Periodic traffic transmission optimization for model distributed training

Abstract

The invention discloses a model distributed training-oriented periodic traffic transmission optimization method, and belongs to the technical field of network communication. The method comprises the following steps: firstly, constructing a network flow model by collecting round-trip time of a link, a maximum link utilization rate and exit queue length change information of a switch; secondly, processing the collected flow signals by using a local Fourier transform method based on a sliding window; then, according to the processed flow signal frequency spectrum, the period composition of the flow signals in the path is analyzed, and local flow signals and non-local flow signals are separated to facilitate subsequent modeling calculation; and finally, establishing a periodic flow peak shifting transmission model based on local and non-local flow signal waves, and calculating to obtain an optimal peak shifting time shift value of a minimized target function. The local flow is delayed to be transmitted by using the transmission time shift value calculated by the method, so that peak shifting transmission of the flow can be effectively realized, the problems of sudden congestion and overflow of a buffer area of an exchange area are relieved, and the link utilization rate and the data transmission efficiency are remarkably improved.

Description

Technical Field

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

[0002] In recent years, the scale of distributed training models has grown exponentially, from initially tens of thousands of parameters to hundreds of billions or even trillions of parameters today. In line with this trend, accelerating distributed training by expanding the scale of computing nodes has become a common practice. In distributed training, nodes need to frequently synchronize gradients or updated parameters during each iteration. This synchronization process requires transferring large amounts of data between nodes within milliseconds, repeated multiple times until the model converges. However, as distributed training systems continue to scale, the increase in the number of nodes and the amount of communication between nodes leads to increased global communication complexity. Communication time is becoming an increasingly significant portion of overall training time, severely impacting the scalability of distributed training.

[0003] To reduce the communication overhead of distributed training, many researchers have conducted research in this area. Some have achieved this by improving transmission protocols. However, most transmission protocols are designed primarily for traditional data center networks and are unable to effectively adapt to the unique traffic patterns of distributed training. Problems such as network congestion and switch buffer overflows caused by traffic bursts still exist. Some research has also focused on periodic traffic in distributed training, but this has not effectively addressed the issue of insufficient link utilization, reducing communication efficiency. Therefore, how to leverage periodic traffic in distributed training to improve communication efficiency is an urgent issue for those skilled in the art. Summary of the Invention

[0004] Based on the shortcomings of existing methods, the present invention discloses a periodic traffic transmission optimization method for model distributed training, which staggers transmission traffic by calculating time shift values, alleviates network congestion and switch buffer overflow problems, improves link utilization, and thus improves communication efficiency.

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

[0006] Collect the 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] The sliding window local Fourier transform method is used to process network traffic signals.

[0009] According to the processed traffic signal spectrum, the periodic characteristics of the traffic signal in the path are analyzed, and the local traffic signal and the non-local traffic signal are separated.

[0010] A periodic traffic peak-shifting transmission model is established according to the local traffic signal and the non-local traffic signal, and an optimal time shift value of the local traffic is calculated.

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

[0012] The specific collection steps are as follows: If the increase in the length of the switch egress queue is less than or equal to 0, the egress queue length changes to , calculate the link utilization of the export link as ,in, is the line speed of the egress link; if the egress queue length growth is greater than 0, then record ; After comparison and calculation, the link utilization value is stored in the switch register and called after the data packet arrives.

[0013] The data packet uses a low-overhead method to carry MPU information. Initially , each time it passes through a hop 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 a periodic traffic transmission optimization method for model distributed training provided by the present invention, the specific steps of constructing a network signal model include:

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

[0016] First, use the exponentially weighted moving average to calculate the two-dimensional values. Smoothing is performed to remove noise data. The formula is: ,in, is the smoothing coefficient.

[0017] Secondly, the RTT is normalized using the formula: ,in, The maximum acceptable RTT threshold to be set. RTT measurements exceeding this value will be assigned to , This is the time required for a round trip of idle traffic. The MPU information itself is a normalized value and does not need to be processed.

[0018] Finally, the RTT and MPU signals are processed uniformly, and the formula is:

[0019] According to a periodic traffic transmission optimization method for model distributed training provided by the present invention, the traffic signal is processed using a local Fourier transform method based on a sliding window, and the specific steps include:

[0020] According to the sliding window and step size, the originally continuous and long signal is divided into many shorter signal segments, and these signal segments are processed by Fourier transform one by one to extract the periodic information of the traffic.

[0021] The size of the sliding window is initially set to After several rounds of iteration, using the rule of thumb To set the window size ,in, is an empirical parameter, is the standard deviation of RTT and MPU signal.

[0022] The portion of the flow signal that overlaps with the window function is selected to perform fast Fourier transform, and then the window slides a step size to continue the same operation until the entire continuous signal is analyzed, where the step size is set to half of the window size.

[0023] According to a periodic traffic transmission optimization method for model distributed training provided by the present invention, the analysis of traffic period characteristics is specifically as follows:

[0024] After obtaining the spectrum of each flow according to the above steps, sort it according to the time series and select the multiple frequencies with the largest amplitude ,pass Calculate the period of the flow signal ,Then, the sending rate of the local traffic is tracked and the ,local traffic period is determined over multiple periods.

[0025] According to a periodic traffic transmission optimization method for model distributed training provided by the present invention, the specific steps of separating local traffic signals from non-local traffic signals include:

[0026] First, remove the local traffic frequency portion from the path spectrum obtained in the above steps and set it to 0. Second, use the inverse Fourier transform to restore the signal wave of the non-local traffic, here only the real part is taken. Finally, subtract the non-local traffic portion from the path traffic signal wave to obtain the local traffic signal wave.

[0027] According to a method for optimizing periodic traffic transmission for distributed model training provided by the present invention, the specific steps of establishing a periodic traffic peak-shifting transmission model to calculate the optimal time shift value of local traffic include:

[0028] The non-local flow signal and the local flow signal are converted into two sine waves. and Description, expressed as: ,in, and are the angular frequencies of wave A and wave B respectively, and is their initial phase, and is their amplitude, represents the time vector.

[0029] The signal wave of the path flow can be expressed as a superposition wave , specifically: ,in, Indicates the phase difference.

[0030] because ,use Obtain and The period of the least common multiple of , define the function as In a cycle The average amplitude within is:

[0031] The goal is to find the phase difference that minimizes the average amplitude , first of all, because It is a nonlinear function, using numerical gradient To approximate the objective function about Then, use the gradient descent method to update , expressed as: ,in, represents the learning rate, Indicates the number of iteration steps; finally, the optimal phase difference is obtained Then, the optimal time shift value of the local traffic can be obtained. .

[0032] In general, the present invention has the following beneficial effects:

[0033] 1. The periodic traffic transmission optimization method for model distributed training provided by the present invention can effectively solve the problems of path burst congestion and buffer overflow caused by synchronous transmission of periodic traffic in distributed training.

[0034] 2. The periodic traffic transmission optimization method for model distributed training provided by the present invention delays the transmission of local traffic by calculating the optimal time shift value of the local traffic, which can effectively improve link utilization and thus accelerate communication. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is an overall flow chart of an embodiment of the present invention.

[0036] Figure 2 Schematic diagram of synchronous transmission of two tasks in an embodiment of the present invention.

[0037] Figure 3 This is a schematic diagram of staggered transmission of two tasks in an embodiment of the present invention.

[0038] Figure 4 Schematic diagram of the iteration time in an embodiment of the present invention. DETAILED DESCRIPTION

[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0040] The core of the present invention is to provide a periodic traffic transmission optimization method for model distributed training, which performs staggered traffic transmission through the calculated optimal time shift value to alleviate congestion and buffer overflow, improve link utilization, and thus speed up communication time.

[0041] In order 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 implementation methods.

[0042] Figure 1 A flowchart of a method for optimizing periodic traffic transmission for distributed model training is provided in an embodiment of the present invention, such as Figure 1 As shown, the following steps are included:

[0043] S1: Collect the transmission status information of each link in the network to represent the network transmission status.

[0044] The round trip time (RTT), maximum link utilization (MPU), and switch egress queue length change information are collected to characterize network transmission conditions.

[0045] In the path, if the increase in the switch egress queue length is less than or equal to 0, the egress queue length changes to , calculate the link utilization of the export link as ,in, is the line speed of the egress link; if the egress queue length growth is greater than 0, then record ; After comparison and calculation, the link utilization value is stored in the switch register and called after the data packet arrives.

[0046] The data packet uses a low-overhead method to carry MPU information. Initially , each time it passes through a hop 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.

[0047] S2: Constructing a network traffic signal model based on the transmission status information.

[0048] After receiving the data packet, the destination end copies the MPU information to the returned ACK and transmits it back to the source end. After receiving the ACK, the source end integrates the MPU and RTT information.

[0049] Specifically, we first use the exponentially weighted moving average to calculate the two-dimensional values. Smoothing is performed to remove noise data. The formula is: ,in, is the smoothing coefficient.

[0050] Secondly, the RTT is normalized using the formula: ,in, The maximum acceptable RTT threshold to be set. RTT measurements exceeding this value will be assigned to , This is the time required for a round trip of idle traffic. The MPU information itself is a normalized value and does not need to be processed.

[0051] Finally, the RTT and MPU signals are processed uniformly, and the formula is:

[0052] S3: Use the sliding window local Fourier transform method to process network traffic signals.

[0053] According to the sliding window and step size, the originally continuous and long signal is divided into many shorter signal segments, and these signal segments are processed by Fourier transform one by one to extract the periodic information of the traffic.

[0054] The sliding window size is initially set to After several rounds of iteration, using the rule of thumb To set the window size ,in, is an empirical parameter, is the standard deviation of RTT and MPU signal.

[0055] The portion of the flow signal that overlaps with the window function is selected to perform fast Fourier transform, and then the window slides a step size to continue the same operation until the entire continuous signal is analyzed, where the step size is set to half of the window size.

[0056] S4: Analyze the periodic characteristics of the traffic signal in the path according to the processed traffic signal spectrum, and separate the local traffic signal from the non-local traffic signal.

[0057] After obtaining the spectrum of each flow according to the above steps, sort it according to the time series and select the multiple frequencies with the largest amplitude ,pass Calculate the period of the flow signal ; Then, based on tracking the sending rate of the local traffic, the local traffic period is determined in multiple periods.

[0058] 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 portion from the above spectrum and mark it as 0; second, use the inverse Fourier transform to restore the signal wave of non-local traffic, here only the real part is taken; finally, subtract the non-local traffic portion from the path traffic signal wave to obtain the local traffic signal wave.

[0059] S5: Based on the local traffic signal and non-local traffic signal obtained in S3, a periodic traffic peak-shifting transmission model is established to calculate the optimal time shift value of the local traffic.

[0060] The non-local flow signal and the local flow signal are converted into two sine waves. and Description, expressed as: ,in, and are the angular frequencies of wave A and wave B respectively, and is their initial phase, and is their amplitude, represents the time vector.

[0061] The signal wave of the path flow can be expressed as a superposition wave , specifically: ,in, Indicates the phase difference.

[0062] because ,use Obtain and The period of the least common multiple of , define the function as In a cycle The average amplitude within is:

[0063] The goal is to find the phase difference that minimizes the average amplitude , first of all, because It is a nonlinear function, using numerical gradient To approximate the objective function about Then, use the gradient descent method to update , expressed as: ,in, represents the learning rate, Indicates the number of iteration steps; finally, the optimal phase difference is obtained Then, the optimal time shift value of the local traffic can be obtained. .

[0064] Specifically, Figure 2 Two ResNet50 tasks are shown and Simulation results of synchronous computational communication and sharing the same link. Due to competition between tasks for path bandwidth, each task only receives 48Gbps of available bandwidth, and there is a significant idle period between the traffic signal peaks of the two tasks. Figure 3 Demonstrates the application of the method of the present invention to achieve delayed transmission tasks The results show that both tasks can achieve 96Gbps bandwidth utilization during peak transmission periods, while effectively reducing the idle time between peaks, which improves link utilization. Figure 4 Comparing the performance of two transmission scenarios over 100 iterations, we observed that the optimized staggered transmission scenario (Scenario 2) achieved a 1.26x speedup compared to the original transmission scenario (Scenario 1). Case simulations show that the transmission time shift values ​​calculated using this method can effectively achieve staggered multi-stream transmission, thereby alleviating network congestion and improving the communication efficiency of distributed training.

[0065] The relevant technical solutions are the same as those in the above embodiment and will not be described in detail here.

[0066] The above is a detailed introduction to the periodic traffic transmission optimization method for distributed model training provided by the present invention. The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same and similar parts between the various embodiments can be referenced to each other.

Claims

1. A periodic traffic transmission optimization method for model distributed training, characterized in that: The following steps are involved: S1: Collects the transmission status information of each link in the network to represent the network transmission status; S2: Building a network traffic signal model based on the transmission state information; S3: Uses sliding window local Fourier transform method to process network traffic signals; S4: Analyze the periodic characteristics of the traffic signal in the path according to the processed traffic signal spectrum, and separate the local traffic signal from the non-local traffic signal; S5: Establish a periodic traffic peak-shifting transmission model based on the local traffic signal and the non-local traffic signal, and calculate the optimal time shift value of the local traffic.

2. The periodic traffic transmission optimization method for model distributed training according to claim 1 is characterized in that: The link information collected in step S1 includes round trip time (RTT), maximum link utilization (MPU), and change in switch egress queue length; In the path, if the increase in the switch egress queue length is less than or equal to 0, the egress queue length changes to , calculate the link utilization of the export link as ,in, is the line speed of the egress link; if the egress queue length growth is greater than 0, then record ;After comparison and calculation, the link utilization value is stored in the switch register and called after the data packet arrives; The data packet uses a low-overhead method to carry MPU information. Initially , each time it passes through a hop 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 model distributed training according to claim 1 is characterized in that: The specific steps of constructing the network traffic signal model in step S2 are: After receiving the data packet, the destination end copies the MPU information into the returned ACK and transmits it back to the source end. After receiving the ACK, the source end integrates the MPU and RTT information; Specifically, we first use the exponentially weighted moving average to calculate the two-dimensional values. Smoothing is performed to remove noise data. The formula is: ， in, is the smoothing coefficient; secondly, the RTT is normalized using the formula: ， in, The maximum acceptable RTT threshold to be set. RTT measurements exceeding this value will be assigned to , The time required for a round trip of empty traffic is 200,000, while the MPU information itself is a normalized value and does not need to be processed. Finally, the RTT and MPU signals are processed uniformly. The formula is: 。 4. The periodic traffic transmission optimization method for model distributed training according to claim 1 is characterized in that The specific steps of using Fourier transform to process the flow signal in step S3 are: The original continuous and long signal is divided into many shorter signal segments according to the sliding window and step size. These signal segments are processed by Fourier transform one by one to extract the periodic information of the traffic flow. The sliding window size is initially set to After several rounds of iteration, using the rule of thumb To set the window size ,in, is an empirical parameter, is the standard deviation of RTT and MPU signals; The portion of the flow signal that overlaps with the window function is selected to perform fast Fourier transform, and then the window slides a step size to continue the same operation until the entire continuous signal is analyzed, where the step size is set to half of the window size.

5. The periodic traffic transmission optimization method for model distributed training according to claim 1 is characterized in that: In the step S4: After obtaining the spectrum of each flow according to the above steps, sort it according to the time series and select the multiple frequencies with the largest amplitude. ,pass Calculate the period of each flow signal ; Then, the sending rate of the local traffic is tracked to determine the local traffic period in multiple periods; For subsequent modeling, it is necessary to separate the local traffic and non-local traffic in the complete signal spectrum. First, remove the local traffic frequency portion from the above spectrum and mark it as 0. Secondly, the inverse Fourier transform is used to restore the signal wave of non-local traffic, where only the real part is taken; finally, the non-local traffic part is subtracted from the path traffic signal wave to obtain the local traffic signal wave.

6. The periodic traffic transmission optimization method for model distributed training according to claim 1 is characterized in that: In the step S5: The non-local flow signal and the local flow signal are converted into two sine waves. Description, expressed as: , in, and are the angular frequencies of wave A and wave B respectively, and is their initial phase, and is their amplitude, represents the time vector; The signal wave of the path flow can be expressed as a superposition wave , specifically: ， in, Represents the phase difference; due to ,use Obtain and The period of the least common multiple of , 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 , first of all, because It is a nonlinear function, using numerical gradient Approximately calculate the objective function about Then, use the gradient descent method to update , expressed as: ， in, represents the learning rate, Indicates the number of iteration steps; finally, the optimal phase difference is obtained Then, the optimal time shift value of the local traffic can be obtained. .

Images