Search

CN-122019447-A - Server and server design method

CN122019447ACN 122019447 ACN122019447 ACN 122019447ACN-122019447-A

Abstract

The invention relates to the technical field of computer servers, in particular to a server and a server design method, wherein the server comprises a basic framework unit, a heterogeneous computing module, an intelligent storage module, a configurable network module, an intelligent management control unit and a system firmware layer which are interconnected through a standardized interface, the management control unit drives computing, storing and network resources to carry out collaborative dynamic reconfiguration through real-time monitoring and AI prediction, and the design method comprises six systemization stages of demand analysis, modularized design, resource virtualization, dynamic reconfiguration mechanism design, energy efficiency optimization and reliability design. The method solves the problems of low utilization rate, poor expansibility and low energy efficiency caused by the static configuration of the traditional server resources, realizes the accurate supply of the resources according to the needs, remarkably improves the utilization rate of the resources, the energy efficiency ratio and the cross-scene adaptability, and can be widely applied to the fields of cloud computing, high-performance computing, edge computing and the like.

Inventors

  • LU LIPENG

Assignees

  • 深圳市东方神龟科技有限公司

Dates

Publication Date
20260512
Application Date
20260227

Claims (9)

  1. 1. The server is characterized by comprising The base frame unit is used for providing standardized physical installation interfaces, power supply and heat dissipation base environments for all modules; at least one computing resource module, which is a hot-pluggable heterogeneous computing unit, supports the selective configuration of an x86, ARM or RISC-V architecture processor, and integrates an out-of-band management controller; The storage resource module is a hot-pluggable storage unit, supports an NVMe, SAS or SATA hybrid interface, and has the functions of storage resource pooling and intelligent data layering based on access characteristics; At least one network interconnection module, which is a hot pluggable network unit, supports the selective configuration of Ethernet, infiniBand or RoCE network protocols, and integrates a software defined network controller; The management control unit comprises a central resource scheduler, an AI load prediction engine and a dynamic reconfiguration controller; The system firmware layer is used for providing a unified hardware abstraction interface, safe starting and firmware hot upgrading support for upper-layer software; the computing resource module, the storage resource module and the network interconnection module are respectively connected with the basic frame unit through standardized electrical and mechanical interfaces and are in data interaction with the management control unit and the system firmware layer through a system bus.
  2. 2. The server of claim 1, wherein the central resource scheduler is configured to monitor resource status of each module in real time; the AI load prediction engine is used for predicting future resource demands based on historical load data; The dynamic reconfiguration controller is used for cooperatively executing calculation, storage and dynamic configuration adjustment of network resources by sending instructions to the calculation resource module, the storage resource module and the network interconnection module according to the prediction result and the real-time state; The AI load prediction engine adopts a long-short-term memory neural network algorithm to construct a prediction model, the prediction precision is not lower than 94%, and the load peak value can be predicted in advance by 300 milliseconds to 5 seconds.
  3. 3. The server and server design method according to claim 1, wherein the intelligent data layering function of the storage resource module automatically migrates data between NVMe SSD, SAS SSD and SATA HDD based on the configured access frequency threshold and access delay threshold, and the data migration process has an influence on the input/output operation performance of the foreground service of less than 5%.
  4. 4. The server and server design method according to claim 1, wherein the software defined network controller of the network interconnection module is based on the OpenFlow protocol and is capable of completing dynamic reconfiguration of the network topology in no more than 50 milliseconds.
  5. 5. The server and the server design method according to claim 1, wherein in the dynamic configuration adjustment process of the resource executed by the dynamic reconfiguration controller, service is not interrupted, and the application performance jitter is less than 3%.
  6. 6. The server design method is characterized by comprising the following steps: s1, a demand analysis stage, namely collecting resource demand characteristics of a target workload on calculation, storage and a network, establishing a multidimensional resource demand model, and determining constraint conditions of performance, power consumption and cost; S2, defining standardized electrical interfaces, mechanical dimensions and heat dissipation specifications of computing, storage and network modules, and designing a low-delay communication protocol and a hot plug management mechanism between the modules; S3, designing a hardware resource abstraction layer, mapping physical resources into a virtual resource pool, and providing a uniform resource scheduling interface and a multi-tenant resource isolation mechanism; S4, a dynamic reconfiguration mechanism design stage, namely deploying a resource monitoring probe to acquire a resource state in real time, establishing a resource demand prediction model based on a machine learning algorithm, and designing an online resource migration algorithm supporting service continuity; s5, in the energy efficiency optimization design stage, a power consumption-utilization ratio relation model of each module is established, and a dynamic voltage frequency adjustment strategy and a pre-dispatching energy-saving mechanism based on load prediction are designed; s6, in the reliability design stage, N+1 redundancy configuration is implemented for the key module, a fault prediction and health management module is designed, and a fault quick isolation and automatic recovery mechanism is realized.
  7. 7. The method according to claim 6, wherein in step S4, the resource demand prediction model is trained by using a long-short-term memory neural network algorithm, the training learning rate ranges from 0.001 to 0.01, and the iteration number ranges from 1000 to 5000.
  8. 8. The method according to claim 6, wherein in step S4, the online resource migration algorithm includes a memory pre-copy technique for migration of computing tasks, an incremental migration technique for migration of storage data, and a session hold technique for migration of network connections.
  9. 9. The method of claim 6, wherein in step S5, the dynamic voltage frequency adjustment strategy is to adjust the operating frequency to 50% of the reference frequency when the CPU utilization is lower than 20%, and to increase the operating frequency to 115% of the reference frequency when the CPU utilization is higher than 80%.

Description

Server and server design method Technical Field The invention relates to the technical field of computer servers, in particular to a server and a server design method. Background With the popularization of cloud computing, big data and artificial intelligence application, the workload facing a data center is increasingly diversified and dynamic, the traditional server generally adopts an immobilized hardware configuration, the proportion of computing, storing and network resources is determined when leaving a factory, and the flexible adaptation to different application scenes or the requirement change of the same application at different stages is difficult, so that two problems are generally caused, namely, when the load is low, a large amount of resources are idle, the overall utilization rate is low, and when the load is peak, the performance is limited because a certain type of resources become bottlenecks. In the prior art, some modularized or expandable server schemes exist, some servers support the expansion of computing capacity by adding a CPU or a memory board or the expansion of storage capacity by adding a hard disk backboard, however, the schemes are limited to the longitudinal expansion of single type resources, lack of the horizontal collaborative reconstruction capability of three large resources of cross computing, storage and network, the adjustment of resource configuration is often performed manually by an administrator, the response speed is slow, and prospective resource scheduling cannot be realized. Accordingly, a server and a server design method are provided by those skilled in the art to solve the above-mentioned problems. Disclosure of Invention In order to solve the technical problems, the invention provides a server and a server design method, which are used for solving the technical problems of low utilization rate, inflexible expansion, poor scene suitability, delayed response of resource allocation and the like caused by the static allocation of the traditional server resources. The server and the server design method comprise a basic framework unit, at least one computing resource module, at least one storage resource module, at least one network interconnection module, a management control unit and a system firmware layer; the base frame unit is used for providing standardized physical installation interfaces, power supply and heat dissipation base environments for all modules; The computing resource module is a hot-pluggable heterogeneous computing unit, supports the selective configuration of an x86, ARM or RISC-V architecture processor, and is integrated with an out-of-band management controller; The storage resource module is a hot-pluggable storage unit, supports an NVMe, SAS or SATA hybrid interface, and has the functions of storage resource pooling and intelligent data layering based on access characteristics; The network interconnection module is a hot pluggable network unit, supports the selective configuration of Ethernet, infiniBand or RoCE network protocols, and integrates a software defined network controller; The management control unit comprises a central resource scheduler, an AI load prediction engine and a dynamic reconfiguration controller, wherein the central resource scheduler is used for monitoring the resource state of each module in real time, the AI load prediction engine is used for predicting the future resource demand based on historical load data, and the dynamic reconfiguration controller is used for executing the calculation, storage and dynamic configuration adjustment of network resources in a coordinated manner by sending instructions to the calculation resource module, the storage resource module and the network interconnection module according to the prediction result and the real-time state; the system firmware layer is used for providing a unified hardware abstraction interface, safe starting and firmware hot upgrading support for upper software; the computing resource module, the storage resource module and the network interconnection module are respectively connected with the basic frame unit through standardized electrical and mechanical interfaces and are in data interaction with the management control unit and the system firmware layer through a system bus. Preferably, the AI load prediction engine adopts a long-term and short-term memory neural network algorithm to construct a prediction model, the prediction precision is not lower than 94%, and the load peak value can be predicted 300 milliseconds to 5 seconds in advance. Preferably, the intelligent data layering function of the storage resource module automatically migrates data among the NVMe SSD, the SAS SSD and the SATA HDD based on the configured access frequency threshold and access delay threshold, and the influence of the data migration process on the input/output operation performance of the foreground service is less than 5%. Preferably, the software defined network c