WO2025143409A1 - Multipath transmission apparatus and method - Google Patents

Abstract

The present disclosure relates to a multipath transmission apparatus, which may comprise: an interface module for integrating different libraries; an automatic path collection module which automatically detects and collects available data transmission paths in a platform, and which selects optimal data transmission paths from among the available data transmission paths so as to generate a set of paths; a data transmission module for transmitting data through the optimal data transmission paths; and a dynamic data classification module for dynamically readjusting an optimal data transmission path by monitoring the data transmission module in real time.

Description

Multipath transmission device and method

The following embodiments relate to a multi-path transmission device and method thereof.

As data-intensive applications such as big data analytics, machine learning, and scientific simulations run in data centers or high-performance computing (HPC) environments,1 the importance of efficient network operations increases.

In data center networks (DCNs), Remote Direct Memory Access (RDMA) is one of the promising networking technologies for data-intensive applications that require high bandwidth and very low latency. RDMA supports zero-copy read/write operations by implementing transfer logic in hardware NICs (Network Interface Cards), which can directly transfer data from the memory of one computing node to the memory of another computing node.

A multi-path transmission device according to one embodiment may include an interface module that integrates different libraries, an automatic path collection module that automatically detects and collects data transmission paths available within a platform and selects optimal data transmission paths among the available data transmission paths to generate a path set, a data transmission module that transmits data through the optimal data transmission paths, and a dynamic data classification module that monitors the data transmission module in real time to dynamically readjust the optimal data transmission path.

The above different libraries may include the NVIDIA Collective Communications Library (NCCL) and the Remote Direct Memory Access (RDMA) API (Application Programming Interface) library, and the interface module may optimize multi-GPU data transmission by transferring the number of GPUs to be used and IP information determined within the NCCL to the RDMA API library.

The above interface module can divide the data and transmit it through different paths only when the size of the data is greater than a threshold value.

The above automatic route collection module can detect network topology changes in real time and perform route search functions.

The above automatic route collection module can generate a number of virtual IP addresses within the server, transmit a search packet, and then compare the route information to which the search packet was transmitted with existing route information to select the optimal data transmission route.

The above automatic route collection module may include a new route in the route set if the route through which the search packet is transmitted is a new route.

The above dynamic data classification module can detect the network bandwidth and transmission delay of the data transmission module and dynamically readjust the optimal data transmission path.

The above dynamic data classification module can monitor performance indicators for the optimal data transmission paths, adjust the data transmission amount for each path, and optimize the transmission performance of the optimal data transmission paths.

An electronic device according to one embodiment may include a memory storing instructions and a processor, wherein the instructions, when executed by the processor, cause the electronic device to integrate different libraries, automatically detect and collect data transmission paths available within a platform, select optimal data transmission paths among the available data transmission paths to generate a path set, transmit data through the optimal data transmission paths, and monitor the data transmission in real time to dynamically readjust the optimal data transmission path.

A multi-path transmission method according to one embodiment may include a step of integrating different libraries, a step of automatically detecting and collecting data transmission paths available within a platform, a step of selecting optimal data transmission paths among the available data transmission paths to generate a path set, a step of transmitting data through the optimal data transmission paths, and a step of monitoring the data transmission in real time to dynamically readjust the optimal data transmission path.

FIG. 1 is a schematic diagram illustrating RDMA according to one embodiment.

FIG. 2 is a schematic diagram illustrating a multi-path transmission device according to one embodiment.

FIG. 3 is a schematic flowchart illustrating a multi-path transmission method according to one embodiment.

FIG. 4 is a block diagram illustrating an electronic device according to one embodiment.

Specific structural or functional descriptions of the embodiments are disclosed for illustrative purposes only and may be implemented in various forms. Accordingly, the actual implemented form is not limited to the specific embodiments disclosed, and the scope of the present disclosure includes modifications, equivalents, or alternatives included in the technical idea described in the embodiments.

Although the terms first or second may be used to describe various components, such terms should be construed only for the purpose of distinguishing one component from another. For example, a first component may be referred to as a second component, and similarly, a second component may also be referred to as a first component.

When it is said that a component is "connected" to another component, it should be understood that it may be directly connected or connected to that other component, but there may also be other components in between.

Singular expressions include plural expressions unless the context clearly indicates otherwise. In this specification, the terms "comprises" or "has" and the like are intended to specify the presence of a described feature, number, step, operation, component, part, or combination thereof, but should be understood to not preclude the presence or addition of one or more other features, numbers, steps, operations, components, parts, or combinations thereof.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art. Terms defined in commonly used dictionaries should be interpreted as having a meaning consistent with the meaning they have in the context of the relevant art, and will not be interpreted in an idealized or overly formal sense unless explicitly defined herein.

Hereinafter, embodiments will be described in detail with reference to the attached drawings. In describing with reference to the attached drawings, identical components are given the same reference numerals regardless of the drawing numbers, and redundant descriptions thereof will be omitted.

FIG. 1 is a schematic diagram illustrating RDMA according to one embodiment.

One or more of the blocks or combinations of blocks of FIG. 1 may be implemented by a special purpose hardware-based computer performing a particular function, or by a combination of special purpose hardware and computer instructions.

Remote Direct Memory Access (RDMA), according to one embodiment, may be a method for enabling direct memory access from the memory of one computer to the memory of another computer without using a computer operating system, a central processing unit, or a cache. RDMA enables high throughput, low latency networking, which may be particularly advantageous in data centers and high-performance computing environments.

RDMA enables networking without memory copying. RDMA allows data (or messages) to be transferred directly from one system's memory to another, thus reducing the need for data copying between buffers.

RDMA can be kernel bypassed. RDMA can bypass the operating system kernel, reducing context switching, interrupts, and central processing unit overhead.

RDMA can be hardware-based transport. RDMA can be managed and executed by a network adapter (RDMA NIC or RNIC), reducing the workload on the central processing unit.

RDMA can support a variety of network fabrics. RDMA can operate over a variety of network fabrics, such as InfiniBand, Ethernet (RDMA over RoCE), and Internet Wide Area RDMA Protocol (iWARP).

RDMA can have low latency and high efficiency. RDMA can significantly reduce latency and increase data transfer efficiency by reducing the number of data copies and context switches.

RDMA can be used in applications that require high data throughput and low latency, such as large database transactions, high-performance computing applications, and storage area networks. The ability to offload work from the central processing unit and minimize latency can be important in data-intensive computing environments.

FIG. 1 illustrates a detailed architecture and operational flow of RDMA in a network environment. In one embodiment, the main components of RDMA may be Queue Pairs (QPs) and Memory Regions (MRs). A QP, which consists of a transmit and receive queue, may be essential for initiating and managing RDMA communication between two hosts. A memory region may be a specific region of memory prepared to be directly accessed through an RDMA operation. Setting up a QP may include setting up a communication channel between hosts having transmit and receive queues. In an MR, a host may identify a specific memory region that can be directly accessed by an RDMA operation. In Work Queue Elements (WQEs), data transfer commands are queued in the QP, and these commands can specify the characteristics of the task, such as transmitting or receiving data. The Network Interface Card (NIC) can process the command by performing a memory transfer directly according to the instructions in the WQE. After completing the task, the NIC can report back through Completion Queue Elements (CQEs) to indicate the status of the RDMA task.

More specifically, FIG. 1 illustrates an overview of RDMA. It can show the process of setting up and executing an RDMA task between two hosts. The main elements include setting up a queue pair (QP) between the hosts, where each QP can consist of a transmit and receive queue. A memory region (MR) can be defined for direct access by a network interface card (NIC). A task can be initiated by posting a work queue element (WQE) to the QP. The NIC can then process this WQE and transmit data according to the instructions provided. The completion of this task can be indicated by a completion queue element (CQE). This setup can show the direct high-speed data transfer capability of RDMA, which reduces latency and increases the efficiency of data transfer between networked computers by bypassing the operating system and CPU.

Below, multipath transmission using RDMA is described.

Multipath transmission may be essential to fully utilize the various network links existing between servers in a data center. Multipath transmission can be utilized for the following two purposes: Multipath transmission is used to utilize a bypass route due to a failure of a specific network link, and data transmission capacity can be increased by utilizing multiple paths simultaneously.

However, since all of the existing technologies were developed based on the classic TCP/IP (Transmission Control Protocol/Internet Protocol)-based network, it may be impossible to apply them to RDMA, a high-performance data transmission technology.

Unlike TCP/IP, RDMA has all of its communication processes embedded in hardware, so it may be impossible to change the transmission logic of RDMA within the system kernel (OS). Therefore, in order to apply multi-path transmission technology to RDMA, 1) new hardware (RNIC, RDMA-only NIC card) must be developed, or 2) the technology must be embedded in the User-level library where the RDMA API is implemented. In the case of 1), it may be difficult to apply it to hardware that has already been purchased, and a method like 2) may require the development of a multi-path transmission technology. However, in the case of conventional technologies, they are technologies that mainly consider RDMA that performs data transmission targeting CPU Memory, and when applied to GPU-based RDMA communication (GPUDirect RDMA) recently used in artificial intelligence, etc., they may operate inefficiently or be difficult to utilize effectively.

The following description may be about a multi-path transmission technology that can be efficiently utilized for various platforms that operate based on high-performance computing utilizing GPUs.

FIG. 2 is a schematic diagram illustrating a multi-path transmission device according to one embodiment.

The description with reference to Fig. 1 can be equally applied to Fig. 2, and overlapping content can be omitted. One or more blocks and combinations of blocks of Fig. 2 can be implemented by a special purpose hardware-based computer performing a specific function, or a combination of special purpose hardware and computer instructions.

A multi-path transmission device according to one embodiment may include an interface module, an automatic path collection module, a data transmission module, and a dynamic data classification module. The multi-path transmission device may transmit data in multiple paths based on RDMA from a GPU. The multi-path transmission device may transmit data by efficiently linking between different libraries, automatically collecting available paths, and dynamically classifying data.

The term "module" can mean, for example, a unit comprising one or more of hardware, software, or firmware. "Module" can be used interchangeably with, for example, the terms unit, logic, logical block, component, or circuit. A "module" can be the smallest unit of an integrally formed component or a portion thereof. A "module" can also be the smallest unit or a portion thereof that performs one or more functions. A "module" can be implemented mechanically or electronically. For example, a "module" can include at least one of an application-specific integrated circuit (ASIC) chip, field-programmable gate arrays (FPGAs), or a programmable-logic device that performs certain operations, known or to be developed in the future.

The interface module can integrate different libraries, which can include the NVIDIA Collective Communications Library (NCCL) and the Remote Direct Memory Access (RDMA) Application Programming Interface (API) library.

NCCL is a library provided by NVIDIA that can support high-performance and scalable collective communication in a multi-GPU environment. NCCL is optimized for multi-GPU, supports collective communication operations, large-scale clusters, and is compatible with various network interfaces.

RDMA API can provide a programming interface for Remote Direct Memory Access. By using NCCL and RDMA API together, data transfer and collective operations between multi-GPUs can be efficiently processed in high-performance computing environments.

The interface module can optimize multi-GPU data transmission by transmitting the number of GPUs to be used and IP information determined within the NCCL to the RDMA API library. The interface module can divide the data and transmit them through different paths only when the size of the data is greater than or equal to a threshold value. For example, the interface module can perform multi-path transmission by dividing the data into multiple small pieces of data and transmitting them through different paths only when the data to be transmitted is larger than a threshold value (e.g., 1 MB).

The interface module can multiplex the paths through which data is transmitted by assigning different src IPs to the data by utilizing the ECMP characteristics and virtual IPs. ECMP (Equal-Cost Multi-Path) is an equal-cost multi-path routing method. ECMP can distribute traffic by using multiple paths with the same cost (e.g., delay time, number of hops) among multiple paths to the same destination. ECMP can reduce network congestion and improve overall network efficiency by transmitting traffic through multiple paths. The src IP (Source IP Address) can mean the source address of a data packet in a network, i.e., the 'source IP address'. The interface module can diversify the src IPs of data by utilizing the characteristics of ECMP and virtual IPs, thereby allowing data packets to be transmitted through multiple different paths. This can improve network load distribution, increase transmission efficiency, and provide higher data transmission reliability.

The interface module may choose to transmit data via a single path, since using multiple paths may be inefficient when the data is sufficiently small. Since data using a single path may be interrupted by data using multiple paths, data using a single path may be given a higher priority to prevent performance degradation.

The interface module can perform the role of converting and transmitting the information provided by NCCL into a form that the RDMA API can understand and utilize. The interface module can first collect information related to GPU usage from NCCL. This information can include the total number of GPUs to be used and the IP address assigned to each GPU.

The interface module can convert the collected information into a format that the RDMA API can understand. The interface module can pass this information to the RDMA API so that the RDMA can establish an efficient data transmission path between each GPU. Based on the information received from the interface module, the RDMA API can establish direct memory access to each GPU through the network. It can minimize network delay and optimize transmission speed during data transmission.

The interface module can monitor whether the interoperability between NCCL and RDMA API is smooth. The interface module can support optimization of data transmission and communication between NCCL and RDMA API.

The automatic path collection module can automatically detect and collect data transmission paths available within the platform, and select optimal data transmission paths among the available data transmission paths to create a path set. Examples of platforms include high-performance computing environments, cloud computing infrastructures, large-scale data centers, or enterprise environments with complex network architectures. The automatic path collection module can determine the optimal path by considering various factors such as network bandwidth, delay time, error rate, and congestion status according to the characteristics of each platform.

The automatic path collection module can detect changes in network topology in real time and perform a path search function. The network topology can mean the deterioration or logical structure of the network. For example, in a high-performance computing environment, the network topology can change over time. In a high-performance computing environment, changes in the network topology can occur for various reasons, such as the addition of new equipment (e.g., the addition of a switch), the removal of existing equipment (e.g., the removal of a server), or the change of network connections. The automatic path collection module can monitor the network in real time to continuously monitor the current status of the network, and detect and identify changes in the topology. After this, the automatic path collection module can recalculate the condensation and optimize the path to select the optimal data transmission path. This can be updated to the data transmission module so that data can be transmitted through the new path.

The automatic route collection module can create multiple virtual IP addresses within the server, send probe packets, and then compare the route information to which the probe packets were sent with the existing route information to select the optimal data transmission route.

The automatic route collection module can include a new route in the route set if the route along which the probe packet was sent is a new route.

In selecting a src IP for multiplexing the path of data, since the selection method according to the IP is different for each switch device, it may be difficult to confirm which path each IP address goes to before actually sending it. Therefore, the automatic route collection module can automatically find the number of available paths in the platform and the set of src IP addresses that should be used for each path. Here, the automatic route collection module creates a number of virtual IPs in the server, and transmits a search packet with the corresponding virtual IP through the path tracking function, and then compares the path information to which the corresponding packet was transmitted with the existing path information to determine whether it is a new path that did not exist before. If it is a new path, the automatic route collection module can add it to the existing path set and utilize the corresponding src IP address for multiple path transmission.

More specifically, the automatic route collection module can generate a virtual IP address. The virtual IP address can be used to experimentally explore various data transmission paths on the network. The virtual IPs can be implemented in software without being assigned to an actual network device. The automatic route collection module can transmit a probe packet. The probe packet can use the generated virtual IP address to determine which path will be used in an actual network situation through the network. The automatic route collection module can compare and analyze the path information. The automatic route collection module can compare the path through which the probe packet is transmitted with existing paths to discover a new path and verify the validity of the new path. The automatic route collection module can add a new path whose validity has been verified to the existing path set. The automatic route collection module can select a src IP for multi-path transmission. The automatic route collection module can select an effective src IP address for each path by considering the path selection that may be different for each network switch device. The automatic route collection module can perform a path search function by periodically performing the above-mentioned process.

The data transmission module can transmit data through optimal data transmission paths selected by the automatic path collection module.

The dynamic data classification module can monitor the data transmission module in real time and dynamically readjust the optimal transmission path. The dynamic data classification module can monitor network traffic patterns, bandwidth usage, or delay time in real time to understand the current status of the network.

The dynamic data classification module can detect the network bandwidth and transmission delay of the data transmission module and dynamically readjust the optimal data transmission path. When the dynamic data classification module detects changes in network congestion or bandwidth, it can readjust the optimal data transmission path on its own, and can also adjust the optimal data transmission path in conjunction with the automatic path collection module.

The dynamic data classification module can monitor performance indicators for optimal data transmission paths to adjust the amount of data transmitted per path and optimize the transmission performance of optimal data transmission paths.

The dynamic data classification module can multiplex paths according to the size of the data when the network conditions are poor and the available bandwidth is insufficient. For example, even for data of the same size, if the network conditions are good, it can be sent through a single path, but if the network conditions are poor, the paths may need to be multiplexed to divide the data into several small pieces and transmit them. The dynamic data classification module can determine whether to multiplex paths using the method of the mathematical expression 1 below. The method of determining whether to multiplex paths of the dynamic data classification module is not limited to the mathematical expression described, and various mathematical methods may be applied.

[Mathematical formula 1]

size_th = max(1byte, 1MB * (current_throughput / max_throughput)

The dynamic data classification module can dynamically set the threshold value (size_th) for data segmentation based on the ratio of the current throughput to the maximum throughput of the network. When the performance of the network deteriorates (e.g., when the current throughput is lower than the maximum throughput), the threshold value for data segmentation can be reduced to more actively utilize path multiplexing. For example, when the network is congested, the threshold value (e.g., from 1 MB to 0.5 MB) can be reduced to transmit smaller data through multiple paths, thereby alleviating network congestion. That is, while the multipath transmission device previously divided data with a size of 1 MB or more into multiple paths for multipath transmission, when a network status abnormality occurs, the dynamic data classification device can reduce the threshold value to 0.5 MB, thereby dividing data with a size of 0.5 MB or more into multiple paths for multipath transmission.

FIG. 3 is a schematic flowchart illustrating a multi-path transmission method according to one embodiment.

The description referring to FIG. 1 and FIG. 2 can be equally applied to FIG. 3, and any duplicate content can be omitted.

The operations of FIG. 3 may be performed in the order and manner illustrated, but the order of some operations may be changed or some operations may be omitted without departing from the spirit and scope of the illustrated embodiment. A plurality of operations illustrated in FIG. 3 may be performed in parallel or simultaneously.

For convenience of explanation, steps (310 to 340) are described as being performed using the multi-path transmission device (200) illustrated in FIG. 2. However, these steps (310 to 340) may be utilized via any other suitable electronic device and within any suitable system.

In step (310), the multi-path transmission device can integrate different libraries. For example, the multi-path transmission device can optimize multi-GPU data transmission by transmitting the number of GPUs to be used and IP information determined within the NCCL to the RDMA API library. The multi-path transmission device can divide the data and transmit it through different paths only when the size of the data is greater than or equal to a threshold value.

In step (320), the multi-path transmission device can automatically detect and collect available data transmission paths within the platform, and select optimal data transmission paths among the available data transmission paths to generate a path set. For example, the multi-path transmission device can detect a network topology change in real time to perform a path search function. The multi-path transmission device can generate a plurality of virtual IP addresses within the server, transmit a search packet, and then compare the path through which the search packet was transmitted with existing path information to select an optimal data transmission path. If the path through which the search packet was transmitted is a new path, the multi-path transmission device can include the new path in the path set.

In step (330), the multi-path transmission device can transmit data through optimal data transmission paths.

In step (340), the multi-path transmission device can monitor data transmission in real time and dynamically readjust the optimal data transmission path. For example, the multi-path transmission device can detect the network bandwidth and transmission delay of the data transmission module and dynamically readjust the optimal data transmission path. The multi-path transmission device can monitor performance indicators for the optimal data transmission paths and adjust the data transmission amount for each path and optimize the transmission performance of the optimal data transmission paths.

FIG. 4 is a block diagram illustrating an electronic device according to one embodiment.

One or more of the blocks and combinations of blocks of FIG. 4 may be implemented by a special purpose hardware-based computer performing a specific function, or a combination of special purpose hardware and computer instructions. The contents described with reference to FIGS. 1 to 3 may be equally applied to FIG. 4. For example, an electronic device (400) according to one embodiment may include a multi-path transmission device (200).

As shown in FIG. 4, the electronic device (400) may include a memory (410) and a processor (420). The electronic device (400) may further include a communication module, and the communication module may include a transmitter and a receiver.

An electronic device (400) according to one embodiment may include a memory (410) and a processor (420) connected to the memory (410) via a system bus or other suitable circuitry.

The electronic device (400) may store program code in memory (410). The memory (410) according to one embodiment may include one or more physical memory devices, such as local memory or one or more bulk storage devices. In this case, the local memory may include a random access memory (RAM) or other volatile memory device that is generally used while actually executing the program code. The bulk storage device may be implemented as a hard disk drive (HDD), a solid state drive (SSD), or other non-volatile memory device.

As the executable program code stored in the memory (410) is executed by the electronic device (400), various operations described in the present disclosure can be performed by the processor (420). For example, the memory (410) can store program code for causing the processor (420) to perform one or more operations described in FIGS. 1 to 3.

Depending on the particular type of device being implemented, the electronic device (400) may include fewer components than those illustrated or additional components not illustrated in FIG. 4. Additionally, one or more of the components may be incorporated into, or otherwise form part of, another component.

The processor (420) according to one embodiment is a hardware configuration that performs overall control functions for controlling operations of the electronic device (400). For example, the processor (420) may control the electronic device (400) overall by executing programs stored in the memory (410) within the electronic device (400). The processor (420) may be implemented as a central processing unit (CPU), a graphics processing unit (GPU), an application processor (AP), a neural processing unit (NPU), etc., provided within the electronic device (400), but is not limited thereto.

The processor (420) can integrate different libraries, automatically detect and collect data transmission paths available within the platform, select optimal data transmission paths among the available data transmission paths to create a path set, transmit data through the optimal data transmission paths, and monitor data transmission in real time to dynamically readjust the optimal data transmission path.

The embodiments described above may be implemented as hardware components, software components, and/or a combination of hardware components and software components. For example, the devices, methods, and components described in the embodiments may be implemented using a general-purpose computer or a special-purpose computer, such as, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), a programmable logic unit (PLU), a microprocessor, or any other device capable of executing instructions and responding to them. The processing device may execute an operating system (OS) and software applications running on the OS. In addition, the processing device may access, store, manipulate, process, and generate data in response to the execution of the software. For ease of understanding, the processing device is sometimes described as being used alone, but those skilled in the art will appreciate that the processing device may include multiple processing elements and/or multiple types of processing elements. For example, a processing device may include multiple processors, or a processor and a controller. Other processing configurations, such as parallel processors, are also possible.

The software may include a computer program, code, instructions, or a combination of one or more of these, which may configure a processing device to perform a desired operation or may independently or collectively command the processing device. The software and/or data may be permanently or temporarily embodied in any type of machine, component, physical device, virtual equipment, computer storage medium or device, or transmitted signal waves, for interpretation by the processing device or for providing instructions or data to the processing device. The software may also be distributed over network-connected computer systems and stored or executed in a distributed manner. The software and data may be stored on a computer-readable recording medium.

The method according to the embodiment may be implemented in the form of program commands that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program commands, data files, data structures, etc., alone or in combination, and the program commands recorded on the medium may be those specially designed and configured for the embodiment or may be those known to and available to those skilled in the art of computer software. Examples of the computer-readable recording medium include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, magneto-optical media such as floptical disks, and hardware devices specially configured to store and execute program commands such as ROMs, RAMs, flash memories, etc. Examples of program commands include not only machine language codes generated by a compiler, but also high-level language codes that can be executed by a computer using an interpreter, etc.

The hardware devices described above may be configured to operate as one or more software modules to perform the operations of the embodiments, and vice versa.

Although the embodiments have been described with limited drawings as described above, those skilled in the art can apply various technical modifications and variations based on them. For example, even if the described techniques are performed in a different order than the described method, and/or the components of the described system, structure, device, circuit, etc. are combined or combined in a different form than the described method, or are replaced or substituted by other components or equivalents, appropriate results can be achieved.

Therefore, other implementations, other embodiments, and equivalents to the claims are also included in the scope of the claims described below.

Claims (11)

In a multi-path transmission device, Interface module that integrates different libraries; An automatic path collection module that automatically detects and collects data transmission paths available within the platform and selects optimal data transmission paths among the available data transmission paths to create a path set; A data transmission module for transmitting data through the above optimal data transmission paths; and A dynamic data classification module that monitors the above data transmission module in real time and dynamically readjusts the optimal data transmission path. A multi-path transmission device comprising: In the first paragraph, The above different libraries are NVIDIA Collective Communications Library (NCCL) and Remote Direct Memory Access (RDMA) Application Programming Interface (API) libraries Including, The above interface module A multi-path transmission device that optimizes multi-GPU data transmission by transmitting the number of GPUs to be used and IP information determined within the above NCCL to the above RDMA API library. In the second paragraph, The above interface module A multi-path transmission device that divides the data and transmits it through different paths only when the size of the data is greater than a threshold value. In the first paragraph, The above automatic route collection module A multi-path transmission device that detects network topology changes in real time and performs a path search function. In the first paragraph, The above automatic route collection module A multi-path transmission device that creates multiple virtual IP addresses within a server, transmits a search packet, and then compares the route information along which the search packet was transmitted with existing route information to select the optimal data transmission route. In paragraph 5, The above automatic route collection module A multi-path transmission device that includes the new path in the path set if the path along which the above search packet is transmitted is a new path. In the first paragraph, The above dynamic data classification module A multi-path transmission device that detects the network bandwidth and transmission delay of the above data transmission module and dynamically readjusts the optimal data transmission path. In the first paragraph, The above dynamic data classification module A multi-path transmission device that monitors performance indicators for the above optimal data transmission paths, adjusts the data transmission amount for each path, and optimizes the transmission performance of the above optimal data transmission paths. In electronic devices, Memory for storing instructions; and Processor Including The above instructions, when executed by the processor, cause the electronic device to: Integrate different libraries, Automatically detects and collects data transmission paths available within the platform, selects optimal data transmission paths among the available data transmission paths, and creates a set of paths. Transmitting data through the above optimal data transmission paths, An electronic device that monitors the above data transmission in real time and dynamically readjusts the optimal data transmission path. In a multi-path transmission method, Steps to integrate different libraries; A step of automatically detecting and collecting data transmission paths available within a platform, and selecting optimal data transmission paths among the available data transmission paths to generate a path set; A step of transmitting data through the above optimal data transmission paths; and A step of monitoring the above data transmission in real time and dynamically readjusting the optimal data transmission path. A multi-path transmission method comprising: A computer program stored on a computer-readable recording medium for executing the method of claim 10 in combination with hardware.

Images