power output index under different hardware resource combinations, as described in steps 4051 to 4054.
[0109] This invention achieves a precise mapping from theoretical peak computing power to actual effective computing power by constructing a cascade analysis of computational discreteness, communication bottlenecks, and performance collapse. It transforms abstract network communication overhead and hardware differences into measurable computing power loss indicators, ensuring that subsequent computing resource allocation schemes are based on the actual collaborative capabilities of hardware combinations rather than theoretical peak values. This avoids task timeouts due to overestimating collaborative efficiency or resource waste due to underestimating hardware potential, thereby improving the resource scheduling accuracy and overall output efficiency of computing clusters in complex parallel scenarios.
[0110] Optionally, the processes of steps 4051 to 4054 include: Step 4051: Based on the sum of the theoretical peak computing power of the candidate hardware resource combinations and the total duration of the future time window, calculate the lossless collaborative computing power output value under ideal conditions to obtain the theoretical benchmark collaborative computing power value.
[0111] Optionally, the evaluation device, based on the sum of the theoretical peak computing power of the candidate hardware resource combination and the total duration of the future time window, under ideal conditions assuming that all nodes can operate at full capacity without any communication blockage or synchronization waiting, multiplies the sum of the theoretical peak computing power by the total duration of the future time window to obtain the maximum theoretical computational workload that the combination can complete within the future time window, i.e., the theoretical benchmark collaborative computing power value.
[0112] Step 4052: Based on the actual collaborative delay duration sequence and the total duration of the future time window, calculate the effective collaborative time ratio to obtain the collaborative efficiency loss coefficient within the future time window.
[0113] Optionally, the evaluation device receives the actual collaborative blocking duration sequence and the total duration of the future time window. Based on the actual waiting time caused by node speed differences and network congestion at each theoretical collaborative interaction moment within the future time window included in the actual collaborative blocking duration sequence, the actual collaborative blocking duration sequence is accumulated to obtain the total actual collaborative blocking duration. That is, the total time that computing nodes are idle and waiting due to collaborative synchronization within the entire future time window. Subsequently, the total actual collaborative blocking duration is subtracted from the total duration of the future time window to obtain the net effective computation time. Then, the net effective computation time is divided by the total duration of the future time window to obtain the effective collaborative time percentage. This percentage reflects the proportion of time the hardware combination actually uses to execute computational tasks after deducting all collaborative blocking. Finally, the effective collaborative time percentage is subtracted from 1 to obtain the collaborative efficiency loss coefficient. This coefficient quantifies the proportion of overall time resources wasted due to collaborative blocking and is between 0 and 1; a larger value indicates lower collaborative efficiency.
[0114] Step 4053: Based on the theoretical benchmark collaborative computing power value and collaborative efficiency loss coefficient, perform nonlinear collapse calculation of collaborative efficiency to obtain the preliminary computing power output rate under different hardware resource combinations.
[0115] Optionally, the evaluation device receives the theoretical benchmark collaborative computing power value and the collaborative efficiency loss coefficient, and performs a nonlinear collapse reduction of the collaborative efficiency. The specific reduction process is as follows: First, the serial bottleneck coefficient γ of the intelligent computing power service is obtained. Then, the corrected effective collaborative efficiency factor is calculated using the formula: (1 - collaborative efficiency loss coefficient) raised to a preset power to obtain the nonlinear efficiency factor. This preset power is greater than 1 (e.g., 1.5 or 2, obtained based on historical data fitting) to reflect the nonlinear collapse effect. Then, the nonlinear efficiency factor is multiplied by (1 - γ) and the serial bottleneck coefficient γ is added. Here, (1 - γ) + γ is used to ensure that the computing power output is not lower than the serial benchmark under extremely high losses. Finally, the theoretical benchmark collaborative computing power value is divided by the total duration of the future time window to obtain the theoretical peak computing power rate, which is then multiplied by the corrected effective collaborative efficiency factor to obtain the preliminary computing power output rate. This rate reflects the actual computing power level that the hardware combination is expected to provide after considering nonlinear losses and serial bottleneck constraints.
[0116] Step 4054: Based on the resource scheduling granularity constraints of the initial computing power output rate and candidate hardware resource combinations, standard performance normalization is performed to obtain the effective computing power output index.
[0117] Optionally, the evaluation device receives resource scheduling granularity constraints on the initial computing power output rate and candidate hardware resource combinations. These resource scheduling granularity constraints refer to the minimum unit restrictions (such as a minimum allocation of one GPU core, minimum memory page size, etc.) imposed by the computing cluster's operating system or virtualization platform when allocating resources.
[0118] Next, the evaluation device obtains a preset standardized correction factor, which is an empirical value determined based on the cluster's current resource fragmentation rate, virtualization overhead, and background daemon utilization rate, and is typically less than 1. The initial computing power output rate is multiplied by the standardized correction factor to obtain the corrected average computing power value. This reflects the expected average computing power in a real system environment, after deducting microsystem losses.
[0119] Finally, the evaluation device rounds down or performs nearest-neighbor alignment on the corrected average computing power value based on resource scheduling granularity constraints. That is, it maps the value to an integer multiple of the nearest scheduling granularity that does not exceed the value, thus obtaining the effective computing power output index.
[0120] This invention achieves precise conversion from theoretical peak computing power to usable computing power by establishing a multi-level conversion model of ideal benchmark, linear loss, nonlinear collapse, and discretization correction. The generated effective computing power output index not only eliminates the inefficiencies caused by communication and synchronization but also accommodates the physical constraints of the underlying scheduling system. This ensures that resource scheduling decisions meet both business performance requirements and the actual operating rules of the cluster, significantly improving the utilization efficiency of computing resources and the reliability of business operations.
[0121] Furthermore, the computing power demand assessment device for intelligent computing power business scenarios provided by the present invention will be described below. The computing power demand assessment device for intelligent computing power business scenarios described below can be referred to in correspondence with the computing power demand assessment method for intelligent computing power business scenarios described above.
[0122] Optional, refer to Figure 2 , Figure 2 This is a schematic diagram of the computing power demand assessment device for intelligent computing power service scenarios provided by the present invention. The computing power demand assessment device for intelligent computing power service scenarios includes: The storage status prediction module 210 is used to perform runtime storage status evolution analysis based on the concurrent request timing characteristics and data processing length distribution of intelligent computing power services, and to obtain the storage resource occupancy status of future time windows. The scenario-based load fusion module 220 is used to fuse computing loads based on storage resource occupancy status, available storage resource boundaries of computing clusters, and basic load of business processing to obtain a scenario-based total computing load including storage loss. The security boundary delineation module 230 is used to delineate the security resource boundary based on the scenario-based total computing power load and the heterogeneous hardware topology of the computing power cluster, and obtain the candidate hardware resource combination that meets the business continuity requirements. The collaborative efficiency evolution module 240 is used to perform communication collaborative efficiency evolution analysis based on the network interconnection bandwidth characteristics of candidate hardware resource combinations and computing power clusters, and obtain the effective computing power output index under different hardware resource combinations. The timeliness adaptation decision module 250 is used to generate a target computing power resource allocation scheme based on the effective computing power output index and the response timeliness constraints of intelligent computing power services.
[0123] The embodiments of the present invention ultimately achieve precise adaptation based on the dynamic coupling relationship between storage and computing power during business operation and the actual collaborative efficiency of hardware, avoiding resource idleness or mismatch caused by static peak configuration, and improving the overall utilization efficiency and resource scheduling accuracy of computing power clusters in intelligent computing power business scenarios.
[0124] Please see Figure 3 , Figure 3 An embodiment diagram of an electronic device provided in accordance with the present invention. For example... Figure 3 As shown, an electronic device 300 provided in this embodiment of the invention includes a memory 310, a processor 320, and a computer program 311 stored in the memory 310 and executable on the processor 320. When the processor 320 executes the computer program 311, it performs the following steps: Based on the concurrent request timing characteristics and data processing length distribution of intelligent computing services, runtime storage state evolution analysis is performed to obtain the storage resource occupancy status of future time windows. The computing load is fused based on the storage resource occupancy status, the available storage resource boundary of the computing cluster, and the basic load of business processing to obtain the scenario-based total computing load including storage loss. Based on the scenario-based total computing power load and the heterogeneous hardware topology of the computing power cluster, security resource boundaries are defined to obtain candidate hardware resource combinations that meet business continuity requirements. Based on the network interconnection bandwidth characteristics of candidate hardware resource combinations and computing power clusters, we conduct communication collaboration efficiency evolution analysis to obtain the effective computing power output index under different hardware resource combinations. Based on the effective computing power output index and the response time constraints of intelligent computing power services, a target computing power resource allocation scheme is generated.
[0125] Please see Figure 4 , Figure 4 An embodiment diagram of a computer-readable storage medium provided in accordance with an embodiment of the present invention is shown. Figure 4As shown, this embodiment provides a computer-readable storage medium 400 on which a computer program 311 is stored. When the computer program 311 is executed by a processor, it performs the following steps: Based on the concurrent request timing characteristics and data processing length distribution of intelligent computing services, runtime storage state evolution analysis is performed to obtain the storage resource occupancy status of future time windows. The computing load is fused based on the storage resource occupancy status, the available storage resource boundary of the computing cluster, and the basic load of business processing to obtain the scenario-based total computing load including storage loss. Based on the scenario-based total computing power load and the heterogeneous hardware topology of the computing power cluster, security resource boundaries are defined to obtain candidate hardware resource combinations that meet business continuity requirements. Based on the network interconnection bandwidth characteristics of candidate hardware resource combinations and computing power clusters, we conduct communication collaboration efficiency evolution analysis to obtain the effective computing power output index under different hardware resource combinations. Based on the effective computing power output index and the response time constraints of intelligent computing power services, a target computing power resource allocation scheme is generated.
[0126] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the computing power demand assessment method for intelligent computing power service scenarios provided by the above methods. The method includes: Based on the concurrent request timing characteristics and data processing length distribution of intelligent computing services, runtime storage state evolution analysis is performed to obtain the storage resource occupancy status of future time windows. The computing load is fused based on the storage resource occupancy status, the available storage resource boundary of the computing cluster, and the basic load of business processing to obtain the scenario-based total computing load including storage loss. Based on the scenario-based total computing power load and the heterogeneous hardware topology of the computing power cluster, security resource boundaries are defined to obtain candidate hardware resource combinations that meet business continuity requirements. Based on the network interconnection bandwidth characteristics of candidate hardware resource combinations and computing power clusters, we conduct communication collaboration efficiency evolution analysis to obtain the effective computing power output index under different hardware resource combinations. Based on the effective computing power output index and the response time constraints of intelligent computing power services, a target computing power resource allocation scheme is generated.
[0127] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0128] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.
Claims
1. A method for assessing computing power requirements in intelligent computing power service scenarios, characterized in that, include: Based on the concurrent request timing characteristics and data processing length distribution of intelligent computing services, runtime storage state evolution analysis is performed to obtain the storage resource occupancy status of future time windows. Based on the storage resource occupancy status, the available storage resource boundary of the computing power cluster, and the basic load of business processing, the computing power load is fused to obtain the scenario-based total computing power load including storage loss. Based on the scenario-based total computing power load and the heterogeneous hardware topology of the computing power cluster, security resource boundaries are defined to obtain candidate hardware resource combinations that meet business continuity requirements. Based on the network interconnection bandwidth characteristics of the candidate hardware resource combinations and computing power clusters, an evolution analysis of communication collaboration efficiency is performed to obtain the effective computing power output index under different hardware resource combinations. Based on the effective computing power output index and the response time constraints of intelligent computing power services, a target computing power resource allocation scheme is generated.
2. The computing power demand assessment method for intelligent computing power service scenarios according to claim 1, characterized in that, The process of obtaining the storage resource usage status for future time windows includes: Based on the arrival timestamp sequence of concurrent requests for intelligent computing services and the data scale parameters corresponding to each request, a multi-stage storage access mode is analyzed to obtain the storage access step size feature sequence of each request at different processing stages. Based on the storage access step size feature sequence of each request and the preset minimum access granularity parameter of the storage medium, storage access alignment efficiency analysis is performed to obtain the theoretical storage access utilization of each request. The concurrent access conflict index within the future time window is obtained by superimposing the concurrent request arrival timestamp sequence and the expected new data block corresponding to each request on the concurrent storage bandwidth demand. Based on the theoretical storage access utilization rate and the concurrent access conflict index, the degree of storage locality disruption is comprehensively calculated to obtain the storage access failure risk factor at each moment within the future time window; Based on the storage access failure risk factor, the arrival timestamp sequence of concurrent requests for intelligent computing services, and the estimated basic processing time for each request, latency inflation correction and lifecycle reconstruction are performed to obtain the storage resource occupancy status for future time windows.
3. The computing power demand assessment method for intelligent computing power service scenarios according to claim 2, characterized in that, The process of performing latency inflation correction and lifecycle reconstruction based on the storage access failure risk factor, the arrival timestamp sequence of concurrent requests for intelligent computing services, and the estimated basic processing time for each request, to obtain the storage resource occupancy status for future time windows, includes: Based on the arrival timestamp sequence of concurrent requests for intelligent computing services and the estimated basic processing time for each request, the basic resource holding period is constructed to obtain the basic resource occupation range for each request. Based on the storage access failure risk factor and the basic resource occupancy range, resource residency latency is increased to obtain the additional resource residency latency for each request caused by the reduction in storage access efficiency. Based on the basic resource occupancy range and the additional resource residency delay, the resource lifecycle is reconstructed to obtain the corrected resource occupancy in each time dimension within the future time window. Based on the corrected resource occupancy and the time dimension sequence of the future time window, a resource occupancy envelope is generated to obtain the storage resource occupancy status; the storage resource occupancy status is a status curve showing the changing trend of storage resources within the future time window and the accompanying processing efficiency decay effect.
4. The computing power demand assessment method for intelligent computing power service scenarios according to claim 1, characterized in that, The obtained scenario-specific total computing load, including storage loss, includes: Based on the storage resource occupancy status and the available storage resource boundary of the computing power cluster, the critical state of storage resources is identified to obtain the high-pressure period of storage resources within the future time window. Based on the dynamic change characteristics of the storage resources during the high-pressure period, the storage medium allocation dispersion is deduced to obtain the local loss index of storage access in each high-pressure period. Based on the local loss index of storage access and the basic access performance parameters of the storage medium, the additional time consumption of a single data interaction is estimated to obtain the average memory access latency increment in each high-pressure time period. Based on the model structure characteristics and operator data dependency strength corresponding to the intelligent computing power service, a storage latency sensitivity analysis is performed to obtain the storage latency amplification coefficient. Based on the average memory access latency increment and the storage latency amplification coefficient, the effective computational stagnation time is converted to obtain the equivalent computational stagnation time in each high-pressure period. Based on the equivalent computation standby time and the basic load of business processing, the computing power demand is reconstructed to obtain the scenario-based total computing power load including storage loss.
5. The computing power demand assessment method for intelligent computing power service scenarios according to claim 4, characterized in that, The process of reconstructing computing power demand based on the equivalent computational downtime and the basic workload of business processing yields a scenario-specific total computing power load that includes storage losses, including: Based on the theoretical total computational load in the basic workload of business processing and the total duration of the future time window, the total pure computing power demand under ideal conditions is calculated to obtain the theoretical baseline computing power value. Based on the equivalent calculation stagnation time and the theoretical benchmark computing power value, an effective computing power loss back-calculation analysis is performed to obtain the equivalent compensation computing power value; The computing power demand is fused based on the theoretical benchmark computing power value and the equivalent compensation computing power value to obtain the scenario-based total computing power load.
6. The computing power demand assessment method for intelligent computing power service scenarios according to claim 1, characterized in that, The network interconnection bandwidth characteristics include network interconnection topology and network transmission bottleneck characteristics; The method for obtaining the effective computing power output index under different hardware resource combinations includes: Based on the node computing performance of each node in the candidate hardware resource combination and the expected load distribution characteristics of intelligent computing services, the dispersion of computing capabilities between nodes is analyzed to obtain the dispersion of cluster computing speed within the future time window. Based on the parallel computing topology of intelligent computing services and the network interconnection topology of computing power clusters, the collaborative data interaction time is identified to obtain the theoretical collaborative interaction time within the future time window. Based on the cluster computing speed dispersion and the theoretical collaborative interaction time, the arrival time difference between nodes is extrapolated to obtain the theoretical collaborative delay time for each collaborative interaction time. Based on the theoretical collaborative blocking duration and the network transmission blocking characteristics of the computing power cluster, the communication congestion amplification effect analysis is performed to obtain the actual collaborative blocking duration at each collaborative interaction moment. Based on the actual collaborative blocking time and the candidate hardware resource combinations, the collaborative efficiency collapse calculation is performed to obtain the effective computing power output index under different hardware resource combinations.
7. The computing power demand assessment method for intelligent computing power service scenarios according to claim 6, characterized in that, The step of performing collaborative performance collapse calculation based on the actual collaborative stall duration and the candidate hardware resource combinations to obtain the effective computing power output index under different hardware resource combinations includes: Based on the sum of the theoretical peak computing power of candidate hardware resource combinations and the total duration of future time windows, the lossless collaborative computing power output value under ideal conditions is calculated to obtain the theoretical benchmark collaborative computing power value. Based on the actual cooperative delay duration sequence and the total duration of the future time window, the effective cooperative time ratio is calculated to obtain the cooperative efficiency loss coefficient within the future time window. Based on the theoretical benchmark collaborative computing power value and the collaborative efficiency loss coefficient, the collaborative efficiency nonlinear collapse calculation is performed to obtain the preliminary computing power output rate under different hardware resource combinations. Based on the resource scheduling granularity constraints of the preliminary computing power output rate and candidate hardware resource combinations, standard performance normalization is performed to obtain the effective computing power output index.
8. A computing power demand assessment device for intelligent computing power business scenarios, characterized in that, The computing power demand assessment method for intelligent computing power service scenarios as described in any one of claims 1 to 7; The computing power demand assessment device for intelligent computing power service scenarios includes: The storage status prediction module is used to perform runtime storage status evolution analysis based on the concurrent request timing characteristics and data processing length distribution of intelligent computing power services, and to obtain the storage resource occupancy status of future time windows. The scenario-based load fusion module is used to fuse computing loads based on the storage resource occupancy status, the available storage resource boundary of the computing power cluster, and the basic load of business processing, to obtain the scenario-based total computing load including storage loss. The security boundary delineation module is used to delineate security resource boundaries based on the scenario-based total computing power load and the heterogeneous hardware topology of the computing power cluster, and obtain candidate hardware resource combinations that meet business continuity requirements. The collaborative efficiency evolution module is used to perform communication collaborative efficiency evolution analysis based on the candidate hardware resource combinations and the network interconnection bandwidth characteristics of the computing power cluster, and to obtain the effective computing power output index under different hardware resource combinations. The timeliness adaptation decision module is used to generate a target computing power resource configuration scheme based on the effective computing power output index and the response timeliness constraints of intelligent computing power services.
9. An electronic device, characterized in that, include: Memory, used to store computer software programs; A processor is configured to read and execute the computer software program, wherein when the processor executes the computer software program, it implements the computing power demand assessment method for intelligent computing power service scenarios as described in any one of claims 1 to 7.
10. A non-transitory computer-readable storage medium, characterized in that, The storage medium stores a computer software program, which, when executed by a processor, implements the computing power demand assessment method for intelligent computing power service scenarios as described in any one of claims 1 to 7.
