Hybrid congestion control method and system based on RTT and INT for data center network

Abstract

The invention discloses a data center network-oriented hybrid congestion control method and system based on RTT and INT, and relates to the field of computer network transmission control. The invention aims to realize high-precision network congestion control with lower data center network load cost, and when facing network environments with different conditions, the host end and the network nodes can timely sense the condition of the network and select different strategies in a targeted manner, so that the rate of a sender is adjusted to an ideal state. According to the invention, a congestion control scheme based on mixing of a time delay signal and a network telemetry signal is adopted. Network information is analyzed, the state of the network is judged, it is ensured that the algorithm can achieve overall control over the general condition of the network with low network overhead through a control means based on time delay, and it is ensured that the algorithm accurately controls the complex condition of the network through a control means based on network telemetering; and the overall performance of data transmission of the data center network is improved through flexible switching between the two.

Description

Technical Field

[0001] The present invention relates to the field of computer network transmission control, and in particular to a hybrid congestion control algorithm based on RTT and INT for data center networks. Background Art

[0002] Data center networks are key infrastructure platforms for providing online services and applications on the Internet today. A large number of distributed applications are deployed in these data center networks. From daily application services such as online search, social shopping, and video games to large language models, large-scale machine learning systems, high-performance computing clusters, and distributed file systems, they all rely on data center networks to store and transmit data. However, the rapid development of software and hardware technologies in recent years and the rapid changes in application requirements and operation modes of data center networks have brought huge challenges to congestion control technologies in data center networks.

[0003] To solve the congestion problem in data center networks, common algorithms include congestion control based on explicit congestion notification (ECN), congestion control based on round-trip time (RTT), and congestion control based on network telemetry (INT). These algorithms use different congestion feedback signals to adjust network congestion, but they all have some problems.

[0004] The congestion control algorithm based on ECN signals has the problem of complex setting of the marking threshold value. If the marking threshold value is too small, the congestion control algorithm will overreact to changes in network status, resulting in insufficient bandwidth utilization. However, if the threshold value is too high, the algorithm will cause frequent congestion and continuous packet loss. In addition, the algorithm has great problems in dealing with burst traffic and mixed traffic of large and small flows. When dealing with specific traffic patterns such as high bursts, how to effectively utilize idle bandwidth without affecting the transmission of the main traffic is still a considerable challenge. Therefore, ECN is not selected as the congestion signal, but RTT and INT are selected to assist each other in congestion detection and control.

[0005] The congestion control algorithm based on delay information can ensure the correct assessment of network congestion status at a relatively low cost. However, it has problems such as accurate measurement, clock synchronization, and noise interference. The basic delay of a data center is generally a few microseconds or tens of microseconds. Accurately measuring the queuing delay requires high-precision clock synchronization. On the other hand, the delay of some host-side hardware processing can easily introduce noise. The measurement and response of RTT may not be sensitive enough under variable network conditions, and the sending rate cannot be adjusted in a timely and accurate manner. Therefore, RTT, as a congestion signal, is low-cost and simple to design, but it is difficult to make a precise assessment and judgment of the network. Its performance depends to a large extent on the accurate measurement of network and host delays. Therefore, it is necessary to use more precise congestion signals to diagnose network conditions when RTT cannot be accurately adjusted.

[0006] The INT-based congestion control algorithm uses real-time congestion signals actively inserted by the switch, so the information is accurate and rich, and it can achieve more accurate congestion perception compared to ECN and delay. However, it has the problem of high cost. The original INT-based solution will add INT headers when sending data packets regardless of the network conditions and require switches to insert congestion information, which will undoubtedly occupy part of the bandwidth for normal data transmission and impose a certain burden on the network. Therefore, in summary, it is chosen to comprehensively utilize the advantages of the two congestion signals RTT and INT to make up for each other's disadvantages, use RTT signals to ensure normal network transmission, and selectively enable INT signals to detect network anomalies, so as to achieve flexible adaptation to complex network conditions at a lower cost. Summary of the invention

[0007] The technical problems to be solved by the present invention are:

[0008] The present invention proposes a hybrid congestion control method and system based on RTT and INT for data center networks, aiming to achieve high-precision network congestion control at a lower data center network load cost. When facing network environments in different situations, the host side and network nodes can timely perceive the network status and select different strategies in a targeted manner, thereby adjusting the sender's rate to an ideal state, while ensuring high network throughput and reducing network delay and flow completion time as much as possible.

[0009] The technical solution adopted by the present invention to solve the above technical problems is:

[0010] A hybrid congestion control method based on RTT and INT for a data center network, the implementation process of the method is as follows:

[0011] Construct a congestion switching mechanism (congestion control algorithm) based on the network operation status. First, identify the network status, then evaluate the network congestion and implement real-time congestion mechanism switching. Specifically:

[0012] Collect real-time traffic information at the host end, detect and evaluate the network congestion status by statistically analyzing the traffic information, use RTT as a congestion detection signal to control the data transmission process when the network is not congested or slightly congested (under normal circumstances), and when the severity or duration of congestion exceeds the threshold, select to enable the INT signal to notify the switch to participate in the congestion perception and regulation. After a period of adjustment until the network status is observed to return to normal, stop using the INT control signal and switch to the original congestion control based on the delay signal RTT;

[0013] Use the hybrid congestion control mechanism of RTT and INT to perform congestion detection and rate control:

[0014] In the congestion control based on delay signal, the sender will receive the response data packet from the receiver and calculate the delay information therein. Then, based on the comparison between the current measured value and the set threshold, it will choose to additively increase the network rate to improve the network bandwidth utilization in the case of insufficient network throughput. When the network is congested, it will calculate the ratio of the basic round-trip delay to the product of the measured delay and the gradient to dynamically adjust the sending rate of the host end.

[0015] In congestion control based on network telemetry signals, the sender chooses to enable the INT tag to notify the switch to detect the network status. After detecting the tag, the switch judges the network congestion status and inserts the congestion information into the data packet. After that, the sender receives the control data packet sent from the receiver, parses the network congestion information therein, and based on the bottleneck link model, uses the bottleneck queue length in the INT information to calculate the number of flying bytes, uses the historical bottleneck queue information to calculate the gradient of the queue length, uses the ratio of the bandwidth delay product to the number of flying bytes as the parameter for perceiving the overall network status, and uses the ratio of the sending rate to the queue gradient as the parameter for perceiving network changes. The former ensures the algorithm's grasp of the overall network, and the latter improves the algorithm's sensitivity to network change trends. Finally, the two are weighted in combination with the hop count information of the traffic, so as to balance the throughput of long-distance traffic and the delay of short-distance traffic, and improve the algorithm's ability to handle high-speed bursts and complex mixed traffic patterns.

[0016] The present invention has the following beneficial technical effects:

[0017] The hybrid congestion control algorithm for data center networks based on delay and network telemetry provided by the present invention realizes high-precision network congestion control at a lower data center network load cost. When faced with network environments of different situations, the host side and network nodes can timely perceive the network status and select different strategies in a targeted manner, thereby adjusting the sender's rate to an ideal state, while ensuring high network throughput, reducing network delay and flow completion time as much as possible.

[0018] The present invention adopts a congestion control scheme based on a mixture of time-delay signals and network telemetry signals. The network information is analyzed and the network status is judged. The control method based on time delay ensures that the algorithm can achieve overall control of the general network conditions with low network overhead. The control method based on network telemetry ensures that the algorithm can accurately control the complex network conditions. The overall performance of data transmission in the data center network is improved by the flexible switching between the two. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Hybrid congestion control framework diagram;

[0020] Figure 2 Ternary congestion state transition diagram;

[0021] Figure 3 Schematic diagram of delay-based congestion control mechanism;

[0022] Figure 4 Schematic diagram of congestion control mechanism based on network telemetry. DETAILED DESCRIPTION

[0023] The invention is a hybrid congestion control algorithm for data center networks based on latency and network telemetry, which needs to be completed by the sender host, switch, and receiver host. The sender is responsible for collecting traffic information and analyzing the network operation status to switch the congestion control mechanism, and further uses the congestion control algorithm based on RTT signal and INT signal to adjust the sender's sending rate to alleviate the network congestion. The receiver is responsible for returning the received network information to the sender. The switch is responsible for accurately detecting the operation status of the network and passing the congestion information to the host side, thereby assisting the host side to achieve precise control of the network. The overall solution consists of several parts: traffic information collection, network status assessment, congestion mechanism switching, latency signal control, and network telemetry detection, as follows.

[0024] 1. Traffic information collection: When the sender receives a data packet, it classifies the data packet into three categories, namely normal data packets, response data packets, and control data packets, according to the marks carried by the data packet, and processes them separately. For response data packets, analyze and record the timestamp information in them. For control data packets, count and analyze information such as routing queues and queuing delays in them. For normal data packets, no additional processing is performed.

[0025] 2. Network condition assessment: After collecting information, assess the network condition. Specifically, divide the operating condition of the network into three states: no congestion, mild congestion, and severe congestion. In the no-congestion state, the network throughput is less than or equal to the network load, and the queue length in the switch is zero. Only need to simply increase the sending rate. In the mild-congestion state, the network throughput is slightly greater than the network load, and there is packet accumulation in the switch but it never exceeds the queue threshold. Just moderately adjust the sending rate. In the severe-congestion state, the network throughput is much greater than the network load, and new arriving data packets cannot be accepted in the switch buffer. At this time, it is necessary to accurately measure the network state to achieve more precise control. The sender analyzes the current and historical traffic information collected in step 1 to drive the network to switch between these three states. The conversion of the three states is considered from two aspects: the severity and duration of congestion. The system will pre-set a low latency threshold t_min and a high latency threshold t_max. After the sender receives a data packet, calculate its round-trip latency information t_current, and then compare it with the two set thresholds. The low threshold evaluates whether the network is congested, while the high threshold evaluates the severity of network congestion. In addition, set a counter S0 to accumulate in the congested state and clear the counter when the network returns to normal to count the duration of network congestion. Then it will also be compared with the pre-set threshold S1. The network defaults to start from the no-congestion state. If t_current < t_min, it will continuously remain in the no-congestion state. If t_min ≤ t_current ≤ t_max and S0 ≤ S1, it will be converted to the mild-congestion state. If t_current > t_max or S0 > S1, it will be converted to the severe-congestion state. When in the severe-congestion state, judge whether the collected routing queue length qlen is all 0. If not, continuously remain in the current congested state until the switch empties the queue to 0 before it can be converted to the no-congestion state.

[0026] 3. Congestion mechanism switching: switch the congestion control mechanism based on the results of network evaluation. When there is no congestion or light congestion, the round-trip delay information is used to adjust the sending rate of the network host. At this time, the sender will mark INT as disabled when sending the data packet, thereby informing the switch to stop using network telemetry detection. When the switch receives the corresponding data packet, it will check the mark and stop inserting INT information to achieve the purpose of reducing network load costs. When the network is severely or continuously congested, the sender will mark INT as enabled when sending the data packet, thereby starting the use of network telemetry detection. After receiving the corresponding data packet, the switch along the way will measure the status of the network and insert the obtained information into the data packet, thereby achieving more accurate perception and control of the network status.

[0027] 4. Delay signal control: Use delay signals to control the sending rate. For general network situations, use congestion control based on RTT signals, which is divided into congestion detection and rate control. The sender can calculate the round-trip delay RTT of this data by using the difference between the timestamp carried by the received ACK data packet and its arrival time: RTT = t current -ack timestamp For the RTT measurement value obtained, after obtaining the difference based on the historical measurement value, the exponentially weighted average is converted into a smoothed difference: rtt diff =(1-α)×rtt diff +α×(rtt current -rtt last ), and then calculate the gradient value rtt of the delay gradient =rtt diff / rtt_min. When detecting congestion, the measured RTT value is compared with the pre-set threshold. If it is lower than the threshold, the sending rate is adjusted to increase additively: rate = rate current +rate ai ; If it is higher than the threshold, calculate the congestion factor Then the sending rate is dynamically adjusted according to the congestion factor: rate = α × (f × rate current +β)+(1-α)×rate current .

[0028] 5. Network telemetry detection. For situations where network congestion is severe or lasts for a long time, in-band network telemetry is used to detect the network status more accurately. The network telemetry detection part is implemented by the sender, the switch supporting the INT function, and the receiver. When the sender sends a data packet, it adds the INT header field to the message header; then, in the process of transmitting the data packet from the sender to the receiver, each switch on the data link adds metadata information to record the current status of the switch according to the INT flag field carried by the data packet, including the queue length (qlen) of the egress port, the transmission bytes (txBytes), and the link bandwidth capacity (B) and other information; when the receiver receives the data packet, it copies all the INT information recorded by the switch to the control data packet sent back to the sender; then the sender can make a more accurate assessment of the network congestion status based on this information and adjust the window more accurately. The sender is responsible for starting and controlling. When sending data, it adds the INT header field to the data packet header to start congestion detection to implement a selective switching congestion mechanism. When receiving data, it is necessary to parse the corresponding INT header to obtain the corresponding information, and then calculate the congestion degree and adjust the sending rate based on this information in the data packet. The switch is responsible for the perception of network congestion. When receiving a data packet, it chooses whether to perform congestion detection based on whether the data packet carries the INT tag. After receiving the INT tag, the switch will count the network information and monitor the queue length. After that, the switch needs to write the current network information including the queue length into the data packet, and finally forward the data packet to the next switch. All switches along the way operate in this way until the data packet is transmitted to the receiving end. The receiving end is responsible for information feedback. After parsing the INT information, it is responsible for copying the INT information into the generated data packet and sending it to the sender.

[0029] Network telemetry detection needs to handle situations where congestion is severe or lasts for a long time. After the sender analyzes the status of the network nodes, it uses the network status as a basis, starting from the changing trend of the network status and the characteristic requirements of the traffic, to more accurately assess the degree of network congestion and make accurate rate control. Combining the current network status with the historical network status, the gradient information of the switch queue can be calculated to enhance the host's perception of the network change trend. Combining the location information, the characteristics and requirements of traffic at different distances can be balanced in a targeted manner, thereby improving the algorithm's dynamic adaptation to sudden changes in traffic and long- and short-distance mixed traffic. Specifically, the sender traverses the queue lengths of the switches along the way, selects the maximum value as the bottleneck queue length, and then calculates the number of network flight bytes based on the transmission rate and basic delay, which is used as the basic parameter to measure the network status: Inflight = qlen + T × txRate, and then calculates the difference between the current bottleneck queue and the last bottleneck queue: qlen gra =(α×(qlen curr -qlenlast ) / time_diff+(1-α)×qlen gra ), the ratio of the two measured values ​​to their target values ​​can be used as a measure of congestion The former is the ratio of bandwidth delay product to the number of flying bytes to control the overall network status, and the latter is the ratio of sending rate to queue change rate to perceive the trend of network changes. Then, the hop count information is used as a coefficient to perform a weighted average of the two to obtain the adjustment parameter. This achieves a balance between traffic at different distances, making the traffic closer to the network more responsive to changes, while achieving low latency for short-distance traffic without compromising the throughput of long-distance traffic. Finally, this parameter is used to guide the sending rate for more accurate updates. The adjusted rate is determined by the current sending rate and the adjustment parameter u. Rate update formula: Rate = γ × (Rate current ×u+β)+(1-γ)×Rate current .

[0030] It has been verified that the method proposed in the present invention solves the technical problem proposed in the present invention. The method of the present invention has been verified through practical application to verify the technical effect and practicality claimed by the present invention. The method of the present invention has been verified through simulation experiments and practical applications to verify the technical effect claimed by the present invention: the advantages of the two congestion signals RTT and INT are selected to make up for each other's disadvantages, the RTT signal is used to ensure the normal transmission of the network, and the INT signal is selectively enabled to detect network anomalies, thereby achieving flexible adaptation to complex network conditions at a lower cost.

[0031] The algorithm (method) proposed in the present invention is the underlying technical core of the present invention, and various products can be derived based on the algorithm.

[0032] Based on the algorithm (method) proposed in the present invention, a hybrid congestion control system based on RTT and INT for data center networks is developed using a programming language. The system has program modules corresponding to the steps of the above-mentioned technical solution, and executes the steps of the above-mentioned hybrid congestion control method based on RTT and INT for data center networks during runtime.

[0033] The computer program of the developed system (software) is stored on a computer-readable storage medium, and the computer program is configured to implement the steps of the above-mentioned hybrid congestion control method based on RTT and INT for data center networks when called by a processor. That is, the present invention is materialized on a carrier to become a computer program product.

[0034] A hybrid congestion control device based on RTT and INT for a data center network, the device comprising at least one processor and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor so that the at least one processor can execute the above-mentioned hybrid congestion control method based on RTT and INT for a data center network, by selecting RTT and INT to assist each other in performing congestion detection and control.

[0035] Various implementations of the systems and techniques described herein can be realized in digital electronic circuit systems, integrated circuit systems, dedicated ASICs (application specific integrated circuits), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0036] The computer programs (also referred to as programs, software, software applications, or codes) of the present invention include machine instructions for programmable processors, and these computer programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used in the present invention, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, device, and / or device (e.g., disk, optical disk, memory, programmable logic device PLD) for providing machine instructions and / or data to a programmable processor, including a machine-readable medium that receives machine instructions as machine-readable signals. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0037] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this application can be executed in parallel, sequentially or in different orders, as long as the expected results of the technical solution disclosed in this application can be achieved, they are all within the scope of protection of the present invention.

Claims

1. A hybrid congestion control method based on RTT and INT for data center networks, characterized in that: The implementation process of the method is: Construct a congestion switching mechanism based on the network operation status. First, identify the network status, then evaluate the network congestion and implement real-time congestion mechanism switching. Specifically: Collect real-time traffic information at the host end, detect and evaluate the network congestion status by statistically analyzing the traffic information, use RTT as a congestion detection signal to control the data transmission process when the network is not congested or slightly congested, and when the severity or duration of congestion exceeds the threshold, select to enable the INT signal to notify the switch to participate in the congestion perception and regulation. After a period of adjustment until the network status is observed to return to normal, stop using the INT control signal and switch to the original congestion control based on the delay signal RTT; Use the hybrid congestion control mechanism of RTT and INT to perform congestion detection and rate control: In the congestion control based on delay signal, the sender will receive the response data packet from the receiver and calculate the delay information therein. Then, based on the comparison between the current measured value and the set threshold, it will choose to additively increase the network rate to improve the network bandwidth utilization in the case of insufficient network throughput. When the network is congested, it will calculate the ratio of the basic round-trip delay to the product of the measured delay and the gradient to dynamically adjust the sending rate of the host end. In congestion control based on network telemetry signals, the sender chooses to enable the INT tag to notify the switch to detect the network status. After detecting the tag, the switch judges the network congestion status and inserts the congestion information into the data packet. After that, the sender receives the control data packet sent from the receiver, parses the network congestion information therein, and based on the bottleneck link model, uses the bottleneck queue length in the INT information to calculate the number of flying bytes, uses the historical bottleneck queue information to calculate the gradient of the queue length, uses the ratio of the bandwidth delay product to the number of flying bytes as the parameter for perceiving the overall network status, and uses the ratio of the sending rate to the queue gradient as the parameter for perceiving network changes. The former ensures the algorithm's grasp of the overall network, and the latter improves the algorithm's sensitivity to network change trends. Finally, the two are weighted in combination with the hop count information of the traffic, so as to balance the throughput of long-distance traffic and the delay of short-distance traffic, and improve the algorithm's ability to handle high-speed bursts and complex mixed traffic patterns.

2. According to claim 1, a hybrid congestion control method based on RTT and INT for data center network, characterized in that: The hybrid congestion control method is based on the cooperation of the sender host, the switch, and the receiver host; the sender is responsible for collecting traffic information and analyzing the network operation status to switch the congestion control mechanism, and further uses the congestion control algorithm based on the RTT signal and the INT signal to adjust the sender's sending rate to alleviate the network congestion; the receiver is responsible for returning the received network information to the sender; The switch is responsible for accurately detecting the operating status of the network and transmitting congestion information to the host, thereby assisting the host in achieving precise control of the network.

3. According to claim 2, a hybrid congestion control method based on RTT and INT for data center networks is characterized in that: In the construction of a congestion switching mechanism based on network operating status, the process of traffic information collection is as follows: When the sender receives a data packet, the data packet is classified into three categories, namely normal data packets, acknowledgment data packets, and control data packets, according to the marks carried by the data packet and processed separately. For acknowledgment data packets, the timestamp information therein is analyzed and recorded. For control data packets, information such as routing queues and queuing delays therein is counted and analyzed. For normal data packets, no additional processing is performed.

4. A hybrid congestion control method based on RTT and INT for data center networks according to claim 1, 2 or 3, characterized in that: In the construction of a congestion switching mechanism based on network operating status, after collecting information, the network status is evaluated. The process of network status evaluation is as follows: Specifically, the operating status of the network is divided into three states: no congestion, mild congestion, and severe congestion. In the state of no congestion, the network throughput is less than or equal to the network load, and the queue length in the switch is zero. Only the sending rate needs to be simply increased. In the state of mild congestion, the network throughput is slightly greater than the network load, and there is packet accumulation in the switch but it never exceeds the queue threshold. The sending rate can be moderately adjusted. In the state of severe congestion, the network throughput is much greater than the network load, and new arriving data packets cannot be accepted continuously in the switch buffer. At this time, the network status needs to be accurately measured to achieve more precise control. The sender analyzes the current and historical traffic information collected in step 1 and drives the network to switch between these three states. The conversion of the three states is considered from two aspects: the severity and duration of congestion. The system will pre-set a low latency threshold t_min and a high latency threshold t_max. After the sender receives a data packet, it calculates its round-trip latency information t_current, and then compares it with the two set thresholds. Among them, the low threshold evaluates whether the network is congested, and the high threshold evaluates the severity of network congestion. A counter S0 is set to accumulate in the congested state and cleared when the network returns to normal, so as to count the duration of network congestion. Then it will also be compared with the pre-set threshold S1. The network is default to start from the state of no congestion. If t_current < t_min, it will continue to be in the state of no congestion. If t_min ≤ t_current ≤ t_max and S0 ≤ S1, it will be converted to the state of mild congestion. If t_current > t_max or S0 > S1, it will be converted to the state of severe congestion. When in the state of severe congestion, it is judged whether the collected routing queue length qlen is all 0. If it is not 0, it will continue to be in the current congested state until the switch empties the queue to 0 before it can be converted to the state of no congestion.

5. According to claim 4, a hybrid congestion control method based on RTT and INT for data center network is characterized in that: In building a congestion switching mechanism based on the network operation status, the congestion mechanism switching process is: switching the congestion control mechanism through the results of network evaluation. When there is no congestion or light congestion, the round-trip delay information is used to adjust the sending rate of the network host. At this time, the sender will mark INT as disabled when sending a data packet, thereby informing the switch to stop using network telemetry detection. When the switch receives the corresponding data packet, it will check the mark and stop inserting INT information to achieve the purpose of reducing network load costs. When the network is severely or continuously congested, the sender will mark INT as enabled when sending a data packet, thereby starting the use of network telemetry detection. After receiving the corresponding data packet, the switch along the way will measure the network status and insert the obtained information into the data packet, thereby achieving more accurate perception and control of the network status.

6. The hybrid congestion control method based on RTT and INT for data center network according to claim 5, characterized in that: In the hybrid congestion control mechanism using RTT and INT for congestion detection and rate control, the delay signal is used to control the sending rate. For general network conditions, the congestion control based on RTT signal is used, which is divided into two parts: congestion detection and rate control. The sender can calculate the round-trip delay RTT of this data by using the difference between the timestamp carried in the received ACK data packet and its arrival time: RTT = t current -ack timestamp For the RTT measurement value obtained, after obtaining the difference based on the historical measurement value, the exponentially weighted average is converted into a smoothed difference: rtt diff =(1-α)×rtt diff +α×(rtt current -rtt last ), and then calculate the gradient value rtt of the delay gradient =rtt diff / rtt_min. When detecting congestion, the measured RTT value is compared with the pre-set threshold. If it is lower than the threshold, the sending rate is adjusted to increase additively: rate = rate current +rate ai ; If it is higher than the threshold, calculate the congestion factor Then the sending rate is dynamically adjusted according to the congestion factor: rate = α × (f × rate current +β)+(1-α)×rate current .

7. The hybrid congestion control method based on RTT and INT for data center network according to claim 6, characterized in that: In the hybrid congestion control mechanism using RTT and INT for congestion detection and rate control, in-band network telemetry is used to more accurately detect network conditions when network congestion is severe or lasts for a long time. The network telemetry detection part is implemented by the sender, the switch supporting the INT function, and the receiver. When the sender sends a data packet, it adds the INT header field to the message header; Then, in the process of transmitting the data packet from the sender to the receiver, each switch on the data link adds metadata information to record the current status of the switch according to the INT flag field carried by the data packet, including the queue length (qlen) of the egress port, the transmission bytes (txBytes) and the link bandwidth capacity (B) and other information; when the receiver receives the data packet, it copies all the INT information recorded by the switch to the control data packet sent back to the sender; then the sender can make a more accurate assessment of the network congestion status based on this information and adjust the window more accurately. The sender is responsible for starting and controlling. When sending data, the INT header field is added to the header of the data packet to start congestion detection to implement a selective switching congestion mechanism. When receiving data, the corresponding INT header needs to be parsed to obtain the corresponding information, and then the congestion level is calculated and the sending rate is adjusted based on this information in the data packet. The switch is responsible for sensing network congestion. When receiving a data packet, it chooses whether to perform congestion detection based on whether the data packet carries the INT tag. After receiving the INT tag, the switch will count network information and monitor the queue length. After that, the switch needs to write the current network information including the queue length into the data packet, and finally forward the data packet to the next switch. All switches along the way operate in this way until the data packet reaches the receiving end. The receiving end is responsible for information feedback. After parsing the INT information, it is responsible for copying the INT information into the generated data packet and sending it to the sender. Network telemetry detection needs to handle situations where congestion is severe or lasts for a long time. After the sender parses the status of the network nodes, it evaluates the degree of network congestion more accurately and makes accurate rate control based on the network status, based on the changing trend of the network status and the characteristic requirements of the traffic. Combining the current network status with the historical network status, the gradient information of the switch queue can be calculated to enhance the host's perception of the network change trend. Combined with the location information, the characteristics and requirements of traffic at different distances can be balanced in a targeted manner, thereby improving the algorithm's dynamic adaptability to sudden changes in traffic and long- and short-distance mixed traffic.

8. The hybrid congestion control method based on RTT and INT for data center network according to claim 7, characterized in that: The specific process of the rate control is: Specifically, the sender traverses the queue lengths of the switches along the way, selects the maximum value as the bottleneck queue length, and then calculates the number of network flight bytes based on the transmission rate and basic delay, using this as the basic parameter to measure the network status: Inflight = qlen + T × txRate, and then calculates the difference between the current bottleneck queue and the last bottleneck queue: qlen gra =(α×(qlen curr -qlen last ) / time_diff+(1-α)×qlen gra ), the ratio of the two measured values ​​to their target values ​​can be used as a measure of congestion The former is the ratio of bandwidth delay product to the number of flying bytes to control the overall network status, and the latter is the ratio of sending rate to queue change rate to perceive the trend of network changes. Then, the hop count information is used as a coefficient to perform a weighted average of the two to obtain the adjustment parameter. This achieves a balance between traffic at different distances, making the traffic closer to the network more responsive to changes, while achieving low latency for short-distance traffic without compromising the throughput of long-distance traffic. Finally, this parameter is used to guide the sending rate to be updated more accurately. The adjusted rate is determined by the current sending rate and the adjustment parameter u; rate update formula: Rate = γ × (Rate current ×u+β)+(1-γ)×Rate current .

9. A hybrid congestion control system based on RTT and INT for data center networks, characterized by: The system has a program module corresponding to the steps of any one of claims 1 to 8 above, and executes the steps in the hybrid congestion control method based on RTT and INT for data center networks when running.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps of the hybrid congestion control method based on RTT and INT for a data center network as described in any one of claims 1 to 8 when called by a processor.

Images