Performance evaluation method and device of storage server, equipment, medium and product

Abstract

The invention discloses a performance evaluation method and device of a storage server, equipment, a medium and a product, and relates to the technical field of framework performance evaluation of the storage server, and the method comprises the steps: analyzing the deviation between the actual measurement performance data of a to-be-tested hard disk and the reference performance data, and when the deviation is greater than a deviation threshold value, determining the performance of the to-be-tested hard disk; fusing the current monitoring data reflecting the real-time power consumption state of each hardware component of the storage server, and obtaining the probability that the performance of the storage server deviates through a pre-trained performance prediction model so as to capture architecture-level implicit association; and constructing a performance deviation feature library of the hardware architecture dimension, and generating suggestions for the hardware architecture through a pre-trained performance optimization model based on the feature library. Therefore, the problem that a related evaluation method neglects a hardware architecture bottleneck can be solved, and systematic technical support is provided for performance optimization of the storage server.

Description

Technical Field

[0001] This application relates to the technical field of storage server architecture performance evaluation, and particularly to a performance evaluation method, device, equipment, medium and product for a storage server. Background Technique

[0002] With the popularization of new information technology (IT) and the continuous increase in performance requirements for enterprise applications, the input / output operations per second (IOPS) performance and low latency of traditional storage have become difficult to meet the needs and have become a bottleneck for enterprise business development. The serious imbalance in the development of storage and computing performance further exacerbates this problem. Although the performance of processors and memory has been rapidly improved under the promotion of "Moore's Law", the improvement of storage performance is slow, which makes storage performance a key factor restricting the development of IT.

[0003] Related technologies use Serial Advanced Technology Attachment Solid State Drives (SATA SSDs) or Serial Attached SCSI Solid State Drives (SAS SSDs) to replace ordinary hard disks, which improves business performance to a certain extent. Among them, solid state drives (SSDs) based on the Non-Volatile Memory express (NVMe) standard have become a new direction to solve the bottleneck of server storage performance. However, in the field of performance evaluation of all-NVME storage server architectures, NVMe SSDs still have many deficiencies. Currently, the Flexible I / O (fio) benchmark tool is mainly used to evaluate the performance of all-NVMe storage servers, which only focuses on the read and write performance of the NVMe SSD itself and does not consider whether there are performance bottlenecks at the hardware architecture level. Summary of the Invention

[0004] This application provides a performance evaluation method, device, equipment, medium and product for a storage server to at least solve the problem of single-dimensional performance evaluation of traditional storage servers in related technologies.

[0005] The present application provides a method for evaluating the performance of a storage server, including: obtaining the performance deviation of a hard disk to be tested of the storage server, where the performance deviation is the deviation between the measured performance data of the hard disk to be tested and the benchmark performance data; in the case where the deviation is greater than the deviation threshold, determining a performance prediction result according to the current monitoring data of the storage server and a pre-trained performance prediction model, where the performance prediction result is used to characterize the probability of a performance deviation of the storage server; constructing a performance deviation feature library based on the performance prediction result; and generating an optimization suggestion according to the performance deviation feature library and a pre-trained performance optimization model.

[0006] The present application also provides a device for evaluating the performance of a storage server, including: an acquisition module, configured to obtain the performance deviation of a hard disk to be tested of the storage server, where the performance deviation is the deviation between the measured performance data of the hard disk to be tested and the benchmark performance data; a performance prediction module, configured to, in the case where the deviation is greater than the deviation threshold, determine a performance prediction result according to the current monitoring data of the storage server and a pre-trained performance prediction model, where the performance prediction result is used to characterize the probability of a performance deviation of the storage server; a feature library construction module, configured to construct a performance deviation feature library based on the performance prediction result; a performance optimization module, configured to generate an optimization suggestion according to the performance deviation feature library and a pre-trained performance optimization model.

[0007] The present application also provides an electronic device, including: a memory, configured to store a computer program; and a processor, configured to implement the steps of any of the above methods for evaluating the performance of a storage server when executing the computer program.

[0008] The present application also provides a computer-readable storage medium, where a computer program is stored in the computer-readable storage medium, and where the computer program, when executed by a processor, implements the steps of any of the above methods for evaluating the performance of a storage server.

[0009] The present application also provides a computer program product, including a computer program, where the computer program, when executed by a processor, implements the steps of any of the above methods for evaluating the performance of a storage server.

[0010] This application analyzes the deviation between the measured performance data and the benchmark performance data of the hard disk to be tested. When the deviation is greater than the deviation threshold, it fuses the current monitoring data reflecting the real-time power consumption status of each hardware component of the storage server, and through a pre-trained performance prediction model, obtains the probability of performance deviation of the storage server to capture architecture-level implicit associations; constructs a performance deviation feature library in the dimension of the hardware architecture, and based on this feature library, generates suggestions for the hardware architecture through a pre-trained performance optimization model. Therefore, it can solve the problem that related evaluation methods ignore the hardware architecture bottleneck and provide systematic technical support for the performance optimization of all-NVME storage servers. Brief Description of the Drawings

[0011] To more clearly illustrate the embodiments of this application, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of this application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0012] Figure 1 It is a schematic diagram of the architecture of the server test end provided by the embodiment of this application; Figure 2 It is a schematic flowchart of a performance evaluation method for a storage server provided by the embodiment of this application; Figure 3 It is a schematic diagram of the structure of a performance evaluation device for a storage server provided by the embodiment of this application; Figure 4 It is a schematic diagram of the structure of an electronic device provided by the embodiment of this application. Detailed Embodiments

[0013] The following will clearly and completely describe the technical solutions in the embodiments of this application with reference to the drawings in the embodiments of this application. Obviously, the described embodiments are only some embodiments of this application, rather than all embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the protection scope of this application.

[0014] It should be noted that in the description of this application, the terms "include", "comprise" or any other variant thereof are intended to cover a non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. The terms "first", "second", etc. in this application are used to distinguish similar objects, rather than to describe a specific order or sequence.

[0015] To more clearly illustrate the embodiments of the present application, the following will briefly introduce the technical terms required in the embodiments: The Non-volatile Memory Host Controller Interface Specification / Non-volatile Storage Standard (NVMe) is a host controller interface and storage protocol designed specifically for high-speed solid-state drives. It utilizes the low latency and high bandwidth characteristics of the Peripheral Component Interconnect Express (PCIe) high-speed serial expansion bus to significantly improve data read and write speeds. At the same time, it supports multi-queue parallel processing to meet the storage efficiency requirements of high-performance computing, data centers, and other scenarios.

[0016] A Solid State Drive (SSD) is a non-volatile storage device based on flash chips or dynamic random access memory. It does not require mechanical components and realizes data read and write through electronic signals.

[0017] The Flexible I / O (fio) is an open-source storage performance testing tool used to test the read and write performance of devices such as disks and SSDs.

[0018] Input / Output Operations Per Second (IOPS) is used to measure the ability of a storage device to process I / O requests.

[0019] On-Line Transaction Processing (OLTP) is a database application mode that supports high-frequency and short transaction operations.

[0020] To enable those skilled in the art of the present technology to better understand the solution of the present application, the following will further elaborate on the present application in conjunction with the accompanying drawings and specific implementation manners.

[0021] In combination with the specific application environment architecture or specific hardware architecture on which the execution of the performance evaluation method of the storage server depends, the specific application environment architecture or specific hardware architecture will be described herein.

[0022] The specific application environment architecture on which the execution of the performance evaluation method of the storage server depends includes: a system monitoring and data collection layer, an algorithm running and model deployment layer, and an application interaction and management layer.

[0023] The system monitoring and data acquisition layer includes: a real-time monitoring system, a reference data management module, and a data transmission protocol. The real-time monitoring system needs to be deployed in the storage server operating system or the embedded management system to continuously collect the measured performance data of the hard disk and the current monitoring data. The reference data management module stores the reference performance data of different hard disk models and supports dynamic updating of the reference library to adapt to new model hardware. The real-time transmission and parsing of hardware status data are achieved through the data transmission protocol to ensure the accuracy and timeliness of data acquisition.

[0024] The algorithm operation and model deployment layer includes a machine learning framework, a distributed computing environment, and data storage and caching. The machine learning framework runs pre-trained performance prediction models and optimization models, and supports Graphics Processing Unit (GPU) acceleration to improve the computing efficiency of complex algorithms. If the scale of the storage server cluster is large, it is necessary to rely on a distributed architecture to achieve multi-node data parallel processing and shorten the performance evaluation cycle. A time-series database is used to store historical performance data and current monitoring data, and is combined with a Remote Dictionary Server (Redis) to cache frequently accessed model parameters and intermediate calculation results.

[0025] The application interaction and management layer includes a management console and an Application Programming Interface (API). The management console provides a visual interface to display the performance deviation analysis results, performance prediction probabilities, and optimization suggestions, and supports administrators to manually trigger evaluations or set timed evaluation strategies. An open data interface is provided for upper-layer business systems to call to achieve the linkage between performance evaluation and business processes.

[0026] The specific hardware architecture on which the execution of the performance evaluation method of the storage server depends includes: core computing components, storage and data acquisition hardware, and network and expansion interfaces.

[0027] The core computing components include a central processing unit, a GPU acceleration card, and a memory configuration. The central processing unit needs to have multiple cores and a high main frequency to support parallel processing of a large amount of performance data and model inference calculations. A GPU card is installed to accelerate matrix operations and model training. The memory configuration ensures the cache space during large-data-volume calculations and avoids calculation interruptions caused by insufficient memory.

[0028] The storage and data acquisition hardware includes the hard disk array to be tested, a current monitoring module, and a hardware monitoring chip. The hard disk array to be tested supports the mixed insertion of hard disks of multiple interface types, and is configured with a Redundant Array of Independent Disks (RAID) controller to achieve unified management of hard disk performance and data redundancy. The current monitoring module is built with a high-precision current sensor and an Analog-to-Digital Converter (ADC) sampling chip, and is deployed on the hard disk backplane or power module to collect the working current data of each hard disk in real time. The hardware monitoring chip obtains the server hardware status through management chips such as the Baseboard Management Controller, and assists in analyzing the associated factors of performance deviation.

[0029] The network and expansion interface includes a high-speed network interface and a PCIe expansion slot. An Ethernet interface is configured to support high-speed transmission of performance data and data synchronization between clusters, meeting the centralized evaluation requirements of large-scale storage nodes. A PCIe expansion slot is reserved for installing a GPU acceleration card, a high-speed memory card, or a dedicated data acquisition card to improve the system scalability.

[0030] The above performance evaluation method realizes a full-process closed loop from hardware data acquisition, model analysis to optimization suggestions through a software-hardware collaborative architecture design. Its effectiveness highly depends on the accuracy of the monitoring hardware, the computing power of the computing resources, and the compatibility of the system architecture.

[0031] As Figure 1 shown, it is a schematic diagram of the architecture of the server test end of the performance evaluation method for a storage server. In some embodiments, the collection module 101 is used to obtain the hardware information of the hard disk to be tested. The processing module 102 is used to group the hard disks to be tested according to the hardware information, determine the hard disk group to which the hard disk to be tested belongs, and then randomly select a marked hard disk from the hard disk group; the execution module 103 is used to perform a benchmark value test on the marked hard disk to obtain benchmark performance data, and then perform a distributed concurrent load test on all the hard disks to be tested to obtain measured performance data. Subsequently, the deviation between the measured performance data and the benchmark performance data is calculated. When the deviation is greater than the deviation threshold, according to the current monitoring data of the storage server and a pre-trained performance prediction model, a performance prediction result is determined, and performance deviation analysis is performed. The monitoring module 104 is used to obtain the current monitoring data. The processing module 102 is also used to construct a performance deviation feature library based on the performance prediction result, and generate an optimization suggestion according to the performance deviation feature library and a pre-trained performance optimization model.

[0032] Embodiments of the present application provide a performance evaluation method for a storage server. In combination with the execution process of the performance evaluation method for the storage server, the method is described in detail.

[0033] As Figure 2As shown in the figure, a method for evaluating the performance of a storage server provided by an embodiment of the present application includes the following steps S201 to S204: S201. Obtain the performance deviation of the hard disk to be tested of the storage server.

[0034] Among them, the storage server is an NVMe storage server, and the hard disk to be tested is an NVMe SSD. The performance deviation is the deviation between the measured performance data and the reference performance data of the hard disk to be tested. The measured performance data is obtained by performing a distributed concurrent load test on the hard disk to be tested, and the reference performance data is obtained by performing a reference value test on some marked hard disks in the hard disk to be tested.

[0035] The distributed concurrent load test is a test method that simulates requests from multiple users or multiple tasks to the hard disk in a real business in a distributed hard disk environment to verify the performance, stability, and fault tolerance of the hard disk in a high-concurrency scenario. The distributed concurrent load test includes, but is not limited to, sequential read and write and random mixed load. Sequential read and write is to linearly read and write continuous large chunks of data, and random mixed load is to randomly combine and read and write small chunks of data. The reference value test is a standardized test method for measuring the hard disk data transfer rate. Its core is to quantify the amount of data that the hard disk can process per unit time by simulating a specific workload to evaluate the performance of the hard disk. The reference performance data includes, but is not limited to, memory bandwidth, cross-node bandwidth, IOPS, throughput, and latency. Exemplarily, the reference value test can use test tools such as fio to perform a full-link test under steady state.

[0036] In some embodiments, before performing step S201, first select marked hard disks from the hard disks to be tested, including obtaining the hardware information of the hard disks to be tested, then grouping the hard disks to be tested according to the hardware information, determining the hard disk group to which the hard disks to be tested belong, and then randomly selecting marked hard disks from the hard disk groups.

[0037] Among them, the hardware information of the hard disks to be tested includes at least one of, but is not limited to, hard disk model, hard disk bandwidth information, and hard disk switch information. Hard disk groups are divided according to this hardware information, and different hardware information corresponds to different hard disk groups. Specifically, when the hard disk models, hard disk bandwidth information, and hard disk switch information of any two hard disks to be tested are the same, it means that the hardware attributes of these two hard disks to be tested are the same, and it is determined that these two hard disks to be tested belong to the same hard disk group.

[0038] After grouping the hard disks to be tested, randomly select at least one hard disk from different hard disk groups and mark a test label, indicating that the hard disk is a marked hard disk.

[0039] The above embodiments achieve the automatic classification of hard disks to be tested through standardized grouping, randomly select marked hard disks for benchmark value testing, avoid testing all hard disks to be tested one by one, reduce the test volume of benchmark value testing to the order of the number of groups, improve the efficiency of benchmark value testing, and ensure the representativeness of test results.

[0040] Based on the above embodiments, after randomly selecting marked hard disks from the hard disk grouping, first perform benchmark value testing on the marked hard disks to obtain benchmark performance data, and then perform a distributed concurrent load test on all hard disks to be tested to obtain measured performance data. Subsequently, calculate the deviation between the measured performance data and the benchmark performance data.

[0041] The above embodiments obtain benchmark performance data by performing benchmark value testing on marked hard disks, which is used as a comparison standard for subsequent testing. Perform a distributed concurrent load test on hard disks to be tested to verify the performance of hard disks to be tested under the actual business load model, avoid the deviation between single-point testing and the real scenario, and at the same time expose possible bottlenecks of hard disks to be tested in a distributed environment. Calculate the deviation between the measured performance data and the benchmark performance data to locate hard disks with hardware failures, locate configuration problems, and topological bottlenecks.

[0042] Optionally, after performing benchmark value testing on the marked hard disks to obtain benchmark performance data, obtain at least one of the read / write ratio (R / W Ratio), access mode, data block size distribution (Block Size Distribution), and queue depth corresponding to the service, and construct a test template for the distributed concurrent load test, which is used to simulate the business load. The access mode is the random / sequential access ratio. Adjust the performance metrics according to the business scenario, and the performance metrics include but are not limited to throughput, latency, IOPS, Quality of Service (QoS). After waiting for the benchmark value testing to complete, perform a distributed concurrent load test on all hard disks to be tested according to the test template to generate measured performance data.

[0043] Exemplarily, for On-Line Transaction Processing (OLTP) type services, the proportion of random access is 70%, the data block size is 4K - 64K, and the queue depth is 256. A test template for the distributed concurrent load test is constructed by combining parameters to simulate the high-frequency and fragmented I / O requirements of database transactions. Correspondingly, performing a distributed concurrent load test on the hard disks to be tested is to select the OLTP template from the template library according to the read / write ratio, access mode, and data block size distribution of the actual business, and test all hard disks to be tested through a multi-node concurrent method, covering scenarios such as sequential read / write and random mixing, simulating multi-client concurrent access in the real business, recording metrics such as throughput, latency, and IOPS during the test, and generating measured performance data.

[0044] In the above embodiments, by constructing a test template for distributed concurrent load testing, the deviation between traditional benchmark tests (such as pure sequential or pure random) and real business is avoided. Through the mapping of business characteristics, test parameters, and performance indicators, the performance bottleneck of the disk under test in the actual business can be more accurately discovered, providing data support for the optimization of the storage system.

[0045] After calculating the deviation between the measured performance data and the benchmark performance data, the deviation is compared with a deviation threshold. Optionally, there is a default value for the deviation threshold (such as ±10%), which can be dynamically adjusted based on the performance prediction results output by the subsequent performance prediction model.

[0046] In some embodiments, first, working condition parameters are obtained. The working condition parameters include, but are not limited to, historical performance data, environmental parameters of the storage server, and business load characteristics. The environmental parameters include, but are not limited to, temperature and humidity, and the business load characteristics include, but are not limited to, read / write request frequency and data size distribution. Then, based on the working condition parameters and combined with each performance prediction result, the non-linear relationship between these data and the deviation threshold is determined. When the performance prediction result output by the performance prediction model indicates a deviation, the Gradient Boosting Decision Tree (GBDT) algorithm is used to learn the non-linear relationship between the working condition parameters and the deviation threshold, and the deviation threshold is recalculated based on this non-linear relationship. This enables the deviation threshold to adapt to the performance fluctuations of the storage server under different working conditions.

[0047] S202. In the case where the deviation is greater than the deviation threshold, based on the current monitoring data of the storage server and the pre-trained performance prediction model, determine the performance prediction result.

[0048] Among them, the current monitoring data includes, but is not limited to, NVMe current and Process Identifier (PID) current information. The current monitoring data can reflect the real-time power consumption status of each hardware component of the storage server.

[0049] The pre-trained performance prediction model is constructed based on the Long Short-Term Memory (LSTM) network with the Support Vector Regression (SVR) algorithm. It is obtained by jointly training the sample current data and the sample deviation characteristics. By combining the current data with the architecture topology, the performance evaluation is extended from single-point testing to full-link bottleneck location.

[0050] The performance prediction result is used to characterize the probability of performance deviation of the storage server. The performance prediction result includes memory bandwidth, node bandwidth, number of CPU cores, CPU operating frequency, and node distribution. If there is a bottleneck in memory bandwidth, node bandwidth, number of CPU cores, CPU operating frequency, or unreasonable node distribution, any of the above situations indicates a performance deviation of the storage server.

[0051] In some embodiments, the current monitoring data of the storage server to be monitored is monitored, and the current monitoring data is normalized. Optionally, the current monitoring data is normalized according to the following formula (1): X is normalized as follows: (1) In formula (1), X norm is the current monitoring data after normalization, X is the original current monitoring data, X min is the minimum value in the current monitoring data, X max is the maximum value in the current monitoring data.

[0052] When the deviation is greater than the deviation threshold, it indicates abnormal disk performance, and then the performance of the storage server is evaluated. The current monitoring data of the storage server is input into a pre-trained performance prediction model, and the performance prediction model performs performance prediction to obtain a performance prediction result.

[0053] In the case where the deviation in this application is greater than the deviation threshold, in addition to using the current monitoring data to perform performance prediction on the disk to be tested, an alarm can also be triggered, and the hardware performance indicators of the storage server can be obtained to analyze the reasons for the possible performance deviation.

[0054] In some embodiments, when the deviation is greater than the deviation threshold, the hardware performance metrics of the storage server are obtained, and performance deviation analysis is performed on the storage server based on the hardware performance metrics. Among them, the hardware performance metrics include, but are not limited to: Central Processing Unit (CPU) operating frequency, CPU utilization rate, memory bandwidth, memory usage capacity, CPU - to - CPU communication bandwidth, Uncore frequency of the processor, Intel Input / Output (IIO) bandwidth, NVME write data volume, NVME Garbage Collection (GC) time window, IO bandwidth information, NVME operating temperature, PID process information. The CPU - to - CPU communication bandwidth includes Ultra Path Interconnect (UPI) or Socket - to - Socket Global Memory Interface (XGMI). The CPU operating frequency, CPU utilization rate, memory bandwidth, and memory usage capacity are used to locate system resource bottlenecks; the CPU - to - CPU communication bandwidth, Uncore frequency, and IIO bandwidth are used to analyze topology bottlenecks; the NVME write data volume, NVME GC time window, IO bandwidth information, and NVME operating temperature are used to diagnose hardware health; the PID process information is used to locate the impact of abnormal processes on performance.

[0055] Through the multi - dimensional hardware performance metrics, the above - mentioned embodiments jointly analyze the performance deviation of the storage server, can eliminate misjudgment of a single metric, identify hardware link bottlenecks, and discover potential risks in advance.

[0056] According to the comparison result between the deviation of the hard disk to be tested and the deviation threshold, the current values of the hard disk to be tested corresponding to different comparison results are obtained respectively. It can be understood that the current values of the hard disks with deviations greater than the deviation threshold are obtained respectively, and the current values of the hard disks with deviations less than or equal to the deviation threshold are obtained, and then according to the magnitude of the current values, it is determined whether it is necessary to detect the hardware link of the hard disk. If it is necessary, detection is performed. When it is detected that the hardware link is abnormal, repair suggestions for the hardware link are output. If it is not necessary to perform hardware link detection, it is detected whether the hard disk is in an over - temperature state and / or a garbage collection state. When it is detected that the hard disk is in an over - temperature state, heat dissipation adjustment suggestions for the hard disk are output; when it is detected that the hard disk is in a garbage collection state, a wait instruction is output to indicate waiting for the hard disk to complete the garbage collection task, and testing is performed after exiting the garbage collection state.

[0057] For ease of description, the following will take any two different hard disks in the hard disks to be tested, the first hard disk and the second hard disk as examples to introduce the performance deviation analysis process.

[0058] Obtain the first current value I of the first hard disk Abnormal and the second current value I of the second hard disk Normal , the deviation of the first hard disk is greater than the deviation threshold, and the deviation of the second hard disk is less than or equal to the deviation threshold. If the first current value I Abnormal is greater than or equal to the second current value I Normal , it indicates that there may be a problem with the hardware link of the first hard disk body, then perform a hardware link detection. Specifically, detect the PCIE bandwidth, Uncore frequency, and IIO bandwidth of the first hard disk. If any information has a problem, for example, PCIE 4.0 is reduced to PCIE 3.0, it is determined that there is a hardware link abnormality in the first hard disk, and corresponding hardware link repair suggestions are output.

[0059] Exemplarily, a decrease in the Uncore frequency or insufficient IO bus bandwidth will lead to a decrease in the communication efficiency between the CPU and the NVMe, increase the device load current, and then output hardware link repair suggestions to avoid ineffective troubleshooting and shorten the fault location time.

[0060] Through the numerical comparison of the first current value and the second current value in the above embodiments, the current abnormality is converted into a quantifiable judgment condition, avoiding the subjectivity of manual judgment and improving the accuracy of fault identification; when the current is abnormal, check the PCIE bandwidth, Uncore frequency, and IIO bandwidth of the first hard disk, reduce the cost of manual troubleshooting for each device, improve the stability of the NVMe storage system from the hardware bottom layer, and ensure business continuity.

[0061] In the case of no abnormality in the hardware link detection, detect whether the first hard disk is in an over-temperature state and / or a garbage collection state. If the first hard disk is in an over-temperature state, output the heat dissipation adjustment suggestion for the first hard disk. If the first hard disk is in the garbage collection (Garbage Collection, GC) state, output a wait instruction to indicate waiting for the first hard disk to complete the garbage collection task before performing the test. Specifically, according to the NVMe write data volume, NVMe over-temperature, and NVMe GC time window, detect whether the first hard disk is in an over-temperature state and / or a garbage collection state.

[0062] If there is no abnormality in the hardware link detection, external factors such as the motherboard, Switch chip, and cable can be excluded, and the fault can be located to the NVMe hard disk itself. When the temperature is too high, the controller may trigger a frequency reduction protection mechanism, resulting in performance fluctuations and current abnormalities. At this time, the performance can be quickly restored through heat dissipation adjustment. During the GC process, the NVMe hard disk will clean up invalid data in the background, consuming additional computing power and power consumption, which may cause a short-term increase in current. If GC occurs frequently or runs for a long time and the current abnormality persists, prompting to wait for GC to end can avoid misjudgment.

[0063] The above embodiment focuses on the internal problems of the hard disk on the basis of eliminating hardware link problems. When internal abnormalities such as overtemperature or GC are detected, targeted solutions are directly output without manual intervention and analysis, thereby improving the real-time performance of fault handling.

[0064] If the first current value is less than the second current value, it indicates that the test script needs to be analyzed, especially the number of threads, queue depth, etc. Specifically, it is determined whether the CPU utilization of the storage server is greater than or equal to a preset utilization, such as 90%. If so, it indicates that the CPU operating frequency does not meet the standard, and a suggestion for adjusting the CPU performance mode is output, or a suggestion for reducing parallel threads is output.

[0065] For example, a suggestion of enabling high-performance mode and removing power consumption restrictions is output, or a suggestion of reducing IO concurrency is output.

[0066] The above embodiment locates the upstream CPU problem through the current anomaly, avoiding the waste of NVMe performance. In addition, it ensures that the NVMe benchmark test data is authentic and reliable, providing an accurate basis for model selection and performance tuning.

[0067] If the utilization rate of the central processing unit is less than the preset utilization rate, the intra-node bandwidth of the node to which the first hard disk belongs is analyzed. Specifically, the node to which the first hard disk belongs is first determined, and then it is determined whether the sum of the hard disk bandwidths of the node is less than or equal to the intra-node bandwidth. Here, the internal bandwidth of a node is compared with the sum of the bandwidths of the hard disks belonging to the node. If not, it means that the intra-node bandwidth is insufficient, and a bandwidth adjustment suggestion is output.

[0068] Optionally, calculate the node memory bandwidth according to the following formula (2): BW numa : (2) In formula (2), bw men is the total memory bandwidth, N numa is the number of nodes.

[0069] The above embodiments focus on the implicit shortage of bandwidth resources within NUMA, and compare whether the bandwidth is sufficient by quantitative numerical value, thereby shortening the operation and maintenance troubleshooting time, which is conducive to preventive identification and systematic optimization of hardware resource bottlenecks.

[0070] When the sum of the hard disk bandwidths is less than or equal to the bandwidth within the node, it indicates that the bandwidth within the node meets the bandwidth requirements of all the hard disks within the node. Then, it is determined whether the first hard disk corresponds to the central processing unit (CPU) core of the remote node. Optionally, obtain the running status information of the hard disk to be tested, including but not limited to: PID process information, the number of CPU cores used by the PID, CPU NUMA distribution, and NVMEM NUMA affiliation information. Based on the running status information, the affiliation mapping relationship between each hard disk to be tested and the central processing unit can be generated. Determine whether the first hard disk corresponds to the CPU core of the remote node according to the affiliation mapping relationship. The affiliation mapping relationship reflects the serial number of the CPU core used by each hard disk to be tested and the node to which the CPU core belongs. The node to which the CPU core belongs includes the local node and the remote node. If the CPU core used by the hard disk belongs to the local node, it is determined that the test of the hard disk by the island is a local process. If the CPU core used by the hard disk belongs to the remote node, it is determined that the test of the hard disk is a remote process.

[0071] Through the correlation analysis of the running status information of the hard disk to be tested in the above embodiments, the affiliation mapping relationship is generated, which can present the physical path of the I / O operation. Through the affiliation mapping relationship, it can be quickly determined whether the insufficient bandwidth is caused by remote processing, avoiding blind troubleshooting of hardware failures.

[0072] When the first hard disk corresponds to the CPU core of the remote node, it indicates that there is cross-node processing in the storage server. Then, it is determined whether the inter-node bandwidth of the storage server is greater than or equal to the preset bandwidth. The inter-node bandwidth can be the UPI bandwidth or the XGMI bandwidth. The preset bandwidth can be 90% of the CPU bandwidth. If the inter-node bandwidth is greater than or equal to the preset bandwidth, it indicates that there is a bottleneck in the inter-node bandwidth. Then, output suggestions for adjusting the inter-node bandwidth, including replacing the CPU with a higher inter-node bandwidth, or adjusting the node distribution, or adjusting the number of hard disks within the node.

[0073] When cross-node processing occurs during the read and write of the first hard disk, monitor whether the inter-node bandwidth of the storage server is greater than or equal to the preset bandwidth to accurately locate the saturation state of inter-node communication and avoid performance failures caused by hidden bandwidth bottlenecks.

[0074] Optionally, calculate the number of hard disks that the inter-node bandwidth can support according to the following formula (3) N NVMe : (3) In formula (3), N UPI is the number of inter-node channels, BW UPI is the inter-node bandwidth, BW NVMeis the hard disk bandwidth baseline value. According to the calculated number of hard disks, adjust the number of hard disks within the node to be less than or equal to this number of hard disks.

[0075] The above embodiments quantitatively calculate the relationship between the bandwidth between nodes and the number of hard disks, predict the impact of the number of hard disks on the bandwidth between nodes, avoid bandwidth exhaustion caused by over-deployment, are beneficial to solving performance problems caused by insufficient cross-node bandwidth, and also provide a quantifiable technical standard for server hardware selection and storage resource allocation.

[0076] S203. Construct a performance deviation feature library based on the performance prediction results.

[0077] In some embodiments, during the process of constructing the performance deviation feature library, first perform feature extraction on the performance prediction results to obtain performance deviation features, and calculate the correlation strength between the performance deviation features; at the same time, establish a semantic association relationship between the performance deviation features and the abnormal state parameters. Then construct the performance deviation feature library based on the correlation strength and the semantic association relationship.

[0078] Among them, the performance deviation features include but are not limited to current anomaly features, time domain features, and frequency domain features. The abnormal state parameters include the fault type and the reason for performance degradation.

[0079] Optionally, perform semantic understanding and feature extraction on the performance test results through Natural Language Processing (NLP) and Graph Neural Network (GNN) to obtain performance deviation features to automatically expand the feature library; use GNN to calculate the correlation strength between performance deviations and analyze the correlation relationship between different performance deviation features. Use knowledge graph technology to establish a semantic association relationship between the performance deviation features and the abnormal state parameters. Then construct the performance deviation feature library according to the correlation strength and the semantic association relationship.

[0080] The above embodiments realize the correlation analysis of cross-dimensional features through knowledge graph modeling, and analyze the hidden associations between features through graph neural networks, which is beneficial to discovering composite fault modes that cannot be covered by traditional rules.

[0081] S204. Generate optimization suggestions according to the performance deviation feature library and the pre-trained performance optimization model.

[0082] Among them, the pre-trained performance optimization model is constructed based on a generative adversarial network or a Transformer model. The performance feature deviations in the performance feature deviation library are input into the pre-trained performance optimization model to obtain optimization suggestions. The optimization suggestions include hardware maintenance measures, software configuration adjustments, and business scheduling strategies, etc. Hardware maintenance measures such as replacing faulty components, software configuration adjustments such as optimizing system parameters, business scheduling strategies such as adjusting load distribution, etc., and the present application does not specifically limit this. The optimization suggestions can be adjusted into natural language form and presented to the user.

[0083] Exemplarily, the optimization suggestions include: if the bottleneck is in the PCIe3.0 bus, it is recommended to migrate the SSD to the PCIe4.0 slot, or split multiple SSDs to different bus channels; if the CPU interrupt handling latency is too high, it is recommended to bind the SSD interrupt to a specific CPU core.

[0084] In some embodiments, the performance optimization model converges based on sample data training. The sample training data includes historical deviation features and historical solutions, and there is a mapping relationship between the historical deviation features and the historical solutions. The performance optimization model learns this mapping relationship to generate a three-dimensional optimization solution including hardware, software, and business scheduling.

[0085] The above embodiments use a pre-trained performance optimization model to automatically generate optimization suggestions by capturing the temporal correlation and implicit association of deviation features through deep learning, realizing the high efficiency, accuracy, and adaptability upgrade of performance optimization.

[0086] In summary, the present application analyzes the deviation between the measured performance data and the benchmark performance data of the hard disk to be tested. When it is greater than the deviation threshold, it fuses the current monitoring data reflecting the real-time power consumption status of each hardware component of the storage server, and through a pre-trained performance prediction model, obtains the probability of performance deviation of the storage server to capture the architecture-level implicit association; constructs a performance deviation feature library in the dimension of the hardware architecture, and based on this feature library, generates suggestions for the hardware architecture through a pre-trained performance optimization model. It can solve the problem that related evaluation methods ignore hardware architecture bottlenecks and provide systematic technical support for the performance optimization of all-NVME storage servers.

[0087] Through the description of the above embodiments, those skilled in the art can clearly understand that the method according to the above embodiments can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware, but in many cases, the former is a better implementation method.

[0088] The embodiments of the present application also provide a performance evaluation device for a storage server, as Figure 3 shown, the device includes: An acquisition module 301, configured to acquire a performance deviation of a hard disk to be tested of a storage server, where the performance deviation is a deviation between measured performance data of the hard disk to be tested and reference performance data; A performance prediction module 302, configured to, when the deviation is greater than a deviation threshold, determine a performance prediction result according to current monitoring data of the storage server and a pre-trained performance prediction model, where the performance prediction result is used to characterize the probability that the performance of the storage server has a deviation; A feature library construction module 303, configured to construct a performance deviation feature library based on the performance prediction result; A performance optimization module 304, configured to generate an optimization suggestion according to the performance deviation feature library and a pre-trained performance optimization model.

[0089] As an optional implementation manner provided in an embodiment of the present application, the acquisition module 301 is specifically configured to: select a marked hard disk from the hard disks to be tested; perform a reference value test on the marked hard disk to obtain reference performance data; perform a distributed concurrent load test on the hard disks to be tested to obtain measured performance data; calculate a deviation between the measured performance data and the reference performance data.

[0090] As an optional implementation manner provided in an embodiment of the present application, the acquisition module 301 is specifically configured to: acquire hardware information of the hard disk to be tested; determine a hard disk group to which the hard disk to be tested belongs according to the hardware information; randomly select a marked hard disk from the hard disk group.

[0091] As an optional implementation manner provided in an embodiment of the present application, the feature library construction module 303 is specifically configured to: perform feature extraction on the performance prediction result to obtain performance deviation features; calculate an association strength between the performance deviation features; establish a semantic association relationship between the performance deviation features and abnormal state parameters; construct a performance deviation feature library based on the association strength and the semantic association relationship.

[0092] As an optional implementation manner provided in an embodiment of the present application, the apparatus further includes a deviation analysis module, configured to: when the deviation is greater than a deviation threshold, acquire a hardware performance index of the storage server; perform performance deviation analysis on the storage server according to the hardware performance index.

[0093] As an optional implementation manner provided in an embodiment of the present application, the apparatus further includes a deviation analysis module, configured to: when the deviation of a first hard disk is greater than a deviation threshold, acquire a first current value of the first hard disk; when the deviation of a second hard disk is less than or equal to the deviation threshold, acquire a second current value of the second hard disk; where the first hard disk and the second hard disk are any two different hard disks among the hard disks to be tested; if the first current value is greater than or equal to the second current value, detect whether a hardware link of the first hard disk is abnormal; if so, output a hardware link repair suggestion for the first hard disk.

[0094] As an alternative implementation provided by the embodiments of the present application, the device further includes a deviation analysis module, configured to: if the hardware link of the first hard disk is normal, detect whether the first hard disk is in an over-temperature state and / or a garbage collection state; in the case where the first hard disk is in an over-temperature state, output a heat dissipation adjustment suggestion for the first hard disk; in the case where the first hard disk is in a garbage collection state, output a waiting instruction, and the waiting instruction is used to indicate waiting for the first hard disk to complete the garbage collection task before performing a test.

[0095] As an alternative implementation provided by the embodiments of the present application, the device further includes a deviation analysis module, configured to: if the first current value is less than the second current value, determine whether the utilization rate of the central processing unit of the storage server is greater than or equal to a preset utilization rate; if so, output a suggestion for adjusting the performance mode of the central processing unit or a suggestion for reducing parallel threads.

[0096] As an alternative implementation provided by the embodiments of the present application, the device further includes a deviation analysis module, configured to: if the utilization rate of the central processing unit of the storage server is less than the preset utilization rate, determine the node to which the first hard disk belongs; determine whether the sum of the hard disk bandwidths of the node to which the first hard disk belongs is less than or equal to the intra-node bandwidth of the node to which the first hard disk belongs; if not, output a bandwidth adjustment suggestion.

[0097] As an alternative implementation provided by the embodiments of the present application, the device further includes a deviation analysis module, configured to: if the sum of the hard disk bandwidths is less than or equal to the intra-node bandwidth, determine whether the first hard disk corresponds to a central processing unit core of a remote node; if so, determine whether the inter-node bandwidth of the storage server is greater than or equal to a preset bandwidth; in the case where the inter-node bandwidth is greater than or equal to the preset bandwidth, output a suggestion for adjusting the inter-node bandwidth.

[0098] As an alternative implementation provided by the embodiments of the present application, the deviation analysis module is specifically configured to: obtain the running state information of the hard disk to be tested; generate an attribution mapping relationship between the node to which the hard disk to be tested belongs and the central processing unit according to the running state information; and determine whether the first hard disk corresponds to a central processing unit core of a remote node according to the attribution mapping relationship.

[0099] For the description of the features in the corresponding embodiments of the performance evaluation device of the storage server, reference may be made to the relevant description in the corresponding embodiments of the performance evaluation method of the storage server, which will not be elaborated here one by one.

[0100] The embodiments of the present application further provide an electronic device, as Figure 4 shown, including a memory 401 and a processor 402. The memory 401 stores a computer program, and the processor 402 is configured to run the computer program to execute the steps in any of the above embodiments of the performance evaluation method of the storage server.

[0101] Embodiments of the present application also provide a computer-readable storage medium storing a computer program, where the computer program is configured to execute the steps in any of the above-described embodiments of the performance evaluation method for a storage server when running.

[0102] In an exemplary embodiment, the above computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs such as USB flash drives, read-only memory (ROM for short), random access memory (RAM for short), external hard drives, magnetic disks, or optical discs.

[0103] Embodiments of the present application also provide a computer program product. The above computer program product includes a computer program, and when the computer program is executed by a processor, it implements the steps in any of the above-described embodiments of the performance evaluation method for a storage server.

[0104] Embodiments of the present application also provide another computer program product, including a non-volatile computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, it implements the steps in any of the above-described embodiments of the performance evaluation method for a storage server.

[0105] Those skilled in the art can further realize that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. To clearly illustrate the interchangeability of hardware and software, the components and steps of the examples have been generally described according to their functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

[0106] The above has introduced in detail a performance evaluation method, device, equipment, medium, and product for a storage server provided by the present application. Specific examples are used in this article to elaborate on the principle and implementation manner of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application. It should be noted that for those of ordinary skill in the art in this technical field, without departing from the principle of the present application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the protection scope of the claims of the present application.

Claims

1. A performance evaluation method for a storage server, characterized in that, Including: Obtain the performance deviation of the hard disk to be tested of the storage server, where the performance deviation is the deviation between the measured performance data and the benchmark performance data of the hard disk to be tested; When the deviation is greater than the deviation threshold, determine a performance prediction result according to the current monitoring data of the storage server and a pre-trained performance prediction model, where the performance prediction result is used to characterize the probability of performance deviation of the storage server; Construct a performance deviation feature library based on the performance prediction result; Generate an optimization suggestion according to the performance deviation feature library and a pre-trained performance optimization model.

2. The method according to claim 1, wherein The obtaining the performance deviation of the hard disk to be tested of the storage server includes: Select a marked hard disk from the hard disks to be tested; Perform a benchmark value test on the marked hard disk to obtain the benchmark performance data; Perform a distributed concurrent load test on the hard disks to be tested to obtain the measured performance data; Calculate the deviation between the measured performance data and the benchmark performance data.

3. The method according to claim 2, characterized in that, The selecting a marked hard disk from the hard disks to be tested includes: Obtain the hardware information of the hard disks to be tested; Determine the hard disk group to which the hard disks to be tested belong according to the hardware information; Randomly select the marked hard disk from the hard disk group.

4. The method according to claim 1, wherein The constructing a performance deviation feature library based on the performance prediction result includes: Extract features from the performance prediction result to obtain performance deviation features; Calculate the association strength between the performance deviation features; Establish a semantic association relationship between the performance deviation features and abnormal state parameters; Construct the performance deviation feature library based on the association strength and the semantic association relationship.

5. The method according to claim 1, wherein The method further includes: When the deviation is greater than the deviation threshold, obtain the hardware performance index of the storage server; Perform a performance deviation analysis on the storage server according to the hardware performance index.

6. The method according to claim 1, characterized in that, The method further includes: When the deviation of the first hard disk is greater than the deviation threshold, obtain the first current value of the first hard disk; When the deviation of the second hard disk is less than or equal to the deviation threshold, obtain the second current value of the second hard disk; where the first hard disk and the second hard disk are any different hard disks among the hard disks to be tested; If the first current value is greater than or equal to the second current value, detect whether the hardware link of the first hard disk is abnormal; If so, output a hardware link repair suggestion for the first hard disk.

7. The method according to claim 6, wherein The method further includes: If the hardware link of the first hard disk is normal, detect whether the first hard disk is in an over-temperature state and / or a garbage collection state; When the first hard disk is in the over-temperature state, output a heat dissipation adjustment suggestion for the first hard disk; When the first hard disk is in the garbage collection state, output a waiting instruction, where the waiting instruction is used to indicate waiting for the first hard disk to complete the garbage collection task before testing.

8. The method according to claim 6, wherein The method further includes: If the first current value is less than the second current value, determine whether the utilization rate of the central processing unit of the storage server is greater than or equal to a preset utilization rate; If so, output a suggestion to adjust the performance mode of the central processing unit or a suggestion to reduce the number of parallel threads.

9. The method according to claim 8, characterized in that The method further includes: If the utilization rate of the central processing unit of the storage server is less than the preset utilization rate, determine the node to which the first hard disk belongs; Judge whether the sum of the hard disk bandwidths of the node to which the first hard disk belongs is less than or equal to the in-node bandwidth of the node to which the first hard disk belongs; If not, output a bandwidth adjustment suggestion.

10. The method according to claim 9, wherein The method further includes: If the sum of the hard disk bandwidths is less than or equal to the in-node bandwidth, judge whether the first hard disk corresponds to the central processing unit core of a remote node; If so, judge whether the inter-node bandwidth of the storage server is greater than or equal to the preset bandwidth; When the inter-node bandwidth is greater than or equal to the preset bandwidth, output a suggestion to adjust the inter-node bandwidth.

11. The method according to claim 10, characterized in that, The judging whether the first hard disk corresponds to the central processing unit core of a remote node includes: Obtain the running state information of the hard disk to be tested; Generate an attribution mapping relationship between the node to which the hard disk to be tested belongs and the central processing unit according to the running state information; Judge whether the first hard disk corresponds to the central processing unit core of a remote node according to the attribution mapping relationship.

12. A performance evaluation device for a storage server, characterized in that, Includes: An acquisition module, configured to acquire a performance deviation of a hard disk to be tested of a storage server, where the performance deviation is a deviation between the measured performance data of the hard disk to be tested and the reference performance data; A performance prediction module, configured to, when the deviation is greater than a deviation threshold, determine a performance prediction result according to the current monitoring data of the storage server and a pre-trained performance prediction model, where the performance prediction result is used to characterize the probability that the performance of the storage server has a deviation; A feature library construction module, configured to construct a performance deviation feature library based on the performance prediction result; A performance optimization module, configured to generate an optimization suggestion according to the performance deviation feature library and a pre-trained performance optimization model.

13. An electronic device, characterized in that, Includes: A memory, configured to store a computer program; A processor, configured to implement the steps of the performance evaluation method of the storage server according to any one of claims 1 to 11 when executing the computer program.

14. A computer-readable storage medium, characterized in that, A computer program is stored in the computer-readable storage medium, where the computer program, when executed by a processor, implements the steps of the performance evaluation method of the storage server according to any one of claims 1 to 11.

15. A computer program product, comprising a computer program, characterized in that, The computer program, when executed by a processor, implements the steps of the performance evaluation method of the storage server according to any one of claims 1 to 11.

Images