High-efficiency computer software system resource scheduling method, system, equipment and medium

Abstract

The invention discloses a high-efficiency computer software system resource scheduling method, system and device and a medium, and the method comprises the steps: deploying a hardware counter in an operating system, and monitoring the CPU utilization rate, memory occupation, disk I / O, network bandwidth, the use condition of data aggregation system resources and the resource demand of each process or thread in real time; based on monitoring data, analyzing a current resource supply and demand state by adopting an advanced algorithm, and predicting a future resource demand trend; according to an analysis result, a resource allocation strategy is dynamically adjusted, it is ensured that a high-priority task obtains needed resources in time, and meanwhile fairness of a low-priority task is considered; data feedback after scheduling execution is collected, a scheduling algorithm is continuously optimized, and the scheduling efficiency is improved; the system, the equipment and the medium are used for realizing the high-efficiency computer software system resource scheduling method. The problems that an existing computer system is low in resource utilization rate, poor in key task stability and high in operation and maintenance cost are solved.

Description

Technical Field

[0001] The present invention belongs to the field of computer software systems, and particularly relates to a high-performance computer software system resource scheduling method, system, device and medium. Background Art

[0002] With the improvement of computer system complexity and the explosive growth of diversified application scenarios, such as cloud computing, edge computing, and real-time data processing, the resource scheduling mechanism has become the core factor of system performance. However, there are still significant deficiencies in multi-dimensional resource collaborative scheduling, dynamic load adaptation, and service quality guarantee in the existing technologies, which are specifically manifested in the following problems:

[0003] 1. Limitations of the static resource allocation mechanism

[0004] (1) Rigidity of the fixed quota strategy: Traditional schedulers adopt a static allocation mode based on time slice rotation and cannot perceive the changes in real-time resource requirements of tasks.

[0005] (2) Lack of coordination of multiple resource types: Existing solutions mostly optimize a single resource (such as CPU or memory) independently, ignoring the coupling relationship between resources.

[0006] (3) Imbalance between I / O and computing resources: When a compute-intensive task occupies a large amount of CPU, it may cause the data processing pipeline to block due to insufficient reserved I / O bandwidth.

[0007] (4) Memory-storage cascade bottleneck: When memory is insufficient, disk swapping is frequently triggered, but traditional schedulers do not consider disk I / O load in the memory recycling policy, exacerbating system jitter.

[0008] 2. Response delay and real-time defects

[0009] (1) The long-tail delay problem is prominent: Scheduling algorithms based on historical average load are difficult to cope with instantaneous load spikes, resulting in a significant increase in tail latency.

[0010] (2) Insufficient guarantee for real-time tasks: Although existing real-time scheduling algorithms can ensure hard real-time tasks, they lack support for the mixed scheduling of soft real-time tasks and batch tasks.

[0011] 3. Singularity and adaptive defects of scheduling strategies

[0012] (1) Mechanical nature of the rule engine: Mainstream scheduling systems rely on manually preset rules and cannot dynamically adapt to changes in load characteristics.

[0013] (2) Lack of prediction and learning ability: Existing technologies mostly adopt reactive scheduling and lack a resource demand prediction mechanism based on machine learning.

[0014] The above problems will cause the computer system to face the problems of low resource utilization rate, poor stability of critical tasks, and high operation and maintenance costs. Summary of the Invention

[0015] To overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a high-performance computer software system resource scheduling method, system, device and medium, which solves the problems of low resource utilization rate, poor stability of critical tasks, and high operation and maintenance costs faced by the existing computer system.

[0016] The present invention is realized through the following technical solutions:

[0017] A high-performance computer software system resource scheduling method includes the following steps:

[0018] S1. Deploy hardware counters in the operating system to monitor the usage of CPU utilization rate, memory occupancy, disk I / O, network bandwidth and data aggregation system resources in real time, as well as the resource requirements of each process or thread;

[0019] S2. Based on the monitoring data obtained in step S1, use an advanced algorithm to analyze the current resource supply and demand status and predict the future resource demand trend;

[0020] S3. According to the analysis result of step S2, dynamically adjust the resource allocation strategy to ensure that high-priority tasks can obtain the required resources in time, while taking into account the fairness of low-priority tasks;

[0021] S4. According to the resource allocation strategy dynamically adjusted in step S3, collect the data feedback after the scheduling execution, continuously optimize the scheduling algorithm, and improve the scheduling efficiency;

[0022] Among them, step S2 is specifically:

[0023] Predict the resource requirements within the next 3 - 7 ms based on the time series analysis ARIMA model;

[0024] Use the LSTM neural network to learn the periodic task pattern and daily peak load;

[0025] Generate dynamic priority weights according to the task type and the user-defined service level agreement;

[0026] According to the generated dynamic priority weights, use a red-black tree to store the process resource requirements to achieve fast query of data complexity;

[0027] Use the resource status matrix to record the current system resource distribution, in the format of a multi-dimensional tensor CPU core × memory block × disk queue, to obtain the future resource demand trend.

[0028] Further, in the step S1, the CPU usage rate is tracked through / proc / stat and process scheduler events, the memory occupancy is determined by monitoring / proc / meminfo and the resident memory of the process, the disk I / O utilization rate is obtained by using the blktrace tool to record the read / write latency and throughput of the block device request queue, the network bandwidth is captured by the Netfilter framework to count the network interface data packets, and the data aggregation uses time window sharding to compress the original data.

[0029] Further, the step S3 is specifically as follows:

[0030] 1) According to the obtained future resource demand trend, high-priority tasks can preempt the resources of low-priority tasks, but at the same time, a minimum guarantee quota is set; a real-time queue and a time-sharing queue are used in parallel;

[0031] 2) Reserve physical cores and memory pages for critical processes through cgroups v2, and the disk I / O bandwidth allocation adopts the token bucket algorithm to limit the burst traffic.

[0032] Further, the step S4 is specifically as follows:

[0033] The performance metric library stores historical scheduling data;

[0034] Based on the PPO algorithm, with the system throughput maximization and the tail latency minimization as the reward function, continuously update the scheduling parameters;

[0035] Run the resource allocation strategy after updating the scheduling parameters and the resource allocation strategy without updating the scheduling parameters in parallel, and select the better resource allocation strategy through the T-test;

[0036] Implement the replacement of the scheduling algorithm without restart through eBPF CO-RE.

[0037] Further, in the step S1, the field programmable gate array technology is used to preprocess the monitoring data.

[0038] Further, in the step S3, the nanosecond-level task scheduling and resource management are realized by integrating a scheduling decision chip, and the large-scale container cluster scheduling is realized by using a GPU-accelerated parallel task migration engine.

[0039] Further, the historical scheduling data is specifically the task completion time and the standard deviation of resource utilization rate.

[0040] An efficient computer software system resource scheduling system, including:

[0041] Monitoring Module: Deploy lightweight probes in the operating system to monitor the CPU usage rate, memory occupancy, disk I / O, network bandwidth, and the usage of data aggregation system resources in real time, as well as the resource requirements of each process or thread;

[0042] Analysis Module: Based on the monitoring data obtained by the monitoring module, adopt advanced algorithms to analyze the current resource supply and demand status and predict the future resource demand trend. Specifically:

[0043] Predict the resource demand within the next 3 - 7 ms based on the ARIMA model of time series analysis;

[0044] Use the LSTM neural network to learn periodic task patterns and daily peak loads;

[0045] Generate dynamic priority weights according to the task type and the user - defined service - level agreement;

[0046] According to the generated dynamic priority weights, use a red - black tree to store the process resource requirements to achieve fast query of data complexity;

[0047] Use the resource status matrix to record the current system resource distribution in the format of multi - dimensional tensor CPU cores × memory blocks × disk queues to obtain the future resource demand trend;

[0048] Scheduling Module: According to the analysis results of the analysis module, dynamically adjust the resource allocation strategy to ensure that high - priority tasks can obtain the required resources in a timely manner, while taking into account the fairness of low - priority tasks;

[0049] Feedback Module: Collect data feedback after scheduling execution, continuously optimize the scheduling algorithm, and improve the scheduling efficiency.

[0050] A high - performance computer software system resource scheduling device, including:

[0051] Memory: Store a computer program for a high - performance computer software system resource scheduling method, which is a computer - readable device;

[0052] Processor: Used to implement a high - performance computer software system resource scheduling method when executing the computer program.

[0053] A computer - readable storage medium stores a computer program, and when the computer program is executed by a processor, it can implement a high - performance computer software system resource scheduling method.

[0054] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0055] The present invention deploys hardware counters in the operating system. Through precise resource demand prediction and dynamic scheduling, it reduces resource waste and improves resource utilization. It uses advanced algorithms to analyze the resource supply and demand status, ensures that critical tasks are executed first, guarantees the normal operation of other tasks, enhances the user experience, shortens the task response time, improves the overall system processing speed, and optimizes the system performance.

[0056] The device of the present invention is designed modularly, which is convenient for integration with other software systems and is easy to expand functions according to requirements, being easy to integrate and expand.

[0057] The present invention realizes the reasonable and efficient scheduling of resources through hardware-level monitoring acceleration, multi-modal resource demand prediction models, and dynamic feedback scheduling strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 is the flowchart of the method of the present invention.

[0059] Figure 2 is the schematic diagram of the key hardware design structure of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0060] The following further elaborates on the present invention with specific embodiments, which are explanations rather than limitations of the present invention.

[0061] The embodiment of the present invention discloses a method for scheduling resources of a high-performance computer software system, and proposes a "perception-prediction-decision-optimization" closed-loop scheduling architecture. Through hardware-level monitoring acceleration, multi-modal resource demand prediction models, and dynamic feedback scheduling strategies, it breaks through the bottleneck of the existing technology. The specific innovations include: 1. Heterogeneous resource unified measurement model: abstracts resources such as CPU, memory, and I / O into "resource units" that can be quantitatively replaced, solving the problem of collaborative allocation of multiple resources. 2. Online-offline hybrid prediction mechanism: combines lightweight time series analysis (short-term decision-making) with deep learning models (long-term trends) to achieve microsecond-level demand prediction. 3. Preemptive hierarchical scheduling: designs a preemptive strategy with weighted backfilling, taking into account the immediate response of high-priority tasks and the fairness of low-priority tasks. It should be noted that this embodiment can be specifically implemented in the following two scenarios:

[0062] (1) Cloud computing scenario:

[0063] Deploy this device in the Kubernetes cluster, and integrate the scheduling module with kube-scheduler.

[0064] Through the monitoring module, identify GPU video memory fragmentation, the analysis module triggers the video memory compression algorithm (such as NVIDIA MPS), and the scheduling module merges scattered ML tasks onto the same GPU.

[0065] (2) Edge computing scenario:

[0066] The device is embedded in the Internet of Things gateway, and the analysis module adjusts the bit rate of the video analysis task according to the network jitter prediction.

[0067] Use lightweight scheduling strategies (such as DQN algorithm) to adapt to low computing power environments.

[0068] The following details the specific implementation process of an efficient computer software system resource scheduling device disclosed in the embodiments of the present invention.

[0069] Refer to Figure 1 As shown, an efficient computer software system resource scheduling method disclosed in the embodiments of the present invention includes:

[0070] S1. Real-time monitor the usage of system resources such as CPU, memory, disk I / O, etc., and the resource requirements of each process or thread. Specifically, the following implementation technologies are included:

[0071] (1) Sensor agent: Deploy hardware counters in the operating system kernel to capture the following data in real time:

[0072] CPU usage rate: Through / proc / stat and process scheduler event tracking;

[0073] Memory occupancy: Monitor / proc / meminfo and the resident memory of the process;

[0074] Disk I / O: Use the blktrace tool to record the read / write latency and throughput of the block device request queue;

[0075] Network bandwidth: Grab network interface packet statistics through the Netfilter framework;

[0076] Data aggregation: Adopt time window sharding to compress the original data and transmit it to the analysis module through shared memory.

[0077] (2) Hardware support:

[0078] Hardware counter: Used for high-precision performance monitoring.

[0079] FPGA-accelerated data preprocessing unit to reduce CPU overhead.

[0080] In this embodiment, the monitoring module samples the CPU cache miss rate at intervals of 100 μs.

[0081] S2. Based on the monitoring data, use advanced algorithms to analyze the current resource supply and demand status and predict the future resource demand trend. Specifically, the following implementation technologies are included:

[0082] (1) Algorithm implementation:

[0083] ① Resource demand prediction:

[0084] Short-term prediction: Predict the resource demand within the next 5ms based on time series analysis (ARIMA model).

[0085] Long-term trend: Use LSTM neural network to learn the periodic task pattern (daily peak load).

[0086] ② Dynamic priority calculation:

[0087] Generate dynamic priority weights according to the task type (real-time task, batch task) and user-defined SLA (Service Level Agreement).

[0088] Formula: Priority = α·(Deadline Urgency) + β·(Resource Demand) + γ·(UserTier)

[0089] ③ Data structure:

[0090] Use a red-black tree to store the process resource demand and achieve fast query with O(log n) complexity.

[0091] The resource status matrix (RSM) records the current system resource distribution in the format of a multi-dimensional tensor (CPU core × memory block × disk queue).

[0092] In this embodiment, the analysis module detects that the LSTM predicted memory demand of the database process will increase by 30% in the next cycle.

[0093] S3. According to the analysis results, dynamically adjust the resource allocation strategy to ensure that high-priority tasks can obtain the required resources in a timely manner, while taking into account the fairness of low-priority tasks. The specific implementation techniques are as follows:

[0094] (1) Dynamic scheduling strategy:

[0095] ① Preemptive allocation:

[0096] High-priority tasks can preempt the resources of low-priority tasks, but set a "minimum guarantee quota" to prevent starvation.

[0097] Adopt a two-level queue: a real-time queue (FIFO) and a time-sharing queue (CFS completely fair scheduler).

[0098] ② Resource reservation mechanism:

[0099] Reserve physical cores and memory pages for critical processes (implemented through `cgroups v2`).

[0100] The disk I / O bandwidth allocation adopts the token bucket algorithm to limit the burst traffic.

[0101] (2) Hardware acceleration:

[0102] Integrated scheduling decision chip (ASIC), supporting nanosecond-level context switching.

[0103] GPU-accelerated parallel task migration engine for large-scale container cluster scheduling.

[0104] In this embodiment, the scheduling module immediately reclaims 200MB of memory from the low-priority log service and binds it to the NUMA node of the database process.

[0105] S4. Collect data feedback after scheduling execution, continuously optimize the scheduling algorithm, and improve the scheduling efficiency. The specific implementation technologies are as follows:

[0106] (1) Closed-loop optimization system:

[0107] Performance metric library: Stores historical scheduling data (task completion time, standard deviation of resource utilization).

[0108] Reinforcement learning model: Based on the PPO algorithm, with system throughput maximization and tail latency minimization as the reward function, continuously optimize the scheduling parameters.

[0109] A / B test engine: Runs new and old scheduling strategies in parallel and selects a better solution through T-test.

[0110] (2) Hot update mechanism:

[0111] Realize dynamic algorithm replacement without restart through eBPF CO-RE technology.

[0112] In this embodiment, the feedback module discovers that this scheduling reduces the query latency by 15%, and updates the weight parameters of the reinforcement learning model.

[0113] Specifically, as Figure 2 shown, an efficient computer software system resource scheduling method further includes key hardware designs:

[0114] 10. Heterogeneous computing units, including:

[0115] 101. CPU cluster (x86 / ARM), used for general computing.

[0116] 102. NPU (neural network processor), dedicated to the prediction model inference of the analysis module.

[0117] 103. RDMA network card, realizing cross-node resource status synchronization.

[0118] 11. Resource isolation controller:

[0119] 111. Cache partitioning based on Intel RDT (Resource Director Technology).

[0120] 112. Memory Bandwidth Arbiter (MBA) to prevent memory-intensive tasks from interfering.

[0121] An efficient computer software system resource scheduling system, comprising:

[0122] Monitoring module: Deploy lightweight probes in the operating system to monitor the CPU usage rate, memory occupancy, disk I / O, network bandwidth, and the usage of data aggregation system resources in real time, as well as the resource requirements of each process or thread;

[0123] Analysis module: Based on the monitoring data obtained by the monitoring module, use advanced algorithms to analyze the current resource supply and demand status and predict future resource demand trends. Specifically:

[0124] Predict resource requirements within the next 3 - 7 ms based on the ARIMA model of time series analysis;

[0125] Use the LSTM neural network to learn periodic task patterns and daily peak loads;

[0126] Generate dynamic priority weights according to the task type and user-defined service level agreement;

[0127] According to the generated dynamic priority weights, use a red-black tree to store the resource requirements of processes to achieve fast query of data complexity;

[0128] Use a resource status matrix to record the current system resource distribution in the format of a multi-dimensional tensor CPU core × memory block × disk queue to obtain future resource demand trends;

[0129] Scheduling module: According to the analysis results of the analysis module, dynamically adjust the resource allocation strategy to ensure that high-priority tasks can obtain the required resources in a timely manner while taking into account the fairness of low-priority tasks;

[0130] Feedback module: Collect data feedback after scheduling execution to continuously optimize the scheduling algorithm and improve scheduling efficiency.

[0131] An efficient computer software system resource scheduling device, comprising:

[0132] Memory: Store a computer program for an efficient computer software system resource scheduling method, which is a computer-readable device;

[0133] Processor: Used to implement an efficient computer software system resource scheduling method when executing the computer program.

[0134] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it can implement a high-performance computer software system resource scheduling method.

[0135] The present invention has been described in detail with reference to the above embodiments. Those of ordinary skill in the art should understand that: it is still possible to modify the specific implementation manners of the present invention or make equivalent replacements, and any modification or equivalent replacement that does not depart from the spirit and scope of the present invention shall be covered by the scope required by the present invention.

Claims

1. An efficient computer software system resource scheduling method, characterized in that, It includes the following steps: S1. Deploy hardware counters in the operating system to monitor the CPU usage rate, memory occupancy, disk I / O, network bandwidth, and the usage of data aggregation system resources in real time, as well as the resource requirements of each process or thread; S2. Based on the monitoring data obtained in step S1, use advanced algorithms to analyze the current resource supply and demand status and predict the future resource demand trend; S3. According to the analysis results of step S2, dynamically adjust the resource allocation strategy to ensure that high-priority tasks can obtain the required resources in a timely manner while taking into account the fairness of low-priority tasks; S4. According to the resource allocation strategy dynamically adjusted in step S3, collect the data feedback after scheduling execution, continuously optimize the scheduling algorithm, and improve the scheduling efficiency; Among them, step S2 is specifically: Predict the resource requirements within the next 3 - 7 ms based on the ARIMA model of time series analysis; Use the LSTM neural network to learn the periodic task patterns and daily peak loads; Generate dynamic priority weights according to the task type and the user-defined service level agreement; According to the generated dynamic priority weights, use a red-black tree to store the process resource requirements to achieve fast query of data complexity; Use the resource status matrix to record the current system resource distribution in the format of multi-dimensional tensor CPU cores × memory blocks × disk queues to obtain the future resource demand trend.

2. An efficient computer software system resource scheduling method according to claim 1, characterized in that, In step S1, the CPU usage rate is tracked through / proc / stat and process scheduler events, the memory occupancy is determined by monitoring / proc / meminfo and the resident memory of the process, the disk I / O utilization rate is obtained by using the blktrace tool to record the read / write latency and throughput of the block device request queue, the network bandwidth is obtained by capturing network interface packet statistics through the Netfilter framework, and data aggregation uses time window sharding to compress the original data.

3. An efficient computer software system resource scheduling method according to claim 1, characterized in that Step S3 is specifically: 1) According to the obtained future resource demand trend, high-priority tasks can preempt the resources of low-priority tasks, but at the same time set a minimum guarantee quota; use real-time queues and time-sharing queues in parallel; 2) Reserve physical cores and memory pages for critical processes through cgroups v2, and use the token bucket algorithm for disk I / O bandwidth allocation to limit burst traffic.

4. An efficient computer software system resource scheduling method according to claim 1, characterized in that, Step S4 is specifically: The performance metric library stores historical scheduling data; Based on the PPO algorithm, use the system throughput maximization and tail latency minimization as the reward function to continuously update the scheduling parameters; Run the resource allocation strategy with updated scheduling parameters and the resource allocation strategy without updated scheduling parameters in parallel, and select the better resource allocation strategy through the T-test; Implement the replacement of the scheduling algorithm without restart through eBPF CO-RE.

5. An efficient computer software system resource scheduling method according to claim 1, characterized in that In step S1, the field programmable gate array technology is used to preprocess the monitoring data.

6. An efficient computer software system resource scheduling method according to claim 1, characterized in that, In step S3, the nanosecond-level task scheduling and resource management are realized through the integrated scheduling decision chip, and the large-scale container cluster scheduling is realized by using the GPU-accelerated parallel task migration engine.

7. An efficient computer software system resource scheduling method according to claim 4, characterized in that The specific historical scheduling data is the task completion time and the standard deviation of resource utilization.

8. An efficient computer software system resource scheduling system, characterized in that, It includes: Monitoring Module: Deploy lightweight probes in the operating system to monitor the CPU usage rate, memory occupancy, disk I / O, network bandwidth, and the usage of data aggregation system resources in real time, as well as the resource requirements of each process or thread; Analysis Module: Based on the monitoring data obtained by the monitoring module, adopt advanced algorithms to analyze the current resource supply and demand status and predict the future resource demand trend. Specifically: Predict the resource demand within the next 3 - 7 ms based on the ARIMA model of time series analysis; Use the LSTM neural network to learn periodic task patterns and daily peak loads; Generate dynamic priority weights according to the task type and the user-defined service level agreement; According to the generated dynamic priority weights, use a red-black tree to store the process resource requirements to achieve fast query of data complexity; Utilize the resource status matrix to record the current system resource distribution in the format of multi-dimensional tensor CPU cores × memory blocks × disk queues to obtain the future resource demand trend; Scheduling Module: According to the analysis results of the analysis module, dynamically adjust the resource allocation strategy to ensure that high-priority tasks can obtain the required resources in a timely manner while taking into account the fairness of low-priority tasks; Feedback Module: Collect data feedback after scheduling execution and continuously optimize the scheduling algorithm to improve the scheduling efficiency.

9. An efficient computer software system resource scheduling device, characterized in that, Including: Memory: Store the computer program of an efficient computer software system resource scheduling method described in any one of claims 1 - 7, which is a computer-readable device; Processor: Used to implement an efficient computer software system resource scheduling method described in any one of claims 1 - 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that, A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it can implement an efficient computer software system resource scheduling method described in any one of claims 1 - 7.

Images