CN-121979579-A - Training system deployment adaptation platform of cloud multi-core in information creation environment

CN121979579ACN 121979579 ACN121979579 ACN 121979579ACN-121979579-A

Abstract

The invention relates to the technical fields of computer system architecture and cloud computing, and provides a cloud multi-core training system deployment and adaptation platform in a credit-creating environment. The platform comprises an instruction analysis module, a time sequence modeling module, a deployment engine, a self-adaptive execution module and a runtime regulation and control module. The method comprises the steps of pre-scanning a target training system through an instruction analysis module to generate a high-risk instruction fingerprint library, establishing a standard time sequence model through a time sequence modeling module, adaptively selecting a primary execution, a lightweight hot patch replacement execution or an on-demand binary translation execution strategy according to the high-risk instruction fingerprint library when the system is in operation, reducing performance cost, and sending a time sequence compensation command to a resource scheduler when time sequence drift is detected by a time sequence regulation and control module when the system is in operation, wherein the time sequence regulation and control module monitors key event time sequence indexes and compares the key event time sequence indexes with the standard time sequence model. The invention effectively solves the problems of poor compatibility, large performance loss and inconsistent running time sequence of the practical training system under the heterogeneous CPU architecture.

Inventors

CHEN HAITAO
CAI JIANLIN
HE XIAODONG
WEI CHEN

Assignees

广州民航职业技术学院

Dates

Publication Date: 20260505
Application Date: 20260126

Claims (10)

1. The utility model provides a real standard system deployment adaptation platform of cloud multicore under the creation environment which characterized in that includes: The system comprises an instruction analysis module, a target training system and a data processing module, wherein the instruction analysis module is used for pre-scanning binary files of the target training system when the target training system is deployed to a created cloud environment containing multiple heterogeneous central processor architectures, and generating a high-risk instruction fingerprint library, wherein the high-risk instruction fingerprint library is used for representing instruction fragments with possibly different compatibility in the multiple heterogeneous central processor architectures; the system comprises a time sequence modeling module, a target training system, a time sequence modeling module and a time sequence analysis module, wherein the time sequence modeling module is used for establishing a standard time sequence model of the target training system; The deployment engine is used for deploying the target training system, the time sequence monitoring agent and the elastic instruction translator into a target operation container of the created cloud environment; The self-adaptive execution module is used for monitoring an instruction execution flow according to the high-risk instruction fingerprint library through the elastic instruction translator when the target practical training system runs, and adaptively selecting an execution strategy according to the actual instruction supporting capability of a central processing unit architecture where a current running container is located when the instruction to be executed is detected to belong to the high-risk instruction fingerprint library, wherein the execution strategy comprises native execution, lightweight hot patch replacement execution and binary translation execution as required; The system comprises a target training system, a time sequence monitoring agent, a time sequence control module and a time sequence compensation module, wherein the time sequence monitoring agent is used for monitoring a time sequence index of a key event in operation of the target training system, comparing the time sequence index of the key event with a standard time sequence model, and sending a time sequence compensation command to a resource scheduler of the created cloud environment when detecting that time sequence drift exists between the time sequence index of the key event and the standard time sequence model so as to adjust the resource quota of the target running container, so that the target training system can run time sequence alignment on different central processor architectures.
2. The platform of claim 1, further comprising: The resource management module is used for carrying out resource pooling on the physical nodes in the created cloud environment and identifying the central processor architecture type of each physical node; The deployment engine is also used for receiving a deployment request aiming at the target practical training system, and selecting a proper physical node in the created cloud environment according to the resource demand and the architecture affinity requirement in the deployment request so as to create the target operation container.
3. The platform of claim 2, wherein the deployment engine is configured to, when a particular central processor architecture type is not specified in the deployment request: analyzing historical operation performance data of the target practical training system, and determining an optimal performance architecture type; and the physical nodes of the optimal performance architecture type are preferentially selected for deployment.
4. The platform of claim 1, wherein the information creation cloud environment comprises a server cluster based on at least two heterogeneous central processor architectures of ARM architecture, MIPS architecture, alpha architecture, and X86 architecture.
5. The platform of claim 1, wherein the instruction analysis module is specifically configured to: disassembling the binary file, and extracting an instruction sequence; Comparing the instruction sequence with a preset cross-architecture compatibility database, wherein the cross-architecture compatibility database records instruction differences and simulation overheads among different central processing unit architectures; And recording the addresses and feature codes of the instruction fragments marked as having difference or high simulation overhead in the cross-architecture compatibility database in the high-risk instruction fingerprint database.
6. The platform according to claim 1, wherein the adaptive execution module is configured to, when adaptively selecting an execution policy: judging whether the current CPU architecture supports the instruction to be executed in a native way or not; if the native support is provided, selecting the native execution; If the instruction to be executed is not supported by the original, judging whether the instruction to be executed can be covered by a preset lightweight hot patch Ding Moban; If the instruction to be executed can be covered, selecting the lightweight hot patch to be replaced and executed, wherein the replacement execution comprises dynamically generating a local instruction sequence adapting to the current central processing unit architecture, and redirecting an entry address of the instruction to be executed to the local instruction sequence; if not, selecting the on-demand binary translation to execute.
7. The platform of claim 1, wherein the timing modeling module is specifically configured to: selecting one of the plurality of heterogeneous central processor architectures as the reference architecture; And running the target training system on the reference framework, and recording a key event time stamp under a standard load, wherein the key event comprises the submission of a rendering frame, the completion of a physical engine calculation step length or the transmission of a network data packet.
8. The platform of claim 1, wherein the key event timing indicator comprises an air traffic radar refresh period in a civil aviation simulation training scenario or a simulated aircraft attitude update period.
9. The platform of claim 1, wherein the timing compensation command includes instructions for adjusting a cpu time slice priority, a memory bandwidth, or an I/O bandwidth of the target run container.
10. The platform of claim 9, wherein the runtime regulation module, when adjusting the resource quota of the target runtime container, is specifically configured to: calculating the required resource quota adjustment amount according to the amplitude and the direction of the time sequence drift; And applying the resource quota adjustment amount by calling a resource control interface of the virtualization layer to accelerate or slow down the execution speed of the target training system until the key event time sequence index returns to be within a preset tolerance threshold range.

Description

Training system deployment adaptation platform of cloud multi-core in information creation environment Technical Field The invention relates to the technical fields of computer system architecture, cloud computing and cross-platform compatibility, in particular to a training system deployment and adaptation platform of a cloud multi-core in a credit-creation environment. Background Along with the comprehensive promotion of the information technology application innovation (creation) strategy in China, the construction of an autonomous and controllable information technology system becomes a key task. At the cloud computing infrastructure level, a "one-cloud-multi-core" architecture has grown and become the dominant trend. The core feature of the architecture is to integrate computing resources of various different Instruction Set Architectures (ISA) under a unified cloud management platform, such as a domestic architecture based on ARM, MIPS, alpha and the like, and an X86 architecture which needs to be compatible at a specific stage. Such deeply heterogeneous computing environments, while enhancing the flexibility and autonomy of the infrastructure, also present unprecedented challenges to migration, deployment and operation of upper-level applications. Especially in the field of professional training, such as civil aviation flight simulation, rail traffic scheduling simulation, complex industrial control system training and the like, the core training system of the system has the characteristics of high computation density, strict real-time requirements and complex sequential logic. A significant portion of these systems are long-term accumulated stock systems developed and optimized based on a particular single CPU architecture (e.g., X86) in depth. The system is directly migrated to a heterogeneous information creation cloud environment, and a plurality of complex technical barriers are faced. First is the cross-architecture instruction compatibility problem. Because different CPU architectures employ different instruction sets, binary files compiled for one architecture cannot be directly executed on another architecture. Existing solutions mainly include source code level adaptation and binary translation. Source code level adaptation requires that source code be acquired and modified to accommodate the target architecture, which is not feasible for many closed source or historic legacy training systems, and is labor intensive and costly to maintain. Although the Dynamic Binary Translation (DBT) technology has better compatibility, the traditional full-scale translation strategy can introduce significant performance overhead, the translation process occupies a large amount of computing resources, and the severe requirements of a practical training system on real-time response are difficult to meet. And secondly, the problem of the consistency of the operation time sequence of the cross-architecture. Even if the basic running compatibility problem is solved, the running performance and time sequence performance of the same training system on different hardware nodes are greatly different due to the inherent differences of different CPU architectures in terms of instruction execution efficiency, micro-architecture design (such as pipeline and cache mechanism) and the unpredictable delay introduced by the translation process. Such timing drift is unacceptable for a timing sensitive training system. For example, in a distributed training scenario requiring multi-node cooperative work, if the running speeds of simulation environments of different nodes are inconsistent, the deviation of the simulation state and the training scenario cannot be synchronized, and the accuracy and the effectiveness of the training are seriously affected. Therefore, how to realize efficient deployment and low-overhead operation of the practical training system in the 'one-cloud-multi-core' information creation cloud environment and ensure consistency of cross-architecture operation time sequences is a key technical problem to be solved in the current information creation migration process. Disclosure of Invention The invention aims to provide a training system deployment adaptation platform of a cloud multi-core in a credit creation environment, which is used for solving the problems that in the prior art, the performance cost is high, the compatibility adaptability is limited and the consistency of the cross-architecture operation time sequence cannot be ensured in the cross-architecture application migration. The embodiment of the invention provides a training system deployment adaptation platform of a cloud multi-core in a credit creation environment, which comprises the following components: The system comprises an instruction analysis module, a target training system and a data processing module, wherein the instruction analysis module is used for pre-scanning binary files of the target training system when the target trai