Multi-intelligent computing center interconnection scheme based on wavelength sharing ring and implementation method

Abstract

The invention provides a multi-intelligent computing center interconnection scheme based on a wavelength sharing ring, and aims to improve the communication efficiency and the resource utilization rate among intelligent computing centers so as to support efficient model parallelization training. The method comprises three parts: an internet based on a shared wavelength ring is constructed, and each center is equipped with a reconfigurable optical add-drop multiplexer to realize flexible uplink and downlink operation of wavelength resources; on the basis of an interaction mechanism of a triple (time label, duration and topological connection), a model controller and a network controller can efficiently cooperate, and network resources are dynamically configured according to training task requirements; through a cooperation mechanism, close cooperation between a model training process and a network dynamic control process is ensured, and effective distribution and utilization of resources are realized. According to the invention, through sharing the wavelength ring and dynamic resource configuration, the data transmission efficiency between the intelligent computing centers is improved, the waste of network resources is greatly reduced, and an efficient and flexible interconnection solution is provided for model parallelization training.

Description

Technical Field

[0001] This invention relates to the field of intelligent computing center network technology, and in particular to a multi-intelligent computing center interconnection scheme and implementation method based on wavelength sharing ring. Background Technology

[0002] In an era of rapid development driven by data and artificial intelligence, data centers capable of providing powerful computing capabilities—Intelligent Computing Centers (ICCs)—have become the core infrastructure supporting large-scale model training and applications. In recent years, with the rapid rise and widespread application of large-scale artificial intelligence models, the demand for computing power in ICCs has exploded, driving the construction scale of ICCs towards the level of over ten thousand computing cards (such as GPUs). However, the current expansion of computing power in a single ICC faces limitations in physical space and rigid constraints in power supply capabilities, which severely restricts further breakthroughs and development in large-scale model technology.

[0003] To overcome this bottleneck, the industry has begun to explore interconnecting multiple intelligent computing centers through high-performance networks, using parallel computing technology to rationally distribute complex models to various intelligent computing centers, and realize cross-center distributed training to significantly improve the scale and complexity of the models that can be processed.

[0004] In achieving this goal, the following key technical issues need to be addressed:

[0005] 1) Ensuring high reliability and stability of data transmission between various computing centers during model training and reducing the risk of training interruption due to network failures is the primary concern.

[0006] 2) How to build a seamless connection and interaction between the training model and the high-performance network to ensure that information such as model data and parameter updates can be transmitted efficiently and accurately is the key to improving the overall training efficiency.

[0007] 3) Designing an effective collaboration mechanism between training tasks and high-performance networks to achieve dynamic optimization of resource allocation and ensure efficient execution of the distributed training process is an important way to improve the overall performance of the system. Summary of the Invention

[0008] To overcome the shortcomings of the existing technologies, this invention designs a multi-computing center interconnection scheme based on a wavelength-shared ring, specifically including: 1. a performance interconnection network based on a shared wavelength ring; 2. an interaction mechanism between the model controller and the network controller based on the {time tag, duration, topology connection} triple; 3. a collaborative mechanism between the model training process and the network dynamic control process.

[0009] The technical solution of the present invention is as follows:

[0010] On the one hand, this invention provides a multi-computing center interconnection scheme based on a wavelength-shared ring, characterized in that it includes:

[0011] An interconnection network based on a shared wavelength ring consists of multiple ICCs interconnected through a wavelength sharing ring. Each ICC is connected to a wavelength sharing ring via wavelength division multiplexing technology and is equipped with at least one ROADM device for performing optical signal off-path and up-path operations within a predetermined wavelength range.

[0012] The interaction mechanism between a MC and a NC involves the MC generating traffic transmission configuration information containing a triplet of {time tag, duration, topology connection}. The NC then performs dynamic wavelength allocation based on this traffic transmission configuration information, generates network configuration information, and achieves this through the following steps:

[0013] MC determines the time label and duration of the model training task, as well as the topology of the required transmission traffic;

[0014] The MC sends the traffic transmission configuration information to the NC;

[0015] Based on the traffic transmission configuration information, the NC allocates at least one dedicated wavelength to each ICC and allocates shared wavelengths according to traffic requirements;

[0016] A collaborative mechanism for communication between MC and NC to achieve coordination between the model training process and the network dynamic control process, specifically including:

[0017] The NC distributes network configuration information to each ICC;

[0018] Each ICC adjusts its ROADM device based on the received network configuration information to perform data transmission;

[0019] After data transmission is complete, each ICC releases its corresponding wavelength resources.

[0020] Furthermore, the aforementioned internet also includes:

[0021] A unidirectional wavelength-sharing ring structure that supports at least N wavelength channels, each with a predetermined bandwidth;

[0022] Each ICC dynamically accesses and de-accesses the N wavelength channels via a ROADM device. The specific wavelengths for access and de-access are determined by the traffic transmission configuration information generated by the MC.

[0023] Furthermore, the collaborative mechanism between MC and NC also includes:

[0024] The NC monitors the data transmission status of the ICC and adjusts the wavelength allocation strategy according to the network conditions.

[0025] On the other hand, the present invention also provides a method for implementing the above-mentioned interconnection scheme of multiple intelligent computing centers based on wavelength sharing rings, characterized by comprising the following steps:

[0026] Step 1. Configure the intelligent computing center, reconfigurable optical add-drop multiplexer equipment, model controller, and network controller;

[0027] Step 2. The model controller transmits the traffic transmission configuration information generated by the training task to the network controller;

[0028] Step 3. The network controller performs dynamic wavelength allocation based on the traffic transmission configuration information and generates network configuration information;

[0029] Step 4. The network controller distributes the dynamic wavelength allocation scheme to each intelligent computing center;

[0030] Step 5. Each intelligent computing center completes the configuration according to the network configuration information and performs data transmission;

[0031] Step 6. After each intelligent computing center completes data transmission, it releases the relevant transmission resources.

[0032] Compared with the prior art, the beneficial effects of the present invention are:

[0033] (1) The interconnection scheme can provide reliability for parallel training of the model. The ring structure adopted by the interconnection scheme has link protection capability. When a certain intelligent computing center (or the working unit that performs the training task) fails, the NC can stably start the fault-tolerant scheme on the ring according to the transmission resources.

[0034] (2) This interconnection scheme enables effective interaction between the training model and the high-performance network. The interconnection scheme adopts an interaction mechanism between the model controller and the network controller based on the triplet {time label, duration, topology connection}. The network controller can sense the transmission requirements of the training model and perform corresponding network configuration, thereby realizing effective interaction between the training model and the high-performance network.

[0035] (3) This interconnection scheme enables collaborative work between the training task and the high-performance network. In this interconnection scheme, the MC can share the traffic transmission information required for the training task with the NC in advance, thus promoting collaborative work between the training task and the high-performance network. Attached Figure Description

[0036] Figure 1 Interconnection of Multi-Computing Centers Based on Wavelength Sharing Ring

[0037] Figure 2Interaction mechanism between model controller, network controller and network devices

[0038] Figure 3 System workflow diagram Detailed Implementation

[0039] The present invention will now be described in detail with reference to the accompanying drawings and embodiments, but this should not be construed as limiting the scope of protection of the present invention.

[0040] I. Performance Interconnect Networks Based on Wavelength Sharing Rings

[0041] (1) Interconnection method: such as Figure 1 As shown, the various Intelligent Computing Centers (ICCs) are interconnected in a ring configuration, forming a wavelength-sharing ring. Each ICC on the ring uses WDM (Wavelength Division Multiplexing) communication. WDM technology can transmit multiple optical signals of different wavelengths in the same optical fiber, thus greatly improving the fiber's transmission capacity. The wavelength-sharing ring supports the transmission of {L = N + M} wavelengths, where N is the number of dedicated wavelengths and M is the number of shared wavelengths.

[0042] (2) Transmission method: The wavelength sharing ring structure uses a unidirectional method to transmit the traffic of each intelligent computing center. The traffic here includes training sample data, data involved in model parameter calculation, etc.

[0043] (3) Equipment Configuration: Each Intelligent Computing Center (ICC) is equipped with a ROADM (Reconfigurable Optical Add-Drop Multiplexer) device. The ROADM device can perform drop-through and add-through operations on wavelengths on the ring, realizing flexible wavelength resource scheduling. The drop-through and add-through methods of the ROADM are determined by the structure of the ROADM used. Each Intelligent Computing Center uses the ROADM device to drop-through and add-through wavelength resources on the ring as needed. This on-demand operation is determined by the parallel training method, that is, dynamically allocating wavelength resources according to the needs of the training task, inserting or splitting optical signals of different wavelengths into the transmission path, and realizing flexible wavelength resource scheduling.

[0044] II. Interaction Mechanism between Model Controller and Network Controller Based on the {Time Tag, Duration, Topology Connection} Triple

[0045] (1) The MC generates traffic transmission configuration information: such as Figure 2 As shown, the MC (Model Controller) generates traffic transmission configuration information {t} based on the training task (which includes training one or more models). k ,δ k ,P k}, where t k δ represents the time label of the model training state k. k P represents the duration of the model training state k. k This represents the topology connection of the traffic required to transmit in model training state k.

[0046] (2) NC performs dynamic wavelength allocation: The NC (Network Controller) transmits configuration information {t} based on the traffic generated by the MC. k ,δ k ,P k} Perform dynamic wavelength allocation and generate network configuration information {λ i ,δ i Each intelligent computing center is allocated at least one dedicated wavelength to ensure connectivity, and shared wavelengths are allocated to it according to the traffic requirements of different intelligent computing centers.

[0047] (3) NC sends network configuration information: The NC sends the network configuration information generated by the dynamic wavelength allocation to each intelligent computing center. After each intelligent computing center completes the data transmission, it releases the relevant transmission resources.

[0048] III. The Coordination Mechanism Between Model Training and Network Dynamic Control

[0049] (1) As Figure 2 As shown, MC and NC share the traffic transmission configuration information generated by the training task, so that NC can obtain the transmission requirements in advance and generate network configuration information.

[0050] (2) Figure 1 As shown, NC makes full use of the transmission resources on the shared wavelength ring by coordinating and sharing them.

[0051] (3) MC and NC can use such a collaborative mechanism to enable the efficient operation of the model training process and the network dynamic control process.

[0052] IV. Implementation Steps

[0053] Step 301: Configure the intelligent computing center and ROADM equipment, using an interconnection method based on a wavelength-shared ring. For example... Figure 1 As shown, N ICC201s, denoted as ICC1, ICC2, ..., ICCN, are each equipped with a corresponding ROADM device (105) and interconnected using a wavelength-sharing ring-based interconnection method (203). The ring has a total of L resources, L = N + M wavelengths (202): N dedicated wavelengths, denoted as λ1, λ2, ..., λ... N And M shared wavelengths, denoted as λ s1 , λ s2 、…、λ sMConfigure NC104 and MC103.

[0054] Step 302: MC103 transmits the traffic transmission configuration information 101 generated by the training task to NC104, such as... Figure 2 As shown.

[0055] Step 303: NC104 performs dynamic wavelength allocation based on traffic transmission configuration information 101 to generate network configuration information 102.

[0056] Step 304: NC104 distributes the dynamic wavelength allocation scheme to each intelligent computing center 201, allocating λ1 and λ2 to ICC1. s1 Assign λ2 to ICC2 and λ to ICCN. N , λ s2 and λ s3 ,like Figure 1 As shown.

[0057] Step 305: Each intelligent computing center 201 completes the configuration according to the network configuration information 102 and performs data transmission.

[0058] Step 306: After each intelligent computing center completes data transmission, it releases the relevant transmission resources.

[0059] In summary, this embodiment of the multi-computing center interconnection scheme based on wavelength-shared rings achieves efficient collaborative operation of the model training process and the network dynamic control process by optimizing transmission methods, equipment configuration, information sharing, and resource coordination. This scheme helps improve the overall performance of intelligent computing centers, accelerates model training speed, and provides strong support for the development of artificial intelligence technology.

Claims

1. A multi-computing center interconnection scheme based on wavelength-shared rings, characterized in that, include: An interconnection network based on a shared wavelength ring consists of multiple ICCs interconnected through a wavelength sharing ring. Each ICC is connected to a wavelength sharing ring via wavelength division multiplexing technology and is equipped with at least one ROADM device for performing optical signal off-path and up-path operations within a predetermined wavelength range. The interaction mechanism between a MC and a NC involves the MC generating traffic transmission configuration information containing a triplet of {time stamp, duration, topology connection}, and the NC performing dynamic wavelength allocation based on this traffic transmission configuration information to generate network configuration information. This is achieved through the following steps: MC determines the time label and duration of the model training task, as well as the topology of the required transmission traffic; The MC sends the traffic transmission configuration information to the NC; Based on the traffic transmission configuration information, the NC allocates at least one dedicated wavelength to each ICC and allocates shared wavelengths according to traffic requirements; A collaborative mechanism for communication between the MC and NC to synchronize the model training process and the network dynamic control process, specifically including: The NC distributes network configuration information to each ICC; Each ICC adjusts its ROADM device based on the received network configuration information to perform data transmission; After data transmission is complete, each ICC releases its corresponding wavelength resources.

2. The multi-computing center interconnection scheme based on wavelength sharing ring according to claim 1, characterized in that, The aforementioned internet also includes: A unidirectional wavelength-sharing ring structure that supports at least N wavelength channels, each with a predetermined bandwidth; Each ICC dynamically accesses and de-accesses the N wavelength channels via a ROADM device. The specific wavelengths for access and de-access are determined by the traffic transmission configuration information generated by the MC.

3. The multi-computing center interconnection scheme based on wavelength sharing ring according to claim 1, characterized in that, The aforementioned collaboration mechanism between MC and NC also includes: The NC monitors the data transmission status of the ICC and adjusts the wavelength allocation strategy according to the real-time network conditions. MC adjusts the execution plan of the model training task based on the network status feedback from NC.

4. A method for implementing the multi-computing center interconnection scheme based on wavelength-shared rings as described in any one of claims 1-3, characterized in that, Includes the following steps: Step 1. Configure the intelligent computing center, reconfigurable optical add-drop multiplexer equipment, model controller, and network controller; Step 2. The model controller transmits the traffic transmission configuration information generated by the training task to the network controller; Step 3. The network controller performs dynamic wavelength allocation based on the traffic transmission configuration information and generates network configuration information; Step 4. The network controller distributes the dynamic wavelength allocation scheme to each intelligent computing center; Step 5. Each intelligent computing center completes the configuration according to the network configuration information and performs data transmission; Step 6. After each intelligent computing center completes data transmission, it releases the relevant transmission resources.

Images