Hybrid domain AI high-performance network flow inhibition congestion control method

Abstract

The invention discloses a mixed domain AI high-performance network flow congestion suppression control method, which relates to the technical field of computer networks, and comprises the following steps: a congestion detection step: introducing data packet transmission round-trip time and a switch queue length as congestion detection signals, and a switch queue length management coefficient lambda and a data packet transmission round-trip time management coefficient theta are set. The method comprises the following specific steps: acquiring the current queue length Q of the switch, the data packet round-trip time RTT and the reference round-trip time baseRTT; and when the switch queue length Q is less than the queue length lower limit Kmin or greater than the queue length upper limit Kmax, calculating the data packet congestion marking probability P according to a formula. According to the invention, on the basis of the ECN, RTT indexes which can be collected without special support of network equipment are introduced as congestion detection signals together, and Q and RTT management coefficients are set, so that the response sensitivity of the network to congestion is further improved.

Description

Technical Field

[0001] The present invention relates to the field of computer network technology, and in particular to a hybrid domain AI high-performance network flow suppression congestion control method. Background Art

[0002] In artificial intelligence computing scenarios, especially during large-scale model training, network congestion can lead to prolonged computing time or even computing failure. Mixed training scenarios within and between domains of the intelligent computing center present characteristics such as large network traffic scale, long transmission time, and delay sensitivity. Traditional TCP / IP network protocols are difficult to meet the high bandwidth, low latency, and high reliability requirements of applications in the era of artificial intelligence, as well as the cross-domain interconnection requirements in large-scale computing scenarios.

[0003] Although remote direct memory access (RDMA) technology has been widely used in cloud data centers and intelligent computing centers, such as the Ethernet-based RoCEv2 protocol, RoCEv2 is sensitive to jitter in the underlying network. The Go-back-N retransmission mechanism it adopts will trigger large-scale data retransmission when packet loss occurs in the network, resulting in a sharp decline in network performance.

[0004] Existing congestion control methods, such as explicit congestion notification (ECN), priority-based flow control (PFC), DCTCP, DCQCN, HPCC, TIMELY, etc., have problems such as untimely response to network congestion, insufficiently precise adjustment of the sending rate, and dependence on special network equipment support. They are unable to effectively deal with network congestion in artificial intelligence training scenarios. Summary of the Invention

[0005] The purpose of the present invention is to provide a hybrid domain AI high-performance network flow suppression congestion control method to solve the problems raised in the above background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a hybrid domain AI high-performance network flow suppression congestion control method, comprising the following steps:

[0007] Congestion detection steps: Introduce the packet transmission round-trip time and the switch queue length as congestion detection signals, and set the switch queue length management coefficient λ and the packet transmission round-trip time management coefficient θ. The specific steps are as follows: Obtain the current switch queue length Q, the packet round-trip time RTT, and the base round-trip time baseRTT. When the switch queue length Q is less than the queue length lower limit Kmin or greater than the queue length upper limit Kmax, calculate the packet congestion marking probability P according to the formula:

[0008]

[0009] Furthermore, it also includes: Congestion resolution steps: When the number of real-time data flows in the network changes, FSCC will determine the latest sending end rate based on the network bandwidth bandwidth and the real-time flow number flow_act_num, that is, the latest theoretical convergence rate of the sending end. At the same time, an empirical value is set, which takes a value between 0 and 1, to redundancy a certain bandwidth for the sending end convergence rate to process the data accumulated at the switch.

[0010] Furthermore, in the congestion detection step, λ and θ take values ​​between 0 and 1 and are set according to the sensitivity of the application scenario to the network delay.

[0011] Furthermore, in the congestion detection step, the network congestion state is determined according to the calculated probability P: if P≥1, congestion is determined to occur; if P≤0, no congestion is determined to occur; otherwise, the data packet is ECN marked with the probability P.

[0012] Furthermore, when determining the reduction of the number of real-time data flows flow_act_num, a judgment period ω is introduced. That is, if a certain sending end address still has no data to send after ω judgments, it is determined that the data flow has stopped sending data. ω controls the rate adjustment accuracy of the congestion control algorithm.

[0013] Furthermore, the value of ω is slightly larger than the maximum number of theoretical data flows in the network.

[0014] The present invention provides a hybrid domain AI high-performance network flow suppression congestion control method, which has the following features:

[0015] Beneficial effects:

[0016] 1. Based on ECN, this application introduces the RTT indicator that can be collected without special support from network equipment as a congestion detection signal, and sets Q and RTT management coefficients to further improve the network's response sensitivity to congestion.

[0017] 2. This application triggers congestion processing by sensing changes in the number of flows, and indicates the bandwidth convergence value of the network sending end link through the real-time number of flows, thereby obtaining a faster congestion response and solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a schematic diagram of the congestion detection process of the present invention. DETAILED DESCRIPTION

[0019] The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0020] See also Figure 1The present invention provides a technical solution: a hybrid domain AI high-performance network flow suppression congestion control method, comprising the following steps:

[0021] Congestion detection steps: Introduce the packet transmission round-trip time and the switch queue length as congestion detection signals, and set the switch queue length management coefficient λ and the packet transmission round-trip time management coefficient θ. The specific steps are as follows: Obtain the current switch queue length Q, the packet round-trip time RTT, and the base round-trip time baseRTT. When the switch queue length Q is less than the queue length lower limit Kmin or greater than the queue length upper limit Kmax, calculate the packet congestion marking probability P according to the formula:

[0022]

[0023] In practical applications, first, according to the specific artificial intelligence training scenario and network environment, set the appropriate queue length lower limit Kmin, queue length upper limit Kmax, switch queue length management coefficient λ, and packet transmission round-trip time management coefficient θ. For example, in scenarios sensitive to latency, λ and θ can be set to larger values ​​to increase sensitivity to changes in queue length and RTT.

[0024] The current queue length Q, packet round-trip time RTT, and base round-trip time baseRTT of the switch are obtained in real time through the network management interface. BaseRTT is the theoretical network delay, which can be calculated based on device characteristics.

[0025] When the network is running, the switch queue length Q and the packet round-trip time RTT are collected in real time. If Q is less than Kmin or greater than Kmax, the congestion marking probability P is calculated according to the above formula. If P ≥ 1, the network is determined to be congested and the switch marks the packet with ECN. If P ≤ 0, the network is determined to be non-congested. Otherwise, the packet is ECN-marked with probability P. After the receiver receives the ECN-marked packet, it sends a congestion notification message to the sender.

[0026] Also includes: Congestion resolution steps: When the number of real-time data flows in the network changes, FSCC will determine the latest sender rate based on the network bandwidth bandwidth and the number of real-time flows flow_act_num

[0027]

[0028] That is, the latest theoretical convergence rate of the sender. At the same time, an empirical value is set, ranging from 0 to 1, to redundant a certain bandwidth for the convergence rate of the sender to process the data accumulated at the switch. Among them, FSCC is a hybrid domain AI high-performance network flow suppression congestion control method.

[0029] In the congestion detection step, λ and θ take values ​​between 0 and 1 and are set according to the sensitivity of the application scenario to network delay.

[0030] In the congestion detection step, the network congestion status is determined based on the calculated probability P: if P ≥ 1, congestion is determined to have occurred; if P ≤ 0, no congestion is determined to have occurred; otherwise, the data packet is ECN marked with probability P.

[0031] When determining whether the number of real-time data flows, flow_act_num, is to be reduced, a judgment period ω is introduced to detect whether the sender has stopped sending data. That is, if a sender address still has no data to send after ω judgments, it is considered that the data flow has stopped sending data. ω controls the rate adjustment accuracy of the congestion control algorithm. The value of ω is slightly larger than the maximum theoretical number of data flows in the network. For example, if the theoretical maximum number of data flows is 1000, ω can be set to 1050 to avoid false positives.

[0032] On the sender side, a hash table is maintained to record the sender addresses and counts of active flows. When a packet is received, the sender address is checked to see if it is in flow_act_num. If not, its count is initialized to 0. If it is, the count for that sender address is reset to 0, and the counts for all other sender addresses are incremented by 1. A check interval ω is set, for example, a value slightly larger than the maximum theoretical number of data flows in the network. For sender addresses whose counts reach or exceed ω, the corresponding flow is considered to have stopped sending data and is removed from flow_act_num.

[0033] Based on the classic congestion control algorithm DCQCN, this method introduces the packet transmission round-trip time (RTT) and switch queue length as dual-dimensional congestion detection signals. The congestion marking probability is calculated through a comprehensive formula. This allows congestion detection to not only reflect queue backlogs but also respond to network delay changes in real time. This method is particularly suitable for delay-sensitive application scenarios.

[0034] Compared to the fixed parameter settings in DCQCN, this method sets switch queue length management coefficient λ and packet transmission round-trip time management coefficient θ (both ranging from 0 to 1), which can be adjusted according to the application scenario. For example, in real-time communication scenarios, the value of θ can be increased to prioritize RTT changes, while in file transfer scenarios, the value of λ can be increased to focus on queue length management, thereby achieving more precise congestion control.

[0035] DCQCN lacks an adaptive adjustment mechanism when dealing with changes in the number of data flows. However, in the congestion resolution step, this method determines the latest convergence rate of the sender based on the network bandwidth and the number of real-time flows. It also uses an empirical value α (between 0 and 1) to set redundant bandwidth to handle switch backlog data, making the rate adjustment more closely aligned with actual network load changes.

[0036] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only used to help understand the method of the present invention and its core ideas. The above are only preferred implementation methods of the present invention. It should be pointed out that due to the limitations of textual expression, there are objectively infinite specific structures. For ordinary technicians in this technical field, without departing from the principles of the present invention, they can make several improvements, modifications or changes, and can also combine the above technical features in an appropriate manner; these improvements, modifications, changes or combinations, or the direct application of the inventive concept and technical solution to other occasions without improvement, should be regarded as the scope of protection of the present invention.

Claims

1. A hybrid domain AI high-performance network flow congestion control method, characterized by: The following steps are involved: Congestion detection steps: Introduce the packet transmission round-trip time and the switch queue length as congestion detection signals, and set the switch queue length management coefficient λ and the packet transmission round-trip time management coefficient θ. The specific steps are as follows: Obtain the current switch queue length Q, the packet round-trip time RTT, and the base round-trip time baseRTT. When the switch queue length Q is less than the queue length lower limit Kmin or greater than the queue length upper limit Kmax, calculate the packet congestion marking probability P according to the formula:

2. A hybrid domain AI high-performance network flow congestion control method according to claim 1, characterized in that: Also includes: Congestion resolution steps: When the number of real-time data flows in the network changes, FSCC determines the latest sender rate based on the network bandwidth and the number of real-time flows (flow_act_num). This is the theoretically latest sender convergence rate. An empirical value is also set, ranging from 0 to 1, to allow for a certain amount of redundant bandwidth for the sender convergence rate to process data accumulated at the switch.

3. A hybrid domain AI high-performance network flow congestion control method according to claim 2, characterized in that: In the congestion detection step, λ and θ take values ​​between 0 and 1 and are set according to the sensitivity of the application scenario to the network delay.

4. A hybrid domain AI high-performance network flow suppression congestion control method according to claim 3, characterized in that: In the congestion detection step, the network congestion state is determined according to the calculated probability P: if P ≥ 1, it is determined that congestion occurs; if P ≤ 0, it is determined that no congestion occurs; Otherwise, the packet is ECN marked with probability P.

5. A hybrid domain AI high-performance network flow suppression congestion control method according to claim 4, characterized in that: When determining the reduction of the number of real-time data flows flow_act_num, a judgment period ω is introduced. That is, if a certain sending end address still has no data to send after ω judgments, it is determined that the data flow has stopped sending data. ω controls the rate adjustment accuracy of the congestion control algorithm.

6. A hybrid domain AI high-performance network flow congestion control method according to claim 5, characterized in that: The value of ω is slightly larger than the maximum theoretical number of data flows in the network.

Images