A method and system for evaluating the running efficiency of a high-performance computing cluster

Abstract

The application relates to the field of efficiency evaluation, and discloses a high-performance computing cluster operation efficiency evaluation method and system, which are used for providing support for optimization and management of the cluster. The method comprises the following steps: determining ideal state distribution of each job in the cluster based on a preset ideal state model of different job types; collecting actual resource usage data of the jobs occupying computing nodes and converting the data into a probability distribution; calculating information divergence of the probability distribution and the ideal state distribution to obtain a multi-dimensional divergence value set; generating a job comprehensive efficiency score by weighting and fusing the multi-dimensional divergence value set according to preset weights associated with the job types; and aggregating the job comprehensive efficiency score according to resource occupation ratios to obtain a cluster overall efficiency index. Key bottleneck nodes are identified based on the contribution of nodes in information divergence calculation. Historical time series databases are formed by continuously collecting data, trend analysis is performed, a report is generated, and the weight configuration or model selection strategy is dynamically adjusted according to the report, which is applied to subsequent evaluation periods. The application realizes comprehensive, accurate and dynamic management of the cluster operation efficiency.

Description

Technical Field

[0001] This invention relates to the field of efficiency evaluation, and in particular to a method and system for evaluating the operating efficiency of a high-performance computing cluster. Background Technology

[0002] In today's digital age, high-performance computing clusters have become a key infrastructure driving the development of numerous fields such as scientific research, engineering simulation, and data analysis. With the increasing complexity of scientific computing tasks and the explosive growth of data volume, extremely high demands are placed on the computing power, processing speed, and operational efficiency of high-performance computing clusters. High-performance computing clusters typically consist of a large number of computing nodes connected via a network, capable of processing large-scale computing tasks in parallel to achieve computing performance far exceeding that of a single computer.

[0003] However, high-performance computing clusters face many challenges during operation. Different types of computing jobs have different resource requirements and operating modes. Some scientific computing jobs may have extremely high requirements for the computing power of the computing cores, while other data analysis jobs may focus more on memory access speed and capacity.

[0004] The computing nodes in the cluster differ in hardware configuration and operating status. As the cluster runs for longer, nodes may experience failures or performance degradation. The allocation and management of cluster resources also face dynamic and complex issues. How to rationally allocate computing resources according to the actual needs of the job and avoid resource waste and local bottlenecks is the key to improving the overall operating efficiency of the cluster.

[0005] Therefore, we propose a method and system for evaluating the operating efficiency of high-performance computing clusters to address the aforementioned problems. Summary of the Invention

[0006] This invention provides a method and system for evaluating the operating efficiency of high-performance computing clusters, which can provide support for cluster optimization and management.

[0007] The first aspect of this invention provides a method for evaluating the operational efficiency of a high-performance computing cluster. The method includes: determining the ideal state distribution for each currently running job in the cluster based on a preset ideal state model corresponding to different job types; collecting actual resource usage data of the computing nodes occupied by each job, and converting the usage data of each node within the same job into a probability distribution; calculating the information divergence between the converted probability distribution and the corresponding ideal state distribution for each job to obtain a multidimensional divergence value set for that job; fusing the multidimensional divergence value set according to preset weights associated with the job type to generate a comprehensive job efficiency score; aggregating the comprehensive efficiency scores of all jobs based on the resource proportion of each job to obtain an overall cluster efficiency index; and identifying key bottleneck nodes based on the contribution of each node in the information divergence calculation.

[0008] Optionally, in a first implementation of the first aspect of the present invention, the historical resource probability distribution sequence for each resource dimension over multiple consecutive evaluation periods is obtained; the changing pattern of the historical resource probability distribution sequence is analyzed to generate state transition description data; and based on the state transition description data, the calculated basic divergence value is corrected to obtain the corrected divergence value.

[0009] Optionally, in the second implementation of the first aspect of the present invention, the method includes: configuring a preset importance weight for each resource dimension according to the job type; using the corresponding weight to perform a weighted summation of each divergence value in the multidimensional divergence value set for each job to generate a comprehensive efficiency score for the job; calculating the total amount of resources occupied by each job in the cluster, and calculating the proportion of its resource amount to the total resources of the cluster to obtain the resource proportion data of each job; and using the resource proportion data of each job as the weight, performing a weighted average of the comprehensive efficiency scores of all jobs to generate an overall cluster efficiency index.

[0010] Optionally, in a third implementation of the first aspect of the present invention, the total amount of various types of computing resources occupied by each job and the total available amount of various types of computing resources in the cluster are obtained; resource weight coefficients are assigned to various types of computing resources according to preset importance coefficients corresponding to different resource types; using the resource weight coefficients, the total amount of heterogeneous resources occupied by each job is converted into standardized resource equivalents, and the total resource equivalents of the cluster are calculated; resource proportion data is determined according to the ratio of the standardized resource equivalents of each job to the total standardized resource equivalents of the cluster.

[0011] Optionally, in a fourth implementation of the first aspect of the present invention, the dimensionless resource equivalent is set as... ,but: ; in, This represents the total number of CPU cores in the cluster. This represents the total memory capacity. and The weights of processor and memory in the global resource evaluation; This represents the actual number of CPU cores used by job j. This represents the actual memory usage of job j.

[0012] Optionally, in the fifth implementation of the first aspect of the present invention, the method includes: calculating the contribution of each occupying node to the information divergence value in each resource dimension based on the resource probability distribution and ideal state distribution of each job; summarizing and sorting the contribution data generated by all jobs in the cluster; selecting a number of node-dimension combinations with the highest contribution according to preset rules as a set of key bottleneck nodes; and generating an efficiency bottleneck diagnosis report based on the set of key bottleneck nodes.

[0013] Optionally, in the sixth implementation of the first aspect of the present invention, the method further includes: continuously collecting and storing the overall efficiency index of the cluster and the comprehensive efficiency score of each operation according to a preset period to form a historical efficiency time series database; performing trend analysis based on the historical efficiency time series database to generate an efficiency trend analysis report; dynamically adjusting the weight configuration data used to generate the comprehensive efficiency score of the operation, or adjusting the model selection strategy used to determine the ideal state distribution, based on the efficiency trend analysis report; applying the adjusted strategy to subsequent evaluation periods, and generating cluster operation status early warning information based on the efficiency trend analysis report.

[0014] Optionally, in the seventh implementation of the first aspect of the present invention, the overall efficiency index sequence of the cluster within a preset time window is extracted from the historical efficiency time series database; the stationarity of the sequence is tested; if the sequence is non-stationary, its difference sequence is calculated; if the sequence is stationary, the original sequence is used; the information entropy value of the difference sequence or the original sequence is calculated as a stability quantification index; and an efficiency trend analysis report is generated by combining the stability quantification index with the statistical characteristics of the sequence.

[0015] A second aspect of this invention provides a high-performance computing cluster operation efficiency evaluation system, comprising: a state determination module, used to determine the ideal state distribution of each currently running job in the cluster based on a preset ideal state model corresponding to different job types; a data acquisition and conversion module, used to collect the actual resource usage data of the computing nodes occupied by each job, and convert the usage data of each node within the same job into a probability distribution; an information divergence module, used to calculate the information divergence between the converted probability distribution and the corresponding ideal state distribution for each job, and obtain a multidimensional divergence value set for the job; an efficiency scoring module, used to fuse the multidimensional divergence value set according to a preset weight associated with the job type to generate a comprehensive job efficiency score, and aggregate the comprehensive efficiency scores of all jobs according to the resource proportion of each job to obtain an overall cluster efficiency index; and a node identification module, used to identify key bottleneck nodes based on the contribution of each node in the information divergence calculation.

[0016] The mechanism of this invention is as follows: the abstract problem of cluster operation efficiency is transformed into a computable, decomposable, and traceable probability distribution difference measurement problem, thereby improving the accuracy, multidimensionality, and dynamic adaptability of efficiency evaluation. Beneficial effects: It can provide accurate efficiency evaluation for different types of operations, making the evaluation results more in line with the actual operation, which helps to allocate resources reasonably according to the characteristics of the operation, improve resource utilization efficiency, and avoid resource waste or unreasonable allocation caused by one-size-fits-all evaluation. It can adapt to dynamic changes in cluster operation status, cluster size expansion, increase in job types, and fluctuations in node performance. Through dynamic adjustment, it ensures that the accuracy and reliability of the evaluation method remain at a high level, providing effective support for continuous cluster optimization. It can accurately identify key bottleneck nodes in the cluster, providing cluster administrators with clear targets for performance optimization. Administrators can take targeted measures against these key bottleneck nodes, such as upgrading hardware and optimizing software configurations, to quickly and effectively resolve performance issues in the cluster and improve the overall operating efficiency of the cluster. The system extracts the overall cluster efficiency index sequence within a preset time window from the historical efficiency time series database, performs stationarity tests and calculates information entropy, and generates an efficiency trend analysis report based on the statistical characteristics of the sequence. Based on the report, it generates cluster operation status early warning information, enabling early detection of abnormal trends in cluster operation status and timely issuance of warnings. This helps administrators take proactive measures to prevent potential performance problems, ensure stable cluster operation, and reduce the risk of business interruption and data loss due to cluster failures. Attached Figure Description

[0017] Figure 1This is a schematic diagram of an embodiment of the high-performance computing cluster operation efficiency evaluation method in this invention.

[0018] Figure 2 This is a schematic diagram of another embodiment of the high-performance computing cluster operation efficiency evaluation method in this invention.

[0019] Figure 3 This is a schematic diagram of one embodiment of the high-performance computing cluster operation efficiency evaluation system in this invention.

[0020] Figure 4 This is a schematic diagram of one embodiment of the high-performance computing cluster operation efficiency evaluation device in this invention. Detailed Implementation

[0021] This invention provides a method and system for evaluating the operational efficiency of high-performance computing clusters, providing support for cluster optimization and management. The terms "first," "second," "third," "fourth," etc. (if present)," in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0022] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the high-performance computing cluster operation efficiency evaluation method in this invention includes: 101. Based on the preset ideal state model corresponding to different job types, determine the ideal state distribution for each job currently running in the cluster.

[0023] It is understood that the executing entity of this invention can be a high-performance computing cluster operation efficiency evaluation system, or it can be a terminal or a server; the specific implementation is not limited here. This embodiment of the invention will be described using a server as an example.

[0024] It should be noted that a certain high-performance computing cluster is currently running two jobs: Job A is a fluid dynamics and meteorological simulation, and Job B is a massive medical image feature extraction.

[0025] The system internally pre-defines three basic statistical intervals for resource utilization: low load (0 to 30%), medium load (30% to 70%), and high load (70% to 100%). The general objective basis for determining the ideal state model corresponding to these different job types is as follows: the system pre-extracts sample data of similar jobs that have been completed in the past, assuming no hardware bottlenecks and excellent operational efficiency scores. It then statistically and normally processes the actual utilization rate distribution of each resource, using the statistically obtained probability mean of each interval as the baseline for the ideal state distribution of that type of job.

[0026] Based on the above methods, the ideal processor state distribution for computationally and memory-intensive tasks (fluid dynamics) is expected to be: low load 5%, medium load 15%, and high load 80%. Since meteorological simulation involves massive grid data and has extremely high memory bandwidth requirements, the ideal memory state distribution is set to: low load 5%, medium load 15%, and high load 80%.

[0027] For input / output intensive tasks (massive image processing), the preset ideal processor distribution expectation is: low load 30%, medium load 50%, high load 20%; the ideal disk read / write distribution expectation is: low load 10%, medium load 20%, high load 70%.

[0028] After the evaluation device is triggered, the server accurately classifies job A as computationally and memory-intensive and job B as input / output intensive based on the scheduler's metadata. Then, feature mapping is performed: the corresponding probabilistic expectation model is directly extracted from the library to establish ideal baselines for processor and memory usage for job A; and ideal baselines for disk read / write and processor usage for job B.

[0029] 102. Collect the actual resource usage data of the computing nodes occupied by each job, and convert the usage data of each node within the same job into a probability distribution.

[0030] It should be noted that task A occupies ten nodes, while task B occupies five nodes. During the actual data collection phase, the system dynamically adjusts the monitoring granularity based on the characteristics of the task. For task A, the system collected data at a rate of once per minute over the past hour, totaling 600 data points (ten nodes multiplied by sixty minutes) on real-time processor and memory usage for task A.

[0031] For job B, considering it is an input / output intensive job, the system selects an appropriate sampling strategy based on the job type, using event-driven or kernel tracing methods to collect disk read / write data, in order to avoid high-frequency polling interfering with I / O performance. During the peak ten-minute period, the system effectively collected disk read / write and processor utilization characteristic data at corresponding frequencies through event tracing aggregation.

[0032] During the probability distribution transformation phase, the system analyzes this massive amount of discrete data: for the processor metrics of task A, out of 600 records, 510 are under high load, 60 under medium load, and 30 under low load. The actual processor state distribution is calculated to be: 85% high load, 10% medium load, and 5% low load. Similarly, its memory monitoring data is processed in parallel to transform it into the actual memory distribution (40% low load, 50% medium load, and 10% high load).

[0033] For the disk read / write metrics of assignment B, the actual distribution was statistically determined to be: 60% high load, 30% medium load, and 10% low load (corresponding to 360 high, 180 medium, and 60 low read / write records). Similarly, the processor monitoring data was processed in parallel to convert it into the actual processor distribution (60% low load, 30% medium load, and 10% high load).

[0034] The complex monitoring records were clearly transformed into an actual state probability distribution aligned with step 101.

[0035] 103. For each task, calculate the information divergence between its transformed probability distribution and the corresponding ideal state distribution to obtain the multidimensional divergence value set for that task. Specifically, in this embodiment of the invention, KL divergence (Kullback-Leibler Divergence) is used as the calculation method for information divergence. The actual probability distribution is set as... The corresponding ideal state distribution is For those with The probability distribution of each interval (e.g., low, medium, and high intervals). =3), the formula for calculating the information divergence of a certain resource dimension is: This calculation formula can quantitatively measure the degree to which actual resource usage deviates from the ideal state.

[0036] It should be noted that the server is currently performing divergence calculations on both Job A and Job B: For Job A: In terms of processors, its actual distribution (high 85%, medium 10%, low 5%) closely matches the ideal distribution, and the divergence algorithm output value is 0.01 (extremely low, representing extremely high efficiency). In terms of memory, since the actual distribution (low 40%, medium 50%, high 10%) falls far short of the ideal high load expectation, the calculated divergence value is 0.05.

[0037] For Assignment B: In terms of disk read / write operations, the actual high-load percentage was slightly lower than expected, resulting in a divergence value of 0.08. In terms of processors, the expected scheduling was moderate (ideal: low 30%, medium 50%, high 20%), but the actual distribution (low 60%, medium 30%, high 10%) indicates that a large number of nodes were in a low-load idle state. This severe distribution inversion led to a high penalty value from the divergence algorithm, resulting in a divergence value of 0.35.

[0038] After quantization, the system encapsulates the data, obtaining the multidimensional divergence value set for Job A as: {Processor divergence: 0.01, Memory divergence: 0.05}; and for Job B as: {Disk read / write divergence: 0.08, Processor divergence: 0.35}. This step accurately and objectively outputs the resource efficiency deviation.

[0039] 104. Based on the preset weights associated with the job type, the multidimensional divergence value set of each job is weighted and fused to generate a comprehensive job efficiency score. Based on the resource proportion of each job, the comprehensive efficiency scores of all jobs are aggregated to obtain the overall efficiency index of the cluster.

[0040] It should be noted that the cluster currently runs a total of fifteen nodes (ten for node A and five for node B).

[0041] In the multidimensional divergence weighted fusion stage: For Task A, a pre-set weighting for both system call computation and memory intensive tasks was applied (processor 80%, memory 20%). The calculated comprehensive divergence value was 0.018 (0.01 x 80% + 0.05 x 20%). Using a nonlinear mapping algorithm, the comprehensive efficiency score for Task A was calculated to be 98.2.

[0042] For Assignment B: Input / output intensive weights are applied (disk read / write 70%, processor 30%). The overall divergence value is 0.161 (0.08 x 70% + 0.35 x 30%). After mapping and conversion, the overall efficiency score for Assignment B is 83.9.

[0043] During the overall cluster metrics aggregation phase: the system evaluation showed that task A accounted for approximately 66.7% of the resources, while task B accounted for approximately 33.3%.

[0044] Weighted aggregation was performed: the score of job A (98.2) was multiplied by 66.7%, and the score of job B (83.9) was multiplied by 33.3%, resulting in an overall cluster efficiency score of 93.4. The system thus objectively demonstrates that despite local bottlenecks, the cluster as a whole remains in a highly efficient operating state supported by large-scale, high-efficiency operations.

[0045] 105. Based on the contribution of each node in the information divergence calculation, analyze and identify the key bottleneck nodes that lead to a decrease in the overall efficiency index of the cluster.

[0046] It should be noted that the 600 original sampling records of the processor in step 102 are retrieved, split by node (120 records per node), and the actual load distribution of each node is calculated separately to generate Diagnostic Table 1: Table 1: Node number Low load duration percentage Percentage of medium load time Percentage of high load time Node-level divergence contribution assessment Node running status diagnosis Node 11 33.3% 50.0% 16.7% Extremely low In line with expectations, operating normally Node 12 33.3% 50.0% 16.7% Extremely low In line with expectations, operating normally Node 13 33.3% 50.0% 16.7% Extremely low In line with expectations, operating normally Node 14 100% 0% 0% Extremely high Serious anomaly, processor idling Node 15 100% 0% 0% Extremely high Serious anomaly, processor idling Note: Nodes 11, 12, and 13 each contributed 40 low-load, 60 medium-load, and 20 high-load records; nodes 14 and 15 each contributed 120 low-load records (absolute zero values ​​were processed using smoothing techniques at the underlying divergence calculation level). After summing the five nodes, the data perfectly corresponds to the actual distribution data used in step 103 when calculating the processor dimension divergence, i.e., an overall low-load distribution of 60% (360 records), a medium-load distribution of 30% (180 records), and a high-load distribution of 10% (60 records).

[0047] According to Table 1 above, nodes 11 to 13 are in good condition, while nodes 14 and 15 are in a low-load idle state throughout the entire cycle. Based on this, the system accurately determines that nodes 14 and 15 are the key bottleneck nodes causing the low efficiency of the processor in Job B and lowering the overall score. Maintenance personnel can directly intervene to check whether these two nodes are experiencing network congestion.

[0048] Please see Figure 2 Another embodiment of the high-performance computing cluster operating efficiency evaluation method in this invention includes: 201. Based on the preset ideal state model corresponding to different job types, determine the ideal state distribution for each job currently running in the cluster.

[0049] Specifically, a model library containing at least two different types of ideal state models is pre-established; the corresponding ideal state model type is configured for different types of jobs, forming a mapping relationship between job type and model type; the job type of the job running in the current cluster is identified; the mapping relationship is queried according to the job type to determine the ideal state model type selected for the job; and the ideal state distribution corresponding to the job is generated based on the determined model type.

[0050] It should be noted that this scenario assumes a national-level high-performance computing center is running multiple research tasks. In the initial setup phase, the server established two core models in the database: First, there's the peak performance preference model, designed specifically for jobs requiring massive floating-point operations; second, there's the memory and throughput preference model, designed specifically for jobs that frequently load massive amounts of data into memory for comparison. The administrator configured the mapping relationships, mapping molecular dynamics simulations to the former and genome sequence alignment to the latter.

[0051] During the online identification phase, the evaluation system is connected to the Slurm job scheduling queue in real time. Two new jobs have been started on the current cluster: Job A (molecular dynamics simulation) and Job B (genome sequence alignment). After parsing the tags, the system accurately matches the pre-defined mapping relationships.

[0052] Based on the selected models, the system instantiates and generates their multidimensional ideal state distribution (divided into three intervals: low load 0-30%, medium load 30-70%, and high load 70-100%). For task A (computation preference), it is expected that the processor will be at low load for 5%, medium load for 15%, and high load for 80% (reflecting extremely high computational saturation); it is expected that the memory will be at low load for 20%, medium load for 50%, and high load for 30%.

[0053] For job B (memory throughput preference), it is expected that its processor scheduling is moderate, with a low load of 20%, a medium load of 60%, and a high load of 20%; it is expected that its memory will remain in a resident state for a long time, with a low load of 5%, a medium load of 15%, and a high load of 80%.

[0054] Through the standardized pipeline described above, the evaluation system has successfully established an accurate evaluation baseline for the currently operating heterogeneous operations.

[0055] 202. Collect the actual resource usage data of the computing nodes occupied by each job, and convert the usage data of each node within the same job into a probability distribution.

[0056] Specifically, the actual utilization rate data of each computing node in the cluster across multiple resource dimensions is collected, along with the identification and type information of the running jobs. Based on the identification information of the running jobs, each job is associated with the computing node it occupies. For each job, the actual utilization rate data of the computing node it occupies across multiple resource dimensions is extracted. The utilization rate data of each resource dimension extracted for each job is normalized to generate the resource probability distribution of the job across multiple resource dimensions.

[0057] It should be noted that this scenario continues from the supercomputing center. Currently, job A is assigned to nodes 01 to 10, and job B is assigned to nodes 11 to 15.

[0058] During the data collection phase, the system sets a 10-minute evaluation cycle, and the node agent program collects data at a high frequency of once every 10 seconds (i.e., each node generates 60 performance records within the cycle).

[0059] During the extraction and splitting phase, job A (10 nodes) extracted a total of 600 consecutive records; job B (5 nodes) extracted a total of 300 consecutive records. The system strictly isolates these data according to processor and memory dimensions.

[0060] During the normalization phase, the system performs a frequency check: For job A, across 600 records, 510 were under high load, 60 under medium load, and 30 under low load. The normalized actual probability distribution is: 85% high load, 10% medium load, and 5% low load. The memory-based normalization distribution is: 30% high load, 50% medium load, and 20% low load.

[0061] Regarding the processor dimension of assignment B (key data correction), an analysis of its 300 records revealed that 156 records were under low load, 105 under medium load, and 39 under high load. Normalization yielded the following actual probability distribution: low load 52%, medium load 35%, high load 13%. At the memory dimension, 240 records were under high load, with a normalized distribution of: high load 80%, medium load 15%, and low load 5%.

[0062] 203. For each task, calculate the information divergence between its transformed probability distribution and the corresponding ideal state distribution to obtain the set of multidimensional divergence values ​​for that task.

[0063] Specifically, for each task, the resource probability distribution and corresponding ideal state distribution for each resource dimension are obtained. For each resource dimension, the information divergence between the resource probability distribution and the ideal state distribution is calculated to obtain the divergence value for that dimension. The divergence values ​​of each resource dimension are combined into a multidimensional divergence value set. Further, the historical resource probability distribution sequence for each resource dimension over multiple consecutive evaluation periods is obtained. The change patterns of the historical sequences are analyzed to generate state transition description data. Based on the state transition description data, the calculated basic divergence value is corrected to obtain the corrected divergence value.

[0064] It should be noted that the system currently performs information divergence calculation and correction for jobs A and B. The actual distribution of job A closely matches the ideal model, and the algorithm yields a processor base divergence value of 0.02 and a memory base divergence value of 0.01. For job B, its memory distribution meets expectations (divergence 0.01); however, its actual processor low load is as high as 52% (ideally only 20%), a significant deviation, resulting in an abnormally high processor base divergence value of 0.65.

[0065] Looking back at historical data from the past 6 periods, the historical sequence of task A shows a continuous steady-state calculation, requiring no correction. The divergence set is confirmed as: {processor 0.02, memory 0.01}.

[0066] For job B, the historical sequence shows a periodic I / O characteristic of writing a global hash table to disk every 30 minutes. During this phase, processor waiting for I / O is compliant. Accordingly, the system triggers a correction mechanism, introducing a tolerance correction coefficient associated with this periodic characteristic. (In this embodiment, it is set) =0.69), according to the formula: The base penalty of 0.65 will be reduced.

[0067] Key logic decision: The system corrects it to 0.45 (i.e., 0.65). (0.69≈0.45). However, the system's preset normal trough divergence threshold is 0.20. The corrected 0.45 is still significantly excessive, sending a clear signal to the evaluation system—although job B is currently in a legitimate I / O shift period, the depth and breadth of its processor idleness far exceed the reasonable range of the software logic, inevitably hiding real physical node congestion. This correction acknowledges the business characteristics while accurately preserving the afterimage of the fault. The divergence set output for job B is: {processor 0.45, memory 0.01}.

[0068] 204. Based on the preset weights associated with the job type, the multidimensional divergence value set of each job is weighted and fused to generate a comprehensive job efficiency score. Based on the resource proportion of each job, the comprehensive efficiency scores of all jobs are aggregated to obtain the overall efficiency index of the cluster.

[0069] Specifically, based on the job type, a preset importance weight is configured for each resource dimension. For each job, the corresponding weight is used to perform a weighted summation of the divergence values ​​in its multidimensional divergence value set to generate a comprehensive efficiency score for the job. The total resource consumption of each job in the cluster is calculated, and its resource consumption is used as the proportion of the total cluster resources to obtain the resource proportion data for each job. Using the resource proportion data of each job as the weight, the comprehensive efficiency scores of all jobs are weighted and averaged to generate an overall cluster efficiency index. Further, the total amount of various types of computing resources consumed by each job and the total available amount of various types of computing resources in the cluster are obtained. Based on the preset importance coefficients corresponding to different resource types, resource weight coefficients are assigned to various types of computing resources. Using the resource weight coefficients, the total amount of heterogeneous resources consumed by each job is converted into standardized resource equivalents, and the total resource equivalent of the cluster is calculated. Based on the ratio of the standardized resource equivalent of each job to the total standardized resource equivalent of the cluster, the resource proportion data is determined.

[0070] It should be noted that the weighting for assignment A is (processor 80%, memory 20%), and the weighted divergence is... After index mapping, Assignment A's overall score is 98.2. Assignment B's weights are (processor 30%, memory 70%), and its weighted divergence is... After index mapping, the overall score for Assignment B is 86.8 points.

[0071] The total available resources of the cluster are set at 1000 CPU cores and 4000GB of memory. Regarding the general selection rules for resource importance weight parameters, the system can adaptively assign weights based on the frequency with which various resources become system bottlenecks in the cluster's global historical operational statistics. In the supercomputing center scenario of this embodiment, since historical data shows that computing resources are more likely to become scarce bottlenecks, processors are assigned a higher resource weight coefficient, preset as processor weight. ,Memory Job A uses 640 cores and 1000GB of memory; Job B uses 160 cores and 2000GB of memory.

[0072] Using the Euclidean resource equivalent formula: ; Substitute the data for precise calculation, and calculate the resource equivalent for task A. .

[0073] Task B Resource Equivalent .

[0074] Total equivalent is This leads to the actual resource usage ratio: Task A accounts for 64.1%, and Task B accounts for 35.9%.

[0075] It should be noted that the current formula structure and weight settings ( =0.6, The percentage calculated using 0.4 (64.1%) is close to the direct CPU core percentage (64%), without absolute weight normalization. This is because the specific scenario in this embodiment is clearly processor computing resources-driven; if the evaluation system wishes to further increase the weight of memory percentage, the system can dynamically adjust... Parameters or a weighted squaring method can be introduced to avoid ambiguity caused by nonlinear weighting.

[0076] 205. Based on the contribution of each node in the information divergence calculation, analyze and identify the key bottleneck nodes that lead to a decrease in the overall efficiency index of the cluster.

[0077] Specifically, for each job, based on the resource probability distribution and ideal state distribution of each resource dimension, the contribution of each occupying node to the information divergence value in each resource dimension is calculated; the contribution data generated by all jobs in the cluster are summarized and sorted; according to preset rules, several node-dimension combinations with the highest contribution are selected as the set of key bottleneck nodes; based on the set of key bottleneck nodes, an efficiency bottleneck diagnosis report is generated.

[0078] It should be noted that, for the 5 computing nodes (nodes 11 to 15) occupied by job B, a drill-down analysis of the divergence contribution at the processor dimension is performed. 300 original independent samples (60 samples per node) from these 5 nodes in step 202 are retrieved, their respective actual probability distributions are generated, and the contribution of each node to the overall abnormal divergence of 0.45 is calculated, generating a descending diagnostic table 2. Table 2 Node number Actual low load percentage and quantity Actual load percentage and quantity Actual high load percentage and number Divergence contribution ranking Node running status diagnosis Node 14 100% (60 records) 0% (0 records) 0% (0 records) First place (extremely high) Serious anomaly, processor deadlocked and idle. Node 15 95% (57 records) 5% (3 records) 0% (0 records) Second place (extremely high) Serious anomaly, processor nearly stopped. Node 11 25% (15 records) 55% (33 records) 20% (12 records) Third place (very low) In line with expectations, operating smoothly Node 12 20% (12 records) 60% (36 records) 20% (12 records) Fourth place (very low) Perfect fit, high-efficiency operation Node 13 20% (12 records) 55% (33 records) 25% (15 records) Fifth place (very low) In line with expectations, operating smoothly The sum of the low-load records of the five nodes (60+57+15+12+12=156 records) perfectly confirms the source of the overall low-load 52% data in step 202.

[0079] The diagnostic results clearly show that, within the same job, nodes 11 to 13 are functioning well, while nodes 14 and 15 exhibit fatal resource idleness. Based on this, the system extracts the top two nodes and generates an efficiency bottleneck diagnostic report: Nodes 14 and 15, belonging to job B, are exhibiting extremely low load idling, completely ruling out the possibility of normal software I / O slowdown. It is highly diagnosed that these two nodes have encountered a failure in the underlying storage parallel file system mount or physical damage to the network card, causing the computing process to fall into data starvation. Please intervene immediately to investigate hardware connectivity.

[0080] 206. Continuously collect and store the overall efficiency indicators of the cluster and the comprehensive efficiency scores of each operation according to the preset cycle to form a historical efficiency time series database; perform trend analysis based on the historical efficiency time series database to generate an efficiency trend analysis report; dynamically adjust the weight configuration data used to generate the comprehensive efficiency scores of operations, or adjust the model selection strategy used to determine the ideal state distribution, based on the efficiency trend analysis report; apply the adjusted strategy to subsequent evaluation cycles, and generate cluster operation status early warning information based on the efficiency trend analysis report.

[0081] Furthermore, the overall efficiency index sequence of the cluster within a preset time window is extracted from the historical efficiency time series database; the stationarity of the sequence is tested; if the sequence is non-stationary, its difference sequence is calculated; if the sequence is stationary, the original sequence is used; the information entropy value of the difference sequence or the original sequence is calculated as a stability quantification index; and an efficiency trend analysis report is generated by combining the stability quantification index with the statistical characteristics of the sequence.

[0082] It should be noted that the overall cluster efficiency index sequence was extracted over the past hour (6 consecutive periods, T1 to T6). The data trend is as follows: Period T1 (97.0 points), Period T2 (85.5 points), Period T3 (85.3 points), Period T4 (85.6 points), Period T5 (85.4 points), Period T6 (85.5 points).

[0083] A stationarity test revealed a step drop in the sequence at T2, followed by a period of sustained low-level fluctuation. The system performed differencing (calculating the difference between adjacent periods), and the difference value after T2 almost reached zero. Subsequently, the system calculated the information entropy: due to the extremely stable low-level deadlock state exhibited after the step drop, with very low disorder, the calculated information entropy value approached zero infinitely.

[0084] Based on common knowledge in the HPC field, random software lag can cause drastic fluctuations in metrics (high entropy value). This kind of step-like drop followed by a stable low level (extremely low entropy value) is a typical characteristic of storage I / O channels being physically saturated or critical network equipment failures causing global congestion.

[0085] Based on this, an efficiency trend forecast report is generated, and a dynamic closed loop is automatically executed: in the next cycle, the tolerance threshold of the system's built-in model for abnormally low load is urgently increased to avoid sending a large number of false alarm spam emails to the scheduler during the confirmation of hardware failure.

[0086] The highest-level red alert was received and pushed to the operations and maintenance team: After an abnormal step drop in overall cluster efficiency, a deadlock has formed at a low level of 85.5. The information entropy is extremely low, and the diagnosis is a global physical bottleneck (combined with the drill-down analysis in step 205, it was confirmed that the storage switch link where nodes 14 / 15 are located is broken).

[0087] The above describes the high-performance computing cluster operation efficiency evaluation method in the embodiments of the present invention. The following describes the high-performance computing cluster operation efficiency evaluation system in the embodiments of the present invention. Please refer to [link / reference]. Figure 3 An embodiment of the high-performance computing cluster operation efficiency evaluation system of the present invention includes: a state determination module 301, used to determine the ideal state distribution of each currently running job in the cluster based on a preset ideal state model corresponding to different job types; a collection and conversion module 302, used to collect the actual resource usage data of the computing nodes occupied by each job, and convert the usage data of each node in the same job into a probability distribution; an information divergence module 303, used to calculate the information divergence between the converted probability distribution and the corresponding ideal state distribution for each job, and obtain a multidimensional divergence value set for the job; an efficiency scoring module 304, used to fuse the multidimensional divergence value set according to a preset weight associated with the job type to generate a comprehensive job efficiency score, and aggregate the comprehensive efficiency scores of all jobs according to the resource proportion of each job to obtain the overall efficiency index of the cluster; and a node identification module 305, used to identify key bottleneck nodes based on the contribution of each node in the information divergence calculation.

[0088] above Figure 3The high-performance computing cluster operation efficiency evaluation system in this embodiment of the invention is described in detail from the perspective of modular functional entities. The high-performance computing cluster operation efficiency evaluation device in this embodiment of the invention is described in detail from the perspective of hardware processing.

[0089] Figure 4 This is a schematic diagram of the structure of a high-performance computing cluster operation efficiency evaluation device provided in an embodiment of the present invention. The device 400 may include: a processor 401, a receiver 402, a transmitter 403, and a memory 404. The receiver 402, transmitter 403, and memory 404 are respectively connected to the processor 401 via a bus. It should be noted that in some possible implementations, the processor 401 and the memory 404 may be integrated together.

[0090] The processor 401 includes one or more processing cores. The processor 401 executes the methods performed by the base station in the random access method provided in this application embodiment by running software programs and modules. The memory 404 can be used to store software programs and modules. Specifically, the memory 404 can store an operating system 4041 and at least one application module 4042 required for a function. The receiver 402 is used to receive communication data sent by other devices, and the transmitter 403 is used to send communication data to other devices.

[0091] The present invention also provides a high-performance computing cluster operation efficiency evaluation device, which includes a memory and a processor. The memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor performs the steps of the high-performance computing cluster operation efficiency evaluation method in the above embodiments.

[0092] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the high-performance computing cluster operating efficiency evaluation method.

[0093] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0094] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0095] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for evaluating the operating efficiency of a high-performance computing cluster, characterized in that, include: Based on the preset ideal state models corresponding to different job types, the ideal state distribution is determined for each job currently running in the cluster, including: A model library containing at least two different types of ideal state models is pre-established. The corresponding ideal state model type is configured for different types of jobs, forming a mapping relationship between job type and model type. The job type of the job running in the current cluster is identified. The mapping relationship is queried according to the job type to determine the ideal state model type selected for the job. Based on the determined model type, the ideal state distribution corresponding to the job is obtained. Collect the actual resource usage data of the computing nodes occupied by each job, and convert the usage data of each node within the same job into a probability distribution; For each task, the information divergence between its transformed probability distribution and the corresponding ideal state distribution is calculated to obtain the multidimensional divergence value set of the task. Based on the preset weights associated with the job type, the multidimensional divergence value set is fused to generate a comprehensive job efficiency score. The comprehensive efficiency scores of all jobs are aggregated according to the resource proportion of each job to obtain the overall efficiency index of the cluster. Based on the contribution of each node in the information divergence calculation, the key bottleneck nodes are identified.

2. The method for evaluating the operating efficiency of a high-performance computing cluster according to claim 1, characterized in that, Obtain the historical resource probability distribution sequence for each resource dimension over multiple consecutive evaluation periods; Analyze the changing patterns of the historical resource probability distribution sequence to generate state transition description data; Based on the state transition description data, the calculated basic divergence value is corrected to obtain the corrected divergence value.

3. The method for evaluating the operating efficiency of a high-performance computing cluster according to claim 1, characterized in that, include: Configure preset importance weights for each resource dimension based on the task type; For each task, the corresponding weights are used to sum the divergence values ​​in its multidimensional divergence value set to generate a comprehensive efficiency score for that task. The total amount of resources used by each job in the cluster is counted, and the proportion of its resource usage to the total cluster resources is calculated to obtain the resource usage data for each job. Using the resource proportion data of each task as the weight, a weighted average is calculated on the comprehensive efficiency scores of all tasks to generate an overall cluster efficiency index.

4. The method for evaluating the operating efficiency of a high-performance computing cluster according to claim 3, characterized in that, Obtain the total amount of various computing resources used by each job, as well as the total available amount of various computing resources in the cluster; Based on the preset importance coefficients corresponding to different resource types, resource weight coefficients are assigned to various computing resources. Using the resource weighting coefficients, the total heterogeneous resources occupied by each job are converted into standardized resource equivalents, and the total resource equivalents of the cluster are calculated. The resource proportion data is determined based on the ratio of the standardized resource equivalent of each task to the total standardized resource equivalent of the cluster.

5. The method for evaluating the operating efficiency of a high-performance computing cluster according to claim 3, characterized in that, Define the dimensionless resource equivalent as ,but: ； in, This represents the total number of CPU cores in the cluster. This represents the total memory capacity. and The weights of processor and memory in the global resource evaluation; This represents the actual number of CPU cores used by job j. This represents the actual memory usage of job j.

6. The method for evaluating the operating efficiency of a high-performance computing cluster according to claim 3, characterized in that, include: For each job, based on the resource probability distribution and ideal state distribution of each resource dimension, calculate the contribution of each occupying node to the information divergence value in each resource dimension. Summarize and sort the contribution data generated by all jobs in the cluster; Based on preset rules, select the node-dimension combination with the highest contribution as the set of key bottleneck nodes; Based on the set of key bottleneck nodes, an efficiency bottleneck diagnostic report is generated.

7. The method for evaluating the operating efficiency of a high-performance computing cluster according to claim 1, characterized in that, Also includes: The overall efficiency index of the cluster and the comprehensive efficiency score of each job are continuously collected and stored according to a preset cycle to form a historical efficiency time series database. Based on the historical efficiency time series database, trend analysis is performed to generate an efficiency trend analysis report; Based on the efficiency trend analysis report, dynamically adjust the weight configuration data used to generate the overall operational efficiency score, or adjust the model selection strategy used to determine the ideal state distribution; The adjusted strategy will be applied to subsequent evaluation cycles, and cluster operation status early warning information will be generated based on the efficiency trend analysis report.

8. The method for evaluating the operating efficiency of a high-performance computing cluster according to claim 7, characterized in that, Extract the cluster overall efficiency index sequence within a preset time window from the historical efficiency time series database; The sequence is subjected to a stationarity test. If the sequence is non-stationary, its difference sequence is calculated. If the sequence is stationary, the original sequence is used. Calculate the information entropy value of the difference sequence or the original sequence as a stability quantification index; By combining the aforementioned stability quantification indicators with the statistical characteristics of the sequence, an efficiency trend analysis report is generated.

9. A high-performance computing cluster operating efficiency evaluation system, characterized in that, include: The state determination module is used to determine the ideal state distribution for each job currently running in the cluster based on the preset ideal state model corresponding to different job types. The data acquisition and conversion module is used to collect the actual resource usage data of the computing nodes occupied by each job, and convert the usage data of each node within the same job into a probability distribution. The information divergence module is used to calculate the information divergence between the transformed probability distribution and the corresponding ideal state distribution for each task, and to obtain the multidimensional divergence value set for the task. The efficiency scoring module is used to fuse the multidimensional divergence value set according to the preset weight associated with the job type to generate a comprehensive job efficiency score, and to aggregate the comprehensive efficiency scores of all jobs according to the resource proportion of each job to obtain the overall efficiency index of the cluster. The node identification module is used to identify key bottleneck nodes based on the contribution of each node in the information divergence calculation.

Images