End-network cooperation rate control method based on explicit congestion notification, medium and equipment

Abstract

The invention provides an end-network cooperative rate control method based on explicit congestion notification, a medium and equipment, and belongs to the field of RDMA rate control. According to the control method, the buffer area of the RLC layer of the gNB serves as a congestion detection point, the queue length of the RLC layer is detected, whether congestion response needs to be made or not is determined according to the queue length, and the congestion response is that a CNP message is actively generated on the gNB and returned to the RH. The RH slows down when receiving the CNP message; the RH updates the congestion feedback coefficient alpha when not receiving the CNP message for a long time; the RH is provided with two counters ntime and nbytes, and the RH is accelerated when any one of the two counters is updated. According to the method, the operation of returning the CNP is directly carried out on the network node sensing the network congestion point, so that a wireless link with long propagation time delay is avoided; and the link data volume is always kept to be up and down fluctuation of the bandwidth delay product, and the congestion influence is reduced to the minimum.

Description

Technical Field

[0001] This invention belongs to the field of RDMA rate control, specifically relating to a method, medium, and device for end-to-end network cooperative rate control based on explicit congestion notification. Background Technology

[0002] In RDMA traffic scenarios, one part is a 5G wireless link, and the other part is a RoCEv2 Ethernet link. This combined wireless and wired scenario presents a unique problem compared to ordinary scenarios: the wireless and wired speeds differ significantly, which negatively impacts traditional rate control methods. For example... Figure 1 As shown.

[0003] If an explicit congestion notification algorithm is used, when the gNB acts as the core congestion detection node, the ECN signal propagation path requires traversing two radio links: downlink to UE and uplink return. The propagation delay of the radio channel is extremely high; a data packet carrying the ECN tag takes 10ms to travel from the gNB, through the air interface radio link to the UE, and back. Even if the UE immediately generates a CNP message upon receiving the ECN tag and transmits it back via the ultra-reliable low-latency communication (ULL) channel specifically configured for 5G networks (i.e., the highest priority channel), the uplink path still experiences a considerable physical layer propagation delay. By the time the CNP finally arrives at the sending end (RH), the delay of the entire control loop far exceeds the allowable range of the RDMA protocol stack.

[0004] Furthermore, in the original DCQCN algorithm, the only action within the network is for the switch to mark the data packets that have experienced congestion with ECN. The location where the receiver needs to wait for the data packets carrying ECN to be transmitted to the receiving end before returning to the CNP to generate the rate control signal is the receiving end. This is unreasonable in scenarios where the propagation delay of the wireless link increases significantly, as it will seriously increase the path delay of rate control and reduce the timeliness of control. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method, medium, and device for end-to-end network cooperative rate control based on explicit congestion notification.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides an end-to-network collaborative rate control method based on explicit congestion notification, which is applied to a wireless RDMA traffic scenario. In the wireless RDMA traffic scenario, there is an Ethernet link of RoCEv2 between a remote server RH, a gateway GW, and a base station gNB, and a 5G wireless link between the gNB and a terminal UE. In the control method, the buffer of the RLC layer of the gNB is used as a congestion detection point to detect the queue length of the RLC layer, and whether to make a congestion response is determined according to the queue length. The congestion response is to actively generate a CNP message on the gNB and return it to the RH.

[0008] Optionally, determining whether to make a congestion response according to the queue length specifically includes: calculating the probability of returning a CNP message according to the queue length of the RLC layer.

[0009] Optionally, the gNB sends a CNP message at regular intervals. [[ID=])

[0010] Optionally, the specific process of the congestion response is as follows:

[0011] Set the queue size K, queue threshold K max ,

[0017] ,

[0016] , , ,

[0013] , , max , min , , ,

[0015] ,

[0014] , and K max , marked probability P max , the minimum time interval N between two CNP messages; record the time when the last CNP message was returned as last_CNP_time, and the current time as current_time;

[0012] When K < K min , no CNP message is generated;

[0013] When K min ≤ K < K max , first calculate the return probability Then determine whether P meets the requirements:

[0014] If the random number r < P, check whether the interval between last_CNP_time and current_time is greater than or equal to N. If so, generate a CNP message and update last_CNP_time; if not, do not generate a CNP message;

[0015] If the random number r ≥ P, do not generate a CNP;

[0016] When K ≥ K max , check whether the interval between last_CNP_time and current_time is greater than or equal to N. If so, generate a CNP message and update last_CNP_time; if not, do not generate a CNP message.

[0017] Optionally, RH slows down when it receives a CNP message; RH updates the congestion feedback coefficient α when it has not received a CNP message for a long time, and α is used to control the aggressiveness of the slowdown; RH is equipped with two counters n. time and n bytes n time This represents a counter that increments by 1 every T time intervals, where n is the number of intervals. bytes This indicates that the counter increments by 1 for every B bytes sent. RH speeds up when either of the two counters is updated.

[0018] Optionally, when RH receives a CNP message, it reduces the speed according to the following formula:

[0019]

[0020] α = α × (1 - g) + g;

[0021] In the formula, SR current This represents the current rate percentage.

[0022] Optionally, if the RH does not receive a CNP message for an extended period, it updates the congestion feedback coefficient α according to the following formula:

[0023] α = α × (1 - g);

[0024] In the formula, g∈(0,1) is the update weight.

[0025] Optionally, RH may increase its speed according to the following formula when either of the two counters is updated:

[0026] When n time >F and n bytes When F > F, where F is a set value, the following formula is used for excess speed increase:

[0027] SR target =min(SR) target +k×Δ HAddI ,1);

[0028] In the formula, SR target For the target rate ratio, k = min(n) time ,n bytes )-F+1,Δ HAddI To increase the step size;

[0029] When n time >F or n bytes When the value is greater than F, the following formula is used for additive rate increase:

[0030] SR target =min(SR) target +Δ AddI,1);

[0031] In the formula, Δ AddI To increase the step size additively;

[0032] When n time ≤F and n bytes When F ≤ F, the target rate ratio is not updated, and the rate recovery is maintained using the following formula:

[0033]

[0034] In the formula, SR current This represents the current rate percentage.

[0035] In a second aspect, the present invention provides a computer-readable storage medium storing a computer program that causes a computer to execute the end-to-end cooperative rate control method based on explicit congestion notification as described in the first aspect.

[0036] Thirdly, the present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the end-to-end cooperative rate control method based on explicit congestion notification as described in the first aspect.

[0037] The beneficial effects of this invention are: (1) Intra-network node speed control: This invention directly places the operation of returning CNP on the network node that senses the network congestion point, avoiding the wireless link that is extended during propagation; (2) End-side response speed control: In this invention, the end side initially sends at line speed, and after receiving CNP, it reduces the speed according to the algorithm, and then gradually increases the speed, always keeping the link data volume fluctuating around the bandwidth-delay product (BDP), minimizing the impact of congestion. At the same time, this invention also takes into account the special characteristics of the wireless and RDMA combined scenario, and provides targeted parameter tuning for the algorithm. Attached Figure Description

[0038] Figure 1 This is a diagram illustrating the long delay in the control path caused by the wireless link.

[0039] Figure 2 This is a flowchart of the rate control method based on explicit congestion notification for end-to-end network coordinated rate control.

[0040] Figure 3 This is a schematic diagram of the 5G protocol stack.

[0041] Figure 4 This is a schematic diagram of the CU-DU separation in a 5G base station architecture.

[0042] Figure 5 This is a diagram illustrating the CNP return probability based on queue length.

[0043] Figure 6 This is a flowchart of the entire end-side speed control algorithm.

[0044] Figure 7 These are experimental results in a no-packet-loss scenario, where (a) is the native DCQCN algorithm and (b) is the end-to-end cooperative rate control method based on explicit congestion notification.

[0045] Figure 8 The results show experimental results with packet loss scenarios, where (a) is the native DCQCN algorithm and (b) is the end-to-end cooperative rate control method based on explicit congestion notification. Detailed Implementation

[0046] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0047] In one embodiment, the present invention proposes an end-to-end network cooperative rate control method based on explicit congestion notification, the process of which is as follows: Figure 2 As shown, the rate control algorithm for end-to-end network collaboration is mainly achieved through rate control by internal network nodes and the response of end-side nodes to rate control signals. The end-to-end network collaborative rate control algorithm consists of two parts:

[0048] 1. Intra-network node rate control: By detecting the length of the RLC layer buffer queue on the gNB node, a probability algorithm is used to determine whether to generate a rate control signal.

[0049] The first step is to select the location for congestion detection. According to the network topology, the intersection of wireless and wired links is the gNB (gear node). Downlink traffic is sent from the wired link to the wireless link. The speed of the wired link is much higher than that of the wireless link. Congestion will occur at this node due to the speed mismatch between the wired and wireless segments. Therefore, it is reasonable to place the detection point on the gNB.

[0050] like Figure 3 As shown, the protocol stack layers with buffers on the gNB consist of SDAP, PDCP, RLC, and MAC layers from top to bottom. In this embodiment, the RLC layer is chosen as the detection point for the following reasons:

[0051] (1) The main function of the SDAP layer is to map the QoS flow of the core network to the Data Radio Bearer (DRB). Its buffer is only set up to temporarily store data packets from the core network or upper layers. After the QoS to DRB mapping is completed, the data packets immediately leave the buffer, and are unaware of the link status of the lower layers. Meanwhile, in actual 5G deployments, there is a CU-DU (Centralized Unit-Distributed Unit) separation architecture. In the CU-DU separation architecture (such as...) Figure 4 Since the SDAP layer is located on the CU, if the channel conditions on the wireless side are to be transmitted to the SDAP layer, the information needs to be transmitted through the CU-DU interface, which greatly increases the additional latency. Therefore, it is obviously inappropriate to place the detection point on the SDAP layer.

[0052] (2) The PDCP layer is mainly used to store data packets from the upper layer until encryption, header compression and other processing are completed. It is only used as a temporary buffer and does not have any perception or response to the status of the lower layer wireless link. Therefore, it is not suitable as a detection point.

[0053] (3) The MAC layer buffer of 5G is the HARQ retransmission buffer, which is used to manage the transmission blocks that will be scheduled to the physical layer. The number of HARQ buffers can be configured according to parameters. Each buffer only stores one transmission block and does not have the concept of a queue, so it cannot be used as a detection point.

[0054] (4) The RLC layer receives PDCP data packets from the upper layer and senses the upper layer data injection rate; at the same time, it senses the scheduling efficiency of the lower layer MAC. The lower layer MAC senses the link capacity and then informs the RLC layer of the amount of data that can be sent this time. The RLC layer assembles the data into transport blocks for transmission based on this information. Therefore, the size of the RLC layer's buffer directly reflects the speed of wired segment data transmission and the current state of the wireless link. The buffer accumulation state generated here is a direct result of the mismatch between wireless and wired rates. Therefore, it is reasonable to choose the RLC layer's buffer as the congestion detection point in this work.

[0055] The detection time is set to each time a data packet is enqueued. The reason for not detecting when a data packet is dequeued is that dequeuing requires waiting for the MAC layer to detect that the lower-layer link status allows transmission before notifying, which has a low frequency and cannot meet the time requirements of rate detection. The rate of enqueuing is closer to the speed of the line segment, so detection is chosen when enqueuing.

[0056] Table 1 returns the CNP algorithm.

[0057]

[0058] Congestion detection refers to detecting the queue length at the RLC layer and determining whether a congestion response is needed based on the queue length. The response involves actively generating a CNP message on the gNB and returning it to the sender. In this embodiment, the detection logic sets a queue threshold K. min and K max The probability of labeling P max The minimum time interval N between two CNP messages is used to calculate the probability of returning a CNP message based on the queue length. The calculation rule is as follows: Figure 5 As shown, a linear relationship exists. However, if the CNP transmission frequency is too high, it will strain link capacity; therefore, CNPs are sent only at intervals. The specific CNP return algorithm is shown in Table 1.

[0059] 2. End-side response rate control: Based on the arrival rate control signal, the transmission rate is quantified and reduced, and parameters are selected specifically to address the problem of high packet loss rate in wireless scenarios.

[0060] This embodiment uses the DCQCN algorithm for rate control in the end-side network nodes, with corresponding improvements for mixed wireless and wired scenarios, and is implemented in the ns-3 RoCEv2 protocol. The core idea is to quickly respond to rate reduction and dynamically increase the rate based on network conditions. At the same time, in order to ensure high throughput and low latency of RDMA, it quickly restores to the original rate, thereby maximizing end-to-end throughput.

[0061] First, some initial settings are required. The initial rate of the sender is:

[0062] S initial =S max ×β start

[0063] Where S max β is the maximum speed of the network card. start This is the scaling factor for the initial rate; the default value is 100%.

[0064] To control the rate reduction and acceleration, some parameters need to be set: the congestion feedback coefficient α is initialized to 1, and the update weight g∈(0,1) has an initial value of 1 / 16. The congestion feedback coefficient α controls the aggressiveness of the rate reduction and needs to be dynamically changed according to the network congestion situation; therefore, a timer 'time' is set for it. α When the timer times out, α is updated. Two rate growth parameters are set for rate recovery: additive increment step size Δ. AddI Superincrement step size Δ HAddI Furthermore, according to the QCN algorithm, in order to quickly recover to normal speed when the rate is very low, two counters need to be set: n time and n bytes , where n timeThis represents a counter that increments by 1 every T time intervals, where n is the number of intervals. bytes This represents a counter that increments by 1 for every B bytes sent.

[0065] There are three times when a response action is required: first, to slow down when a CNP message is received; second, to update α when no CNP message is received for a long time; and third, to perform parameterized speed increase when either of the two counters is updated.

[0066] When the sender receives a CNP congestion message notification, it needs to initiate a rate-down process, remembering the speed before the rate-down for later recovery. Simultaneously, the congestion feedback coefficient α needs to be updated.

[0067]

[0068] α=α×(1-g)+g

[0069] To facilitate decoupling between classes during implementation, this embodiment chooses to control the actual rate ratio rather than directly operating on the absolute speed. The actual rate calculation is then performed during RoCEv2 socket transmission. SR current This represents the current rate percentage.

[0070] timer α Used to detect the arrival of CNP packets. If no CNP packets are received for a long time, it indicates that the network condition is currently good, and α can be gradually reduced according to the following formula:

[0071] α=α×(1-g)

[0072] Decreasing α can control the degree of aggression in deceleration.

[0073] When two counters n time and n bytes When any update occurs (i.e., the time interval since the last update exceeds T or more than B bytes have been sent), the target rate needs to be increased. There are three modes for rate increase: super-increase, additive increase, and fast recovery.

[0074] (1) Overincrement. When n time >F and n bytes When the value is >F, it indicates that data has been continuously and stably transmitted for a relatively long period of time. The following formula can be used for over-incremental speedup:

[0075] SR target =min(SR) target +k×Δ HAddI ,1)

[0076] Among them, SR targetFor the target rate ratio, k = min(n) time ,n bytes )-F+1 allows the step size for rate recovery to change from a fixed value Δ HAddI It transforms into a state where the degree of increase gradually increases over time, thus achieving super-increment.

[0077] (2) Additive increase. When n time >F or n bytes When the value is greater than F, the following formula is used for additive rate increase:

[0078] SR target =min(SR) target +Δ AddI ,1)

[0079] (3) Rapid recovery. When n time ≤F and n bytes When the value is less than or equal to F, there is no need to update the target rate; normal recovery can be maintained.

[0080]

[0081] Figure 6 The flowchart of the entire end-side speed control algorithm is shown.

[0082] Figure 7 (a) and (b) show the native DCQCN algorithm and the end-to-end cooperative algorithm based on explicit congestion signals designed in this embodiment, respectively, under the ideal state of zero packet loss. It can be seen that in the native DCQCN algorithm, the transmitting end rate is indeed suppressed. As shown by the blue wavy line in the figure, the transmitting end slows down (the slope decreases) every time a CNP is received. However, because the CNP needs to pass through two air interfaces, the delay is relatively large, leading to untimely control. It can be seen from the figure that the line segment rate is still higher than the wireless segment rate at intervals, causing queue accumulation. However, in the modified algorithm of this embodiment, the line segment rate is almost always equal to the wireless segment rate, and there is almost no accumulation at intermediate nodes.

[0083] Figure 8(a) and (b) show the native DCQCN algorithm and the end-to-end cooperative algorithm based on explicit congestion signals designed in this embodiment, respectively, under a static packet loss rate of 0.05. The native DCQCN algorithm, at a packet loss rate of 0.05, already allows the GBN algorithm to proceed without sending to the maximum sequence number. By limiting the sending rate, it prevents excessive data accumulation in the buffer, reducing queuing latency. Furthermore, because the receiving end can receive packets faster, the feedback signal arrives slightly sooner, preventing the sending end from remaining in a useless retransmission state for an extended period. The end-to-end cooperative method designed in this embodiment can better suppress the sending end rate, allowing the CNP response logic to be faster. It is worth noting that the breakpoints in the blue lines in the figures are due to the retransmitted packets being lost again, leading to a timeout event. This is a random and unavoidable event, but in the algorithm of this embodiment, the harm of such timeouts can be reduced by reasonably adjusting the sending end rate.

[0084] In another embodiment, the present invention provides a computer-readable storage medium storing a computer program that causes a computer to execute the end-to-end cooperative rate control method based on explicit congestion notification of the foregoing embodiments.

[0085] In another embodiment, the present invention provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the end-to-end cooperative rate control method based on explicit congestion notification of the foregoing embodiment.

[0086] In the embodiments disclosed in this application, a computer storage medium may be a tangible medium that may contain or store programs for use by or in conjunction with an instruction execution system, apparatus, or device. The computer storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of computer storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CDROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0087] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0088] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A terminal-network cooperative rate control method based on explicit congestion notification, applied to a wireless RDMA traffic scenario, wherein the remote server (RH), gateway (GW), and base station (gNB) are connected via a RoCEv2 Ethernet link, and the gNB is connected to the terminal UE via a 5G wireless link; characterized in that: In the control method, the buffer of the RLC layer of the gNB is used as a congestion detection point to detect the queue length of the RLC layer, and whether to make a congestion response is determined according to the queue length. The congestion response is to actively generate a CNP message on the gNB and return it to the RH.

2. The end-to-end network cooperative rate control method based on explicit congestion notification as described in claim 1, characterized in that: The determination of whether to make a congestion response according to the queue length is specifically: calculating the probability of returning a CNP message according to the queue length of the RLC layer.

3. The end-to-end network cooperative rate control method based on explicit congestion notification as described in claim 1, characterized in that: The gNB sends a CNP message at regular intervals.

4. The end-to-end network cooperative rate control method based on explicit congestion notification as described in claim 1, characterized in that: The specific process of the congestion response is as follows: Set the queue size K and queue threshold K of the RLC layer. min and K max The probability of labeling P max The minimum time interval N between two CNP messages; Record the time when the last CNP message was returned as last_CNP_time, and the current time as current_time; When K <K min At this time, no CNP message is generated; When K min ≤K <K max First calculate the return probability. Next, determine whether P meets the requirements: If the random number r < P, check whether the interval between last_CNP_time and current_time is greater than or equal to N. If so, generate a CNP message and update last_CNP_time; if not, do not generate a CNP message; If the random number r ≥ P, do not generate a CNP; When K≥K max When the interval between last_CNP_time and current_time is greater than or equal to N, check if the interval is greater than or equal to N. If it is, generate a CNP message and update last_CNP_time; otherwise, do not generate a CNP message.

5. The end-to-end network cooperative rate control method based on explicit congestion notification as described in claim 1, characterized in that: The RH reduces the speed when receiving a CNP message; the RH updates the congestion feedback coefficient α when it has not received a CNP message for a long time. α is used to control the aggressiveness of speed reduction; RH is configured with two counters n time and n bytes n time This represents a counter that increments by 1 every T time intervals, where n is the number of intervals. bytes This indicates that the counter increments by 1 for every B bytes sent. RH speeds up when either of the two counters is updated.

6. The end-to-end network cooperative rate control method based on explicit congestion notification as described in claim 5, characterized in that: When the RH receives a CNP message, it reduces the speed according to the following formula: α = α × (1 - g) + g; In the formula, SR current This represents the current rate percentage.

7. The end-to-end network cooperative rate control method based on explicit congestion notification as described in claim 5, characterized in that: When the RH has not received a CNP message for a long time, it updates the congestion feedback coefficient α according to the following formula: α＝α×(1-g)； In the formula, g ∈ (0, 1) is the update weight.

8. The end-to-end network cooperative rate control method based on explicit congestion notification as described in claim 5, characterized in that: When either of the two counters in the RH is updated, it increases the speed according to the following formula: When n time >F and n bytes When F > F, where F is a set value, the following formula is used for excess acceleration: SR target =min(SR target +k×Δ HAddI ,1); In the formula, SR target For the target rate proportion, k = min(n) time ,n bytes )-F+1,Δ HAddI To increase the step size; When n time >F or n bytes When the value is greater than F, the following formula is used for additive rate increase: SR target =min(SR target +Δ AddI ,1); In the formula, Δ AddI To increase the step size additively; When n time ≤F and n bytes When F ≤ F, the target rate ratio is not updated, and the rate recovery is maintained using the following formula: In the formula, SR current This represents the current rate percentage.

9. A computer-readable storage medium storing a computer program, characterized in that, The computer program causes the computer to execute the end-to-network collaborative rate control method based on explicit congestion notification according to any one of claims 1-8.

10. An electronic device, characterized in that, It includes: A memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the end-to-network collaborative rate control method based on explicit congestion notification according to any one of claims 1-8.

Images