Method and related device for configuring computing power of intelligent computing center
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA MOBILE GROUP DESIGN INST
- Filing Date
- 2026-02-09
- Publication Date
- 2026-06-09
Smart Images
Figure CN122173273A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of computer technology, and in particular to a method for planning the computing power configuration of an intelligent computing center and related equipment. Background Technology
[0002] With the rapid development of artificial intelligence technology, intelligent computing power has become a core driving force for technological innovation and industrial transformation. As the core infrastructure providing intelligent computing power, the accuracy of the computing power configuration planning of Intelligent Computer Centers (ICCs) directly affects investment efficiency and business support capabilities. In existing technologies, the computing power configuration planning of intelligent computing centers is typically based on the theoretical peak computing power and empirical redundancy provided by hardware manufacturers. However, in practical applications, the effective computing capacity (ECC) of an intelligent computing center (ICC) differs significantly from the theoretical peak due to factors such as load characteristics, memory bandwidth, and software optimization. Therefore, relying solely on theoretical peak computing power and empirical redundancy for planning can easily lead to configuration deviations, resulting in a mismatch between resource allocation and actual needs, causing resource waste or insufficient performance.
[0003] Therefore, how to improve the accuracy of computing power allocation planning in intelligent computing centers is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0004] This application provides a computing power configuration planning method for intelligent computing centers, which solves the problem in the prior art that the resource configuration is mismatched with the actual needs, resulting in resource waste or insufficient performance, due to the reliance on theoretical peak computing power and empirical redundancy provided by hardware manufacturers for resource configuration.
[0005] This application also provides a computing power configuration planning device for a smart computing center, an electronic device, a computer-readable storage medium, and a computer program product.
[0006] The embodiments of this application adopt the following technical solutions: In a first aspect, embodiments of this application provide a method for planning the computing power allocation of a smart computing center, including: Obtain the business load requirements of the target business for the intelligent computing center's computing power configuration planning; Standardized performance metrics for business load requirements are determined based on a pre-built benchmark library, and equivalent performance metrics are determined based on the standardized performance metrics. The benchmark library stores standardized performance data for various standard tasks on different hardware. Based on the equivalent performance indicators and the theoretical benchmark performance indicators of the target hardware under standard tasks, determine the effective computing power requirements of the target service. Based on the effective computing power requirements, a planning scheme for the effective computing power configuration of the intelligent computing center is determined.
[0007] Optionally, the workload requirement information should at least include: the types of tasks involved in the target business; then, standardized performance metrics for the workload requirement information should be determined based on a pre-built benchmark library, including: Based on the benchmark library and the task types involved in the target business, the basic computational load, task type mode coefficients, and benchmark library calibration coefficients of the target business are determined. Based on the basic computational load, task type mode coefficient, and benchmark library calibration coefficient, the standardized performance indicators of business load requirements are determined.
[0008] Optionally, equivalent performance indicators may be determined based on standardized performance indicators, including: Determine the ratio of the number of tasks in each task type within the target business to the total number of tasks in the target business; Determine the business demand for each task type based on business load demand information. The equivalent performance indicators of the target business are determined by weighted aggregation based on the proportion of task quantity, business demand, standardized performance indicators, and hardware efficiency calibration parameters.
[0009] Optionally, determine the hardware efficiency calibration parameters, including: Obtain the target model utilization rate for the target business; Obtain the utilization rate of reference models for standard tasks from the benchmark library; The hardware efficiency calibration parameters are determined based on the ratio between the utilization rate of the target model and the utilization rate of the reference model.
[0010] Optionally, based on equivalent performance metrics and the theoretical benchmark performance metrics of the target hardware under standard tasks, the effective computing power requirements of the target service are determined, including: Determine the theoretical peak computing power and global adjustment factor of the target hardware. The global adjustment factor is used to compensate for the unmodeled performance loss introduced by the difference between the operation of standard tasks and target services. Calculate the ratio between the equivalent performance index and the theoretical benchmark performance index; The effective computing power requirement is obtained by multiplying the ratio, the theoretical peak computing power of the target hardware, and the global adjustment factor.
[0011] Optionally, before determining the effective computing power requirements of the target service based on equivalent performance metrics and the theoretical benchmark performance metrics of the target hardware under standard tasks, the method further includes: Obtain at least one standard task corresponding to the target business; The theoretical benchmark business requirements corresponding to each standard task are obtained from the benchmark library. The theoretical benchmark business requirements are derived from the benchmark test results in the benchmark library. Based on the theoretical baseline service demand, standardized performance data, and hardware efficiency calibration parameters of each standard task, the theoretical baseline throughput corresponding to each standard task is determined. The theoretical benchmark throughput corresponding to each standard task is aggregated to obtain the theoretical benchmark performance index of the target hardware under the standard task.
[0012] Optionally, the method further includes: Determine the theoretical peak computing power requirement of the target service, and determine the planned peak computing power requirement of the target service based on the theoretical peak computing power requirement and the hardware redundancy coefficient; Determine the peak computing power configuration plan for the intelligent computing center based on the planned peak computing power requirements; The deviation between the determined peak computing power configuration plan and the effective computing power configuration plan; When the deviation is greater than the preset deviation threshold, the computing power configuration planning scheme update operation is executed repeatedly until the deviation between the updated peak computing power configuration planning scheme and the updated effective computing power configuration planning scheme is less than or equal to the preset deviation threshold. The operation of updating the computing power configuration planning scheme includes: Adjust the hardware redundancy factor and / or the hardware efficiency calibration parameters used to determine the equivalent performance metrics; The peak computing power configuration planning scheme and computing power configuration planning scheme are updated based on the adjusted hardware redundancy coefficient and / or hardware efficiency calibration parameters.
[0013] Optionally, determine the theoretical peak computing power requirement of the target service, including: Determine the computational load, business scale, and time constraint coefficients of the target business model; Based on the model's computational load, business scale, time constraint coefficient, and preset reserved computing power margin, determine the theoretical peak computing power requirement of the target business.
[0014] Secondly, embodiments of this application provide a computing power configuration planning device for an intelligent computing center, including an information acquisition module, an indicator determination module, a demand determination module, and a scheme determination module, wherein: The information acquisition module is used to acquire the business load requirements information of the target business for the intelligent computing center's computing power configuration planning; The indicator determination module is used to determine standardized performance indicators of business load requirements based on a pre-built benchmark library, and to determine equivalent performance indicators based on the standardized performance indicators; wherein, the benchmark library stores standardized performance data of various standard tasks on different hardware. The demand determination module is used to determine the effective computing power requirements of the target business based on the equivalent performance indicators and the theoretical benchmark performance indicators of the target hardware under standard tasks. The scheme determination module is used to determine the effective computing power configuration planning scheme of the intelligent computing center based on the effective computing power requirements.
[0015] Thirdly, embodiments of this application provide an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the computing power configuration planning method for the intelligent computing center as described above.
[0016] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the computing power configuration planning method for an intelligent computing center as described above.
[0017] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the computing power configuration planning method for a smart computing center as described above.
[0018] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects: The method provided in this application obtains the service load demand information of the target service and converts it into standardized performance indicators using a pre-built benchmark library, thereby obtaining equivalent performance indicators characterizing the overall load level of the target service. Simultaneously, by introducing theoretical benchmark performance indicators of the target hardware under standard tasks to calibrate the equivalent performance indicators, the effective computing power demand reflecting the impact of actual load characteristics, memory bandwidth constraints, and software optimization differences on the target hardware can be determined. In this way, the effective computing power demand no longer relies solely on the theoretical peak computing power and empirical redundancy provided by hardware manufacturers, but is based on the mapping relationship between standard task performance data and theoretical benchmark performance indicators. This improves the accuracy of the effective computing power configuration planning scheme determined based on the effective computing power demand and reduces resource waste or performance insufficiency caused by mismatch between resource allocation and actual demand. Attached Figure Description
[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A schematic diagram illustrating the implementation process of a computing power allocation planning method for an intelligent computing center, provided in an embodiment of this application; Figure 2A schematic diagram illustrating the implementation process of a method for determining the effective computing power requirements of a target service, provided in an embodiment of this application; Figure 3 A schematic diagram illustrating the implementation process of a peak computing power configuration planning scheme provided in this application embodiment; Figure 4 This application provides a schematic diagram of the specific structure of a computing power configuration planning device for an intelligent computing center. Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0021] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.
[0022] To address the problem in existing technologies where resource allocation relies on theoretical peak computing power and empirical redundancy provided by hardware manufacturers, resulting in resource waste or insufficient performance due to mismatch between resource allocation and actual needs, this application provides a computing power allocation planning method for intelligent computing centers.
[0023] The execution subject of this method can be various types of computing devices, or it can be an application or app installed on the computing device. The computing device can be a user terminal such as a mobile phone, tablet computer, or smart wearable device, or it can be a server.
[0024] For ease of description, this application uses a server as the execution subject of the method in its embodiments to illustrate the method. Those skilled in the art will understand that this embodiment uses a server as an example to describe the method, which is merely an illustrative example and does not limit the scope of protection of the corresponding claims.
[0025] Specifically, the implementation flow of the method provided in this application embodiment is as follows: Figure 1 As shown, it includes the following steps: Step 102: Obtain the business load requirements information of the target business for the intelligent computing center's computing power configuration planning.
[0026] In this embodiment of the application, the target service may consist of at least one type of sub-task.
[0027] Business load requirements information may include the set of task types involved in the target business, the business requirements of each task type, such as throughput requirements, arrival rate, concurrency, etc., as well as optional quality of service constraints, such as time and elasticity requirements, accuracy requirements, etc.
[0028] The task type involved in the target business refers to the category of artificial intelligence tasks to which the target business belongs, such as image classification tasks, object detection tasks, machine translation tasks, speech recognition tasks, large language model (LLM) training tasks or inference tasks, etc.
[0029] Optionally, the types of tasks involved in the target business can be determined based on information such as business configuration, product requirements, and project proposals provided by the business side; or, the types of tasks involved in the target business can be inferred from the system side, for example, from job queues / scheduling logs, model repository metadata, and component types of training / inference pipelines.
[0030] The business requirements for each task type can be understood as the amount of business that needs to be processed per unit of time (e.g., per second). For example, the number of images (images / s), text tokens (tokens / s), or training samples (samples / s) that need to be processed per second. Among these, the training task for the target business usually involves backpropagation, and its computational cost is usually much greater than that of forward inference for the same model.
[0031] The precision requirements of the target service, i.e., the computational precision required for the operation of the target service, such as FP32, FP16, BF16, INT8, etc., vary significantly in the number of floating-point operations (FLOPS) under different precisions, and a unified conversion needs to be performed in subsequent steps.
[0032] The time and flexibility requirements of the target business can include, for example, whether the target business is continuous operation or project-based delivery, planning cycle, etc., and whether the demand for computing power is constant, peak type, or elastic scaling characteristics.
[0033] In one alternative implementation, when the target business is not yet fully clear, it can be recorded as a business type to be identified, and the task type and demand can be supplemented later by log sampling, load test playback or business-side configuration parameters.
[0034] Step 104: Determine standardized performance metrics for business load requirements based on the pre-built benchmark library, and determine equivalent performance metrics based on the standardized performance metrics.
[0035] The benchmark library stores standardized performance data for various standard tasks on different hardware.
[0036] In this embodiment of the application, the benchmark database is used to establish the mapping relationship between standard tasks, benchmark models, computing hardware and performance indicators.
[0037] Optionally, the benchmark library can be built based on standard load test results data provided by benchmarking organizations. For example, it can use typical AI task loads covered by benchmark reports such as MLPerf as a standard task set, and select representative benchmark models for each type of standard task to be included in the library. At the same time, it can select mainstream computing hardware models as a hardware set, thereby forming a benchmark data set covering mainstream task scenarios, representative models, and mainstream hardware platforms. In other words, the benchmark library can include information such as mainstream AI tasks, top models, mainstream hardware, performance indicators, and model utilization.
[0038] In some embodiments, the benchmark library can be organized as a mapping matrix B as shown below:
[0039] Each record in the mapping matrix B corresponds to a combination of a standard task, a benchmark model, and a computing hardware, and stores the performance index data of this combination under standard load and the model utilization data corresponding to the performance index data.
[0040] Specifically, the mapping matrix B can include at least the fields Task, Model, Hardware, Throughput, SU_rate, and MFU. Among them, Task is used to represent the benchmark workload task classification. For example, when using the MLPerf benchmark, Task can be a standard task covering mainstream artificial intelligence tasks, including image classification, object detection, translation, recommendation, speech recognition, large language models, etc.
[0041] Model is used to represent the benchmark model corresponding to the standard task. When building the benchmark library, TOP model tests can be carried out for a certain type of task, and representative models can be selected for inclusion in the library based on the test results.
[0042] Hardware is used to characterize computing hardware models, such as GPU or NPU models. Specifically, examples include NVIDIA A100, Huawei Ascend 910B, and Cambricon MLU370.
[0043] Throughput is a metric used to characterize the throughput of a standard task when executed on computing hardware. Optionally, in addition to throughput metrics, it may also include score-based metrics to characterize the performance of a standard task on computing hardware.
[0044] In some embodiments, considering that throughput can directly reflect the ability of computing hardware to process standard tasks per unit time, is more intuitive than score-based indicators, and has a stronger linear correlation with hardware computing power, throughput can be preferred as a performance indicator.
[0045] SU_rate is used to characterize the throughput of a standard task, which is the standardized throughput obtained by normalizing the throughput of different standard tasks under their original units. It is used to achieve a unified comparison and subsequent equivalent aggregation across tasks and units.
[0046] MFU is used to characterize model utilization data, which is the measured model utilization under combination. It reflects the effective computing power utilization of computing hardware in the actual execution of standard tasks, and provides a basis for subsequent correction of hardware efficiency differences and software optimization differences.
[0047] Optionally, to ensure the reproducibility and comparability of the benchmark data, the benchmark library can also store test configuration caliber information corresponding to the standard load test results, such as accuracy configuration, hardware scale configuration, and software stack configuration, to avoid misuse or accumulation of deviations due to differences in test caliber.
[0048] In this embodiment of the application, when determining the standardized performance index of the business load requirement information, the basic computing volume, task type mode coefficient and benchmark calibration coefficient of the target business can be determined based on the benchmark library and the task types involved in the target business; then, the standardized performance index of the business load requirement information is determined based on the basic computing volume, task type mode coefficient and benchmark calibration coefficient.
[0049] Among them, the basic computing power of the target business is used to characterize the basic computing intensity of the task types involved in the target business under the unit business processing volume.
[0050] In some embodiments, the basic computational cost can be determined in any of the following ways: Lookup table method: For each task type, since the benchmark library pre-stores its unit computation intensity under the representative model / representative input scale, such as the standard unit corresponding to processing 1 sample / 1 token or the standard unit / second conversion rate corresponding to the throughput per second, the basic computation volume can be directly read by querying the benchmark library according to the task type.
[0051] Statistical induction: When there are multiple representative models of the same task type in the benchmark library, such as TOP3, the unit computation intensity of these models can be weighted averaged or the median can be taken, and the weighted average or median result can be used as the basic computational quantity of the task type to reduce the bias of a single model, such as the bias caused by a certain model being extremely bandwidth sensitive.
[0052] Similarity mapping method: When the task type of the target business does not have a completely identical entry in the benchmark library, similarity can be calculated through task feature tags, such as sequence length, sparse / dense, operator type, etc., and then the similarity can be mapped to the most similar standard task type. Then, its basic computational cost can be read, thereby improving coverage and enhancing feasibility.
[0053] It should be noted that the above-listed methods for determining the basic computational load are merely illustrative examples of the embodiments of this application and do not impose any limitations on the embodiments of this application.
[0054] The task type pattern coefficient characterizes the difference in additional computational / system overhead for the same task type under different computational modes. For example, training typically requires backpropagation and gradient-related computations compared to inference, leading to a significant increase in computational cost. Therefore, the task type pattern coefficient for training tasks can be set to 2.0, while the task type pattern coefficient for inference tasks can be set to 1.0.
[0055] In practice, if the task type is a training task, such as pre-training / fine-tuning / continuous training, the task type mode coefficient can be set to the training mode coefficient, such as 1.0.
[0056] Alternatively, if the task type is a reasoning task, such as online reasoning / batch reasoning / service-based reasoning, the task type mode coefficient can be set to the reasoning mode coefficient, such as 2.0.
[0057] The benchmark calibration coefficient is used to compensate for systematic differences between the standardized performance data in the benchmark library and the actual operation of the target business, such as differences in test configuration (accuracy, cluster size, software stack version), differences in input size (batch / seq length / resolution), and differences in the source of benchmark library data (different testing institutions or different test batches).
[0058] In some embodiments, the standardized performance indicators of the business load requirement information can be determined based on the basic computational load, task type mode coefficient, and benchmark library calibration coefficient, according to the following calculation method:
[0059] in, Standardized performance metrics that represent business load demand information; Indicates the basic computational load; Indicates the task type mode coefficient; This represents the calibration coefficient of the reference library.
[0060] After obtaining the standardized performance metrics, equivalent performance metrics can be further determined based on these metrics. Specifically, the ratio of the number of tasks in each task type within the target service to the total number of tasks in the target service can be determined; the service demand corresponding to each task type can be determined based on service load demand information; and the equivalent performance metrics of the target service can be determined by weighted aggregation based on the task quantity ratio, service demand, standardized performance metrics, and hardware efficiency calibration parameters.
[0061] Among them, the equivalent performance index is used to characterize the overall load intensity of the target business under a unified standard.
[0062] It should be noted that, in the embodiments of this application, determining the equivalent performance index is to unify the business demand of different task types involved in the target business, such as image tasks and large language model tasks, under different original units, into a standard that can be added and benchmarked through standardized performance indexes, and further combine hardware efficiency calibration parameters to correct the differences in effectiveness, thereby forming an overall equivalent index of the business that can be used for subsequent calculation of effective computing power demand.
[0063] The number of tasks included in each task type refers to the number of task instances falling into a certain task type. A task instance can correspond to an inference request, a training job, a batch processing task, or other countable task units. The total number of tasks refers to the total number of task instances included in the target service. In some embodiments, task type labeling and classification statistics can be performed on each task instance based on the service scheduling system, job queue logs, service gateway logs, or service-side configuration list to obtain the number of tasks for each task type. The ratio between the number of tasks for each task type and the total number of tasks can then be calculated as the proportion of the number of tasks for that task type in the target service.
[0064] The task quantity ratio is used to characterize the task structure and contribution weight of the target business. Without the task quantity ratio, the equivalent performance index may implicitly assume equal weighting for each task type during aggregation, potentially leading to overestimation of low-frequency tasks or underestimation of high-frequency tasks, thus reducing the accuracy of subsequent effective computing power requirement calculations. In this embodiment, by introducing the task quantity ratio, the equivalent performance index aligns with the actual task structure of the target business, thereby reducing configuration deviations.
[0065] Business demand is used to characterize the scale of business that a certain task type needs to be processed per unit of time. For example, it can be throughput demand (such as the number of samples processed per second, the number of tokens per second, the number of requests per second, etc.), arrival rate, concurrency, or other quantities that can equivalently reflect the scale of business.
[0066] Hardware efficiency calibration parameters, also known as MFU calibration coefficients, are used to correct for effective computing power deviations caused by differences in hardware efficiency and / or software optimization. In this embodiment, when determining the hardware efficiency calibration parameters, the target model utilization rate of the target service can be obtained first; then, the reference model utilization rate of the standard task can be obtained from the benchmark library; finally, the hardware efficiency calibration parameters can be determined based on the ratio between the target model utilization rate and the reference model utilization rate.
[0067] Specifically, the formula for calculating the hardware efficiency calibration parameters can be expressed as follows:
[0068] in, This indicates the hardware efficiency calibration parameters. Indicates the utilization rate of the target model. This indicates the utilization rate of the reference model.
[0069] It should be noted that, considering practical applications, while directly using measured MFUs as the calibration basis is theoretically more accurate, sufficient measured data is often difficult to obtain during the planning phase, and it places high demands on users and engineering implementation. Alternatively, using fixed coefficients to uniformly correct all services and hardware can be too coarse and fail to reflect the differences in different task types, hardware platforms, and software optimization levels. Therefore, this application provides a compromise: constructing an MFU experience value library. This library pre-stores model utilization experience values or ranges associated with at least one dimension, such as task type, hardware type, and software stack configuration. During the planning phase, the corresponding target model utilization is selected from the MFU experience value library based on the task type and target hardware characteristics of the target service to determine the MFU calibration coefficient value.
[0070] For example, when the target service's optimization level is poor, resulting in low model utilization, the MFU calibration coefficient can be set to greater than 1. This allows for appropriate amplification of demand redundancy during computing power allocation planning, reducing performance risks caused by insufficient effective computing power. Furthermore, to facilitate measurement and engineering implementation, the value of the MFU calibration coefficient can be consistent with the calculation process of the target service's equivalent throughput when determining effective computing power requirements. In other words, the MFU calibration coefficient can be used as a unified calibration parameter in the conversion of equivalent throughput and the derivation of effective computing power requirements, thereby ensuring the consistency and reproducibility of the unified conversion on the demand side and the effectiveness correction on the supply side.
[0071] Among them, the target model utilization rate is used to characterize the effective computing power utilization level when the target business is executed under the target hardware, which can be understood as a measure of the actual effective computing utilization during the model execution process.
[0072] Optionally, the utilization rate of the target model can be obtained using one of the following methods: (1) Conduct trial runs / stress tests on the target business or a representative load equivalent to it on the target hardware, and obtain the utilization rate of the target model through hardware counters, framework performance analysis, or training / inference log statistics; (2) If it is difficult to measure the utilization rate of the target model in the planning stage, the estimated value of the target model utilization rate can be obtained by indexing the pre-built MFU experience value library by task type, model size, hardware type, software stack version, etc., and used as the target model utilization rate.
[0073] Reference model utilization refers to the model utilization data in the benchmark library corresponding to the standard task. For example, taking the mapping matrix B listed above, the MFU in mapping matrix B is the model utilization data.
[0074] In some embodiments, after obtaining the task quantity ratio, business demand, standardized performance indicators, and hardware efficiency calibration parameters based on the above content, this information can be weighted and aggregated in the following manner to obtain the equivalent performance indicators of the target business:
[0075] in, Indicates the equivalent performance index of the target business; Indicates the percentage of tasks; Indicates standardized performance indicators; This indicates the hardware efficiency calibration parameters.
[0076] Step 106: Determine the effective computing power requirements of the target service based on the equivalent performance indicators and the theoretical benchmark performance indicators of the target hardware under standard tasks.
[0077] In some embodiments, the theoretical benchmark performance index of the target hardware under standard tasks can be determined as follows: At least one standard task corresponding to the target service is obtained; the theoretical benchmark service demand corresponding to each standard task is obtained based on a benchmark library, where the theoretical benchmark service demand originates from benchmark test results in the benchmark library; the theoretical benchmark throughput corresponding to each standard task is determined based on the theoretical benchmark service demand, standardized performance data, and hardware efficiency calibration parameters; and the theoretical benchmark throughput corresponding to each standard task is aggregated to obtain the theoretical benchmark performance index of the target hardware under standard tasks.
[0078] Specifically, based on the task types reflected in the workload requirements of the target business, standard tasks in the benchmark library can be mapped one-to-one. For example, when the target business involves task types such as image classification, object detection, and large language modeling, standard tasks of the corresponding task types in the benchmark library can be selected as standard tasks. When there are no completely identical entries for the target business task type in the benchmark library, similarity matching can be further performed based on the task's input features (such as sequence length, resolution, batch size), computational features (such as whether backpropagation is used, whether gradient synchronization exists), and resource-sensitive features (such as bandwidth sensitivity / communication sensitivity) to select standard tasks that are closer to the target business and improve mapping accuracy.
[0079] For each standard task, the corresponding benchmark test entry can be located in the benchmark library, and the benchmark test results in that entry can be read as the theoretical benchmark service requirement. This benchmark test result can be the throughput result of the standard task under specific hardware / specific parameters, such as images / sec, tokens / sec, etc., or it can be a pre-defined representative value in the benchmark library, such as the results of the Top model, the statistical representative value of the Top 3 model results, etc. For situations with multiple models or multiple test batches, preset rules can be used to select the theoretical benchmark service requirement, such as selecting the representative model results from the official benchmark test, or taking a weighted average / median of the Top 3 model results to reduce the bias of a single model.
[0080] Then, based on the theoretical baseline service requirements, standardized performance data, and hardware efficiency calibration parameters of each standard task, the theoretical baseline throughput corresponding to each standard task can be determined.
[0081] Finally, based on the theoretical benchmark throughput corresponding to each standard task, the theoretical benchmark performance indicators of the target hardware under the standard tasks can be obtained by aggregating them in the following manner:
[0082] in, This represents the theoretical baseline performance index of the target hardware under standard tasks; Indicates standard task i The corresponding theoretical benchmark throughput.
[0083] After obtaining the theoretical benchmark performance indicators of the target hardware under standard tasks, the effective computing power requirements of the target service can be determined based on the equivalent performance indicators and the theoretical benchmark performance indicators of the target hardware under standard tasks.
[0084] like Figure 2As shown, in some implementations, the effective computing power requirements of the target service can be determined according to the following steps, based on the equivalent performance indicators and the theoretical benchmark performance indicators of the target hardware under standard tasks: Step 202: Determine the theoretical peak computing power and global adjustment factor of the target hardware. The global adjustment factor is used to compensate for the unmodeled performance loss introduced by the difference between the operation of standard tasks and target services.
[0085] The theoretical peak computing power characterizes the theoretical maximum computing capability of the target hardware, and its unit can be FLOP / s. In practical applications, the theoretical peak computing power can be obtained from publicly available specifications from hardware manufacturers, information reported by hardware device drivers / management interfaces, or preset hardware model configuration tables. When the target hardware is a multi-card server or a multi-node cluster, the theoretical peak computing power of a single card can be aggregated by the number of cards / nodes to obtain the theoretical peak computing power of the target hardware, for example, by summing the total number of cards, thus providing a clear peak reference for subsequent calculations of effective computing power requirements.
[0086] In one alternative implementation, the global adjustment factor can be determined based on historical operating data, trial operation stress test results, or planning experience values; and it can be adjusted later based on actual measurement data feedback. For example, if the relevant efficiency is low for a long period of time in actual operation, the global adjustment factor can be appropriately increased to improve the robustness of the planning to complex system factors.
[0087] Optionally, unmodeled performance overhead may include at least one of the following: communication overhead, I / O overhead, resource scheduling and load balancing overhead, memory / video memory bandwidth limitations, and performance overhead caused by concurrency contention.
[0088] Step 204: Calculate the ratio between the equivalent performance index and the theoretical benchmark performance index.
[0089] This ratio can be understood as the relative multiple of the target business's demand intensity under a unified standard to the target hardware's baseline capability under a standard task standard. When the ratio is greater than 1, it indicates that the target business's demand intensity is higher than the target hardware's baseline capability, requiring a corresponding increase in computing power; when the ratio is less than 1, it indicates that the target business's demand intensity is lower than the target hardware's baseline capability, and the corresponding computing power requirement can be reduced accordingly.
[0090] Step 206: Multiply the ratio, the theoretical peak computing power of the target hardware, and the global adjustment factor to obtain the effective computing power requirement.
[0091] In some embodiments, the effective computing power requirement can be determined as follows:
[0092] in, Indicates the effective computing power requirement; This represents the theoretical peak computing power of the target hardware. This represents the global adjustment factor.
[0093] Step 108: Based on the effective computing power requirements, determine the effective computing power configuration plan for the intelligent computing center.
[0094] Among them, the effective computing power configuration planning scheme is used to output a resource configuration list that can directly guide the construction and procurement of intelligent computing centers. It includes at least one of the following: computing power resource configuration, such as server model / quantity, accelerator model / quantity, and cluster size.
[0095] Optionally, an effective computing power configuration plan can further include infrastructure requirements that match the computing power resource configuration, such as the number of racks, network bandwidth, power supply capacity, cooling capacity, etc.
[0096] In practice, the effective supply capacity of the target hardware can be determined first, that is, the effective computing power capacity that a single server, a single accelerator, or a single node can provide under the standard task caliber consistent with the benchmark library. The effective computing power capacity can be derived from the benchmark performance indicators, model utilization data, and global adjustment factors of the target hardware under the standard task, or obtained from trial operation and actual testing.
[0097] Subsequently, the effective computing power demand and effective computing power capacity are matched and converted to determine the amount of resources required to meet the effective computing power demand. For example, when the effective computing power demand is greater than the effective computing power capacity of a single resource unit, the number of resource units is determined by rounding up, and the number of resources is further corrected according to the preset redundancy strategy (such as N+1 redundancy, expansion reservation ratio). When multiple candidate hardware solutions exist, the selection and combination optimization can be performed based on at least one of the following constraints: cost, energy efficiency, power density, rack space constraints, network topology constraints, and scalability constraints, under the constraint of meeting the effective computing power requirements, so as to output the final effective computing power configuration planning scheme.
[0098] Through the above methods, the embodiments of this application can use effective computing power demand as a planning benchmark to directly map the actual effective demand on the business side into the number of feasible resource configurations and infrastructure support requirements, thereby avoiding resource waste or insufficient performance caused by relying solely on theoretical peak computing power and empirical redundancy, and improving the accuracy of computing power configuration planning and resource utilization efficiency of intelligent computing centers.
[0099] The following examples illustrate how the methods provided in the embodiments of this application are applied in practice.
[0100] 1. Scene Description Assume the target business scenario of the intelligent computing center is an autonomous driving full-stack model training center, which includes image classification tasks, object detection tasks, and large language model tasks.
[0101] Target business requirement type: Image classification task, accounting for 10%, with a business demand of 5,000 img / s; Target detection tasks account for 60% of the total, with a business requirement of 15,000 img / s; Large language model tasks account for 30%, with a business demand of 50,000 tokens / s; 2. Example of building a benchmark library This application's embodiments are based on a benchmark library of MLPerfTrainingv3.1 results, showcasing task types and hardware platform classifications, including the Top 3 models and mainstream hardware NVIDIA / HUAWEI / Cambricon.
[0102] in: Task categories: Covers all 6 task categories in MLPerf; Model selection: The top 3 models officially adopted by MLPerf are selected for each task category; Hardware platform: NVIDIA: A100 / H100 (current mainstream data center GPUs) HUAWEI: Ascend 910B (representing domestically produced NPUs) CAMBricon: MLU370-X8 (Cambricon flagship accelerator card) Performance values: Throughput (units / second), MFU (%), and SU_rate (standard units / second) after equivalent unit conversion.
[0103] The test configuration was uniformly set to an 8-card cluster with FP16 precision.
[0104] Some of the reference library's schematic tables are shown in Tables 1 to 3 below: Table 1. Schematic diagram of the benchmark library (image classification task)
[0105] Table 2. Benchmark Library Diagram (Target Detection Task)
[0106] Table 3. Benchmark Library Diagram (Large Language Model Task)
[0107] SU_rate calculation instructions: The training task SU_rate is improved by 2-3 times (due to increased backpropagation computation). LLM training significantly reduces MFU (gradient synchronization overhead).
[0108] 3. Calculate the effective computing power requirements of the target business: (1) Calculate the equivalent performance index of the target service. In this embodiment, the performance index is taken as the throughput, that is, calculate the equivalent throughput of the target service: (2) Obtain the theoretical baseline performance indicators of the target hardware under standard tasks. Continuing with the previous example, take throughput as an example, that is, obtain the theoretical baseline throughput of the target hardware under standard tasks:
[0109] in, .
[0110] (3) Calculate the effective computing power requirement:
[0111] (4) Effective computing power requirements, output server configuration table.
[0112] Assume that the effective capacity of a single server is 8.09 PFLOPS; Based on the effective computing power requirement: 42.55 PFLOPS; The number of servers can be determined as: Nserver = 42.55 / 8.09 = 5.26 ≈ 6 units; Actual total computing power (effective) = 6 × 8.09 PFLOPS = 48.54 PFLOPS; For ease of comparison later, the computing power of this scheme can be converted into theoretical peak value: Actual total computing power (theoretical) = 6 × 15.8 PFLOPS = 94.8 PFLOPS.
[0113] In one optional implementation, to verify the consistency of the effective computing power configuration planning scheme obtained based on the effective computing power demand and reduce the planning risk caused by parameter value deviations, after obtaining the effective computing power configuration planning scheme, a peak computing power configuration planning scheme can also be generated based on the peak value. Deviation evaluation and closed-loop calibration are then performed on the two configuration planning schemes. This improves the rationality of the planning parameter values and the consistency of the planning results, thereby enhancing the reliability and feasibility of the configuration planning. The specific process is as follows: Figure 3 As shown, it includes the following steps: Step 302: Determine the theoretical peak computing power requirement of the target service, and determine the planned peak computing power requirement of the target service based on the theoretical peak computing power requirement and the hardware redundancy coefficient.
[0114] In this embodiment of the application, when determining the theoretical peak computing power requirement of the target service, the model computing volume, service scale, and time constraint coefficient of the target service can be determined first; then, based on the model computing volume, service scale, time constraint coefficient, and the preset reserved computing power margin, the theoretical peak computing power requirement of the target service can be determined.
[0115] In this embodiment, the theoretical peak computing power requirement is used to characterize the theoretical computing capacity required to meet the peak load of the target business under the peak planning caliber. It can be determined comprehensively based on parameters such as the model computational load, business scale, and time constraint coefficient of the target business. Specifically, the model computational load characterizes the computational intensity corresponding to a unit of business processing volume; the business scale characterizes the amount of tasks to be processed per unit time during the peak period; and the time constraint coefficient characterizes the impact of completion deadlines on the peak computing power requirement under project-based delivery or continuous operation scenarios. The reserved computing power margin, also known as the redundancy coefficient, can be preset according to actual needs.
[0116] In some embodiments, the theoretical peak computing power requirement of the target service The calculation formula can be expressed as follows:
[0117] Optionally, to convert the theoretical peak computing power requirement into a feasible planning framework, after obtaining the theoretical peak computing power requirement of the target business, the planned peak computing power requirement of the target business can be determined based on the theoretical peak computing power requirement and the hardware redundancy coefficient. The hardware redundancy coefficient represents the proportion of hardware redundancy reserved for expansion, sudden fluctuations, and uncertainties, such as reserving 20% expansion space.
[0118] In some embodiments, the formula for calculating the planned peak computing power requirement of the target service can be expressed as follows:
[0119] in, This indicates the planned peak computing power requirement for the target business. This represents the hardware redundancy factor.
[0120] In this embodiment of the application, by introducing a hardware redundancy coefficient, it can be ensured that the planned peak computing power demand can reduce the risk of insufficient peak resources when the business is not yet fully solidified or the peak fluctuation is large.
[0121] Step 304: Determine the peak computing power configuration plan for the intelligent computing center based on the planned peak computing power requirements.
[0122] In this embodiment, the peak computing power configuration planning scheme is used to output a resource configuration list based on peak performance. In specific implementation, candidate hardware specifications (including server model, GPU / accelerator model, theoretical peak computing power, etc.) can be collected first, and then the planned peak computing power requirement can be converted into the required number of servers / accelerators. Furthermore, the requirements for infrastructure such as racks, power, and cooling can be output, thereby forming a configuration list that can guide procurement and construction.
[0123] Specifically, when collecting candidate hardware specifications, a server selection matrix can be constructed by surveying the current and near-term (planning period) products of major GPU manufacturers (NVIDIA, AMD, Intel) and server OEMs (Dell, HPE, Lenovo, Supermicro, Inspur, Huawei, etc.).
[0124] Secondly, select server configurations. Based on the planned peak total computing power demand, and taking into account factors such as cost, energy efficiency (FLOPS / W), power density, network requirements, future scalability, and vendor support, select one or more main server configurations from the matrix. Finally, the system calculates infrastructure requirements (racks, power, cooling, space) and outputs configuration tables, including a configuration list (server configuration master table, rack configuration table, and infrastructure requirements summary table), providing a direct basis for rack procurement, data center design, and construction.
[0125] The following examples illustrate how the peak computing power configuration planning scheme for intelligent computing centers based on planned peak computing power requirements, as provided in this application, can be applied in practice.
[0126] Assuming that this application embodiment plans a smart computing data center, the business type is not yet fully defined (preliminary judgment indicates it involves image-related AI processing), and the theoretical peak computing power requirement of its target business is estimated based on experience data from similar businesses, then: (1) Calculate the theoretical peak computing power requirement of the target service:
[0127] The values of each core parameter are determined based on the following: Model computational cost: Using a conservative industry experience value, i.e., 0.2 PFLOPS / sample; Business scale: Estimated based on peak load = 200 samples / second; Time constraint factor: Given a normal project cycle, set to = 1.0; Hardware redundancy factor: Since the business model is not fixed, the median value is taken as 2.5.
[0128] (2) Determine the planned peak computing power requirement for the target service based on the theoretical peak computing power requirement and the hardware redundancy coefficient:
[0129] The hardware redundancy factor is set at 0.2 (with 20% capacity reserved for expansion).
[0130] (3) Determine the peak computing power configuration plan of the intelligent computing center based on the planned peak computing power demand, that is, convert the planned peak computing power demand into actual physical resource allocation.
[0131] This application embodiment uses an NVIDIA DGX H100 server (each unit is configured with 8×H100 GPUs): At this point, the peak computing power configuration plan is as follows: The theoretical single-machine computing power is 15.8 PFLOPS (based on FP16 tensor core mixed precision). Server quantity calculation: Nserver = 120 / 15.8 = 7.59 ≈ 8 servers The actual total computing power (theoretical) of the peak computing power configuration plan = 8 × 15.8 PFLOPS = 126.4 PFLOPS Resource allocation conclusions of the peak computing power configuration planning scheme: This solution, configured with 8 DGX H100 servers (a total of 64 H100 GPUs), can meet the following requirements: Covering peak business demand of 200 samples / second (calculated using a conservative model); Provides a 20% hardware expansion redundancy margin; This provides a buffer against the risks arising from uncertainties in the business model over the next 12-18 months.
[0132] Step 306: The deviation between the determined peak computing power configuration plan and the effective computing power configuration plan.
[0133] In this embodiment, the deviation is used to measure the degree of difference between the peak computing power configuration planning scheme and the effective computing power configuration planning scheme. The deviation can be determined according to a preset deviation calculation caliber. For example, the relative difference can be calculated based on the total configuration computing power (or resource dimensions such as the number of servers and GPUs) corresponding to the two schemes, thereby obtaining a deviation value that can be used for threshold judgment.
[0134] The preset deviation threshold can be set according to the planning target, for example, it can usually be set to 10% to 30%.
[0135] In some embodiments, deviation The calculation formula can be expressed as follows:
[0136] in, This represents the peak computing power configuration planning scheme; This represents the effective computing power allocation plan.
[0137] By introducing deviation and preset deviation threshold, the difference between the peak computing power configuration planning scheme and the effective computing power configuration planning scheme can be quantified, enabling the system to automatically enter the parameter correction and scheme update process when the difference is too large, thereby avoiding over-configuration or conservative configuration caused by the long-term separation of the two sets of standards.
[0138] Step 308: When the deviation is greater than the preset deviation threshold, the computing power configuration planning scheme update operation is executed repeatedly until the deviation between the updated peak computing power configuration planning scheme and the updated effective computing power configuration planning scheme is less than or equal to the preset deviation threshold.
[0139] The operation of updating the computing power configuration planning scheme includes: Adjust the hardware redundancy factor and / or the hardware efficiency calibration parameters used to determine the equivalent performance metrics; The peak computing power configuration planning scheme and the effective computing power configuration planning scheme are updated based on the adjusted hardware redundancy coefficient and / or hardware efficiency calibration parameters.
[0140] In some embodiments, when the deviation is greater than a preset deviation threshold, a cyclic update is performed: each round of update performs a computing power configuration planning scheme update operation, and regenerates the peak computing power configuration planning scheme and the effective computing power configuration planning scheme based on the updated parameters, and then recalculates the deviation; until the updated deviation is less than or equal to the preset deviation threshold.
[0141] Optionally, when |δ| > the threshold, an optimization update strategy can be activated, for example: Hardware layer: Switching to high MFU devices, improving communication efficiency, etc.
[0142] Business layer: dynamic weight allocation, model lightweighting, etc.
[0143] To reduce the deviation, repeat the above steps until the planned target is met.
[0144] It should be noted that the activation optimization and update strategies listed above are merely exemplary descriptions of embodiments of this application and do not impose any limitations on embodiments of this application.
[0145] In some embodiments, the computing power configuration planning scheme update operation may include at least one of the following: a) Adjust the hardware redundancy factor; When the deviation indicates that the peak caliber scheme is significantly overestimated or underestimated relative to the effective caliber scheme, the hardware redundancy coefficient can be adjusted to correct the planned peak computing power requirement, thereby updating the peak computing power configuration plan.
[0146] b) Adjust the hardware efficiency calibration parameters used to determine the equivalent performance metrics; Hardware efficiency calibration parameters are used to correct for differences in hardware efficiency and / or software optimization when determining equivalent performance indicators. They can be determined by the ratio of the target model utilization to the reference model utilization. When the target service optimization is poor (low MFU), the calibration coefficient is greater than 1 to appropriately amplify the demand redundancy.
[0147] Therefore, when the deviation indicates that the effective computing power requirement is underestimated or overestimated, the effective computing power requirement and effective computing power configuration planning scheme can be updated by adjusting the hardware efficiency calibration parameters, such as switching the MFU empirical value, or updating the target model utilization estimate and equivalent performance index based on trial operation data.
[0148] c) Update the peak computing power configuration planning scheme and the effective computing power configuration planning scheme based on the adjusted hardware redundancy coefficient and / or hardware efficiency calibration parameters; In each update cycle, the system executes the following: Based on the adjusted hardware redundancy coefficient, the peak computing power requirement is redefined, and the peak computing power configuration plan is updated. The equivalent performance index is re-determined based on the adjusted hardware efficiency calibration parameters, thereby affecting the effective computing power requirement and updating the effective computing power configuration planning scheme. The deviation between the two schemes is then recalculated, and if it is still greater than the threshold, the next round of updates will begin.
[0149] Using the method provided in this application, the cyclic update mechanism is equivalent to establishing a consistency constraint between peak caliber planning and effective caliber planning: on the one hand, the peak caliber path can quickly absorb the planned reserved demand through the hardware redundancy coefficient; on the other hand, the effective caliber path can absorb the differences between hardware efficiency and software optimization through the hardware efficiency calibration parameter. Both converge iteratively under the deviation threshold constraint, ensuring that the final solution neither deviates from the peak caliber commonly used in engineering procurement, nor deviates from the planning deviation with effective computing power as the core objective.
[0150] The following examples illustrate how the methods provided in the embodiments of this application are applied in practice.
[0151] For ease of comparison, the computing power of the effective computing power configuration plan is converted into the theoretical peak value: the actual total computing power (theoretical) of the effective computing power configuration plan = 6 × 15.8 PFLOPS = 94.8 PFLOPS.
[0152] The planning conclusions obtained from the peak computing power configuration planning scheme and the effective computing power configuration planning scheme are compared, evaluated and verified, and the deviation δ is calculated.
[0153] Because the deviation δ in this embodiment is greater than the preset deviation threshold: |δ|>20%; Therefore, the plan needs to be updated; repeat the above steps until the expected goals of the plan are met.
[0154] Specifically, in this embodiment, the parameters of the peak computing power configuration planning scheme can be adjusted. If the business requirements are clearly defined, the hardware redundancy coefficient is set to 0.02. In this case, the updated peak computing power configuration planning scheme is as follows: Total planned computing power requirements: =100×(1+0.02)=102 PFLOPS.
[0155] The theoretical single-machine computing power is 15.8 PFLOPS (based on FP16 tensor core mixed precision). Server count calculation: Nserver = 102 / 15.8 = 6.46 ≈ 7 servers; The actual total computing power (theoretical) of the peak computing power configuration plan = 7 × 15.8 PFLOPS = 110.6 PFLOPS; Recalculate the deviation δ: Deviation in the updated peak computing power configuration planning scheme:
[0156] Resource allocation conclusion: At this point, the deviation |δ| < 20%. Based on the objective, this application adopts an effective computing power configuration planning scheme.
[0157] This plan configures 6 DGX H100 servers (a total of 48 H100 GPUs). The following table compares several computing power configuration plans: Table 4 Comparison of Several Computing Power Configuration Planning Schemes
[0158] This application embodiment generates three schemes through a dual-path computing power planning mechanism: Peak computing power configuration planning scheme (purely theoretical): ① Using the traditional theoretical peak method, a redundancy factor of 20% is set (total demand of 120 PFLOPS). ② Configure 8 DGX H100s (64 GPUs); ③ The actual computing power utilization rate is only 33.66% (126.4 PFLOPS theoretical computing power vs 42.55 PFLOPS effective demand).
[0159] Effective computing power configuration planning scheme: Based on accurate calculations of standard computing power conversion rates, only 6 DGX H100s (48 GPUs) were configured, achieving a utilization rate of 51.16% (48.54 PFLOPS effective computing power ÷ 94.8 PFLOPS theoretical computing power).
[0160] Updated peak computing power configuration planning scheme (theoretical-driven optimized version): The redundancy coefficient was reverse-calibrated using the calculation results of the effective computing power configuration planning scheme (from 20% to 2%); the configuration was reduced to 7 DGX H100s (56 GPUs).
[0161] The computing power demand calculation method based on standard load presented in this application improves planning efficiency and resource utilization by accurately quantifying computing power demand through the construction of an MLPerf benchmark library. Compared with traditional planning techniques, the computing power demand calculation method based on standard load uses data derived from official MLPerf v3.1 test results, covering six major categories of AI tasks, Top 3 models, and mainstream hardware platforms, ensuring the authority and reproducibility of the benchmark library. The peak computing power configuration planning scheme (theoretical peak-driven method) provides safety redundancy during periods of business uncertainty, avoiding the risk of insufficient resources; the effective computing power configuration planning scheme achieves accurate matching of effective computing power through dynamic benchmark mapping after the business is clarified.
[0162] The method provided in this application obtains the service load demand information of the target service and converts it into standardized performance indicators using a pre-built benchmark library, thereby obtaining equivalent performance indicators characterizing the overall load level of the target service. Simultaneously, by introducing theoretical benchmark performance indicators of the target hardware under standard tasks to calibrate the equivalent performance indicators, the effective computing power demand reflecting the actual load characteristics, memory bandwidth constraints, and software optimization differences of the target hardware can be determined. In this way, the effective computing power demand no longer relies solely on the theoretical peak computing power and empirical redundancy provided by hardware manufacturers, but is based on the mapping relationship between standard task performance data and theoretical benchmark performance indicators. This improves the accuracy of the effective computing power configuration planning scheme determined based on the effective computing power demand and reduces resource waste or performance insufficiency caused by mismatch between resource allocation and actual demand.
[0163] To address the problem in existing technologies where resource allocation relies on theoretical peak computing power and empirical redundancy provided by hardware manufacturers, leading to resource waste or insufficient performance due to mismatch between resource allocation and actual needs, this application provides a computing power allocation planning device for intelligent computing centers. A schematic diagram of the device's specific structure is shown below. Figure 4 As shown, it includes an information acquisition module 41, an indicator determination module 42, a demand determination module 43, and a solution determination module 44. The functions of each module are as follows: Information acquisition module 41 is used to acquire the business load requirement information of the target business for the intelligent computing center computing power configuration planning; The indicator determination module 42 is used to determine the standardized performance indicators of business load demand information based on the pre-built benchmark library, and to determine the equivalent performance indicators based on the standardized performance indicators; wherein, the benchmark library stores standardized performance data of various standard tasks on different hardware. The requirement determination module 43 is used to determine the effective computing power requirement of the target service based on the equivalent performance index and the theoretical benchmark performance index of the target hardware under standard tasks. The scheme determination module 44 is used to determine the effective computing power configuration planning scheme of the intelligent computing center based on the effective computing power requirements.
[0164] Optionally, the business load requirement information shall include at least: the types of tasks involved in the target business; this metric determination module is used for: Based on the benchmark library and the task types involved in the target business, the basic computational load, task type mode coefficients, and benchmark library calibration coefficients of the target business are determined. Based on the basic computational load, task type mode coefficient, and benchmark library calibration coefficient, the standardized performance indicators of business load requirements are determined.
[0165] Optionally, the indicator determination module 42 is used for: Determine the ratio of the number of tasks in each task type within the target business to the total number of tasks in the target business; Determine the business demand for each task type based on business load demand information. The equivalent performance indicators of the target business are determined by weighted aggregation based on the proportion of task quantity, business demand, standardized performance indicators, and hardware efficiency calibration parameters.
[0166] Optionally, the computing power configuration planning device of the intelligent computing center also includes a calibration coefficient determination module; the calibration coefficient determination module is used for: Obtain the target model utilization rate for the target business; Obtain the utilization rate of reference models for standard tasks from the benchmark library; The hardware efficiency calibration parameters are determined based on the ratio between the utilization rate of the target model and the utilization rate of the reference model.
[0167] Optional, requirement determination module 43, used for: Determine the theoretical peak computing power and global adjustment factor of the target hardware. The global adjustment factor is used to compensate for the unmodeled performance loss introduced by the difference between the operation of standard tasks and target services. Calculate the ratio between the equivalent performance index and the theoretical benchmark performance index; The effective computing power requirement is obtained by multiplying the ratio, the theoretical peak computing power of the target hardware, and the global adjustment factor.
[0168] Optionally, the computing power configuration planning device of the intelligent computing center also includes a theoretical benchmark performance index determination module; the theoretical benchmark performance index determination module is used for: Obtain at least one standard task corresponding to the target business; The theoretical benchmark business requirements corresponding to each standard task are obtained from the benchmark library. The theoretical benchmark business requirements are derived from the benchmark test results in the benchmark library. Based on the theoretical baseline service demand, standardized performance data, and hardware efficiency calibration parameters of each standard task, the theoretical baseline throughput corresponding to each standard task is determined. The theoretical benchmark throughput corresponding to each standard task is aggregated to obtain the theoretical benchmark performance index of the target hardware under the standard task.
[0169] Optionally, the computing power configuration planning device of the intelligent computing center further includes a computing power configuration planning scheme update module, which includes: The first determining unit is used to determine the theoretical peak computing power requirement of the target service, and to determine the planned peak computing power requirement of the target service based on the theoretical peak computing power requirement and the hardware redundancy coefficient. The second determining unit is used to determine the peak computing power configuration planning scheme of the intelligent computing center based on the planned peak computing power demand. The third determining unit is used to determine the deviation between the peak computing power configuration planning scheme and the effective computing power configuration planning scheme; The update unit is used to repeatedly execute the computing power configuration planning scheme update operation when the deviation is greater than the preset deviation threshold, until the deviation between the updated peak computing power configuration planning scheme and the updated effective computing power configuration planning scheme is less than or equal to the preset deviation threshold. The operation of updating the computing power configuration planning scheme includes: Adjust the hardware redundancy factor and / or the hardware efficiency calibration parameters used to determine the equivalent performance metrics; The peak computing power configuration planning scheme and computing power configuration planning scheme are updated based on the adjusted hardware redundancy coefficient and / or hardware efficiency calibration parameters.
[0170] Optionally, the first determining unit is used for: Determine the computational load, business scale, and time constraint coefficients of the target business model; Based on the model's computational load, business scale, time constraint coefficient, and preset reserved computing power margin, determine the theoretical peak computing power requirement of the target business.
[0171] Using the apparatus provided in this application, the workload requirements of the target service are acquired, and the workload requirements are converted into standardized performance indicators using a pre-built benchmark library, thereby obtaining equivalent performance indicators characterizing the overall workload level of the target service. Simultaneously, by introducing theoretical benchmark performance indicators of the target hardware under standard tasks to calibrate the equivalent performance indicators, the effective computing power requirements reflecting the actual load characteristics, memory bandwidth constraints, and software optimization differences of the target hardware can be determined. In this way, the effective computing power requirements no longer rely solely on the theoretical peak computing power and empirical redundancy provided by hardware manufacturers, but are based on the mapping relationship between standard task performance data and theoretical benchmark performance indicators. This improves the accuracy of the effective computing power configuration planning scheme determined based on the effective computing power requirements, and reduces resource waste or performance insufficiency caused by mismatch between resource allocation and actual needs.
[0172] Figure 5 To illustrate the hardware structure of an electronic device according to various embodiments of this application, the electronic device may include a processor 501 and a memory 502 storing computer program instructions. Specifically, the processor 501 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of this application.
[0173] Memory 502 may include mass storage for data or instructions. For example, and not limitingly, memory 502 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 502 may include removable or non-removable (or fixed) media. Where appropriate, memory 502 may be internal or external to an electronic device. In a particular embodiment, memory 502 may be a non-volatile solid-state memory.
[0174] In one embodiment, memory 502 may be read-only memory (ROM). In one embodiment, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically rewritable ROM (EAROM), or flash memory, or a combination of two or more of these.
[0175] The processor 501 reads and executes computer program instructions stored in the memory 502 to implement any of the computing power configuration planning methods for intelligent computing centers in the above embodiments.
[0176] In one example, the electronic device may also include a communication interface 503 and a bus 510. Wherein, as... Figure 5 As shown, the processor 501, memory 502, and communication interface 503 are connected through bus 510 and complete communication with each other.
[0177] The communication interface 503 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.
[0178] Bus 510 includes hardware, software, or both, that couples components of an electronic device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 510 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, this application contemplates any suitable bus or interconnect.
[0179] Furthermore, in conjunction with the computing power configuration planning method for intelligent computing centers in the above embodiments, this application embodiment can provide a computer-readable storage medium for implementation. This computer-readable storage medium stores computer program instructions; when these computer program instructions are executed by a processor, they implement any of the computing power configuration planning methods for intelligent computing centers in the above embodiments.
[0180] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.
[0181] The above description is merely a specific implementation example of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.
[0182] Secondly, those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0183] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0184] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0185] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0186] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.
[0187] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.
[0188] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.
[0189] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0190] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.
Claims
1. A method for planning the computing power allocation of an intelligent computing center, characterized in that, include: Obtain the business load requirements of the target business for the intelligent computing center's computing power configuration planning; Based on a pre-built benchmark library, standardized performance metrics for the business load requirements are determined, and equivalent performance metrics are determined based on the standardized performance metrics; wherein, the benchmark library stores standardized performance data for various standard tasks on different hardware. Based on the equivalent performance indicators and the theoretical benchmark performance indicators of the target hardware under the standard task, the effective computing power requirements of the target service are determined. Based on the aforementioned effective computing power requirements, an effective computing power configuration plan for the intelligent computing center is determined.
2. The method as described in claim 1, characterized in that, The business load requirement information includes at least: the types of tasks involved in the target business; The standardized performance metrics for determining the business load demand information based on a pre-built benchmark library include: Based on the benchmark library and the task types involved in the target service, determine the basic computational load, task type mode coefficient, and benchmark library calibration coefficient of the target service; Based on the basic computational load, task type mode coefficient, and benchmark library calibration coefficient, the standardized performance indicators of the business load requirement information are determined.
3. The method as described in claim 1 or 2, characterized in that, The step of determining the equivalent performance index based on the standardized performance index includes: Determine the ratio of the number of tasks in each task type involved in the target business to the total number of tasks in the target business; The business demand volume corresponding to each of the task types is determined based on the business load demand information. The equivalent performance index of the target service is determined by weighted aggregation based on the proportion of the number of tasks, the business demand, the standardized performance index, and the hardware efficiency calibration parameters.
4. The method as described in claim 3, characterized in that, Determining the hardware efficiency calibration parameters includes: Obtain the target model utilization rate of the target business; Obtain the reference model utilization rate for the standard task from the benchmark library; The hardware efficiency calibration parameters are determined based on the ratio between the utilization rate of the target model and the utilization rate of the reference model.
5. The method as described in claim 1, characterized in that, The step of determining the effective computing power requirement of the target service based on the equivalent performance index and the theoretical benchmark performance index of the target hardware under the standard task includes: Determine the theoretical peak computing power and global adjustment factor of the target hardware. The global adjustment factor is used to compensate for the unmodeled performance loss introduced by the operational differences between the standard task and the target service. Calculate the ratio between the equivalent performance index and the theoretical benchmark performance index; The effective computing power requirement is obtained by multiplying the ratio, the theoretical peak computing power of the target hardware, and the global adjustment factor.
6. The method according to any one of claims 1 to 5, characterized in that, Before determining the effective computing power requirement of the target service based on the equivalent performance index and the theoretical benchmark performance index of the target hardware under the standard task, the method further includes: Obtain at least one standard task corresponding to the target service; Based on the benchmark library, the theoretical benchmark service requirements corresponding to each of the standard tasks are obtained, and the theoretical benchmark service requirements are derived from the benchmark test results in the benchmark library. Based on the theoretical baseline service demand, the standardized performance data, and the hardware efficiency calibration parameters of each standard task, the theoretical baseline throughput corresponding to each standard task is determined respectively. The theoretical baseline throughput corresponding to each of the standard tasks is aggregated to obtain the theoretical baseline performance index of the target hardware under the standard tasks.
7. The method as described in claim 1, characterized in that, The method further includes: Determine the theoretical peak computing power requirement of the target service, and determine the planned peak computing power requirement of the target service based on the theoretical peak computing power requirement and the hardware redundancy coefficient; Determine the peak computing power configuration plan for the intelligent computing center based on the planned peak computing power requirements; Determine the deviation between the peak computing power configuration planning scheme and the effective computing power configuration planning scheme; When the deviation is greater than the preset deviation threshold, the computing power configuration planning scheme update operation is executed repeatedly until the deviation between the updated peak computing power configuration planning scheme and the updated effective computing power configuration planning scheme is less than or equal to the preset deviation threshold. The computing power configuration planning scheme update operation includes: Adjust the hardware redundancy coefficient and / or the hardware efficiency calibration parameters used to determine the equivalent performance index; The peak computing power configuration planning scheme and the computing power configuration planning scheme are updated based on the adjusted hardware redundancy coefficient and / or hardware efficiency calibration parameters.
8. The method as described in claim 7, characterized in that, Determining the theoretical peak computing power requirement of the target service includes: Determine the computational load, business scale, and time constraint coefficient of the target business model; Based on the computational load of the model, the business scale, the time constraint coefficient, and the preset reserved computing power margin, the theoretical peak computing power requirement of the target business is determined.
9. A computing power allocation planning device for an intelligent computing center, characterized in that, It includes an information acquisition module, an indicator determination module, a demand determination module, and a solution determination module, among which: The information acquisition module is used to acquire the business load requirements information of the target business for the intelligent computing center's computing power configuration planning; The indicator determination module is used to determine the standardized performance indicators of the business load requirement information based on a pre-built benchmark library, and to determine the equivalent performance indicators based on the standardized performance indicators; wherein, the benchmark library stores standardized performance data of various standard tasks on different hardware. The demand determination module is used to determine the effective computing power demand of the target service based on the equivalent performance index and the theoretical benchmark performance index of the target hardware under the standard task. The scheme determination module is used to determine the effective computing power configuration planning scheme of the intelligent computing center based on the effective computing power requirements.
10. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the computing power configuration planning method for the intelligent computing center as described in any one of claims 1 to 8.
11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the computing power configuration planning method for the intelligent computing center as described in any one of claims 1 to 8.
12. A computer program product, characterized in that, It includes a computer program that, when executed by a processor, implements the computing power configuration planning method for the intelligent computing center as described in any one of claims 1 to 8.