Power consumption energy-saving method, device and equipment based on optical module multi-channel convergence technology

Abstract

The invention provides a power consumption saving method, device and equipment based on an optical module multi-channel convergence technology. The power consumption saving method based on the optical module multi-channel convergence technology comprises the following steps: acquiring transmission flow upper limit values of N serializer-deserializer channels, wherein N is a positive integer; obtaining a current traffic value at the convergence port; and comparing the current flow value with the transmission flow upper limit value, and enabling a corresponding number of serializer-deserializer channels according to a comparison result. The embodiment of the invention aims to realize dynamic and accurate control of the number of channels of the multi-channel optical module, and power consumption is remarkably reduced while communication quality is ensured.

Description

Technical Field

[0001] This disclosure relates to the field of supernode computing interconnection communication, specifically to power saving methods, devices, and equipment based on optical module multi-channel convergence technology. Background Technology

[0002] In the design of supernode computer systems, a large number of optical modules are used. To improve the communication bandwidth and energy efficiency of optical modules, most are multi-channel optical module technologies, such as QSFP56-200G and OSFP56-400G. Due to the demands of parallel computing, most current supernode computers use full-mesh interconnect structures to achieve all-to-all communication topologies, resulting in massive GPU interconnect communication requirements. This also heavily relies on multi-channel optical module technology. Examples include Huawei's Cloud Matrix384 supernode computing cluster and the upcoming Dongfang Suanxin DF500-256 supernode computer. The energy consumed by these optical modules can account for 30%-40% of the entire computer's power consumption. Furthermore, optical modules do not support online upgrades to increase supported speeds; this means that once powered on, the configured speed remains unchanged. Initially, to meet the growing communication demands, supernode designs typically plan to use 200G, 400G, or even 800G optical modules. However, these optical modules do not carry data at full speed and bandwidth; in actual communication, low traffic situations often occur. Therefore, based on existing technology, dynamically adjusting the rate and bandwidth of these interfaces to save energy consumption, thereby improving energy efficiency and reducing OPEX, has become a technical problem that needs to be solved.

[0003] Therefore, existing technologies still need to be improved and enhanced.

[0004] It should be noted that the above introduction to the technical background is only for the purpose of providing a clear and complete explanation of the technical solutions of this application and facilitating understanding by those skilled in the art. It should not be assumed that these technical solutions are known to those skilled in the art simply because they have been described in the background section of this application. Summary of the Invention

[0005] To address at least one of the aforementioned problems, as well as one or more other potential problems, this disclosure proposes a power-saving method, equipment, and medium based on multi-channel convergence technology for optical modules. By introducing a flow shaping and prediction module, it solves the problems of excessive power consumption of multi-channel optical modules during low-flow periods, unintelligent dynamic adjustment mechanisms, and communication interruptions caused by channel switching in the prior art.

[0006] To achieve the above objectives, this disclosure provides the following technical solution: In the first aspect of the present disclosure, a power consumption saving method based on optical module multi-channel aggregation technology is proposed, which includes: obtaining the transmission traffic upper limit value of N serializer-deserializer channels, where N is a positive integer; obtaining the current traffic value at the aggregation port; comparing the current traffic value with the transmission traffic upper limit value, and enabling the corresponding number of serializer-deserializer channels according to the comparison result; wherein, the determination step of the number M of enabled serializer-deserializer channels includes: when the current traffic value ≤ (1 / N) * the transmission traffic upper limit value, M = 1; when (K - 1) / N * the transmission traffic upper limit value < the current traffic value ≤ K / N * the transmission traffic upper limit value, M = K, where K is an integer and 1 < K < N; when the current traffic value > (N - 1) / N * the transmission traffic upper limit value, M = N.

[0007] Further, in some embodiments, while enabling M serializer-deserializer channels, the remaining channels of the N serializer-deserializer channels except the M channels are closed.

[0008] Further, in some embodiments, the current traffic value is obtained through a traffic shaping module, and the traffic shaping module includes: multiple FIFO buffers for temporarily storing data to be sent; a water level monitoring unit for monitoring the data volume of each FIFO; a prediction unit for predicting the traffic trend based on the water level data and historical traffic characteristics; a control unit for outputting a channel enable control signal according to the prediction result.

[0009] Further, in some embodiments, the prediction unit uses a deep learning model for traffic prediction, and the deep learning model includes an LSTM or a CNN-LSTM network.

[0010] Further, in some embodiments, the FIFO buffer is provided with a high water level line and a low water level line. When the data volume of any FIFO exceeds the high water level line, enabling a new channel is triggered; when the data volumes of all FIFOs continue to be lower than the low water level line, closing redundant channels is triggered.

[0011] Further, in some embodiments, the optical module adopts OTP storage for fixed configuration, and no PLL reconfiguration is required when enabling channels.

[0012] In the second aspect of the present disclosure, a power consumption saving device based on optical module multi-channel aggregation technology is further proposed, which includes: an initialization module for obtaining the transmission traffic upper limit value of N serializer-deserializer channels, where N is a positive integer; a monitoring module for obtaining the current traffic value at the aggregation port; a monitoring and control module for dynamically enabling or closing the corresponding number of serializer-deserializer channels according to the current traffic value.

[0013] In the third aspect of the present disclosure, a power consumption saving device based on optical module multi-channel aggregation technology is further proposed, which includes: an initialization module for obtaining the upper limit value of the transmission traffic of N serializer-deserializer channels, where N is a positive integer; a monitoring module for obtaining the current traffic value at the aggregation port; a monitoring control module for comparing the current traffic value with the upper limit value of the transmission traffic and controlling the enabling of the corresponding number of serializer-deserializer channels according to the comparison result; wherein, the determination step of the number M of enabled serializer-deserializer channels includes: when the current traffic value ≤ (1 / N) * the upper limit value of the transmission traffic, M = 1; when (K - 1) / N * the upper limit value of the transmission traffic < the current traffic value ≤ K / N * the upper limit value of the transmission traffic, M = K, where K is an integer and 1 < K < N; when the current traffic value > (N - 1) / N * the upper limit value of the transmission traffic, M = N. Further, in some embodiments, the monitoring control module is configured to: while enabling M serializer-deserializer channels, close the remaining channels of the N serializer-deserializer channels except the M channels.

[0014] In the fourth aspect of the present disclosure, an electronic device is further proposed, which includes a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps of the method described above are implemented.

[0015] In the fifth aspect of the present disclosure, a computer-readable storage medium is further proposed, in which at least one instruction, at least one program, a code set or an instruction set is stored, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement the steps of the method described above.

[0016] The present disclosure has the following beneficial effects compared with the prior art:

[0017] (1) In some embodiments, it has a significant energy-saving effect: by finely associating traffic monitoring with channel enabling, the system can maintain only a few channels working during low-load periods, and turn off the optoelectronic components of most idle channels, directly reducing the dynamic power consumption of the optical module. Tests and simulations show that under the typical supernode computing load curve, the method in some embodiments of the present disclosure can reduce the overall power consumption of the optical module by 30% to 50%. In scenarios with obvious communication intermittency (such as between batch processing tasks) or periodic collective operations (such as All-Reduce), the energy-saving effect is more prominent, and in some scenarios, the energy-saving amplitude can exceed 50%. For optical modules with a power consumption accounting for 30% - 40% of the entire machine, this means a significant improvement in the overall energy efficiency of the machine and savings in operating costs (electricity bills).

[0018] (2) In some embodiments, high performance and low latency are guaranteed: This method introduces an intelligent decision-making mechanism based on FIFO water level and prediction algorithm. It is not only reactive but also predictive. The system can activate the backup channel in advance before the traffic substantially impacts the bandwidth bottleneck, effectively avoiding instantaneous congestion, queue overflow and data packet loss caused by channel switching delay, thereby ensuring the stable low-latency communication required by high-performance computing applications. The traffic shaping module itself also plays a smoothing role in burst traffic and improves communication quality.

[0019] (3) In some embodiments, seamless channel switching and uninterrupted communication are possible: By using optical modules configured based on OTP, this method completely avoids the lengthy PLL reconfiguration process required for traditional SerDes channel switching. The enabling and disabling of channels can be quickly switched at the physical layer, which is transparent and lossless to upper-layer protocols and applications, meeting the stringent requirements of continuous high availability of supernode computing systems 24 / 7.

[0020] (4) In some embodiments, it features system-level optimization and ease of integration: This disclosure proposes an end-to-end complete solution from host-side traffic management to physical layer optical module control. It tightly integrates FIFO management, traffic prediction algorithms, and optical module hardware characteristics to form a collaboratively optimized system. This solution can be implemented through software-driven updates or partially or entirely through dedicated hardware logic (such as control units in ASICs or FPGAs), and is easy to integrate and deploy on existing supernode computing platform architectures.

[0021] (5) In some embodiments, it exhibits good scalability and adaptability: the core logic (threshold-based multi-level control) and architecture (monitoring-prediction-control) of the method do not depend on a specific number of channels (N) or single-channel rate. Therefore, it can adapt to multi-channel optical modules ranging from 100G to 800G and even higher rates in the future. The prediction algorithm module can also be continuously trained and optimized based on the communication mode data of the actual deployment environment, possessing good adaptive capabilities. Attached Figure Description

[0022] The above and other features, advantages and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description, wherein: Figure 1 The diagram illustrates the interconnection between the optical module and the GPU module according to some examples; Figure 2 A schematic diagram illustrating the flow determination and serializer-deserializer enable relationship according to some embodiments of this disclosure is shown; Figure 3A flowchart illustrating a power-saving method based on optical module multi-channel convergence technology according to some embodiments of the present disclosure is shown. Figure 4 Schematic diagrams of power-saving devices based on optical module multi-channel convergence technology according to some embodiments of the present disclosure are shown; and In the various figures, the same or corresponding reference numerals indicate the same or corresponding parts. Detailed Implementation

[0023] Embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present disclosure. It should be understood that the accompanying drawings and embodiments of the present disclosure are for illustrative purposes only and are not intended to limit the scope of protection of the present disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0024] In the description of embodiments of this disclosure, the term "comprising" and similar terms should be understood as open-ended inclusion, i.e., "including but not limited to". The term "based on" should be understood as "at least partially based on". The term "one embodiment" or "this embodiment" should be understood as "at least one embodiment". The terms "first", "second", etc., may refer to different or the same objects. Other explicit and implicit definitions may also be included below. It should also be understood that the term "and / or" as used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0025] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, a first state may also be referred to as a second state, and similarly, a second state may also be referred to as a first state. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0026] It should be understood that in the field of server communication, optical modules are often used for interconnection between GPU modules, such as... Figure 1As shown, multi-channel optical modules are frequently used due to space and future bandwidth requirements. The communication paradigm generally involves, for example, multiple sets of SerDes signals from an OAM module entering an OCS optical module for muxing, then being converted from photoelectric signals to optical signals, which are then converted back to SerDes signals in the next OCS optical module before entering another GPU OAM module. Generally, the SerDes interface between OAM modules can be aggregated; SerDes is short for serializer and deserializer. Communication between OAM modules is typically via GPU interfaces. Common interfaces include PCIe, NVLink, and Ethernet. Because the data transmission rate is very high, the SerDes method is now mainly used. For example, in the Ethernet field, LAG (link group aggregation) technology aggregates four 50G PAM4 ports into one 200G Port. When the traffic is less than 75% of 50G (the total 200G only uses 75% of 50G), the other three SerDes can be turned off to save optical module energy consumption. If the traffic starts to increase, the second SerDes is enabled, and so on.

[0027] It should also be understood that currently, there are some technologies that attempt to save energy by dynamically adjusting link rates or aggregating link states. For example, Link Aggregation (LAG) technology in Ethernet can bundle multiple physical ports into a single logical port and perform load balancing among member links based on traffic. Some Energy Efficient Ethernet (IEEE) standards also support entering low-power modes during low traffic. However, these technologies face significant limitations when applied to optical module scenarios: 1) Coarse granularity and delayed response: Traditional energy-saving mechanisms are often based on fixed, periodic detection or simple timeout mechanisms, which cannot provide fine-grained, real-time responses to rapidly changing service traffic (such as sudden high-traffic spikes in All-Reduce communication). Frequent and inaccurate channel switching can lead to performance jitter or packet loss. 2) Lack of intelligent prediction: Existing solutions are mostly reactive adjustments, that is, decisions are made based on the current instantaneous traffic, lacking forward-looking prediction of traffic trends. Before a rapid increase in traffic, it is impossible to prepare additional channels in advance, which may lead to temporary congestion; when traffic decreases, conservative strategies delay channel shutdown, wasting energy. 3) Configuration overhead and interruption issues: For SerDes-based channels, switching operations typically involve physical layer (PHY) reconfiguration, such as phase-locked loop (PLL) relocking. This process can take tens or even hundreds of milliseconds, causing temporary interruptions to the communication link, which is unacceptable for computing tasks with high availability requirements. 4) Low system integration: Existing solutions are mostly fragmented optimizations at the protocol or driver layers, lacking a complete solution that deeply integrates host traffic management, data buffer (FIFO) scheduling, and physical layer optical module control at the system architecture level. Therefore, there is an urgent need in this field for a technical solution that can overcome the above shortcomings and achieve refined, intelligent, and lossless dynamic management of multi-channel optical module energy consumption, in order to significantly improve the overall performance per watt of supernode computing systems and reduce operating costs (OPEX).

[0028] The following will provide a more detailed explanation in conjunction with the accompanying drawings.

[0029] In some embodiments, this is a power-efficient system based on an 8-channel 400G optical module. For example, this system can be applied to an AI training cluster with a fully interconnected architecture. Each compute node is equipped with multiple GPUs, and the nodes are interconnected via 400G QSFP-DD optical modules. Each optical module contains eight independent 56G PAM4 SerDes channels.

[0030] Regarding the hardware platform: The host uses a self-developed acceleration card that supports PCIe 5.0 and high-speed Ethernet. The card integrates a network interface controller (NIC) and the power saving control unit disclosed herein (implemented with FPGA logic). The optical module uses a custom model that supports receiving independent enable commands for each channel via the I2C / MDIO interface, and the internal laser driver configuration is embedded in the OTP.

[0031] Regarding software / firmware configuration: In the host driver, set the total number of channels N=8, the maximum bandwidth per channel to 56Gbps, and the maximum total bandwidth to 448Gbps. Based on the power-saving method for multi-channel aggregation technology of optical modules disclosed herein, set 7 threshold points, corresponding to 1 / 8, 2 / 8, ..., 7 / 8 of the total bandwidth, forming 8 control intervals.

[0032] Regarding the traffic management pipeline: Data packets originating from GPU memory first enter the NIC's transmit queue. This system adds eight parallel transmit FIFOs (8KB deep) after this queue. A scheduler distributes traffic relatively evenly across these eight FIFOs based on the destination address hash of the data packets. Each FIFO is connected to an independent SerDes channel transmit logic.

[0033] Regarding the workflow of the intelligent control unit: The water level monitoring circuit within the control unit reads the current data volume of the eight FIFOs in real time. Further, the prediction subunit executes every 10 microseconds: it acquires the water level sequence of each FIFO over the past 100 microseconds and inputs it into a pre-trained simplified LSTM model (this model has learned the communication patterns of typical training tasks for this cluster). The model outputs a flow rate prediction for the next 20 microseconds (a total of 8 levels, corresponding to the number of channels to be activated). Further still, the control logic compares the prediction results with the currently activated number of channels. If more channels are predicted, an "Enable Channel X" command is immediately sent to the optical module via the I2C bus; if fewer channels are predicted, a conservative timer is set, and a "Disable Channel Y" command is only sent after a low flow rate prediction has persisted for a certain period (e.g., 50 microseconds) to prevent jitter. Furthermore, after receiving the command, the optical module typically completes the lighting or extinguishing of the specified channel laser within 5-20 microseconds, with power consumption varying accordingly.

[0034] It should be understood that the power consumption of optical modules typically accounts for 30-40% of the total power consumption of the device. However, through some embodiments of this disclosure, the power consumption of optical modules can be reduced by 30-50%, and in All-Reduce communication scenarios, it can even be reduced by more than 50%.

[0035] Further, the present disclosure provides a power consumption saving method based on optical module multi-channel aggregation technology. The method 100 includes the following steps: Step 110, obtaining the upper limit value of the transmission traffic of N serializer-deserializer channels, where N is a positive integer; Step 120, obtaining the current traffic value at the aggregation port; Step 130, comparing the current traffic value with the upper limit value of the transmission traffic, and enabling the corresponding number of serializer-deserializer channels according to the comparison result; wherein, the determination step of the number M of enabled serializer-deserializer channels includes: when the current traffic value ≤ (1 / N) * the upper limit value of the transmission traffic, M = 1; when (K-1) / N * the upper limit value of the transmission traffic < the current traffic value ≤ K / N * the upper limit value of the transmission traffic, M = K, where K is an integer and 1 < K < N; when the current traffic value > (N-1) / N * the upper limit value of the transmission traffic, M = N. It should be understood that for "when the current traffic value ≤ (1 / N) * the upper limit value of the transmission traffic, M = 1" in the determination step of the number M of enabled serializer-deserializer channels; it can be considered that when it is detected that the current traffic value is exactly less than or equal to 1 / N times the upper limit value of the transmission traffic, then only 1 serializer-deserializer channel needs to be enabled to carry the current traffic value, and then the remaining serializer-deserializer channels can be just closed, so as to achieve energy saving. Further, for "when (K-1) / N * the upper limit value of the transmission traffic < the current traffic value ≤ K / N * the upper limit value of the transmission traffic, M = K, where K is an integer and 1 < K < N" in the determination step of the number M of enabled serializer-deserializer channels; it can be considered that when it is detected that the current traffic value just exceeds 1 / N times the upper limit value of the transmission traffic (just exceeds the traffic value that can be carried by enabling 1 serializer-deserializer channel, but has not reached the traffic value that can be carried by enabling all N serializer-deserializer channels), new serializer-deserializer channels need to be further enabled, but not all N serializer-deserializer channels need to be enabled. Further, for "when the current traffic value > (N-1) / N * the upper limit value of the transmission traffic, M = N" in the determination step of the number M of enabled serializer-deserializer channels; it can be considered that the current traffic value has already been between (N-1) / N * the upper limit value of the transmission traffic and the upper limit value of the transmission traffic, then all N serializer-deserializer channels can only be enabled. It should also be understood that N is a positive integer greater than or equal to 3, and usually takes the value of 4 or 8.

[0036] It should also be understood that, in some alternative embodiments, this disclosure further provides a power-saving method based on multi-channel aggregation technology of optical modules. The method 200 includes: Step 210: Parameter initialization. The total number of channels N (N is a positive integer, such as 4, 8, etc.) of the multi-channel optical modules used in the system is obtained and set, as well as the nominal transmission traffic limit value of each serializer-deserializer (SerDes) channel (e.g., 50Gbps per channel). The total theoretical bandwidth limit of the aggregation port can then be calculated, which is N times the single-channel limit value. Step 220: Real-time traffic monitoring. At the aggregation port, the current actual transmission traffic value is continuously monitored. This monitoring point can be set between the data link layer and physical layer interface on the host side, and real-time throughput data is obtained through a hardware counter or software polling. Step 230: Traffic comparison and channel control decision. The current traffic value obtained in step 220 is compared with a series of thresholds preset based on the total bandwidth, and a channel enable control command is generated based on the comparison result. The thresholds are logically divided based on the number of channels N: It should be noted that when the current traffic value is less than 1 / N of the total bandwidth limit, it is determined to be a very low load state. The control logic issues an instruction to enable one SerDes channel, while simultaneously suggesting or directly controlling the shutdown of the remaining N-1 channels. When the current traffic value is greater than (N-2) / N of the total bandwidth limit but less than (N-1) / N, it is determined to be a high load but not saturated state. The control logic issues an instruction to enable N-1 SerDes channels. When the current traffic value is greater than (N-1) / N of the total bandwidth limit, it is determined to be a near-saturated or overloaded state. The control logic issues an instruction to enable all N SerDes channels to ensure performance.

[0037] It should also be understood that in actual implementation, more and more refined threshold ranges can be set as needed (for example, corresponding to opening 2, 3...N-2 channels) to form a multi-level control strategy.

[0038] Furthermore, in some embodiments, the method 200 further includes: Step 240: Instruction execution and state synchronization. The channel control instruction generated in step 230 is sent to the drive circuit or management interface of the optical module. The optical module quickly enables or disables the specified physical channel according to the instruction. At the same time, the upper layer of the system (such as the network protocol stack or task scheduler) should be aware of the changes in the currently available bandwidth and adjust the traffic scheduling strategy as necessary.

[0039] Preferably, in some embodiments, in step 220, the acquisition of the real-time traffic value is not a simple instantaneous sampling, but rather a combination of smoothing or prediction of short-term historical data. To achieve this, the system introduces a traffic shaping and prediction module. The core input of this module comes from a set of First-In-First-Out (FIFO) buffers. These FIFOs are located between the system data generation end (such as the GPU memory controller or network protocol processing unit) and the physical layer SerDes interface, and are used to buffer data packets to be sent. Regarding water level monitoring: the traffic shaping module continuously monitors the amount of stored data (i.e., the "water level") of each FIFO. Each FIFO has preset "high water level" and "low water level". Regarding traffic inference and prediction: the module uses algorithms to infer the current and recent total traffic demand based on the water level change rate of each FIFO, the data read rate (usually synchronized with the SerDes transmission clock of the enabled channel), and historical water level patterns. A simple implementation can use statistical methods such as weighted moving averages. A better implementation can integrate lightweight machine learning models, such as Long Short-Term Memory (LSTM) networks or a combination of convolutional neural networks and LSTM (CNN-LSTM), to learn the characteristics of the communication task (such as the type and scale of MPI ensemble operations), thereby more accurately predicting traffic trends. Regarding control signal generation: based on the predicted traffic value (not just the current instantaneous value), the module generates channel control commands. For example, when it is predicted that the traffic will rise and exceed the next threshold within the next few microseconds, the module can issue a command in advance to activate additional channels, avoiding instantaneous congestion caused by channel preparation delays. Conversely, when it is predicted that the traffic will remain below a certain threshold, redundant channels can be decisively shut down.

[0040] Preferably, in some embodiments, to eliminate communication interruptions caused by channel switching, this disclosure specifically selects optical modules that support fixed configurations based on OTP (One-Time Programmable) storage. Unlike traditional SerDes interfaces that require dynamic PLL configuration to adapt to different rates, these optical modules load a preset set of fixed parameters corresponding to the number of channels N and the rate per channel from the OTP upon power-up. When a channel needs to be activated, only an enable signal needs to be sent to the optical module, and the internal circuitry can quickly activate the laser and driver of that channel based on the fixed parameters, without undergoing a time-consuming PLL relocking and parameter negotiation process. This achieves millisecond-level or even microsecond-level channel recovery, virtually imperceptible to upper-layer applications.

[0041] Furthermore, in some embodiments, see Figure 4, the present disclosure further provides a power consumption saving device based on the optical module multi-channel aggregation technology. The device includes: an initialization module for configuring the optical module channel parameters and the traffic threshold. A monitoring module, including a FIFO array and a water level sensor, for capturing the data buffer status in real time. A prediction and control module integrating a prediction algorithm (such as a statistical model or a lightweight AI model), calculating the predicted traffic according to the input of the monitoring module, and generating channel control instructions according to the method described in the first aspect. In addition, the device may further include: a drive interface module for converting the control instructions into signals recognizable by the optical module and sending them down.

[0042] In some alternative embodiments, the present disclosure provides a power consumption saving method based on the optical module multi-channel aggregation technology, including the following steps: obtaining the upper limit value of the transmission traffic of N serializer-deserializer channels, where N is a positive integer. Generally, in most embodiments, N is a positive integer greater than or equal to 3; monitoring the current traffic value at the aggregation port in real time; comparing the current traffic value with a preset traffic threshold, and dynamically enabling or closing the corresponding number of channels according to the comparison result; when the current traffic value ≤ (1 / N) * the upper limit value of the transmission traffic, enabling 1 channel and closing the remaining N - 1 channels; when (K - 1) / N * the upper limit value of the transmission traffic < the current traffic value ≤ K / N * the upper limit value of the transmission traffic, enabling K channels and closing the remaining N - K channels, where K is an integer and 1 < K < N; when the current traffic value > (N - 1) / N * the upper limit value of the transmission traffic, enabling all N channels.

[0043] Preferably, in some embodiments, the traffic monitoring and prediction are realized by setting multiple FIFO buffers and their water level lines, and the traffic feature extraction and prediction are performed in combination with a hash algorithm or a deep learning model (such as LSTM, CNN-LSTM) to improve the accuracy and real-time performance of the channel switching decision.

[0044] Preferably, in some embodiments, the optical module uses OTP (one-time programmable) storage for fixed configuration, and there is no need to reconfigure the PLL when the channel is enabled, achieving millisecond-level recovery and avoiding communication interruption.

[0045] It should also be understood that in some alternative embodiments, a system can be provided, comprising: a Host chip, multiple FIFO buffers, a traffic shaping and prediction module, and a multi-channel optical module. Data enters the FIFO from the system side. The traffic shaping module calculates the real-time traffic based on the FIFO level and read rate, and outputs channel control signals to the optical module driver circuit in conjunction with the prediction model. Specifically, the traffic shaping module includes: a level monitoring unit, a hash scheduling unit, a prediction model unit (optional), and a control logic unit. The level monitoring unit collects the data volume of each FIFO in real time; the hash scheduling unit is responsible for balanced traffic distribution; the prediction model unit uses algorithms such as LSTM to predict traffic trends; and the control logic unit generates channel enable signals based on the level and prediction results. In a further embodiment, the channel enable control logic includes: initializing the number of channels N and the upper limit of the bandwidth of each channel; continuously monitoring the traffic at the aggregation port; determining the required number of channels based on the traffic range; and sending control commands to the optical module to realize dynamic switching of channels. Specifically, in some embodiments, each FIFO is set with two level lines, high and low. When the data volume of a certain FIFO exceeds the high watermark, a new channel is triggered; when the data volume of all FIFOs remains below the low watermark for a certain period of time, redundant channels are shut down. Thus, this disclosure achieves dynamic optimization of optical module power consumption while ensuring communication continuity through the above mechanism. Experiments show that in typical All-Reduce communication scenarios, optical module power consumption can be reduced by 30%-50%, significantly improving overall energy efficiency.

[0046] In some embodiments, this disclosure also provides an electronic device including a memory, a processor, and a computer program stored in the memory. When the processor executes the program, it implements the methods described above.

[0047] In some embodiments, this disclosure provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the methods described above.

[0048] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

[0049] The above description is merely an optional embodiment of this disclosure and is not intended to limit this disclosure. Various modifications and variations can be made to this disclosure by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A power-saving method based on multi-channel convergence technology of optical modules, characterized in that, include: Obtain the maximum transmission throughput of N serializer-deserializer channels, where N is a positive integer; Get the current traffic value at the aggregation port; The current traffic value is compared with the upper limit of the transmission traffic, and a corresponding number of serializer-deserializer channels are enabled based on the comparison result; wherein, the step of determining the number M of serializer-deserializer channels to be enabled includes: When the current traffic value is ≤ (1 / N) * the upper limit of the transmission traffic, M = 1; When (K-1) / N * the upper limit of transmission traffic < the current traffic value ≤ K / N * the upper limit of transmission traffic, M = K, where K is an integer and 1 < K < N; When the current traffic value is greater than (N-1) / N * the upper limit of the transmission traffic, M=N.

2. The method according to claim 1, characterized in that, While enabling M serializer-deserializer channels, the remaining channels of the N serializer-deserializer channels other than the M channels are turned off.

3. The method according to claim 1, characterized in that, The current flow value is obtained through a flow shaping module, which includes: Multiple FIFO buffers are used to temporarily store data to be sent; The water level monitoring unit is used to monitor the data volume of each FIFO. The prediction unit is used to predict flow trends based on water level data and historical flow characteristics. The control unit is used to output a channel enable control signal based on the prediction results.

4. The method according to claim 3, characterized in that, The prediction unit uses a deep learning model to predict traffic flow, and the deep learning model includes an LSTM or CNN-LSTM network.

5. The method according to claim 3, characterized in that, The FIFO buffer has a high water level and a low water level. When the data volume of any FIFO exceeds the high water level, a new channel is enabled. When the data volume of all FIFOs remains below the low water level, redundant channels are disabled.

6. The method according to claim 1, characterized in that, The optical module uses a fixed OTP storage configuration, and the PLL does not need to be reconfigured when the channel is enabled.

7. A power-saving device based on multi-channel convergence technology of optical modules, characterized in that, include: The initialization module is used to obtain the upper limit of the transmission flow of N serializer-deserializer channels, where N is a positive integer; The monitoring module is used to obtain the current traffic value at the aggregation port; The monitoring and control module is used to compare the current traffic value with the upper limit of the transmission traffic value, and control the enabling of a corresponding number of serializer-deserializer channels according to the comparison result; The steps for determining the number M of enable serializer-deserializer channels include: When the current traffic value is ≤ (1 / N) * the upper limit of the transmission traffic, M = 1; When (K-1) / N * the upper limit of transmission traffic < the current traffic value ≤ K / N * the upper limit of transmission traffic, M = K, where K is an integer and 1 < K < N; When the current traffic value is greater than (N-1) / N * the upper limit of the transmission traffic, M=N.

8. The apparatus according to claim 7, characterized in that, The monitoring and control module is configured to enable M serializer-deserializer channels while shutting down the remaining channels of the N serializer-deserializer channels other than the M channels.

9. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, The storage medium stores at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the steps of the method as described in any one of claims 1 to 6.

Images