Optimization of periodic traffic transmission for distributed model training

Abstract

This invention discloses a periodic traffic transmission optimization method for distributed model training, belonging to the field of network communication technology. First, a network traffic model is constructed by collecting information on link round-trip time, maximum link utilization, and changes in switch egress queue length. Second, the collected traffic signals are processed using a sliding window-based local Fourier transform method. Then, based on the processed traffic signal spectrum, the periodic composition of the traffic signals in the path is analyzed, and local and non-local traffic signals are separated to facilitate subsequent modeling calculations. Finally, a periodic traffic peak-shifting transmission model is established based on the local and non-local traffic signal waveforms, and the optimal peak-shifting time shift value that minimizes the objective function is calculated. Using the transmission time shift value calculated by this invention to delay the transmission of local traffic can effectively achieve peak-shifting transmission of traffic, alleviate sudden congestion and switch buffer overflow problems, and significantly improve link utilization and data transmission efficiency.

Description

Technical Field

[0001] This invention relates to the field of network communication technology, and in particular to a periodic traffic transmission optimization method for distributed model training. Background Technology

[0002] In recent years, the scale of distributed training models has grown exponentially, from tens of thousands of parameters initially to hundreds of billions or even trillions of parameters currently. To keep pace with this trend, accelerating distributed training by scaling up the number of computing nodes has become a common practice. In distributed training, nodes need to frequently synchronize gradients or update parameters during each iteration. This synchronization process requires transferring large amounts of data between nodes within milliseconds and repeating multiple iterations until the model converges. However, with the continuous expansion of distributed training systems, the increase in the number of nodes and the amount of communication between them leads to an increase in global communication complexity. Communication time accounts for an increasingly larger proportion of the overall training time, severely impacting the scalability of distributed training.

[0003] To reduce communication overhead in distributed training, many researchers have conducted related studies. Some researchers have attempted to improve transmission protocols, but most of these protocols are designed for traditional data center networks and cannot effectively adapt to the unique traffic patterns of distributed training. Problems such as network congestion and switch buffer overflows due to traffic bursts still exist. Other studies have addressed the periodic traffic in distributed training, but these have not adequately solved the problem of insufficient link utilization, thus reducing communication efficiency. Therefore, how to utilize the periodic traffic in distributed training to improve communication efficiency is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] To address the shortcomings of existing methods, this invention discloses a periodic traffic transmission optimization method for distributed model training. By calculating time shift values, it can stagger traffic transmission to alleviate network congestion and switch buffer overflow problems, improve link utilization, and thus improve communication efficiency.

[0005] To achieve the above objectives, this invention discloses a periodic traffic transmission optimization method for distributed model training, comprising:

[0006] Collect transmission status information of each link in the network to characterize the network transmission status.

[0007] Based on the transmission status information, a network traffic signal model is constructed.

[0008] A sliding window local Fourier transform method is used to process network traffic signals.

[0009] Based on the processed flow signal spectrum, the periodic characteristics of the flow signal in the path are analyzed, and the local flow signal and the non-local flow signal are separated.

[0010] Based on the local traffic signal and the non-local traffic signal, a periodic traffic peak shifting transmission model is established, and the optimal time shift value of the local traffic is calculated.

[0011] According to the periodic traffic transmission optimization method for distributed model training provided by the present invention, the link information includes round-trip time (RTT), maximum link utilization (MPU), and changes in the length of the switch egress queue.

[0012] The specific data collection steps are as follows: If the increase in the switch's egress queue length is less than or equal to 0, then the change in the egress queue length is... Calculate the link utilization rate of the outbound link. ,in, This represents the linear velocity of the outbound link; if the increase in the outbound queue length is greater than 0, then record... After comparison and calculation, the link utilization value is stored in the switch register and retrieved when a data packet arrives.

[0013] The data packet carries MPU information using a low-overhead method, initially... Each time the data packet passes through a switch, the link utilization U stored in the switch is compared with the MPU of the data packet, and the maximum value is assigned to the MPU.

[0014] According to the periodic traffic transmission optimization method for distributed model training provided by the present invention, the specific steps of constructing the network signal model include:

[0015] After receiving the data packet, the destination copies the MPU information into the returned ACK and sends it back to the source. After receiving the ACK, the source integrates the MPU and RTT information.

[0016] First, use an exponentially weighted moving average to analyze the two-dimensional values. The formula for smoothing and removing noisy data is as follows:

[0017] ,in, This is the smoothing coefficient.

[0018] Secondly, the RTT is normalized using the following formula:

[0019] ,in, The maximum acceptable RTT threshold is set; RTT measurements exceeding this value will be assigned a value. , The time required for a round trip of idle traffic is given, while the MPU information itself is a normalized value and does not require processing.

[0020] Finally, the RTT and MPU signals are processed uniformly using the following formula:

[0021]

[0022] According to the periodic traffic transmission optimization method for distributed model training provided by the present invention, the method of processing the traffic signal using a sliding window-based local Fourier transform method includes the following specific steps:

[0023] The original continuous and long signal is divided into many shorter signal segments based on the sliding window and step size. Fourier transform is then performed on each of these signal segments to extract the periodic information of the flow rate.

[0024] The initial size of the sliding window is set to After several rounds of iteration, using rules of thumb To set the window size ,in, These are empirical parameters. It is the standard deviation of the RTT and MPU signals.

[0025] The portion of the flow signal that overlaps with the window function is selected for Fast Fourier Transform (FFT). Then, the window is slid by one step and the same operation is performed until the entire continuous signal has been analyzed. The step size is set to half the window size.

[0026] According to the periodic traffic transmission optimization method for distributed model training provided by the present invention, the analysis of traffic periodic characteristics specifically includes:

[0027] After obtaining the spectrum of each flow based on the above steps, sort them according to the time series and select the frequencies with the largest amplitudes. ,pass Calculate the period of the flow signal Then, the local traffic sending rate is tracked to determine the local traffic cycle over multiple periods.

[0028] According to the periodic traffic transmission optimization method for distributed model training provided by the present invention, the specific steps of separating local traffic signals and non-local traffic signals include:

[0029] First, remove the local flow frequency component from the path spectrum obtained in the above steps, i.e., denote it as 0; second, use inverse Fourier transform to restore the non-local flow signal wave, taking only the real part; finally, subtract the non-local flow component from the path flow signal wave to obtain the local flow signal wave.

[0030] According to the periodic traffic transmission optimization method for distributed model training provided by the present invention, the specific steps for establishing a periodic traffic peak-shifting transmission model to calculate the optimal time-shift value of local traffic include:

[0031] The non-local flow signal and the local flow signal are represented by two sine waves. and The description is expressed as: ,in, and These are the angular frequencies of wave A and wave B, respectively. and It is their initial phase. and It is their amplitude, Represents the time vector.

[0032] The signal wave of path flow can be represented as a superimposed wave. Specifically:

[0033] ,in, This indicates the phase difference.

[0034] because ,use Seek and The period of the least common multiple Define the function as In a cycle The average amplitude within is:

[0035]

[0036] The goal is to find the phase difference that minimizes the average amplitude. Firstly, because It is a nonlinear function, using numerical gradient. To approximate the calculation of the objective function with respect to The gradient is calculated; then, the gradient descent method is used to update the gradient. , is represented as:

[0037] ,in, Indicates the learning rate. This represents the number of iteration steps; finally, the optimal phase difference value is obtained. Then, the optimal time shift value for the local flow can be obtained. .

[0038] In general, the present invention has the following beneficial effects:

[0039] 1. The periodic traffic transmission optimization method for distributed model training provided by this invention can effectively solve the problems of sudden path congestion and buffer overflow caused by synchronous transmission of periodic traffic in distributed training.

[0040] 2. The periodic traffic transmission optimization method for distributed model training provided by this invention delays the transmission of local traffic by calculating the optimal time shift value of local traffic, which can effectively improve link utilization and thus accelerate communication. Attached Figure Description

[0041] Figure 1 This is an overall flowchart of an embodiment of the present invention.

[0042] Figure 2 This is a schematic diagram of the synchronous transmission of two tasks in an embodiment of the present invention.

[0043] Figure 3 This is a schematic diagram of staggered transmission of two tasks in an embodiment of the present invention.

[0044] Figure 4 This is a schematic diagram illustrating the iteration time in an embodiment of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0046] The core of this invention is to provide a periodic traffic transmission optimization method for distributed model training. By calculating the optimal time shift value, peak traffic transmission is carried out to alleviate congestion and buffer overflow, improve link utilization, and thus accelerate communication time.

[0047] To enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0048] Figure 1 A flowchart illustrating a periodic traffic transmission optimization method for distributed model training provided in this embodiment of the invention is shown below. Figure 1 As shown, it includes the following steps:

[0049] S1: Collect transmission status information of each link in the network to characterize the network transmission status.

[0050] Collect information on round-trip time (RTT), maximum link utilization (MPU), and changes in switch egress queue length to characterize network transmission status.

[0051] In the path, if the increase in the switch's egress queue length is less than or equal to 0, then the change in the egress queue length is as follows: Calculate the link utilization rate of the outbound link. ,in, This represents the linear velocity of the outbound link; if the increase in the outbound queue length is greater than 0, then record... After comparison and calculation, the link utilization value is stored in the switch register and retrieved when a data packet arrives.

[0052] The data packet carries MPU information using a low-overhead method, initially... Each time the data packet passes through a switch, the link utilization U stored in the switch is compared with the MPU of the data packet, and the maximum value is assigned to the MPU.

[0053] S2: Based on the transmission status information, construct a network traffic signal model.

[0054] After receiving the data packet, the destination copies the MPU information into the returned ACK and sends it back to the source. After receiving the ACK, the source integrates the MPU and RTT information.

[0055] Specifically, firstly, an exponentially weighted moving average is used to analyze the two-dimensional values. The formula for smoothing and removing noisy data is as follows:

[0056] ,in, This is the smoothing coefficient.

[0057] Secondly, the RTT is normalized using the following formula:

[0058] ,in, The maximum acceptable RTT threshold is set; RTT measurements exceeding this value will be assigned a value. , The time required for a round trip of idle traffic is given, while the MPU information itself is a normalized value and does not require processing.

[0059] Finally, the RTT and MPU signals are processed uniformly using the following formula:

[0060]

[0061] S3: The sliding window local Fourier transform method is used to process network traffic signals.

[0062] The original continuous and long signal is divided into many shorter signal segments based on the sliding window and step size. Fourier transform is then performed on each of these signal segments to extract the periodic information of the flow rate.

[0063] The initial size of the sliding window is set to After several rounds of iteration, using rules of thumb To set the window size ,in, These are empirical parameters. It is the standard deviation of the RTT and MPU signals.

[0064] The portion of the flow signal that overlaps with the window function is selected for Fast Fourier Transform (FFT). Then, the window is slid by one step and the same operation is performed until the entire continuous signal has been analyzed. The step size is set to half the window size.

[0065] S4: Based on the processed flow signal spectrum, analyze the periodic characteristics of the flow signal in the path and separate the local flow signal from the non-local flow signal.

[0066] After obtaining the spectrum of each flow based on the above steps, sort them according to the time series and select the frequencies with the largest amplitudes. ,pass Calculate the period of the flow signal Then, based on the sending rate of the local traffic, the local traffic cycle is determined in multiple cycles.

[0067] For subsequent modeling, it is necessary to separate local traffic from non-local traffic in the complete signal spectrum. First, the local traffic frequency part is removed from the spectrum, i.e., it is recorded as 0. Second, the non-local traffic signal wave is obtained by using inverse Fourier transform, and only the real part is taken here. Finally, the non-local traffic part is subtracted from the path traffic signal wave to obtain the local traffic signal wave.

[0068] S5: Using the local traffic signal and non-local traffic signal obtained from S3, establish a periodic traffic peak shift transmission model and calculate the optimal time shift value of the local traffic.

[0069] The non-local flow signal and the local flow signal are represented by two sine waves. and The description is expressed as: ,in, and These are the angular frequencies of wave A and wave B, respectively. and It is their initial phase. and It is their amplitude, Represents the time vector.

[0070] The signal wave of path flow can be represented as a superimposed wave. Specifically:

[0071] ,in, This indicates the phase difference.

[0072] because ,use Seek and The period of the least common multiple Define the function as In a cycle The average amplitude within is:

[0073]

[0074] The goal is to find the phase difference that minimizes the average amplitude. Firstly, because It is a nonlinear function, using numerical gradient. To approximate the calculation of the objective function with respect to The gradient is calculated; then, the gradient descent method is used to update the gradient. , is represented as:

[0075] ,in, Indicates the learning rate. This represents the number of iteration steps; finally, the optimal phase difference value is obtained. Then, the optimal time shift value for the local flow can be obtained. .

[0076] Specifically, Figure 2 Two ResNet50 tasks were shown. and Simulation results are computed synchronously and communicated and shared on the same link. Due to competition for path bandwidth between tasks, each task only obtains 48Gbps of available bandwidth, and there are significant idle periods between the traffic signal peaks of the two tasks. Figure 3 This invention demonstrates the application of the method of the present invention to implement a delayed transmission task. The results show that both tasks can achieve a bandwidth utilization of 96Gbps during peak transmission periods, while effectively reducing idle periods between peaks, thus improving link utilization. Figure 4By comparing the performance of the two transmission scenarios over 100 iterations, it can be observed that the peak-shifting transmission scenario optimized by the method of this invention (Scenario 2) achieves a speedup of 1.26 times compared to the original transmission scenario (Scenario 1). Simulation results show that the transmission time shift value calculated according to this invention can effectively achieve multi-stream peak-shifting transmission, thereby alleviating network congestion and improving the communication efficiency of distributed training.

[0077] The relevant technical solutions are the same as those in the above embodiments, and will not be repeated here.

[0078] The above provides a detailed description of the periodic traffic transmission optimization method for distributed model training provided by this invention. The various embodiments in the specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably.

Claims

1. A method for optimizing periodic traffic transmission for distributed model training, characterized in that, Includes the following steps: S1: Collect transmission status information of each link in the network to characterize the network transmission status; S2: Based on the transmission status information, after receiving the data packet, the destination end copies the MPU information into the returned ACK and sends it back to the source end. After receiving the ACK, the source end integrates the MPU and RTT information; specifically, it first uses an exponentially weighted moving average to process the two-dimensional values. The formula for smoothing and removing noisy data is as follows: ， in, The first factor is the smoothing coefficient; secondly, the RTT is normalized using the following formula: ， in, The maximum acceptable RTT threshold is set; RTT measurements exceeding this value will be assigned a value. , The time required for a round trip of idle traffic is considered, while the MPU information itself is a normalized value and does not require processing. Finally, the RTT and MPU signals are processed uniformly to construct a network traffic signal model, with the following formula: ； S3: The sliding window local Fourier transform method is used to process network traffic signals; S4: Based on the processed flow signal spectrum, analyze the periodic characteristics of the flow signal in the path and separate the local flow signal from the non-local flow signal; S5: Based on the local traffic signal and the non-local traffic signal, establish a periodic traffic peak shifting transmission model and calculate the optimal time shift value of the local traffic.

2. The periodic traffic transmission optimization method for distributed model training according to claim 1, characterized in that, The link information collected in step S1 includes round-trip time (RTT), maximum link utilization (MPU), and changes in the length of the switch egress queue. In the path, if the increase in the switch's egress queue length is less than or equal to 0, then the change in the egress queue length is as follows: Calculate the link utilization rate of the outbound link. ,in, This represents the linear velocity of the outbound link; if the increase in the outbound queue length is greater than 0, then record... After comparison and calculation, the link utilization value is stored in the switch register and retrieved when a data packet arrives. The data packet carries MPU information using a low-overhead method, initially... Each time the data packet passes through a switch, the link utilization U stored in the switch is compared with the MPU of the data packet, and the maximum value is assigned to the MPU.

3. The periodic traffic transmission optimization method for distributed model training according to claim 1, characterized in that, The specific steps for processing the flow signal using Fourier transform in step S3 are as follows: The original continuous and long signal is divided into many shorter signal segments according to the sliding window and step size. Fourier transform is performed on each of these signal segments to extract the periodic information of the flow rate. The initial size of the sliding window is set to After several rounds of iteration, using rules of thumb To set the window size ,in, These are empirical parameters. It is the standard deviation of the RTT and MPU signals; The portion of the flow signal that overlaps with the window function is selected for Fast Fourier Transform (FFT). Then, the window is slid by one step and the same operation is performed until the entire continuous signal has been analyzed. The step size is set to half the window size.

4. The periodic traffic transmission optimization method for distributed model training according to claim 1, characterized in that, In step S4: After obtaining the spectrum of each flow based on the above steps, sort them according to the time series and select the frequencies with the largest amplitudes. ,pass Calculate the period of each flow signal. ; Then, track the local traffic sending rate and determine the local traffic cycle over multiple cycles; For subsequent modeling, it is necessary to separate local traffic and non-local traffic in the complete signal spectrum. First, remove the local traffic frequency part from the above spectrum, that is, denote it as 0. Secondly, the signal wave of non-local flow is obtained by using inverse Fourier transform, and only the real part is taken here; finally, the non-local flow part is subtracted from the path flow signal wave to obtain the local flow signal wave.

5. The periodic traffic transmission optimization method for distributed model training according to claim 1, characterized in that, In step S5: The non-local flow signal and the local flow signal are represented by two sine waves. The description is expressed as: , in, and These are the angular frequencies of wave A and wave B, respectively. and It is their initial phase. and It is their amplitude, Represents the time vector; The signal wave of path flow can be represented as a superimposed wave. Specifically: ， in, Indicates phase difference; due to ,use Seeking and The period of the least common multiple Define the function as In a cycle The average amplitude within is: ； The goal is to find the phase difference that minimizes the average amplitude. Firstly, because It is a nonlinear function, using numerical gradient. Approximate calculation of the objective function with respect to The gradient is calculated; then, gradient descent is used to update the gradient. , is represented as: ， in, Indicates the learning rate. This represents the number of iteration steps; finally, the optimal phase difference value is obtained. Then, the optimal time shift value for the local flow can be obtained. .

Images