CN-122001828-A - SmartNIC network card multidimensional data acquisition method oriented to heterogeneous computation

CN122001828ACN 122001828 ACN122001828 ACN 122001828ACN-122001828-A

Abstract

The invention provides a SmartNIC network card multidimensional data acquisition method for heterogeneous computation, which relates to the technical field of network management and comprises the steps of taking a SmartNIC network card as a data hub, actively inquiring the running state to obtain M heterogeneous component state data streams, dividing the M heterogeneous component state data streams in an aligned mode across computing components to construct K associated behavior images, mining performance modes to construct a across component behavior map, carrying out interference pattern recognition, carrying out interference root cause quantitative analysis according to K component call chains to obtain an interference root cause component and an interference propagation quantitative link, and carrying out closed-loop resource dynamic balancing of the SmartNIC network card. The invention solves the technical problems that the prior art is usually focused on monitoring single computing resources, but ignores the synergistic effect among a plurality of computing components in a heterogeneous computing environment, so that the efficiency of resource scheduling and task scheduling is low, and the overall performance of the system is further affected.

Inventors

LIU LEZHI
HE XIAO
MA QIANG
WANG YAO
ZHANG HAO

Assignees

翼华科技(北京)股份有限公司

Dates

Publication Date: 20260508
Application Date: 20260407

Claims (10)

1. The method for collecting the SmartNIC network card multidimensional data facing heterogeneous calculation is characterized by comprising the following steps: taking SmartNIC network cards as data hubs, and actively inquiring the running states of the M interconnected heterogeneous computing components through private management interfaces to obtain M heterogeneous component state data streams; According to K life cycle time windows of K task tracks in the historical task log, aligning and dividing the M heterogeneous component state data streams across the computing components to construct K associated behavior portraits; Constructing a cross-component behavior map by performing performance mode mining on the K associated behavior portraits; After the inter-component behavior patterns are identified in an interference mode, carrying out interference root cause quantitative analysis on the interference modes according to K component call chains of the K task tracks to obtain an interference root cause component and an interference propagation quantitative link; and taking the interference source component as a starting point, and carrying out closed-loop resource dynamic balancing of the SmartNIC network card according to the interference propagation quantized link.
2. The method for collecting multi-dimensional data of SmartNIC network card facing heterogeneous computation according to claim 1, wherein after identifying an interference pattern on the inter-component behavior pattern, according to K component call chains of the K task trajectories, performing interference root cause quantization analysis of the interference pattern to obtain an interference root source component and an interference propagation quantization link, the method comprises: Traversing the map edges of the cross-component behavior map by adopting a preset performance anomaly threshold to identify a plurality of weight anomaly edges; performing space-time clustering analysis on the plurality of weight abnormal edges to obtain the interference mode and a corresponding interference time window; Dividing K calling chain fragments from the K component calling chains according to the intersection time periods of the interference time window and the K life cycle time windows; Performing reverse causal backtracking on the K call chain fragments by taking the associated components of the multiple weight abnormal edges as starting points, positioning an interference propagation chain, and taking the link starting points of the interference propagation chain as the interference root components; And extracting the associated edge weight of the interference propagation chain from the cross-component behavior spectrum, and constructing the interference propagation quantized link.
3. The method for collecting multi-dimensional data of SmartNIC network card facing heterogeneous computation according to claim 2, wherein the method comprises performing space-time clustering analysis on the plurality of weight anomaly edges to obtain the interference pattern and a corresponding interference time window, and the method comprises: extracting a plurality of groups of original task execution time information from the M heterogeneous component state data streams according to a plurality of heterogeneous computing component combinations corresponding to the plurality of weight abnormal edges; based on the time overlapping density of the multiple groups of original task execution time information and the topological adjacency of the multiple weight abnormal edges in the cross-component behavior patterns, performing space-time joint clustering on the multiple weight abnormal edges to obtain multiple abnormal edge aggregation sets; matching a plurality of alternative modes in an interference mode feature library according to heterogeneous computing component constitution and component interaction relation associated with the plurality of abnormal edge aggregation sets; Quantifying a plurality of space-time cohesive degree scores according to the original task execution time information composition and the abnormal edge topology distance set associated with the plurality of abnormal edge aggregation sets; positioning the interference patterns in the plurality of alternative patterns according to the descending order of the plurality of space-time cohesive degree scores; and performing task execution and core region positioning according to the weight abnormal edge set corresponding to the interference mode to obtain the interference time window.
4. The method for collecting multi-dimensional data of SmartNIC network card facing heterogeneous computing according to claim 1, wherein the M heterogeneous component state data streams are aligned and partitioned across computing components according to K lifecycle time windows of K task trajectories in a historical task log, and K associated behavior portraits are constructed, the method comprising: component calling time window boundaries are defined on the K task tracks, and the K life cycle time windows are obtained; After the M heterogeneous component state data streams are aligned in time sequence, data point index splicing is carried out in the K life cycle time windows according to the sequence of the K component call chains, and K component-crossing task data time sequence fragments are constructed; And carrying out structural feature extraction on the K cross-component task data time sequence fragments, and outputting the K associated behavior portraits.
5. The method for collecting multi-dimensional data of SmartNIC network card for heterogeneous computing according to claim 4, wherein the structural feature extraction is performed on the K time sequence segments of the task data of the cross-component, and the K associated behavior portraits are output, the method comprising: Decomposing the first cross-component task data time sequence fragments to obtain M single-component behavior data time sequence fragments; performing staged behavior pattern recognition on the M single-component behavior time sequence data fragments to obtain M single-component behavior pattern feature vectors; according to the time sequence relation among the M single component behavior data time sequence fragments, performing time sequence correlation feature statistics of the M single component behavior pattern feature vectors to obtain M component comprehensive feature vectors; And performing behavior pattern vectorization coding on the M component comprehensive feature vectors, and outputting a first associated behavior portrait.
6. The multi-dimensional data collection method of SmartNIC network card for heterogeneous computing according to claim 1, wherein the performance pattern mining is performed on the K associated behavior portraits to construct a cross-component behavior profile, the method comprising: Performing standard performance feature extraction based on the K associated behavior portraits to obtain K task-level multi-dimensional performance vectors, wherein the task-level multi-dimensional performance vectors cover a time sequence dimension vector group, a resource dimension vector group and an interaction dimension vector group; Carrying out component behavior pattern clustering on K task-level multidimensional performance vectors by adopting K-means clustering to obtain P task behavior pattern clusters; using M heterogeneous computing components as topological nodes, establishing directed connection among the nodes according to the calling relation of the K component calling chains, and constructing initial component calling topology; And according to the component interaction rules of the P task behavior pattern clusters, carrying out side relation weighting on the initial component call topology, and completing construction of the cross-component behavior patterns.
7. The heterogeneous computation-oriented SmartNIC network card multidimensional data acquisition method as claimed in claim 1, wherein the closed-loop resource dynamic equalization of the SmartNIC network card is performed according to the interference propagation quantization link with the interference root component as a starting point, the method comprising: performing resource regulation weight calculation according to the interference propagation quantized link to obtain a resource allocation balanced link; And taking the interference source component as a starting point, and carrying out closed-loop resource dynamic balancing of the SmartNIC network card according to the resource allocation balancing link.
8. The method for collecting multi-dimensional data of SmartNIC network cards for heterogeneous computing according to claim 1, wherein the method further comprises: In the process of actively inquiring and obtaining M heterogeneous component original data streams by taking the SmartNIC network card as a data hub, acquiring basic physical index streams of the SmartNIC network card by using a direct-connection sensing module; And carrying out data calibration under observer state awareness on the M heterogeneous component original data streams by adopting the basic physical index stream to obtain the M heterogeneous component state data streams.
9. The heterogeneous computing-oriented SmartNIC network card multidimensional data collection method of claim 5 wherein each component integrated feature vector includes an event causal feature, a resource competition feature, and an efficiency coupling feature.
10. The method for collecting multi-dimensional data of SmartNIC network card for heterogeneous computing according to claim 5, wherein the M component integrated feature vectors are subjected to low-dimensional vectorization encoding based on dimension reduction feature fusion, and the first associated behavior representation is output.

Description

SmartNIC network card multidimensional data acquisition method oriented to heterogeneous computation Technical Field The invention relates to the technical field of network management, in particular to a SmartNIC network card multidimensional data acquisition method oriented to heterogeneous computation. Background SmartNIC (intelligent network card) is used as a key component in a network, and can optimize data transmission, load balancing, task scheduling and the like through an advanced hardware acceleration technology and an advanced network function, particularly in a heterogeneous computing system, the SmartNIC network card is used as a data hub, and the capability of monitoring and optimizing data flow and resource allocation among the heterogeneous computing components in real time becomes an important tool for system optimization. However, the prior art has problems and challenges in practical applications, mainly in how to efficiently collect, integrate and optimize multidimensional data streams, especially in complex multi-computing component environments. In particular, the prior art generally focuses on monitoring a single computing resource, ignoring the synergistic effect between multiple computing components in a heterogeneous computing environment, each having different performance characteristics and resource usage patterns, and how to collect and integrate their state data in real time is a challenge. The system cannot accurately grasp the mutual influence of all computing components in the task execution process, cannot identify resource competition or performance bottleneck among the components, and finally causes low efficiency of resource scheduling and task scheduling, thereby affecting the overall performance of the system. Disclosure of Invention The application provides a SmartNIC network card multidimensional data acquisition method oriented to heterogeneous computing, and aims to solve the technical problems that the prior art is usually focused on monitoring single computing resources, and the collaborative effect among a plurality of computing components in a heterogeneous computing environment is ignored, so that the efficiency of resource scheduling and task scheduling is low, and the overall performance of a system is further influenced. The application discloses a SmartNIC network card multidimensional data acquisition method for heterogeneous computation, which comprises the steps of taking a SmartNIC network card as a data hub, actively inquiring the running states of M interconnected heterogeneous computation components through a private management interface to obtain M heterogeneous component state data streams, dividing the M heterogeneous component state data streams in an aligned manner across the computation components according to K life cycle time windows of K task tracks in a historical task log, constructing K associated behavior portraits, constructing a cross-component behavior map by carrying out performance pattern mining on the K associated behavior portraits, carrying out interference pattern recognition on the cross-component behavior map, carrying out interference root quantitative analysis of the interference patterns according to K component call chains of the K task tracks, obtaining an interference root component and an interference propagation quantized link, and carrying out closed-loop dynamic resource balancing of SmartNIC by taking the interference root component as a starting point according to the interference propagation quantized link. The one or more technical schemes provided by the application have at least the following beneficial effects: The method has the advantages that the SmartNIC network card is used as a data hub, the running states of M heterogeneous computing components can be actively inquired by adopting a private management interface, the state data streams of the computing components can be obtained in real time, the states of task execution and the resource use conditions in the whole heterogeneous computing environment can be comprehensively mastered, the observability and the instantaneity of the system are further enhanced, on the basis of aligning and dividing the state data streams of the M heterogeneous components across the computing components, the K associated behavior portraits are constructed according to K task tracks in a historical task log, the state information of the computing components can be combined to form a global view task behavior analysis, a basis is provided for further optimizing task scheduling and resource allocation, a cross-component behavior map is constructed, the executing modes of tasks among the components and the interaction relations among the components are intuitively shown, the task is effectively identified, the interference source of the task is accurately identified through an assembly quantization chain of the K task tracks, the task source is accurately identified, the interf