A heterogeneous computing power scheduling optimization method and system

Abstract

The application discloses a heterogeneous computing power scheduling optimization method and system, and belongs to the technical field of computer resource management. A multi-dimensional comprehensive capability fingerprint is constructed by collecting static capability parameters and running state parameters of a heterogeneous computing device, and capability grouping and representative device calibration are performed; a task-capability sub-index correlation degree table is established in combination with task static characteristic parameters and task category identification; a task-device predicted baseline running index is derived; target devices are selected in combination with predicted performance, predicted energy consumption and running state during scheduling; and the baseline index, correlation degree and fingerprint are jointly updated by using a running deviation information sliding window and multi-device fingerprint and deviation ordering consistency after running. The application can realize fine matching based on task characteristics among heterogeneous devices, improve scheduling performance and energy efficiency, and has continuous self-adaptive optimization capability.

Description

Technical Field

[0001] This invention relates to the field of computer technology, specifically a method and system for optimizing heterogeneous computing power scheduling. Background Technology

[0002] Existing heterogeneous computing power scheduling strategies typically suffer from the following problems:

[0003] Most strategies allocate tasks based solely on device type, a few single performance metrics, or static configurations, lacking a unified characterization of computing power, storage capacity, communication capabilities, and energy efficiency, making it difficult to accurately reflect the comprehensive differences between different heterogeneous computing devices.

[0004] Task feature identification is not detailed enough, often relying only on task type or simple labels, lacking quantitative descriptions of static features such as task computation intensity, memory access intensity, peripheral input / output intensity, parallelism, and latency sensitivity, making it difficult to accurately match task features with device capabilities.

[0005] The lack of a systematic calibration mechanism for the operation of tasks on different heterogeneous computing devices makes it difficult to build reliable task-device expected baseline operating metrics for different task categories.

[0006] Operational feedback typically only makes local corrections to a small number of parameters, fails to fully utilize the correlation between operational deviation information on multiple heterogeneous computing devices, lacks a joint update mechanism based on the correlation degree of task-capability sub-indicators and device capability fingerprints compared across multiple devices, and has insufficient adaptive capability of scheduling strategies. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a heterogeneous computing power scheduling optimization method, comprising the following steps:

[0008] Step 1: Generate a multidimensional comprehensive capability fingerprint containing multiple capability sub-indicators based on the static capability parameters and operating status parameters of each heterogeneous computing device. Group the heterogeneous computing devices by capability based on the similarity of the multidimensional comprehensive capability fingerprint. Select a representative device in each capability group and record the multidimensional comprehensive capability fingerprint and the correspondence between the capability group and the representative device in the device capability information database.

[0009] Step 2: Generate task static feature parameters and task category identifier based on the task description information of the task to be scheduled. Record the task static feature parameters and task category identifier in the task feature information database. Record the task-capability sub-indicator correlation degree between the task category identifier and the capability sub-indicator in the multi-dimensional comprehensive capability fingerprint according to the preset correlation degree initialization rule, and form a task-capability sub-indicator correlation degree table.

[0010] Step 3: When there is no target task category identifier calibration record in the task feature information database, the task to be scheduled corresponding to the target task category identifier is issued as a calibration task to the representative equipment of each capability group for flow-limited trial operation. Based on the calibration operation monitoring information and the correlation between task and capability sub-indicators, the baseline operation index of task-capability group is generated. Based on the ratio of capability sub-indicators in the multi-dimensional comprehensive capability fingerprint of non-representative equipment and representative equipment in the capability group and the correlation between task and capability sub-indicators, the expected baseline operation index of task-equipment for each heterogeneous computing device is derived and recorded in the task feature information database.

[0011] Step 4: When there are tasks to be scheduled, read the task category identifier of the task to be scheduled from the task feature information database and read the task-device expected baseline operation index of each candidate heterogeneous computing device. Read the operation status parameters of each candidate heterogeneous computing device from the device capability information database. Based on the preset scheduling evaluation rules and combined with the correlation degree of task-capability sub-indicators, calculate the scheduling evaluation value of each candidate heterogeneous computing device based on the task-device expected baseline operation index and operation status parameters. Select the target heterogeneous computing device with the best scheduling evaluation value to execute the task to be scheduled.

[0012] Step 5: Collect the operation monitoring information of the tasks to be scheduled on the target heterogeneous computing devices, calculate the operation deviation information, maintain a sliding window for operation deviation information for each task category identifier and each heterogeneous computing device, and when the sliding windows for operation deviation information of multiple heterogeneous computing devices corresponding to the same task category identifier meet the preset deviation conditions, adjust the expected baseline operation indicators of the corresponding task-device according to the preset update step size, and jointly update the task-capability sub-indicator correlation degree and multi-dimensional comprehensive capability fingerprint based on the consistency between the numerical ranking of capability sub-indicators in the multi-dimensional comprehensive capability fingerprint of multiple heterogeneous computing devices and the numerical ranking of operation deviation information.

[0013] Furthermore, the static capability parameters include at least one of the following: computing throughput capability parameters, storage capacity parameters, storage bandwidth parameters, memory access latency parameters, supported data type parameters, and static power consumption parameters. The operating status parameters include at least one of the following: current load level parameters, task queue length parameters, current power consumption parameters, and current temperature parameters. The preset fingerprint generation rules are used to map the static capability parameters and operating status parameters into multiple capability sub-indicators and combine them to form a multi-dimensional comprehensive capability fingerprint. Capability groups are generated based on the similarity between the multi-dimensional comprehensive capability fingerprints. Within each capability group, representative devices are selected based on the degree of difference between the multi-dimensional comprehensive capability fingerprint and the statistical center value of the capability group's comprehensive capability fingerprint.

[0014] Furthermore, the task description information includes at least one of the following: the business type to which the task belongs, the scale of the task input data, the types of operations included in the task, and the service quality requirements of the task. The task static feature parameters include at least one of the following: computational intensity, memory access intensity, peripheral input / output intensity, parallelizability, and latency sensitivity. The preset correlation initialization rule is used to set the initial correlation degree between the task and the capability sub-indicators in the multi-dimensional comprehensive capability fingerprint according to the task static feature parameters, and write the initial correlation degree between the task and the capability sub-indicator into the task-capability sub-indicator correlation degree table.

[0015] Furthermore, the calibration task is generated by splitting the task to be scheduled into at least one calibration sub-task or by reducing the size of the input data. The calibration operation monitoring information includes at least one of calibration execution time, calibration peak resource usage, and calibration energy consumption. The task-capability group baseline operation index includes at least one of baseline execution time index and baseline energy efficiency index for each capability group.

[0016] Furthermore, the aforementioned task-equipment expected baseline operating parameters are obtained in the following manner:

[0017] For non-representative devices within each capability group, obtain the capability sub-indicator values ​​from the multidimensional comprehensive capability fingerprint of the non-representative devices and the corresponding capability sub-indicator values ​​from the multidimensional comprehensive capability fingerprint of the representative devices. Based on the ratio of capability sub-indicator values ​​and the correlation between task and capability sub-indicators, scale and convert the baseline operation indicators of the task-capability group to obtain the expected baseline execution time indicator and the expected baseline energy efficiency indicator of the task-equipment for non-representative devices. Then, write the expected baseline execution time indicator and the expected baseline energy efficiency indicator of the task-equipment for non-representative devices as part of the expected baseline operation indicators of the task-equipment into the task feature information database.

[0018] Furthermore, the preset scheduling evaluation rules include:

[0019] Based on the baseline operating indicators and operating status parameters of the task-device model, the estimated execution time and energy consumption of the scheduled task on each candidate heterogeneous computing device are estimated. The estimated execution time is corrected according to the task queue length parameter and the current load level parameter, and the estimated energy consumption is corrected according to the current power consumption parameter. Weights are assigned to computing-related capability sub-indicators, storage-related capability sub-indicators, communication-related capability sub-indicators, and energy efficiency-related capability sub-indicators according to the correlation between the task and capability sub-indicators. The scheduling evaluation value is calculated based on the corrected execution time, corrected energy consumption, and weights.

[0020] Furthermore, the aforementioned operation deviation information sliding window is used to record performance deviation information and energy efficiency deviation information corresponding to a preset number of operations. When the performance deviation information or energy efficiency deviation information exceeds a preset deviation condition and reaches a preset number of times within the operation deviation information sliding window, at least one of the corresponding task-equipment expected baseline execution time index and task-equipment expected baseline energy efficiency index is monotonically adjusted according to a preset update step size, generating a new task-equipment expected baseline operation index and writing it into the task feature information database.

[0021] Furthermore, within a preset time period, when the sliding windows of the operation deviation information of the same task category identified on at least two heterogeneous computing devices meet the preset deviation conditions, a comparison device set containing at least two heterogeneous computing devices that meet the preset deviation conditions is constructed. The multidimensional comprehensive capability fingerprint values ​​of each heterogeneous computing device in the comparison device set are sorted for each capability sub-index, and the corresponding operation deviation information values ​​are also sorted. The direction and intensity of the influence between the capability sub-index and the operation deviation information are determined based on the consistency between the multidimensional comprehensive capability fingerprint value sorting and the operation deviation information value sorting. The task-capability sub-index correlation degree corresponding to the capability sub-index whose influence direction is consistent with the sorting is increased, and the task-capability sub-index correlation degree corresponding to the capability sub-index whose influence direction is opposite to the sorting is decreased. After normalizing all task-capability sub-index correlation degrees, they are written into the task-capability sub-index correlation degree table.

[0022] Furthermore, when the sliding window of the operation deviation information of multiple heterogeneous computing devices in the same capability group continuously meets the preset deviation conditions within a preset time period, the heterogeneous computing devices in the same capability group are re-aggregated and divided based on the updated multidimensional comprehensive capability fingerprint to generate a new capability group. In the new capability group, representative devices are selected according to the degree of difference between the multidimensional comprehensive capability fingerprint and the statistical center value of the comprehensive capability fingerprint of the new capability group, and the correspondence between capability groups and representative devices is updated in the device capability information database.

[0023] A heterogeneous computing power scheduling optimization system, applying the aforementioned heterogeneous computing power scheduling optimization method, includes: a device capability management module, a task feature management module, a calibration and baseline derivation module, a scheduling execution module, a data processing module, and a runtime feedback update module; the device capability management module, task feature management module, calibration and baseline derivation module, scheduling execution module, and runtime feedback update module are respectively connected to the data processing module;

[0024] The device capability management module is used to generate multi-dimensional comprehensive capability fingerprints, group capabilities based on the similarity of multi-dimensional comprehensive capability fingerprints, select representative devices in each capability group, and build a device capability information database.

[0025] The task feature management module is used to generate task static feature parameters, task category identifiers, and task-capability sub-indicator correlation degrees, and to construct a task feature information database and a task-capability sub-indicator correlation degree table.

[0026] The calibration and baseline derivation module is used to run calibration tasks on representative devices, generate task-capability grouping baseline operating indicators, and derive the expected baseline operating indicators of tasks and devices based on the proportion of capability sub-indicators in the multi-dimensional comprehensive capability fingerprint and the correlation between task-capability sub-indicators, and write them into the task feature information database.

[0027] The scheduling execution module is used to calculate the scheduling evaluation value of each candidate heterogeneous computing device based on the expected baseline operating indicators of the task-device, operating status parameters, preset scheduling evaluation rules and the correlation degree of the task-capability sub-indicators when there are tasks to be scheduled, and select the target heterogeneous computing device to execute the task to be scheduled.

[0028] The aforementioned operation feedback update module is used to adaptively update the expected baseline operation indicators of the task-equipment, the correlation between the task-capability sub-indicators and the multidimensional comprehensive capability fingerprint based on the sliding window of operation deviation information, the multidimensional comprehensive capability fingerprint and the correlation between the task and capability sub-indicators, and to update the equipment capability information database and the task feature information database.

[0029] The beneficial effects of this invention are:

[0030] By using multi-dimensional integrated capability fingerprints to uniformly characterize the computing power, storage capacity, communication capabilities, and energy efficiency of heterogeneous computing devices, and combining capability grouping and representative device mechanisms, calibration overhead is reduced while maintaining the accuracy of capability description.

[0031] By linking task static feature parameters and task category identifiers with capability sub-indicators in the multi-dimensional comprehensive capability fingerprint, a task-capability sub-indicator correlation table is formed, which is beneficial for selecting capability dimensions that have a greater impact on performance and energy efficiency based on task category.

[0032] By using the calibration results of representative equipment to construct baseline operating indicators for task-capability groups, and then combining the proportion of capability sub-indicators and the correlation between task-capability sub-indicators to derive the expected baseline operating indicators for task-equipment, the baseline indicators can be expanded from a small number of representative equipment to all equipment within the capability group, thereby reducing calibration costs.

[0033] By recording the deviations of multiple runs through a sliding window of operational deviation information, and based on the consistency between the multidimensional comprehensive capability fingerprint ranking and the operational deviation information ranking among multiple heterogeneous computing devices, the positive or negative influence relationship between capability sub-indicators and deviation information is identified. Joint updates are performed on the task-capability sub-indicator correlation and multidimensional comprehensive capability fingerprint to improve the adaptive capability of the scheduling strategy during long-term operation. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating a heterogeneous computing power scheduling optimization method.

[0035] Figure 2 This is a schematic diagram of the device fingerprint generation and grouping process. Detailed Implementation

[0036] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0037] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0038] Example 1

[0039] like Figure 1 As shown, a heterogeneous computing power scheduling optimization method includes the following steps:

[0040] Step 1: Generate a multidimensional comprehensive capability fingerprint containing multiple capability sub-indicators based on the static capability parameters and operating status parameters of each heterogeneous computing device. Group the heterogeneous computing devices by capability based on the similarity of the multidimensional comprehensive capability fingerprint. Select a representative device in each capability group and record the multidimensional comprehensive capability fingerprint and the correspondence between the capability group and the representative device in the device capability information database.

[0041] Step 2: Generate task static feature parameters and task category identifier based on the task description information of the task to be scheduled. Record the task static feature parameters and task category identifier in the task feature information database. Record the task-capability sub-indicator correlation degree between the task category identifier and the capability sub-indicator in the multi-dimensional comprehensive capability fingerprint according to the preset correlation degree initialization rule, and form a task-capability sub-indicator correlation degree table.

[0042] Step 3: When there is no target task category identifier calibration record in the task feature information database, the task to be scheduled corresponding to the target task category identifier is issued as a calibration task to the representative equipment of each capability group for flow-limited trial operation. Based on the calibration operation monitoring information and the correlation between task and capability sub-indicators, the baseline operation index of task-capability group is generated. Based on the ratio of capability sub-indicators in the multi-dimensional comprehensive capability fingerprint of non-representative equipment and representative equipment in the capability group and the correlation between task and capability sub-indicators, the expected baseline operation index of task-equipment for each heterogeneous computing device is derived and recorded in the task feature information database.

[0043] Step 4: When there are tasks to be scheduled, read the task category identifier of the task to be scheduled from the task feature information database and read the task-device expected baseline operation index of each candidate heterogeneous computing device. Read the operation status parameters of each candidate heterogeneous computing device from the device capability information database. Based on the preset scheduling evaluation rules and combined with the correlation degree of task-capability sub-indicators, calculate the scheduling evaluation value of each candidate heterogeneous computing device based on the task-device expected baseline operation index and operation status parameters. Select the target heterogeneous computing device with the best scheduling evaluation value to execute the task to be scheduled.

[0044] Step 5: Collect the operation monitoring information of the tasks to be scheduled on the target heterogeneous computing devices, calculate the operation deviation information, maintain a sliding window for operation deviation information for each task category identifier and each heterogeneous computing device, and when the sliding windows for operation deviation information of multiple heterogeneous computing devices corresponding to the same task category identifier meet the preset deviation conditions, adjust the expected baseline operation indicators of the corresponding task-device according to the preset update step size, and jointly update the task-capability sub-indicator correlation degree and multi-dimensional comprehensive capability fingerprint based on the consistency between the numerical ranking of capability sub-indicators in the multi-dimensional comprehensive capability fingerprint of multiple heterogeneous computing devices and the numerical ranking of operation deviation information.

[0045] Specifically, in one embodiment, the heterogeneous computing power scheduling optimization system of the present invention is deployed in a heterogeneous computing power cluster including multiple heterogeneous computing devices. The system includes:

[0046] Equipment Capability Management Module: Used to collect static capability parameters and operating status parameters, generate multi-dimensional comprehensive capability fingerprints, perform capability grouping and representative equipment selection, and maintain the equipment capability information database.

[0047] Task Feature Management Module: Used to collect task description information of tasks to be scheduled, generate task static feature parameters and task category identifiers, initialize task-capability sub-indicator correlation degree, and maintain task feature information database and task-capability sub-indicator correlation degree table.

[0048] Calibration and Baseline Derivation Module: Used to construct calibration tasks, run calibration tasks on representative equipment, generate task-capability grouping baseline operating indicators, and derive task-equipment expected baseline operating indicators.

[0049] The scheduling execution module is used to calculate the scheduling evaluation value based on the expected baseline operating indicators of the task-device, operating status parameters, and the correlation between the task-capability sub-indicators during the scheduling phase, and to select the target heterogeneous computing device.

[0050] Operational Feedback Update Module: This module is used to collect task operation monitoring information, calculate operational deviation information, maintain a sliding window for operational deviation information, and jointly update the task-device expected baseline operational indicators, the correlation between task-capability sub-indicators, and the multi-dimensional comprehensive capability fingerprint based on the consistency of multi-device sorting.

[0051] Each module of the system can be implemented through software, hardware, or a combination of both.

[0052] In one embodiment, the device capability management module collects parameters in the following manner:

[0053] Collect static capability parameters: including at least one of the following: computing throughput parameters, storage capacity parameters, storage bandwidth parameters, memory access latency parameters, supported data type parameters, and static power consumption parameters.

[0054] Collect operating status parameters, including at least one of the following: current load level, task queue length, current power consumption, and current temperature.

[0055] like Figure 2 As shown, the device capability management module converts the above parameters into capability sub-indicators according to preset fingerprint generation rules. An example of the process is as follows:

[0056] All static capability parameters and operational status parameters are normalized. Normalization can be achieved by configuring minimum and maximum values ​​for each parameter, linearly mapping the original values ​​to a fixed range of values.

[0057] The normalized parameters are combined according to functional dimensions to generate sub-indicators for computing power, storage capacity, communication capacity, and energy efficiency.

[0058] The computing power sub-index can be generated based on the computing throughput parameter and the supported data type parameter. For example, when the computing throughput parameter is higher than the preset throughput threshold and the supported data type range includes the preset type set, the computing power sub-index will be set to a higher level.

[0059] Storage capacity sub-metrics can be generated based on storage capacity parameters, storage bandwidth parameters, and memory access latency parameters. For example, when the storage capacity parameter is higher than the capacity threshold, the storage bandwidth parameter is higher than the bandwidth threshold, and the memory access latency parameter is lower than the latency threshold, the storage capacity sub-metric will be increased by one level.

[0060] The communication capability sub-indicator can be generated based on parameters such as network interface bandwidth and peripheral access speed. When the bandwidth or access speed is higher than the corresponding threshold, the communication capability sub-indicator is improved.

[0061] The energy efficiency capability sub-index can be generated based on static power consumption parameters and power consumption measured under standard load. When the power consumption per unit of computation is lower than the energy efficiency threshold, the energy efficiency capability sub-index is improved.

[0062] Each capability sub-indicator is quantified into a finite number of discrete levels, with level boundaries defined in a configuration file. All capability sub-indicators are combined into a vector in a fixed order to form a multi-dimensional comprehensive capability fingerprint.

[0063] In one embodiment, capability grouping and representative device selection includes:

[0064] Each heterogeneous computing device's multidimensional integrated capability fingerprint is treated as a point in a multidimensional space. Aggregation is performed based on the distance between the multidimensional integrated capability fingerprints. The distance can be measured by a weighted sum of the differences of all capability sub-indicators. When the distance is less than a preset aggregation threshold, the two devices can be classified into the same capability group.

[0065] Multiple capability groups are formed through iterative merging until the distance between any two devices in any capability group is lower than the aggregation threshold, or the number of capability groups reaches a preset number.

[0066] Calculate the statistical central value of the comprehensive capability fingerprint for each capability group. The statistical central value can be obtained by averaging the multidimensional comprehensive capability fingerprints within each group.

[0067] For each device within each capability group, calculate the distance between its multidimensional integrated capability fingerprint and the statistical center value of the capability group's integrated capability fingerprint, and select the device with the smallest distance as the representative device. If multiple representative devices are needed, select the first few in ascending order of distance.

[0068] The equipment capability management module writes multi-dimensional comprehensive capability fingerprints, capability group identifiers, and the correspondence between capability groups and representative equipment into the equipment capability information database.

[0069] In one embodiment, the task feature management module performs the following steps upon receiving a task to be scheduled:

[0070] Read at least one of the following from the task description information: the business type to which the task belongs, the scale of the task input data, the types of operations included in the task, and the service quality requirements of the task.

[0071] The computational intensity is calculated based on the proportion of arithmetic operations in the task's computational types and the size of the task's input data. When the proportion of arithmetic operations is greater than a preset percentage threshold and the size of the task's input data is greater than a preset size threshold, the computational intensity is classified as high-level; otherwise, it is classified as medium-level or low-level.

[0072] Memory access intensity is calculated based on the ratio between the amount of data accessed by the task and the size of the task's input data. When this ratio exceeds a preset memory access threshold, the memory access intensity is classified as high-level.

[0073] The peripheral input / output density is calculated based on the number of times the task accesses the peripheral and the amount of peripheral data read / write. When this indicator is greater than the preset peripheral access threshold, the peripheral input / output density is classified into a high-level category.

[0074] Parallelism is calculated based on the ratio of parallelizable operands to the total number of operands. When this ratio is greater than a preset parallelism threshold, the parallelism is classified as a high-level category.

[0075] Based on the response time constraints in the task's service quality requirements, latency sensitivity is determined. When a task must be completed within a timeframe shorter than a preset response time threshold, the latency sensitivity is classified as high-level.

[0076] The task feature management module combines computational intensity, memory access intensity, peripheral input / output intensity, parallelizability, and latency sensitivity in a predetermined format to form static feature parameters of the task.

[0077] In one embodiment, the task category identifier is formed by sequentially concatenating the business type code to which the task belongs and the static feature parameter level codes of each task. Different combinations of business types and static feature levels correspond to different task category identifiers.

[0078] The task feature management module records the correspondence between task category identifiers and task static feature parameters in the task feature information database, and generates the initial correlation degree between task-capability sub-indicators according to preset correlation degree initialization rules. The initialization rules can be:

[0079] When the computational intensity is at the high level, the initial correlation degree between the task category identifier and the computational capability sub-indicator is set to a high value.

[0080] When the memory access intensity is at a high level, set the initial correlation between the task category identifier and the storage capacity sub-indicator to a high value.

[0081] When the peripheral input / output density is at a high level, the initial correlation degree between the task category identifier and the communication capability sub-indicator is set to a high value.

[0082] When the latency sensitivity is at a high level, the initial correlation between the task category identifier and the computing capability sub-indicator and the communication capability sub-indicator is further improved.

[0083] The task feature management module normalizes the initial correlation degree of all task-capability sub-indicators corresponding to the same task category identifier, so that the sum of all initial correlation degrees is a fixed constant, and writes the normalized result into the task-capability sub-indicator correlation degree table.

[0084] In one embodiment, when there is no calibration record for a certain task category identifier in the task feature information database, the calibration and baseline derivation module constructs and executes the calibration task according to the following steps:

[0085] The task to be scheduled corresponding to this task category is marked as a calibration candidate task.

[0086] Generate the calibration task using one of the following methods:

[0087] The calibration candidate task is split into multiple calibration subtasks, and the static feature parameters of each calibration subtask are consistent with those of the calibration candidate task; the input data size of the calibration candidate task is reduced to a preset ratio so that the computational structure remains unchanged.

[0088] The calibration tasks are executed sequentially on representative devices of each capability group, with the number of concurrent executions limited to reduce external interference.

[0089] Record calibration operation monitoring information during the calibration task operation. The calibration operation monitoring information shall include at least: calibration execution time, calibration peak resource usage, and calibration energy consumption.

[0090] The calibration and baseline derivation module calculates baseline operating indicators for task-capability groups based on calibration operation monitoring information from multiple runs of representative equipment. An example of the process is as follows:

[0091] The baseline execution time index for a capability group is obtained by averaging the execution times of multiple calibrations of representative equipment within the same capability group. Similarly, the baseline energy efficiency index for a capability group is obtained by averaging the energy consumption of multiple calibrations of representative equipment within the same capability group. If some calibration results exceed a preset reasonable range, these outliers are removed when calculating the average. The baseline execution time index and the baseline energy efficiency index are combined to form the task-capability group baseline operational index.

[0092] For non-representative equipment within each capability group, the calibration and baseline derivation module generates the task-equipment expected baseline operating parameters according to the following logic:

[0093] Obtain the value of each capability sub-index in the multidimensional comprehensive capability fingerprint of the non-representative device, and the value of the corresponding capability sub-index in the multidimensional comprehensive capability fingerprint of the representative device.

[0094] For each capability sub-index, the ratio of the value of the non-representative equipment capability sub-index to the value of the representative equipment capability sub-index is calculated, which is used to represent the capability ratio of the non-representative equipment relative to the representative equipment in the corresponding capability dimension.

[0095] For the execution time dimension: Select capability sub-indicators that are highly correlated with execution time from the task-capability sub-indicator correlation table, such as computing capability sub-indicators and storage capability sub-indicators;

[0096] For each relevant capability sub-indicator, an execution time adjustment factor is calculated based on the capability ratio and the correlation between the task and the capability sub-indicator. When the capability ratio is greater than 1 and the correlation between the task and the capability sub-indicator is high, the execution time adjustment factor decreases; when the capability ratio is less than 1 and the correlation between the task and the capability sub-indicator is high, the execution time adjustment factor increases.

[0097] Multiple execution time adjustment factors are weighted and averaged according to the correlation between task and capability sub-indicators to obtain the overall execution time adjustment factor for the non-representative equipment; the overall execution time adjustment factor is used to adjust the baseline execution time index of the task-capability group to obtain the expected baseline execution time index of the task-equipment for the non-representative equipment.

[0098] For the energy efficiency dimension: Select capability sub-indicators that are highly correlated with energy efficiency from the task-capability sub-indicator correlation table, such as energy efficiency capability sub-indicators;

[0099] The energy efficiency adjustment factor is calculated based on the capacity ratio and the correlation between the task and the capacity sub-indicator. When the capacity ratio is greater than 1 and the correlation between the task and the capacity sub-indicator is high, the energy efficiency adjustment factor decreases, indicating that the energy consumption per unit task is reduced. The weighted average result of multiple energy efficiency adjustment factors is used to adjust the baseline energy efficiency index of the task-capacity group to obtain the expected baseline energy efficiency index of the non-representative equipment task.

[0100] The expected baseline execution time of the task-equipment and the expected baseline energy efficiency of the task-equipment constitute the expected baseline operation index of the task-equipment, and are written into the task feature information database.

[0101] In one embodiment, the scheduling execution module performs the following operations when a task arrives:

[0102] Based on the task category identifier of the task to be scheduled, read the expected baseline operating indicators of the task-device for all candidate heterogeneous computing devices from the task feature information database.

[0103] Read the current operating status parameters of the candidate heterogeneous computing devices from the device capability information database, including the current load level parameters, task queue length parameters, and current power consumption parameters.

[0104] Estimate the expected execution time for each candidate heterogeneous computing device:

[0105] The base execution time is based on the task-device expected baseline execution time index; the queuing time is estimated based on the task queue length parameter, for example, by multiplying the task queue length parameter by the average task execution time to obtain the queuing time estimate; the base execution time is amplified or reduced based on the current load level parameter. When the load level parameter is higher than the preset load threshold, the base execution time is amplified proportionally.

[0106] Estimate the projected energy consumption for each candidate heterogeneous computing device: use the task-device projected baseline energy efficiency index as the base energy consumption; adjust the base energy consumption according to the current power consumption parameters, and increase the base energy consumption proportionally when the current power consumption parameters are higher than the preset power consumption threshold.

[0107] The scheduling and execution module assigns weights to capability sub-indicators related to execution time and energy consumption based on the correlation between tasks and capability sub-indicators. When latency sensitivity is at a high level, the weight of execution time in the scheduling evaluation value is increased; when the business is more concerned about overall energy consumption, the weight corresponding to energy consumption is increased.

[0108] The scheduling execution module calculates the scheduling evaluation value based on the revised estimated execution time, estimated energy consumption, and corresponding weights. The scheduling evaluation value can be obtained by weighted summation of the execution time part and the energy consumption part.

[0109] Select the device with the lowest scheduling evaluation value from all candidate heterogeneous computing devices, and assign the task to be scheduled to that device as the target heterogeneous computing device.

[0110] The runtime feedback update module maintains a sliding window of runtime deviation information for each task category identifier and each heterogeneous computing device, and performs the following steps:

[0111] After the task is completed, record the actual execution time, actual energy consumption, and actual resource usage. Calculate the execution time deviation information and energy efficiency deviation information based on the task-equipment expected baseline execution time index and the task-equipment expected baseline energy efficiency index, for example, using the difference or ratio between the actual value and the expected value as the deviation information.

[0112] The execution time deviation information and energy efficiency deviation information are written into the corresponding sliding window. If the number of records in the sliding window exceeds the preset window length, the earliest record is deleted.

[0113] When the number of records with deviation information exceeding the preset deviation condition in the sliding window reaches a preset number, it is considered that the deviation between the expected baseline operating indicators of the task-equipment and the actual situation continues to exist. The operation feedback update module adjusts the expected baseline execution time indicator and the expected baseline energy efficiency indicator of the task-equipment according to the preset update step size, and moves the expected indicators toward the actual average value.

[0114] In one embodiment, the runtime feedback update module performs a joint update using deviation records from multiple devices within a preset time period:

[0115] For each task category identifier, a set of heterogeneous computing devices that meet the preset deviation conditions within a preset time period are collected from the sliding window of operational deviation information to form a comparison device set.

[0116] In the comparison set of equipment, for each capability sub-index, the equipment is sorted from high to low according to the capability sub-index values ​​in the multidimensional comprehensive capability fingerprint, resulting in a capability sub-index sorting sequence.

[0117] In the comparison set of devices, the execution time deviation information or energy efficiency deviation information is sorted from high to low according to the value to obtain the deviation information sorting sequence.

[0118] For each capability sub-index, compare the relative order of equipment pairs in the capability sub-index ranking sequence with that in the deviation information ranking sequence, and statistically analyze the proportion of equipment pairs with the same order and the proportion of equipment pairs with opposite orders:

[0119] When the proportion of devices with consistent sequence is higher than the preset consistency threshold, it is determined that there is a positive influence between this capability sub-indicator and the deviation information.

[0120] When the proportion of devices with reversed order is higher than the preset consistency threshold, it is determined that there is a reverse influence between this capability sub-indicator and the deviation information.

[0121] When neither of the two ratios reaches the consistency threshold, it is considered that the influence of this capability sub-indicator on the deviation information is not significant.

[0122] Based on the proportion of those in the same or opposite order, the influence intensity of each capability sub-indicator is divided into multiple levels, such as high influence, medium influence, and low influence levels.

[0123] For capability sub-indicators determined to have a positive impact, the correlation between the corresponding task and capability sub-indicator is increased according to the level of impact intensity; for capability sub-indicators determined to have a negative impact, the correlation between the corresponding task and capability sub-indicator is decreased according to the level of impact intensity; and no adjustments are made to capability sub-indicators with unclear impact relationships.

[0124] Normalize the correlation degree of all updated task-capability sub-indicators to ensure that the sum of the correlation degrees of all task-capability sub-indicators corresponding to the same task category remains unchanged, and write the normalization result into the task-capability sub-indicator correlation degree table.

[0125] Based on the updated correlation between the mission-equipment expected baseline operating indicators and the mission-capability sub-indicators, the capability sub-indicators that are highly correlated with deviation information in the multidimensional integrated capability fingerprint can be appropriately modified to make the multidimensional integrated capability fingerprint more consistent with actual operating performance.

[0126] When the sliding window of operational deviation information for multiple heterogeneous computing devices within the same capability group repeatedly meets the preset deviation conditions within a preset time period, the operational feedback update module and the device capability management module can perform the following operations:

[0127] Examine the differences between the updated multidimensional integrated capability fingerprints of these heterogeneous computing devices. If the differences exceed the capability grouping re-division threshold, re-execute the aggregation algorithm to split or merge the capability groups to which these devices belong, generating new capability groups.

[0128] Within the new capability group, representative devices are reselected based on the distance between the statistical center value of the new capability group's comprehensive capability fingerprint and the multidimensional comprehensive capability fingerprint of each device.

[0129] The new capability group identifier and new representative equipment information are written into the equipment capability information database, so that subsequent calibration tasks and scheduling decisions are based on the updated capability group structure.

[0130] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A heterogeneous computing power scheduling optimization method, characterized in that, Includes the following steps: Step 1: Generate a multidimensional comprehensive capability fingerprint containing multiple capability sub-indicators based on the static capability parameters and operating status parameters of each heterogeneous computing device. Group the heterogeneous computing devices by capability based on the similarity of the multidimensional comprehensive capability fingerprint. Select a representative device in each capability group and record the multidimensional comprehensive capability fingerprint and the correspondence between the capability group and the representative device in the device capability information database. Step 2: Generate task static feature parameters and task category identifier based on the task description information of the task to be scheduled. Record the task static feature parameters and task category identifier in the task feature information database. Record the task-capability sub-indicator correlation degree between the task category identifier and the capability sub-indicator in the multi-dimensional comprehensive capability fingerprint according to the preset correlation degree initialization rule, and form a task-capability sub-indicator correlation degree table. Step 3: When there is no target task category identifier calibration record in the task feature information database, the task to be scheduled corresponding to the target task category identifier is issued as a calibration task to the representative equipment of each capability group for flow-limited trial operation. Based on the calibration operation monitoring information and the correlation between task and capability sub-indicators, the baseline operation index of task-capability group is generated. Based on the ratio of capability sub-indicators in the multi-dimensional comprehensive capability fingerprint of non-representative equipment and representative equipment in the capability group and the correlation between task and capability sub-indicators, the expected baseline operation index of task-equipment for each heterogeneous computing device is derived and recorded in the task feature information database. Step 4: When there are tasks to be scheduled, read the task category identifier of the task to be scheduled from the task feature information database and read the task-device expected baseline operation index of each candidate heterogeneous computing device. Read the operation status parameters of each candidate heterogeneous computing device from the device capability information database. Based on the preset scheduling evaluation rules and combined with the correlation degree of task-capability sub-indicators, calculate the scheduling evaluation value of each candidate heterogeneous computing device based on the task-device expected baseline operation index and operation status parameters. Select the target heterogeneous computing device with the best scheduling evaluation value to execute the task to be scheduled. Step 5: Collect the operation monitoring information of the tasks to be scheduled on the target heterogeneous computing devices, calculate the operation deviation information, maintain a sliding window for operation deviation information for each task category identifier and each heterogeneous computing device, and when the sliding windows for operation deviation information of multiple heterogeneous computing devices corresponding to the same task category identifier meet the preset deviation conditions, adjust the expected baseline operation indicators of the corresponding task-device according to the preset update step size, and jointly update the task-capability sub-indicator correlation degree and multi-dimensional comprehensive capability fingerprint based on the consistency between the numerical ranking of capability sub-indicators in the multi-dimensional comprehensive capability fingerprint of multiple heterogeneous computing devices and the numerical ranking of operation deviation information.

2. The heterogeneous computing power scheduling optimization method according to claim 1, characterized in that, The static capability parameters include at least one of the following: computing throughput capability parameters, storage capacity parameters, storage bandwidth parameters, memory access latency parameters, supported data type parameters, and static power consumption parameters. The operating status parameters include at least one of the following: current load level parameters, task queue length parameters, current power consumption parameters, and current temperature parameters. The preset fingerprint generation rules are used to map the static capability parameters and operating status parameters into multiple capability sub-indicators and combine them to form a multi-dimensional comprehensive capability fingerprint. Capability groups are generated based on the similarity between the multi-dimensional comprehensive capability fingerprints. Within each capability group, representative devices are selected based on the degree of difference between the multi-dimensional comprehensive capability fingerprint and the statistical center value of the capability group's comprehensive capability fingerprint.

3. The heterogeneous computing power scheduling optimization method according to claim 1, characterized in that, The task description information includes at least one of the following: the business type to which the task belongs, the scale of the task input data, the types of operations included in the task, and the service quality requirements of the task. The static feature parameters of the task include at least one of the following: computational intensity, memory access intensity, peripheral input / output intensity, parallelizability, and latency sensitivity. The preset correlation initialization rule is used to set the initial correlation degree between the task and the capability sub-indicators in the multi-dimensional comprehensive capability fingerprint according to the static feature parameters of the task, and write the initial correlation degree between the task and the capability sub-indicator into the correlation degree table between the task and the capability sub-indicator.

4. The heterogeneous computing power scheduling optimization method according to claim 1, characterized in that, The calibration task is generated by splitting the task to be scheduled into at least one calibration subtask or by reducing the size of the input data. The calibration operation monitoring information includes at least one of calibration execution time, calibration peak resource usage, and calibration energy consumption. The task-capability group baseline operation index includes at least one of baseline execution time index and baseline energy efficiency index for each capability group.

5. The heterogeneous computing power scheduling optimization method according to claim 1, characterized in that, The aforementioned task-equipment expected baseline operating parameters are obtained in the following manner: For non-representative devices within each capability group, obtain the capability sub-indicator values ​​from the multidimensional comprehensive capability fingerprint of the non-representative devices and the corresponding capability sub-indicator values ​​from the multidimensional comprehensive capability fingerprint of the representative devices. Based on the ratio of capability sub-indicator values ​​and the correlation between task and capability sub-indicators, scale and convert the baseline operation indicators of the task-capability group to obtain the expected baseline execution time indicator and the expected baseline energy efficiency indicator of the task-equipment for non-representative devices. Then, write the expected baseline execution time indicator and the expected baseline energy efficiency indicator of the task-equipment for non-representative devices as part of the expected baseline operation indicators of the task-equipment into the task feature information database.

6. The heterogeneous computing power scheduling optimization method according to claim 1, characterized in that, The preset scheduling evaluation rules include: Based on the baseline operating indicators and operating status parameters of the task-device model, the estimated execution time and energy consumption of the scheduled task on each candidate heterogeneous computing device are estimated. The estimated execution time is corrected according to the task queue length parameter and the current load level parameter, and the estimated energy consumption is corrected according to the current power consumption parameter. Weights are assigned to computing-related capability sub-indicators, storage-related capability sub-indicators, communication-related capability sub-indicators, and energy efficiency-related capability sub-indicators according to the correlation between the task and capability sub-indicators. The scheduling evaluation value is calculated based on the corrected execution time, corrected energy consumption, and weights.

7. The heterogeneous computing power scheduling optimization method according to claim 1, characterized in that, The aforementioned operation deviation information sliding window is used to record performance deviation information and energy efficiency deviation information corresponding to a preset number of operations. When the performance deviation information or energy efficiency deviation information exceeds the preset deviation condition and reaches the preset number of times within the operation deviation information sliding window, at least one of the corresponding task-equipment expected baseline execution time index and task-equipment expected baseline energy efficiency index is monotonically adjusted according to the preset update step size, generating a new task-equipment expected baseline operation index and writing it into the task feature information database.

8. The heterogeneous computing power scheduling optimization method according to claim 1, characterized in that, Within a preset time period, when the sliding window of the operation deviation information of the same task category identified on at least two heterogeneous computing devices meets the preset deviation condition, a comparison device set containing at least two heterogeneous computing devices that meet the preset deviation condition is constructed. The multidimensional comprehensive capability fingerprint values ​​of each heterogeneous computing device in the comparison device set on each capability sub-index are sorted, and the corresponding operation deviation information values ​​are sorted. The direction and intensity of the influence between the capability sub-index and the operation deviation information are determined based on the consistency between the multidimensional comprehensive capability fingerprint value sorting and the operation deviation information value sorting. The task-capability sub-index correlation degree corresponding to the capability sub-index whose influence direction is consistent with the sorting is increased, and the task-capability sub-index correlation degree corresponding to the capability sub-index whose influence direction is opposite to the sorting is decreased. After normalizing all task-capability sub-index correlation degrees, they are written into the task-capability sub-index correlation degree table.

9. The heterogeneous computing power scheduling optimization method according to claim 1, characterized in that, When the sliding window of the operation deviation information of multiple heterogeneous computing devices in the same capability group continuously meets the preset deviation conditions within a preset time period, the heterogeneous computing devices in the same capability group are re-aggregated and divided based on the updated multidimensional comprehensive capability fingerprint to generate a new capability group. In the new capability group, representative devices are selected according to the degree of difference between the multidimensional comprehensive capability fingerprint and the statistical center value of the comprehensive capability fingerprint of the new capability group, and the correspondence between capability group and representative device is updated in the device capability information database.

10. A heterogeneous computing power scheduling and optimization system, characterized in that, The heterogeneous computing power scheduling optimization method according to any one of claims 1-9 includes: a device capability management module, a task feature management module, a calibration and baseline derivation module, a scheduling execution module, a data processing module, and a running feedback update module; wherein the device capability management module, the task feature management module, the calibration and baseline derivation module, the scheduling execution module, and the running feedback update module are respectively connected to the data processing module; The device capability management module is used to generate multi-dimensional comprehensive capability fingerprints, group capabilities based on the similarity of multi-dimensional comprehensive capability fingerprints, select representative devices in each capability group, and build a device capability information database. The task feature management module is used to generate task static feature parameters, task category identifiers, and task-capability sub-indicator correlation degrees, and to construct a task feature information database and a task-capability sub-indicator correlation degree table. The calibration and baseline derivation module is used to run calibration tasks on representative devices, generate task-capability grouping baseline operating indicators, and derive the expected baseline operating indicators of tasks and devices based on the proportion of capability sub-indicators in the multi-dimensional comprehensive capability fingerprint and the correlation between task-capability sub-indicators, and write them into the task feature information database. The scheduling execution module is used to calculate the scheduling evaluation value of each candidate heterogeneous computing device based on the expected baseline operating indicators of the task-device, operating status parameters, preset scheduling evaluation rules and the correlation degree of the task-capability sub-indicators when there are tasks to be scheduled, and select the target heterogeneous computing device to execute the task to be scheduled. The aforementioned operation feedback update module is used to adaptively update the expected baseline operation indicators of the task-equipment, the correlation between the task-capability sub-indicators and the multidimensional comprehensive capability fingerprint based on the sliding window of operation deviation information, the multidimensional comprehensive capability fingerprint and the correlation between the task and capability sub-indicators, and to update the equipment capability information database and the task feature information database.

Images