Computing power scheduling system simultaneously supporting NVIDIA card and taftosource card

Abstract

The invention discloses a computing power scheduling system capable of simultaneously supporting NVIDIA and taftosource cards, which comprises a hardware layer, a driving adaptation layer, a computing power scheduling layer and an application interface layer, and bidirectional data interaction is carried out among the layers through a data bus; the hardware layer is used for carrying out hardware state monitoring and data transmission; the driving adaptation layer is used for converting the computing power feedback data into a standard data format and transmitting the computing power feedback data to the computing power scheduling layer; the computing power scheduling layer is used for distributing computing power tasks; and the application interface layer provides an API (Application Program Interface) to be matched with a third-party application program. The method has the beneficial effects that the computing power request instruction can be respectively converted into an instruction format conforming to a CUDA framework and an instruction format conforming to a Bolt framework through the driving adaptation layer, and is sent to the NVIDIA display card module or the tafen-original display card module through the corresponding driving sub-module; therefore, the NVIDIA display card and the tafen-original display card are simultaneously supported, an independent scheduling system does not need to be built, and the hardware and software cost is reduced.

Description

Technical Field

[0001] This invention relates to the field of computer computing power scheduling technology, and in particular to a computing power scheduling system that simultaneously supports NVIDIA and Fuse cards. Background Technology

[0002] With the rapid development of fields such as artificial intelligence, big data processing, and high-performance computing, the demand for computing power in computer hardware is increasing daily. Currently, the mainstream high-performance graphics cards on the market mainly include NVIDIA series graphics cards and Flinttech series graphics cards. Both types of graphics cards are widely used in different application scenarios due to their respective hardware architectures and software ecosystems. However, in existing technologies, computing power scheduling systems are typically designed only for graphics cards of a single brand or architecture. For example, scheduling systems for NVIDIA graphics cards are mostly developed based on the CUDA architecture, while scheduling systems for Flint graphics cards rely on its self-developed Bolt system architecture. The lack of compatibility between these two systems prevents unified scheduling of both types of graphics cards. This necessitates building separate scheduling systems for computing clusters simultaneously deploying NVIDIA and Flint cards, increasing hardware procurement and software maintenance costs, and leading to wasted computing resources and low scheduling efficiency. Furthermore, due to differences in hardware parameters (such as memory capacity, number of computing cores, and power consumption limits) and software interfaces between NVIDIA and Flint cards, existing scheduling systems cannot dynamically allocate computing power based on the characteristics of the two types of graphics cards, failing to meet the differentiated computing power requirements of different tasks and further reducing the overall operating efficiency of the computing cluster. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a computing power scheduling system that simultaneously supports NVIDIA and Fuse cards.

[0004] The objective of this invention is achieved through the following technical solution: a computing power scheduling system that simultaneously supports NVIDIA and Fuse cards, comprising a hardware layer, a driver adaptation layer, a computing power scheduling layer and an application interface layer, wherein each layer interacts bidirectionally with the other through a data bus. The hardware layer is used for hardware status monitoring and data transmission; The driver adaptation layer is used to convert computing power feedback data into a standard data format and transmit it to the computing power scheduling layer; The computing power scheduling layer is used to allocate computing power tasks; The application interface layer provides API interfaces for adaptation to third-party applications.

[0005] Preferably, the hardware layer includes an NVIDIA graphics card module, an Arcturus graphics card module, a central control module, and a power management module; NVIDIA graphics card modules are used to provide computing power based on the CUDA ecosystem; The Flint graphics card module is used to provide computing power based on the Bolt ecosystem; The central control module uses an ARM architecture processor and is electrically connected to the NVIDIA graphics card module and the Flint graphics card module via a PCIe switch. It is used for hardware status monitoring and data transmission of the two types of graphics card modules. The power management module supplies power to the NVIDIA graphics card module, the Flint graphics card module, and the central control module.

[0006] Preferably, the driver adaptation layer is deployed on the central control module, specifically including the NVIDIA driver submodule, the Fuse driver submodule, and the driver conversion submodule; The NVIDIA driver submodule is used to enable communication between the central control module and the NVIDIA graphics card module. The Flint driver submodule is used to enable communication adaptation between the central control module and the Flint graphics card module; The driver conversion submodule is used to convert the computing power request instructions of the application task into instruction formats that conform to the CUDA architecture and those that conform to the Bolt architecture, respectively, and send them to the NVIDIA graphics card module or the Flint graphics card module through the corresponding driver submodule. It also converts the computing power feedback data into a standard data format and transmits it to the computing power scheduling layer.

[0007] Preferably, the computing power scheduling layer includes a task parsing module, a resource monitoring module, a scheduling decision module, and a task allocation module; The task parsing module is used to receive computing power task requests from the application interface layer and parse the computing power requirement parameters of the task to generate task feature tags. The resource monitoring module obtains the hardware status parameters of the NVIDIA graphics card module and the Flint graphics card module in real time through the driver adaptation layer, and establishes a computing power resource status database. The scheduling decision module makes decisions based on task feature tags and a computing resource status database, using a dynamic weighted scheduling algorithm. Based on the decision results of the scheduling decision module, the task allocation module allocates computing tasks to the corresponding graphics card modules through the driver adaptation layer and tracks the task execution progress in real time.

[0008] Preferably, the application interface layer includes a computing power request interface, a task status query interface, and a computing power statistics interface; The computing power request interface supports third-party applications to send computing power task requests via HTTP or RPC protocols and receive task submission feedback; The task status query interface is used to query the execution progress of submitted tasks in real time, the currently assigned graphics card module type, and the estimated task completion time. The computing power statistics interface is used to generate computing power usage reports for the NVIDIA graphics card module and the Flint graphics card module.

[0009] The present invention has the following advantages: 1. This invention can convert computing power request instructions into instruction formats that conform to the CUDA architecture and those that conform to the Bolt architecture through a driver adaptation layer, and send them to the NVIDIA graphics card module or the Flint graphics card module through the corresponding driver submodule. This enables simultaneous support for two different architecture graphics cards, NVIDIA and Flint, without the need to build an independent scheduling system, thus reducing hardware and software costs.

[0010] 2. The computing power scheduling layer of this invention adopts a dynamic weight scheduling algorithm, which can accurately allocate computing power according to task characteristics and real-time status of graphics cards, avoid waste of computing power resources, and improve the overall operating efficiency of the computing cluster. Attached Figure Description

[0011] none. Detailed Implementation

[0012] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.

[0013] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0014] In this embodiment, a computing power scheduling system that simultaneously supports NVIDIA and Fuse cards includes a hardware layer, a driver adaptation layer, a computing power scheduling layer, and an application interface layer, and each layer interacts bidirectionally with the others via a data bus. The hardware layer is used for hardware status monitoring and data transmission; The driver adaptation layer is used to convert computing power feedback data into a standard data format and transmit it to the computing power scheduling layer; The computing power scheduling layer is used to allocate computing power tasks; The application interface layer provides API interfaces for adaptation with third-party applications. Specifically, the application interface layer provides standardized API interfaces, eliminating the need for third-party applications to develop adaptations for different graphics cards, thus lowering the barrier to entry for applications. It also supports computing power statistics functions, facilitating resource management for users. The driver adaptation layer can convert computing power request instructions into instruction formats compliant with the CUDA architecture and those compliant with the Bolt architecture, respectively. These are then sent to the NVIDIA graphics card module or the Flint graphics card module through the corresponding driver submodules, achieving simultaneous support for both NVIDIA and Flint graphics cards with different architectures. This eliminates the need to build a separate scheduling system, reducing hardware and software costs.

[0015] Furthermore, the hardware layer includes an NVIDIA graphics card module, an Arcturus graphics card module, a central control module, and a power management module; NVIDIA graphics card modules are used to provide computing power based on the CUDA ecosystem; The Flint graphics card module is used to provide computing power based on the Bolt ecosystem; The central control module uses an ARM architecture processor and is electrically connected to the NVIDIA graphics card module and the Flint graphics card module via a PCIe switch. It is used for hardware status monitoring and data transmission of the two types of graphics card modules. The power management module supplies power to the NVIDIA graphics card module, the Flint graphics card module, and the central control module. Specifically, the NVIDIA graphics card module uses a GPU chip supporting the CUDA architecture, equipped with dedicated video memory and a PCIe 4.0 interface; the Flint graphics card module uses a GPU chip supporting the Bolt architecture, equipped with dedicated video memory and a PCIe 4.0 interface; the power management module adopts a multi-channel power supply design to ensure power supply stability.

[0016] Furthermore, the driver adaptation layer is deployed on the central control module, specifically including the NVIDIA driver submodule, the Fuse driver submodule, and the driver conversion submodule; The NVIDIA driver submodule is used to enable communication between the central control module and the NVIDIA graphics card module. The Flint driver submodule is used to enable communication adaptation between the central control module and the Flint graphics card module; The driver conversion submodule converts application task computing power request instructions into instruction formats compliant with the CUDA architecture and those compliant with the Bolt architecture, respectively. These converted instructions are then sent to the NVIDIA or Flint graphics card modules via the corresponding driver submodules. The driver also converts computing power feedback data into a standard data format for transmission to the computing power scheduling layer. Specifically, the driver adaptation layer can convert computing power request instructions into instruction formats compliant with the CUDA and Bolt architectures, respectively, and send them to the NVIDIA or Flint graphics card modules via the corresponding driver submodules. This achieves simultaneous support for both NVIDIA and Flint graphics cards with different architectures, eliminating the need for a separate scheduling system and reducing hardware and software costs.

[0017] In this embodiment, the computing power scheduling layer includes a task parsing module, a resource monitoring module, a scheduling decision module, and a task allocation module; The task parsing module is used to receive computing power task requests from the application interface layer and parse the computing power requirement parameters of the task to generate task feature tags. The resource monitoring module obtains the hardware status parameters of the NVIDIA graphics card module and the Flint graphics card module in real time through the driver adaptation layer, and establishes a computing power resource status database. The scheduling decision module makes decisions based on task feature tags and a computing resource status database, using a dynamic weighted scheduling algorithm. The task allocation module, based on the decision results of the scheduling decision module, distributes computing tasks to corresponding graphics card modules through the driver adaptation layer and tracks task execution progress in real time. When a graphics card module fails, unfinished tasks are automatically migrated to another type of graphics card module. Specifically, computing power requirements parameters include computational accuracy requirements, data size, and task priority; hardware status parameters include memory utilization, computing core utilization, and real-time power consumption. The scheduling decision module makes decisions based on task feature tags and a computing power resource status database, using a dynamic weighted scheduling algorithm. Specifically, when a task has high requirements for CUDA ecosystem compatibility, computing power resources of NVIDIA graphics card modules are allocated first; when a task has high requirements for the low-power characteristics of the Bolt architecture, computing power resources of the Flint graphics card module are allocated first; when both types of graphics card modules meet the task requirements, the resources are evenly allocated based on the real-time load rate of the graphics cards to ensure maximum utilization of computing power resources. In other words, the computing power scheduling layer uses a dynamic weighted scheduling algorithm to accurately allocate computing power based on task characteristics and the real-time status of graphics cards, avoiding waste of computing power resources and improving the overall operating efficiency of the computing cluster.

[0018] Furthermore, the application interface layer includes a computing power request interface, a task status query interface, and a computing power statistics interface; The computing power request interface supports third-party applications to send computing power task requests via HTTP or RPC protocols and receive task submission feedback; The task status query interface is used to query the execution progress of submitted tasks in real time, the currently assigned graphics card module type, and the estimated task completion time. The computing power statistics interface is used to generate computing power usage reports for NVIDIA graphics card modules and Flint graphics card modules. Specifically, the computing power statistics interface can generate computing power usage reports for NVIDIA graphics card modules and Flint graphics card modules according to time periods (hours, days, months), including data such as total computing power output, average resource utilization, and number of tasks completed, for users to manage computing power resources and calculate costs.

[0019] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A computing power scheduling system that simultaneously supports NVIDIA and Flint cards, characterized in that: It includes a hardware layer, a driver adaptation layer, a computing power scheduling layer, and an application interface layer, and each layer interacts bidirectionally with the others via a data bus. The hardware layer is used for hardware status monitoring and data transmission; The driver adaptation layer is used to convert the computing power feedback data into a standard data format and transmit it to the computing power scheduling layer; The computing power scheduling layer is used to allocate computing power tasks; The application interface layer provides API interfaces for adaptation to third-party applications.

2. The computing power scheduling system that simultaneously supports NVIDIA and Flint cards according to claim 1, characterized in that: The hardware layer includes an NVIDIA graphics card module, a Fuse graphics card module, a central control module, and a power management module; The NVIDIA graphics card module is used to provide computing power based on the CUDA ecosystem; The Flint graphics card module is used to provide computing power based on the Bolt ecosystem; The central control module uses an ARM architecture processor and is electrically connected to the NVIDIA graphics card module and the Flint graphics card module via a PCIe switch. It is used for hardware status monitoring and data transmission of the two types of graphics card modules. The power management module is used to supply power to the NVIDIA graphics card module, the Flint graphics card module, and the central control module.

3. The computing power scheduling system that simultaneously supports NVIDIA and Flint cards according to claim 2, characterized in that: The driver adaptation layer is deployed on the central control module and specifically includes the NVIDIA driver submodule, the Suiyuan driver submodule, and the driver conversion submodule. The NVIDIA driver submodule is used to enable communication adaptation between the central control module and the NVIDIA graphics card module. The Frostbite driver submodule is used to enable communication adaptation between the central control module and the Frostbite graphics card module; The driver conversion submodule is used to convert the computing power request instructions of the application task into instruction formats that conform to the CUDA architecture and those that conform to the Bolt architecture, respectively, and send them to the NVIDIA graphics card module or the Flint graphics card module through the corresponding driver submodule. It also converts the computing power feedback data into a standard data format and transmits it to the computing power scheduling layer.

4. The computing power scheduling system that simultaneously supports NVIDIA and Flint cards according to claim 3, characterized in that: The computing power scheduling layer includes a task parsing module, a resource monitoring module, a scheduling decision module, and a task allocation module; The task parsing module is used to receive computing power task requests from the application interface layer and parse the computing power requirement parameters of the task to generate task feature tags. The resource monitoring module obtains the hardware status parameters of the NVIDIA graphics card module and the Flint graphics card module in real time through the driver adaptation layer, and establishes a computing power resource status database. The scheduling decision module makes decisions based on task feature tags and computing resource status database using a dynamic weighted scheduling algorithm. The task allocation module allocates computing power tasks to the corresponding graphics card modules through the driver adaptation layer based on the decision result of the scheduling decision module, and tracks the task execution progress in real time.

5. The computing power scheduling system that simultaneously supports NVIDIA and Flint cards according to claim 4, characterized in that: The application interface layer includes a computing power request interface, a task status query interface, and a computing power statistics interface; The computing power request interface supports third-party applications to send computing power task requests via HTTP or RPC protocols and receive task submission feedback; The task status query interface is used to query the execution progress of submitted tasks, the currently assigned graphics card module type, and the estimated task completion time in real time. The computing power statistics interface is used to generate computing power usage reports for the NVIDIA graphics card module and the Flint graphics card module.