A container-based heterogeneous computing resource perception and intelligent scheduling system

Abstract

The present application relates to the field of computer computing power scheduling, especially to a heterogeneous computing resource perception and intelligent scheduling system based on containerization, comprising a probe generation module, a distribution and collection module, a credential generation module, a scheduling decision module and a monitoring and punishment module, the probe generation module is used to generate interference detection probes according to the computing characteristics of the task to be scheduled; the distribution and collection module distributes the interference detection probes to the candidate computing nodes and obtains the hardware contention quantization information of each node; the credential generation module is used to generate performance guarantee credentials according to the quantization information; the scheduling decision module is used to determine the target container and schedule the task according to the credentials and node history records; the monitoring and punishment module is used to monitor the performance deviation and trigger the adjustment operation when there is a breach. The present application reduces the performance loss caused by resource scheduling conflicts, improves the accuracy of the judgment of the real available capacity of the nodes by the scheduling system in the containerized deployment data center, and improves the computing efficiency.

Description

Technical Field

[0001] This invention relates to the field of computer computing power scheduling, and in particular to a containerized heterogeneous computing power resource perception and intelligent scheduling system. Background Technology

[0002] With the widespread deployment of container orchestration technology in data centers, containerized runtime environments have become the fundamental form of support for large-scale distributed task execution. In practice, application scenarios such as artificial intelligence training and inference, high-performance computing, and big data analysis have different specific requirements for computing power. Data centers typically deploy multiple types of computing hardware, such as central processing units (CPUs), graphics processing units (GPUs), and neural network processors. In such environments, containerized scheduling systems need to allocate user-submitted tasks to computing nodes with corresponding computing resources for execution. Since different types of tasks have different requirements for computing power, memory bandwidth, interconnect topology, and other characteristics of computing hardware, and multiple containerized tasks may run simultaneously on the same node sharing underlying hardware resources, resource contention between tasks can affect actual performance. Therefore, how to perceive the resource availability status of each node and the degree of interference between tasks under heterogeneous computing power deployment conditions, and make reasonable scheduling decisions accordingly, is a key technical challenge faced by containerized heterogeneous computing power scheduling systems.

[0003] Chinese Patent Publication No. CN117971492A discloses a method, system, electronic device, and storage medium for cross-cloud heterogeneous computing resource scheduling. The method includes: integrating computing resource nodes with the same system architecture in different regions to build multiple heterogeneous container clusters to form a computing power cluster; determining a core cluster from the multiple heterogeneous container clusters; establishing a communication connection between the core cluster and the multiple heterogeneous container clusters; and connecting the core cluster to the computing power control interface of each heterogeneous container cluster and establishing a computing power control plane to achieve unified scheduling and management of all cross-cloud heterogeneous computing power. This invention clusters computing resources according to different regions and system architectures, providing an architecture solution for computing power integration through a computing power control plane built between the computing power clusters and the core cluster. This significantly improves the utilization efficiency of enterprise internal computing resources and saves the migration costs of existing data assets between data centers and cloud services, while also making it easier for emerging architecture resources to be adopted and accessed within enterprises.

[0004] However, the following problems still exist in the existing technology: 1. In existing technologies, data centers deploying containerized applications determine whether a container is suitable for handling tasks based on the amount of remaining resources reported by the container. However, data centers often run multiple tasks simultaneously, and resource conflicts can arise between tasks on underlying hardware resources such as cache, memory bandwidth, and interconnect buses. These resource conflicts cannot be represented by the amount of remaining resources, which may lead to tasks being assigned to containers that appear to have sufficient resources but actually have internal interference, resulting in low task execution efficiency.

[0005] 2. Existing technologies lack a mechanism for tracking and providing feedback on the actual performance of nodes after making scheduling decisions, and cannot dynamically adjust scheduling tendencies based on the historical performance of containers, resulting in insufficient flexibility in scheduling strategies. Summary of the Invention

[0006] To address this, the present invention provides a containerized heterogeneous computing resource perception and intelligent scheduling system to overcome the problems in the prior art where data centers deploying containerized applications struggle to accurately quantify resource conflicts between tasks, lack tracking and feedback on container fulfillment, resulting in low task execution efficiency and insufficient scheduling strategy flexibility.

[0007] To achieve the above objectives, this invention provides a containerized heterogeneous computing resource awareness and intelligent scheduling system, comprising: The probe generation module is used to respond to container scheduling requests, determine the computational characteristics of the task to be scheduled, and determine interference detection probes based on the computational characteristics. The interference detection probes are used to reproduce the interference mode of the task on the hardware microarchitecture on heterogeneous computing power nodes. The distribution and acquisition module is used to distribute the interference detection probe to several candidate computing nodes and acquire quantitative characterization information of the degree of contention for shared hardware resources collected by each candidate container during the execution of the interference detection probe. A credential generation module, deployed on each of the candidate computing nodes, is used to determine performance guarantee credentials for the task based on the quantization characterization information. The scheduling decision module receives the performance guarantee credentials corresponding to each candidate container, determines the target container from each candidate container based on the performance guarantee credentials and the node's historical performance record, and schedules the task to the target container. The monitoring and penalty module is used to monitor the deviation of the actual performance from the performance guarantee index during task execution, and trigger adjustment operations for the target container based on the cumulative deviation between the actual performance data and the performance guarantee index. The adjustment operation includes correcting the historical performance record of the target container so that the weight of the target container is reduced in subsequent scheduling.

[0008] Furthermore, the process by which the probe generation module responds to a container scheduling request and determines the computational characteristics of the task to be scheduled includes: Respond to the container scheduling request of the task to be scheduled, and obtain the task type and task parameters of the task to be scheduled; The computing power consumption parameters are determined based on the task type and task size. Determine the memory pressure parameters based on the task type and accuracy requirements; The task parameters include accuracy requirements and task size, and the computational features include computing power consumption parameters and memory pressure parameters.

[0009] Furthermore, the probe generation module determines the interference detection probes based on the computational features, including: The computational load intensity of the interference detection probe is determined based on the aforementioned computing power occupancy parameter. The amount of memory accessed by the interference detection probe is determined based on the aforementioned memory pressure parameter. Based on the computational load intensity and the amount of memory access data, an interference detection probe is constructed to reproduce the hardware load of the scheduled task. The interference detection probe is a lightweight payload segment that runs independently of the scheduled task.

[0010] Furthermore, the process by which the distribution and acquisition module distributes the interference detection probe to several candidate computing nodes and obtains quantitative characterization information on the degree of contention for shared hardware resources collected by each candidate container during the execution of the interference detection probe includes: Containers with available computing power resources are selected from the cluster to form a candidate container set; Distribute the interference detection probes to each candidate container in the candidate container set; Obtain the quantitative characterization information generated by each candidate container when executing the interference detection probe.

[0011] Furthermore, the distribution and acquisition module acquires the quantitative characterization information generated by each candidate container when executing the interference detection probe, including: Obtain a set of container-related parameters, including the ready state pause duration of each container during the execution of the interference detection probe, and the reduction ratio of available cache space of each container during the execution of the interference detection probe. Obtain a set of computing power related parameters, including the decrease in computing unit utilization rate of each container during the execution of the interference detection probe, and the decrease in the proportion of available video memory bandwidth of each container during the execution of the interference detection probe. The computational resource contention characterization value is determined based on the container-related parameter set, and the memory access path contention characterization value is determined based on the computing power-related parameter set. The quantitative characterization information includes the computing resource contention characterization value and the memory access path contention characterization value.

[0012] Furthermore, the credential generation module determines the performance guarantee credentials for the task based on the quantified characterization information, including: The comprehensive interference index is obtained by weighted summing of the computing resource contention characterization value and the memory access path contention characterization value; The performance guarantee indicators in the performance guarantee certificate are determined based on the comprehensive interference index. The performance guarantee index is inversely proportional to the comprehensive interference index, and the performance guarantee credential includes the performance guarantee index and the candidate container number.

[0013] Furthermore, the scheduling decision module receives performance guarantee credentials corresponding to each candidate container, and the process of determining the target container from the candidate containers based on the performance guarantee credentials and the node's historical performance records includes: Obtain the historical performance records of the nodes corresponding to each candidate container, and determine the credibility coefficient of each candidate container based on the historical performance records of the nodes; The performance guarantee index in the performance guarantee credential of each candidate container is multiplied by the corresponding credibility coefficient to obtain the adoption score of each candidate container; The candidate container with the highest adoption score is selected as the target container.

[0014] Furthermore, the scheduling decision module obtains the historical performance records of the nodes corresponding to each candidate container, and determines the reliability coefficient of each candidate container based on the historical performance records of the nodes, including: Obtain the confidence ratio of the number of fulfillments to the number of defaults in the historical fulfillment records corresponding to the candidate container; The confidence ratio is mapped to a preset range of values ​​to obtain the confidence coefficient; The confidence coefficient increases as the confidence ratio increases.

[0015] Furthermore, the monitoring and penalty module is used to monitor the deviation of actual performance from the performance guarantee index during task execution, and based on the cumulative deviation between the actual performance data and the performance guarantee index, triggers adjustment operations for the target container, including: During task execution, actual performance data of the task in the target container is collected at preset intervals. Based on the actual performance data and the performance guarantee indicators, the deviation amount is determined; Based on the deviation, determine whether the corresponding preset period meets the preset default conditions; If the number of preset periods in which the preset default conditions are met consecutively exceeds the default threshold, an adjustment operation is triggered for the target container.

[0016] Furthermore, the monitoring and penalty module determines whether the preset period meets the preset default conditions based on the deviation amount, including: If the deviation is greater than or equal to a preset deviation threshold, then the preset default condition is determined to be met. If the deviation is less than the preset deviation threshold, it is determined that the preset default condition is not met.

[0017] Compared with existing technologies, the system of this invention includes a probe generation module for generating interference detection probes based on the computational characteristics of the task to be scheduled; a distribution and acquisition module for distributing the interference detection probes to candidate computing nodes and acquiring hardware contention quantification information of each node; a credential generation module for generating performance guarantee credentials based on the quantification information; a scheduling decision module for determining the target container and scheduling the task based on the credentials and node historical records; and a monitoring and penalty module for monitoring performance deviations and triggering adjustment operations when a violation occurs. This invention reduces performance loss caused by resource scheduling conflicts, improves the accuracy of the scheduling system in containerized data centers in judging the true availability of nodes, and improves computational efficiency.

[0018] In particular, this invention considers the performance conflicts arising between containers regarding shared hardware resources in heterogeneous computing environments, where these conflicts are difficult to quantify and accurately estimate. In reality, the amount of idle resources reported by containers only represents the idle state of resources and cannot intuitively represent the performance constraints that specific types of computing tasks may encounter, such as insufficient cache, memory bandwidth congestion, and limited bus bandwidth. This invention uses a probe generation module to synthesize interference detection probes based on the computational characteristics of the task to be scheduled. A distribution and acquisition module distributes these probes to candidate containers, collecting four types of parameters: ready-state stagnation time, available cache space reduction ratio, computing unit occupancy rate decrease, and available memory bandwidth reduction ratio decrease. Through quantification of computational resource contention and memory access path contention values, the performance guarantee credentials generated by the credential generation module represent the available computing power level of the candidate container for the corresponding task under real concurrent load conditions, rather than just empty theoretical resource quantities.

[0019] In particular, this invention considers the microarchitectural characteristics of modern computing devices, such as superscalar execution, increased instruction pipeline depth, out-of-order execution, and expanded decoding width. These characteristics, while enhancing computing power, make it difficult to accurately predict the actual execution performance of a task on specific hardware using static parameters. Heterogeneous computing hardware of different generations or architectures may exhibit different actual throughput performances for the same computational load. Furthermore, performance contention between containers sharing resources further complicates performance prediction. For large-scale tasks, scheduling based on static resource allocation or theoretical peak performance is highly susceptible to errors in prediction, leading to inappropriate container selection and lower-than-expected task execution efficiency. This invention uses a probe generation module to synthesize interference detection probes for the task to be scheduled. These probes reproduce the task's load on the hardware in a real target environment using lightweight load fragments, thus establishing performance assurance credentials. This invention replaces static prediction of complex hardware behavior with a short-cycle test, enabling the scheduling decision module to make horizontal comparisons based on the measured performance of each candidate container under the same detection conditions and select the target container. This reduces the prediction uncertainty caused by microarchitecture differences and resource contention coupling, and improves the scheduling efficiency of large-scale computing tasks in heterogeneous computing environments.

[0020] In particular, this invention transforms quantified characterization information into performance guarantee credentials through a credential generation module, thereby quantifying the actual usable computing power of containers based on measured interference. Each candidate container provides a performance guarantee credential for its corresponding task based on its current perturbation state; containers with stronger perturbations provide lower performance guarantee credentials. The scheduling decision module receives not the resource availability values ​​of each container, but rather comparable performance commitment data generated by each node under the same probing conditions.

[0021] In particular, when selecting target containers, the scheduling decision module of this invention considers not only the performance guarantee indicators recorded in the performance guarantee certificate, but also the historical performance records of each candidate container. A single performance guarantee certificate represents the instantaneous disturbance state of a candidate container at the probe time, and cannot fully reflect the actual performance of the corresponding container. For containers with a low ratio of fulfillment to default in their historical performance records, even if they provide high performance guarantee indicators in the current probe, the reliability of the performance guarantee indicators should be questioned. The scheduling decision module obtains an adoption score by multiplying the performance guarantee indicators by a reliability coefficient determined based on historical performance records. This allows containers that maintain a high fulfillment rate over a long period to obtain a higher adoption score under the same conditions, while containers that frequently default are unlikely to win in the bidding process, even if they apply for high performance guarantee indicators. This enables the identification of the long-term reliability of containers in the scheduling decision.

[0022] In particular, this invention constructs a closed-loop adjustment mechanism after scheduling through a monitoring and penalty module. During task execution, the monitoring and penalty module collects actual performance data at preset intervals and compares it with the performance guarantee indicators recorded in the performance guarantee certificate. If the number of preset periods in which preset default conditions are met exceeds the default threshold, an adjustment operation is triggered for the target container to correct its historical performance record, reduce its reliability coefficient in subsequent scheduling, and thus improve the overall scheduling reliability of the system. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the containerized heterogeneous computing resource perception and intelligent scheduling system according to an embodiment of the invention. Figure 2 This is a flowchart illustrating the operational steps of an embodiment of the invention. Figure 3 This is a logic diagram for determining whether a corresponding container has idle computing resources, as shown in an embodiment of the invention. Figure 4 This is a logic diagram for determining whether a preset default condition is met, as shown in the embodiment of the invention. Detailed Implementation

[0024] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0025] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0026] Please see Figure 1 The diagram shown is a structural schematic of a containerized heterogeneous computing resource perception and intelligent scheduling system according to an embodiment of the present invention. The containerized heterogeneous computing resource perception and intelligent scheduling system according to an embodiment of the present invention includes: The probe generation module is used to respond to container scheduling requests and reproduce the interference pattern of the task on the hardware microarchitecture on heterogeneous computing power nodes with interference detection probes determined based on the computational characteristics of the task to be scheduled. The distribution and acquisition module is used to distribute the interference detection probe to several candidate computing nodes and acquire quantitative characterization information of the degree of contention for shared hardware resources collected by each candidate container during the execution of the interference detection probe. A credential generation module, deployed on each of the candidate computing nodes, is used to determine performance guarantee credentials for the task based on the quantization characterization information. The scheduling decision module receives the performance guarantee credentials corresponding to each candidate container, determines the target container from each candidate container based on the performance guarantee credentials and the node's historical performance record, and schedules the task to the target container. The monitoring and penalty module is used to monitor the deviation of the actual performance from the performance guarantee index during task execution, and trigger adjustment operations for the target container based on the cumulative deviation between the actual performance data and the performance guarantee index. The adjustment operation includes correcting the historical performance record of the target container so that the weight of the target container is reduced in subsequent scheduling.

[0027] Specifically, the container refers to an isolated environment running on each computing node in a data center where containerized applications are deployed. This isolated environment contains the executable programs, dependent libraries, and resource configuration information required for the scheduled tasks to run, and shares underlying hardware resources with other containers in the data center. Resource isolation and usage quota limits are achieved between containers through the namespace and control group mechanisms of the operating system kernel.

[0028] Specifically, there are no restrictions on the specific forms of the probe generation module, distribution and acquisition module, credential generation module, scheduling decision module, and monitoring and punishment module. They can be composed of logic components, including field-programmable processors, computers, or microprocessors in computers, which will not be elaborated further.

[0029] Specifically, this invention proactively injects interference detection probes into each candidate container before scheduling execution to obtain the actual available computing power level of each candidate container for the scheduled task under the current load environment through actual measurement, rather than relying on static indicators such as idle resource quantity for decision-making. Each candidate container determines the computational resource contention characterization value and memory access path contention characterization value based on the ready-state stagnation time, available cache space reduction ratio, computing unit utilization rate reduction, and memory bandwidth availability ratio reduction collected during probe operation, thereby determining performance guarantee credentials. When horizontally comparing the performance guarantee credentials of each candidate container, the scheduling decision module further introduces a reliability coefficient based on historical performance records, ensuring that the scheduling result takes into account both the current disturbance state of the candidate container and its long-term performance reliability. After task scheduling is completed, the monitoring and penalty module tracks the deviation between actual performance and performance guarantee indicators, using the accumulated deviation as the basis for triggering adjustment operations, correcting the historical performance records of defaulting containers, thereby reducing their probability of being selected in subsequent scheduling. This invention achieves collaborative optimization of resource contention awareness and long-term performance guarantee during task scheduling in data centers deploying containerized applications under heterogeneous computing power environments. This reduces performance loss caused by resource conflicts and improves the accuracy of scheduling decisions and the overall computing efficiency of the system.

[0030] It is understood that the application scenarios of this invention include, but are not limited to, data centers that simultaneously host artificial intelligence training tasks, online inference services, and big data analysis jobs. In such scenarios, different tasks have varying requirements for shared hardware resources such as cache, video memory bandwidth, and computing units, and the load status of each node changes dynamically. The interference detection and credential bidding mechanism provided by this invention is particularly suitable for environments with complex load structures and where actual computing power is difficult to predict accurately in advance. Furthermore, the technical solution of this invention can also be extended to task scheduling under conditions of limited computing resources in edge computing scenarios, as well as heterogeneous computing environments with virtual machines as the running carrier, which will not be elaborated further here.

[0031] Please see Figure 2 The diagram shows the operation steps of an embodiment of the present invention. The operation steps of the containerized heterogeneous computing resource perception and intelligent scheduling system of the present invention include: Step S1: Respond to the container scheduling request, determine the computational feature type of the task to be scheduled, and determine the interference detection probe based on the computational feature; Step S2: Distribute the interference detection probe to several candidate computing nodes and obtain quantitative characterization information on the degree of contention for shared hardware resources collected by each candidate container during the execution of the interference detection probe; Step S3: Determine the performance guarantee credential for the task based on the quantified characterization information; Step S4: Receive the performance guarantee credentials returned by each candidate container, and determine the target container from each candidate container based on the performance guarantee credentials and the node's historical performance record; Step S5: Schedule the task to the target container; Step S6: Monitor the deviation of the actual performance from the performance guarantee index during task execution, and trigger an adjustment operation for the target container based on the cumulative deviation between the actual performance data and the performance guarantee index.

[0032] Specifically, in this embodiment, the probe generation module responds to a container scheduling request and determines the computational characteristics of the task to be scheduled, including: Respond to the container scheduling request of the task to be scheduled, and obtain the task type and task parameters of the task to be scheduled; The computing power consumption parameters are determined based on the task type and task size. Determine the memory pressure parameters based on the task type and accuracy requirements; The task parameters include accuracy requirements and task size, and the computational features include computing power consumption parameters and memory pressure parameters.

[0033] Specifically, the task type refers to the computational resource requirement pattern of the task to be scheduled. Different task types have fundamentally different hardware resource consumption characteristics. For example, artificial intelligence training tasks are characterized by intensive use of computing units and continuous consumption of video memory bandwidth; online inference services aim for low-latency response, with relatively lower computational intensity than training tasks but are more sensitive to interference with shared resources; data analysis tasks exhibit a mixed characteristic of computation and input / output, and their demand for cache capacity and memory bandwidth changes dynamically with the data characteristics.

[0034] Specifically, the precision requirement refers to the precision requirements of floating-point operations during the execution of the task to be scheduled, including low-precision mode or high-precision mode. In this embodiment of the invention, low-precision mode corresponds to data formats such as FP8, FP16, and INT8, while high-precision mode corresponds to data formats such as FP32, FP64, and BF16.

[0035] Specifically, the task size refers to the number of basic processing units contained in the task to be scheduled. Different task types have different basic processing units; artificial intelligence training tasks use samples as basic processing units, high-performance computing tasks use discrete points as basic processing units, and big data analysis tasks use data records as basic processing units. In practical application scenarios, the specific meaning of the basic processing unit is determined according to the task type, which will not be elaborated further here.

[0036] Specifically, precision requirements indicate the degree of memory usage and bandwidth consumption of a task. In low-precision mode, the amount of computational tasks processed within the same time period is larger, and the frequency of memory access requests per unit time is higher, putting greater pressure on memory bandwidth. In high-precision mode, the required memory capacity is larger for the same task size, but the instantaneous bandwidth of the memory access channel is relatively smaller. Task size characterizes the total computing power demand of the scheduled task per unit time. The larger the task size, the longer the task maintains a high computing unit utilization rate during execution, and the greater the possibility of competing for computing resources with other containers.

[0037] Specifically, the computing power occupancy parameter is a quantitative indicator used to characterize the degree and duration of resource consumption by the task to be scheduled. The probe generation module determines the value of the computing power occupancy parameter based on the computational intensity characteristics corresponding to the task type and the task size.

[0038] Specifically, the memory pressure parameter is a quantitative indicator used to characterize the intensity of memory bandwidth and memory capacity resource consumption by the scheduled task. The probe generation module determines the value of the memory pressure parameter based on the memory access mode determined by the task type and the data format specified by the accuracy requirements.

[0039] Specifically, the process by which the probe generation module determines the computing power occupancy parameter includes: According to the task type, the corresponding unit-scale computation coefficient is matched from the preset computation intensity mapping table, where the unit-scale computation coefficient represents the number of floating-point operations required for each basic processing unit. Multiply the unit-scale computation coefficient by the task scale to obtain the computing power occupancy base; The ratio of the computing power occupancy base to the reference operand constant is determined as the computing power occupancy parameter.

[0040] The computation intensity mapping table is predetermined by the actual application scenario of the embodiment based on the typical computation density of different task types on heterogeneous computing hardware.

[0041] Specifically, the reference operand constant is determined as follows: During the system testing phase, the theoretical peak computing power of all computing nodes deployed in the data center is obtained, the average computing power of the nodes with the lowest computing power of 20% is calculated, and the product of the average computing power and the scaling factor is used as a reference operand constant.

[0042] The scaling factor ranges from [0.5, 0.7]. The scaling factor is used to appropriately reduce the average of the node's theoretical peak computing power as a reference operand constant. If the scaling factor is too large, the reference operand constant will be close to the average of the node's theoretical peak computing power, making it difficult for the scheduling decision module to effectively distinguish the relative demand intensity of tasks on computing unit resources based on the computing power occupancy parameter. If the scaling factor is too small, the reference operand constant will be too low, and the overall computing power occupancy parameter will be too large.

[0043] Specifically, the process by which the probe generation module determines the memory pressure parameter includes: The number of storage bytes occupied by each data element is determined according to the accuracy requirements; the number of data elements contained in each basic processing unit is determined according to the task type; the number of storage bytes is multiplied by the number of data elements contained in each basic processing unit and the task size to obtain the video memory space usage. The memory access frequency coefficient is determined based on the task type, and the memory space usage is multiplied by the memory access frequency deviation coefficient to determine the memory pressure base. The ratio of the memory pressure base to the reference memory constant is determined as the memory pressure parameter.

[0044] In this embodiment, the memory access frequency coefficient is determined based on the memory access mode of the task type, and the product of the video memory space occupancy and the memory access frequency coefficient is used as the base of video memory pressure. The memory access frequency coefficient corresponding to sequential memory access tasks is a positive deviation coefficient; the memory access frequency coefficient corresponding to discrete memory access tasks is a negative deviation coefficient; and the memory access frequency coefficient corresponding to mixed memory access tasks is a neutral deviation coefficient.

[0045] Understandably, sequential memory access tasks exhibit continuous and high-density memory bandwidth usage during execution, with frequent and regular memory access requests, resulting in significant competition pressure on the memory path. Therefore, a positive bias coefficient greater than 1 is used to amplify the memory pressure base and characterize the actual competition intensity. Discrete memory access tasks exhibit random and sparse memory access requests, with a lower degree of continuous memory bandwidth usage. Therefore, a negative bias coefficient less than 1 is used to reduce the memory pressure base and avoid overestimating the degree of competition for the memory path. Hybrid memory access tasks combine both sequential and discrete memory access modes, with memory bandwidth usage intensity falling between the two. Therefore, a neutral bias coefficient close to 1 is used.

[0046] In this embodiment of the invention, for continuous memory access tasks, the memory access frequency coefficient ranges from [1.05, 1.25]; for discrete memory access tasks, the memory access frequency coefficient ranges from [0.75, 0.95]; and for mixed memory access tasks, the memory access frequency coefficient ranges from (0.95, 1.05). The above value ranges are preset based on the actual memory bandwidth usage intensity of different memory access modes and can be determined according to the actual situation, which will not be elaborated here.

[0047] Specifically, the sequential memory access task refers to a task whose access to memory space during execution exhibits regular and sequential characteristics; the discrete memory access task refers to a task whose access to memory space exhibits random and discontinuous characteristics; and the hybrid memory access task refers to a task that combines both sequential and discrete memory access modes. The memory access mode of a task is a fundamental attribute, and its actual application scenario is determined based on the task's computational logic and data access patterns, which will not be elaborated further here.

[0048] Specifically, the reference memory constant is determined as follows: During the system testing phase, the memory bandwidth specifications of all computing nodes deployed in the data center are obtained, the average memory bandwidth of the nodes with the lowest 20% memory bandwidth is calculated, and the product of the average memory bandwidth and the scaling factor is used as the reference memory constant.

[0049] Specifically, the process by which the probe generation module determines the interference detection probe based on the computational features includes: The computational load intensity of the interference detection probe is determined based on the aforementioned computing power occupancy parameter. The amount of memory accessed by the interference detection probe is determined based on the aforementioned memory pressure parameter. Based on the computational load intensity and the amount of memory access data, an interference detection probe is constructed to reproduce the hardware load of the scheduled task. The interference detection probe is a lightweight payload segment that runs independently of the scheduled task.

[0050] The process of determining the computational load intensity of the interference detection probe based on the computing power occupancy parameter, as described in the embodiment, includes: The computing power occupancy base is used as the computational load intensity of the interference detection probe, and the computational load intensity represents the total number of operations that the floating-point operation loop embedded in the interference detection probe needs to complete.

[0051] The interference detection probe is configured to continuously execute floating-point arithmetic loops until the total number of operations is reached, and the execution time is determined by the actual computing speed of the computing units in the candidate container.

[0052] The process of determining the amount of memory accessed by the interference detection probe based on the memory pressure parameter, as described in the embodiment, includes: The memory pressure base is used as the memory access data volume of the interference detection probe, where the memory access data volume represents the total number of data bytes that the interference detection probe needs to read and write. The interference detection probe is configured to perform read and write operations on a memory buffer of the same size as the memory access data volume during execution, according to the memory access mode of the task type.

[0053] In this embodiment, the process of constructing an interference detection probe that reproduces the hardware load of the scheduled task based on the computational load intensity and the amount of memory access data includes: A floating-point arithmetic loop is constructed with the total number of operations as the loop limit, and a memory access read / write loop is constructed with the memory access data volume as the buffer size. The floating-point arithmetic loop and the memory access read / write loop are organized into an executable process in a time-division alternating manner. The ratio of computation operations to memory access operations in the process is set according to the computation-to-memory ratio of the task type. The executable process is encapsulated as an independent container image, serving as the interference detection probe.

[0054] It is understood that the interference detection probe is a lightweight load segment that runs independently of the scheduled task. It is used to simulate the resource occupation mode of the scheduled task on the computing unit and memory path when executed in the candidate container, and induce the competition for shared hardware resources on the computing node where the candidate container is located, so that the distribution and acquisition module can collect quantitative characterization information of the degree of competition for shared hardware resources generated by each candidate container during the execution of the interference detection probe.

[0055] Specifically, the distribution and acquisition module distributes the interference detection probe to several candidate computing nodes, and obtains quantitative characterization information on the degree of contention for shared hardware resources collected by each candidate container during the execution of the interference detection probe, including: Containers with available computing power resources are selected from the cluster to form a candidate container set; Distribute the interference detection probes to each candidate container in the candidate container set; Obtain the quantitative characterization information generated by each candidate container when executing the interference detection probe.

[0056] The cluster described in this embodiment of the invention refers to a heterogeneous computing resource pool composed of several computing nodes. Each computing node is configured with one or more types of computing hardware, including at least one of a central processing unit, a graphics processing unit, and a neural network processor. Each computing node runs a container runtime environment, supporting the concurrent execution of multiple containerized tasks.

[0057] Please see Figure 3 As shown, this is a logical judgment diagram for determining whether a corresponding container has idle computing resources according to an embodiment of the present invention. The process of selecting containers with idle computing resources from the cluster to form a candidate container set includes: Obtain the current resource usage status of each container in the cluster. The resource usage status includes the proportion of computing power allocated to the computing hardware of the computing node where the container is located and the proportion of video memory occupied. If the allocated computing power ratio of the computing node where the container is located is lower than the preset computing power idle threshold and the occupied video memory ratio is lower than the preset video memory idle threshold, then the corresponding container is determined to have idle computing power resources. Add all containers with available computing resources to the candidate container set.

[0058] The idle computing power threshold and the idle video memory threshold are preset according to the cluster load characteristics in the actual application scenario of the embodiment. Preferably, the idle computing power threshold can be set to a range of 60% to 80%, and the idle video memory threshold can be set to a range of 50% to 70%.

[0059] Specifically, the distribution and acquisition module acquires the quantitative characterization information generated by each candidate container when executing the interference detection probe, including: Obtain a set of container-related parameters, including the ready state pause duration of each container during the execution of the interference detection probe, and the reduction ratio of available cache space of each container during the execution of the interference detection probe. Obtain a set of computing power related parameters, including the decrease in computing unit utilization rate of each container during the execution of the interference detection probe, and the decrease in the proportion of available video memory bandwidth of each container during the execution of the interference detection probe. The computational resource contention characterization value is determined based on the container-related parameter set, and the memory access path contention characterization value is determined based on the computing power-related parameter set. The quantitative characterization information includes the computing resource contention characterization value and the memory access path contention characterization value.

[0060] In this embodiment, the computational resource contention characterization value is determined based on a standardized stagnation coefficient and a cache reduction coefficient. The standardized stagnation coefficient is determined by the ratio of the ready-state stagnation time to a preset baseline stagnation time, and the cache reduction coefficient is determined by the reduction ratio of available cache space. In this embodiment, the computational resource contention characterization value is obtained by summing the standardized stagnation coefficient and the cache reduction coefficient. The ready-state stagnation time refers to the length of time a candidate container is in a ready state while waiting for computational unit resources during the execution of an interference detection probe, and the available cache space reduction ratio refers to the percentage reduction in the available final-level cache capacity of the candidate container relative to its idle state during the execution of the interference detection probe.

[0061] Understandably, the ready-state stagnation time characterizes the length of time a candidate container cannot be scheduled for execution due to waiting for computing unit resources. This metric is significantly affected by the overall load fluctuations of the node where the container resides, and there may be substantial variance between different detection cycles. The available cache space reduction ratio directly characterizes the degree of competition between the interference detection probe and other containers on the same node for the final-level cache space. The impact of cache capacity reduction on task execution efficiency is more stable and repeatable. Therefore, assigning a higher weight to the cache reduction coefficient allows the computing resource contention metric to focus more on the degree of contention in the cache resource dimension, reducing the interference of ready-state stagnation time measurement fluctuations on the overall judgment.

[0062] In this embodiment, the memory access path contention characterization value is determined based on the utilization rate reduction coefficient and the bandwidth reduction coefficient. The utilization rate reduction coefficient is determined based on the ratio of the utilization rate reduction of the computing unit to a preset baseline utilization rate reduction, and the bandwidth reduction coefficient is determined based on the reduction in the available memory bandwidth ratio. Preferably, the sum of the utilization rate reduction coefficient and the bandwidth reduction coefficient is used to determine the memory access path contention characterization value. The computing unit utilization rate reduction refers to the percentage decrease in the actual utilization rate of the candidate container's computing unit during the execution of the interference detection probe relative to its idle state utilization rate, and the available memory bandwidth ratio reduction refers to the percentage decrease in the actual available memory bandwidth ratio of the candidate container during the execution of the interference detection probe relative to its theoretical peak value.

[0063] Understandably, the decrease in computing unit utilization represents the percentage decrease in the actual utilization of candidate container computing units during probe execution relative to the idle state. This indicator is influenced by both the probe's own computational load intensity and the node scheduling strategy. The decrease in available memory bandwidth directly represents the degree of contention for memory bandwidth between the probe and other tasks on the same node. In heterogeneous computing environments, memory bandwidth is often a key bottleneck resource restricting task execution efficiency, and its contention level typically has a greater impact on actual task performance than fluctuations in computing unit utilization. Therefore, assigning a slightly higher weight to the bandwidth decrease coefficient than the utilization decrease coefficient allows the memory access path contention value to more accurately represent the competition for memory bandwidth resources.

[0064] Specifically, the credential generation module determines the performance guarantee credentials for the task based on the quantified characterization information, including: The comprehensive interference index is obtained by weighted summing of the computing resource contention characterization value and the memory access path contention characterization value; The performance guarantee indicators in the performance guarantee certificate are determined based on the comprehensive interference index. The performance guarantee index is inversely proportional to the comprehensive interference index, and the performance guarantee credential includes the performance guarantee index and the candidate container number.

[0065] In this embodiment, the product of a preset upper limit guarantee index and the comprehensive interference index is determined as the performance guarantee index.

[0066] The upper limit guarantee index in this embodiment of the invention is the maximum value of the system-preset performance guarantee index, representing the highest performance level that a candidate container can promise to the task to be scheduled under an ideal state with no interference. The numerical range of the upper limit guarantee index is [0.95, 0.98), which is preset according to the minimum theoretical performance loss of computing nodes in the data center in the actual application scenario of the embodiment. The reason for setting the upper limit guarantee index below 100% is that even without other containers running concurrently, task execution still suffers from inherent performance degradation such as operating system scheduling overhead, hardware interrupt handling, and virtualization layer loss. Setting the upper limit guarantee index to 100% would deviate from the actually achievable performance commitment range.

[0067] Specifically, the scheduling decision module receives performance guarantee credentials corresponding to each candidate container, and based on the performance guarantee credentials and the node's historical performance records, determines the target container from among the candidate containers, including... Obtain the historical performance records of the nodes corresponding to each candidate container, and determine the credibility coefficient of each candidate container based on the historical performance records of the nodes; The performance guarantee index in the performance guarantee credential of each candidate container is multiplied by the corresponding credibility coefficient to obtain the adoption score of each candidate container; The candidate container with the highest adoption score is selected as the target container.

[0068] Specifically, the scheduling decision module obtains the historical performance records of the nodes corresponding to each candidate container, and determines the credibility coefficient of each candidate container based on the historical performance records of the nodes, including: Obtain the confidence ratio of the number of fulfillments to the number of defaults in the historical fulfillment records corresponding to the candidate container; The confidence ratio is mapped to a preset range of values ​​to obtain the confidence coefficient; The confidence coefficient increases as the confidence ratio increases.

[0069] In this embodiment of the invention, the confidence ratio is the ratio of the number of fulfillments to the number of defaults, characterizing the reliability of a candidate container in fulfilling its performance commitments in historical scheduling. A lower confidence ratio indicates a higher frequency of container defaults and a lower degree of credibility of its performance guarantee credentials; therefore, a smaller confidence coefficient should be assigned to reduce its adoption score in scheduling decisions. Conversely, a higher confidence ratio indicates more stable long-term performance of the container, and a larger confidence coefficient should be assigned to provide positive incentives. Considering the moderating effect of the confidence coefficient on scheduling results, its increase with the confidence ratio should exhibit a diminishing marginal return characteristic; that is, the increase in the confidence coefficient is larger in the low ratio range and slows down in the high ratio range, to reduce the overuse defects caused by containers with extremely high confidence ratios in scheduling.

[0070] In this embodiment of the invention, the correspondence between the confidence ratio and the credibility coefficient includes: When the confidence ratio is in the interval [0, 0.5), the preset range of the confidence coefficient is [0.55, 0.65]; when the confidence ratio is in the interval [0.5, 0.9), the preset range of the confidence coefficient is [0.75, 0.85]; when the confidence ratio is in the interval [0.9, 1.25), the preset range of the confidence coefficient is [0.95, 1.05]; when the confidence ratio is in the interval [1.25, 3), the preset range of the confidence coefficient is [1.15, 1.25]; when the confidence ratio is in the interval [3, +∞), the preset range of the confidence coefficient is [1.45, 1.55]. The confidence ratio intervals and corresponding confidence coefficient ranges in this embodiment are preset based on the historical scheduling statistics of the cluster in the actual application scenario. Different clusters have different task load characteristics and node performance fluctuations. The interval boundaries are adjusted according to actual historical data, or the confidence coefficient value is fine-tuned while keeping the interval boundaries unchanged. This will not be elaborated further.

[0071] Specifically, the monitoring and penalty module monitors the deviation of actual performance from the performance guarantee index during task execution. Based on the cumulative deviation between the actual performance data and the performance guarantee index, it triggers adjustment operations for the target container, including... During task execution, actual performance data of the task in the target container is collected at preset intervals. Based on the actual performance data and the performance guarantee indicators, the deviation amount is determined; Based on the deviation, determine whether the corresponding preset period meets the preset default conditions; If the number of preset periods in which the preset default conditions are met consecutively exceeds the default threshold, an adjustment operation is triggered for the target container.

[0072] The process for determining the default threshold described in the embodiment is as follows: The system acquires the actual performance data of historically scheduled tasks during their execution, statistically analyzes the deviation distribution of each task's actual performance relative to the performance guarantee index over several consecutive preset periods, and defines periods where the deviation exceeds a preset deviation threshold as potential default periods. It then calculates the probability distribution of consecutive potential default periods and determines the default threshold as the smallest integer that makes the cumulative probability of consecutive potential default periods reach a preset confidence level. In this embodiment, the default threshold is set to 3.

[0073] Please see Figure 4 As shown, this is a logic diagram for determining whether a preset breach condition is met according to an embodiment of the present invention. The monitoring and penalty module determines whether the preset period corresponding to the deviation meets the preset breach condition based on the deviation amount, including... If the deviation is greater than or equal to a preset deviation threshold, then the preset default condition is determined to be met. If the deviation is less than the preset deviation threshold, it is determined that the preset default condition is not met.

[0074] Specifically, the deviation threshold is determined as follows: The actual performance data of historically scheduled tasks in the cluster during their execution is obtained. The deviation distribution of the actual performance of each task relative to the performance guarantee index within each preset period is calculated, and the 5th percentile of the deviation distribution is determined as the deviation threshold. In this embodiment of the invention, the deviation threshold is set to 5% of the performance guarantee index.

[0075] Understandably, by periodically monitoring performance and determining defaults during task execution, an adjustment mechanism is built from scheduling decisions to execution feedback. When the actual performance of a target container consistently deviates from the performance guarantee indicators over multiple consecutive preset periods, an adjustment operation is triggered to correct the container's historical performance record, reducing its reliability in subsequent scheduling. This results in a lower adoption score and a reduced probability of it being selected as a target container again. This adjustment mechanism enables the scheduling system to continuously monitor the actual service quality information of each container, using historical performance as a constraint on future scheduling decisions. This avoids repeatedly scheduling tasks to containers that are difficult to fulfill performance commitments in the long term, thus improving the long-term reliability of the scheduling system within a containerized data center.

[0076] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A containerized heterogeneous computing resource perception and intelligent scheduling system, characterized in that, include: The probe generation module is used to respond to container scheduling requests, determine the computational characteristics of the task to be scheduled, and determine interference detection probes based on the computational characteristics. The interference detection probes are used to reproduce the interference mode of the task on the hardware microarchitecture on heterogeneous computing power nodes. The distribution and acquisition module is used to distribute the interference detection probe to several candidate computing nodes and acquire quantitative characterization information of the degree of contention for shared hardware resources collected by each candidate container during the execution of the interference detection probe. A credential generation module, deployed on each of the candidate computing nodes, is used to determine performance guarantee credentials for the task based on the quantization characterization information. The scheduling decision module receives the performance guarantee credentials corresponding to each candidate container, determines the target container from each candidate container based on the performance guarantee credentials and the node's historical performance record, and schedules the task to the target container. The monitoring and penalty module is used to monitor the deviation of actual performance from the performance guarantee index during task execution, and trigger adjustment operations for the target container based on the cumulative deviation between the actual performance data and the performance guarantee index. The adjustment operation includes correcting the historical performance record of the target container so that the weight of the target container is reduced in subsequent scheduling.

2. The containerized heterogeneous computing resource perception and intelligent scheduling system according to claim 1, characterized in that, The probe generation module responds to the container scheduling request and determines the computational characteristics of the task to be scheduled, including: Respond to the container scheduling request of the task to be scheduled, and obtain the task type and task parameters of the task to be scheduled; The computing power consumption parameters are determined based on the task type and task size. Determine the memory pressure parameters based on the task type and accuracy requirements; The task parameters include accuracy requirements and task size, and the computational features include computing power consumption parameters and memory pressure parameters.

3. The containerized heterogeneous computing resource perception and intelligent scheduling system according to claim 2, characterized in that, The probe generation module determines the interference detection probes based on the computational features. The computational load intensity of the interference detection probe is determined based on the aforementioned computing power occupancy parameter. The amount of memory accessed by the interference detection probe is determined based on the aforementioned memory pressure parameter. Based on the computational load intensity and the amount of memory access data, an interference detection probe is constructed to reproduce the hardware load of the scheduled task. The interference detection probe is a lightweight payload segment that runs independently of the scheduled task.

4. The containerized heterogeneous computing resource perception and intelligent scheduling system according to claim 1, characterized in that, The distribution and acquisition module distributes the interference detection probe to several candidate computing nodes, and obtains quantitative characterization information on the degree of contention for shared hardware resources collected by each candidate container during the execution of the interference detection probe, including... Containers with available computing power resources are selected from the cluster to form a candidate container set; Distribute the interference detection probes to each candidate container in the candidate container set; Obtain the quantitative characterization information generated by each candidate container when executing the interference detection probe.

5. The containerized heterogeneous computing resource perception and intelligent scheduling system according to claim 4, characterized in that, The distribution and acquisition module obtains the quantitative characterization information generated by each candidate container when executing the interference detection probe, including... Obtain a set of container-related parameters, including the ready state pause duration of each container during the execution of the interference detection probe, and the reduction ratio of available cache space of each container during the execution of the interference detection probe. Obtain a set of computing power related parameters, including the decrease in computing unit utilization rate of each container during the execution of the interference detection probe, and the decrease in the proportion of available video memory bandwidth of each container during the execution of the interference detection probe. The computational resource contention characterization value is determined based on the container-related parameter set, and the memory access path contention characterization value is determined based on the computing power-related parameter set. The quantitative characterization information includes the computing resource contention characterization value and the memory access path contention characterization value.

6. The containerized heterogeneous computing resource perception and intelligent scheduling system according to claim 5, characterized in that, The credential generation module determines the performance guarantee credentials for the task based on the quantified characterization information. The comprehensive interference index is obtained by weighted summing of the computing resource contention characterization value and the memory access path contention characterization value; The performance guarantee indicators in the performance guarantee certificate are determined based on the comprehensive interference index. The performance guarantee index is inversely proportional to the comprehensive interference index, and the performance guarantee credential includes the performance guarantee index and the candidate container number.

7. The containerized heterogeneous computing resource perception and intelligent scheduling system according to claim 1, characterized in that, The scheduling decision module receives performance guarantee credentials corresponding to each candidate container, and determines the target container from among the candidate containers based on the performance guarantee credentials and the node's historical performance records. Obtain the historical performance records of the nodes corresponding to each candidate container, and determine the credibility coefficient of each candidate container based on the historical performance records of the nodes; The performance guarantee index in the performance guarantee credential of each candidate container is multiplied by the corresponding credibility coefficient to obtain the adoption score of each candidate container; The candidate container with the highest adoption score is selected as the target container.

8. The containerized heterogeneous computing resource perception and intelligent scheduling system according to claim 7, characterized in that, The scheduling decision module obtains the historical performance records of the nodes corresponding to each candidate container, and determines the credibility coefficient of each candidate container based on the historical performance records of the nodes, including: Obtain the confidence ratio of the number of fulfillments to the number of defaults in the historical fulfillment records corresponding to the candidate container; The confidence ratio is mapped to a preset range of values ​​to obtain the confidence coefficient; The confidence coefficient increases as the confidence ratio increases.

9. The containerized heterogeneous computing resource perception and intelligent scheduling system according to claim 8, characterized in that, The monitoring and penalty module monitors the deviation of actual performance from the performance guarantee index during task execution. Based on the cumulative deviation between the actual performance data and the performance guarantee index, it triggers adjustment operations for the target container, including... During task execution, actual performance data of the task in the target container is collected at preset intervals. Based on the actual performance data and the performance guarantee indicators, the deviation amount is determined; Based on the deviation, determine whether the corresponding preset period meets the preset default conditions; If the number of preset periods in which the preset default conditions are met consecutively exceeds the default threshold, an adjustment operation is triggered for the target container.

10. The containerized heterogeneous computing resource perception and intelligent scheduling system according to claim 9, characterized in that, The monitoring and penalty module determines whether the preset default conditions are met based on the deviation amount within a corresponding preset period, including... If the deviation is greater than or equal to a preset deviation threshold, then the preset default condition is determined to be met. If the deviation is less than the preset deviation threshold, it is determined that the preset default condition is not met.

Images