Congestion information inference trust-sensitive congestion control method, device and medium

Abstract

The application discloses a congestion control method and device sensitive to congestion information inference credibility in a data center network and a medium. The method divides the congestion control system state into a region where congestion information inference is credible and a region where congestion information inference is not credible based on a congestion criterion, system inherent inertia represented by queuing delay and its change trend. In the region where congestion information inference is credible, an analytical speed regulation rule is adopted to efficiently utilize current congestion information to guide speed regulation and maintain the system state in the region where congestion information inference is credible as much as possible. In the region where congestion information inference is not credible, a probing speed regulation rule is adopted to only perform coarse-grained directional speed regulation according to congestion information and push the system state to the region where congestion information inference is credible. Experimental results show that the congestion control method has high throughput and low delay, does not need to modify switches and is an economical and effective method suitable for data center networks.

Description

Technical Field

[0001] This invention relates to the field of congestion control in data center networks, and more particularly to a congestion control method, device, and medium based on congestion information to infer credibility sensitivity. Background Technology

[0002] Since its introduction in Vegas, delay signals have consistently played a crucial role in congestion control (CC). Today, mainstream CC solutions such as Copa and BBR still rely heavily on delay signals. In data center environments, thanks to hardware timestamps and a more controllable network environment, the measurement accuracy of delay signals has significantly improved. Compared to the ECN signals used in DCTCP and DCQCN, delay signals can reflect congestion intensity more precisely. Therefore, delay-based CC has become an ideal choice for data centers.

[0003] Numerous latency-based congestion information (CiD) methods have been applied in data centers, all following the same architecture: latency measurement, CiD, and rate adjustment rules. For example, Timely achieves precise latency measurement through hardware timestamps. Its CiD focuses on both queue length (queue latency) measured based on RTT and the queue length variation trend (latency gradient). This CiD directly provides congestion status information to the probing rate adjustment rules, adjusting the transmission rate heuristically and iterating to the desired value. Furthermore, Swift inherits precise latency measurement but only uses the queue length from the CiD results to guide its probing rate adjustment rules, avoiding multiple equilibrium points by eliminating the use of the latency gradient. To address the issue of latency lagging behind ECN in cases of large queue latency, Swift also sets a lower latency target for the probing rate adjustment rules. Additionally, θ-PowerTCP measures both latency and latency gradient, applying aggregated arrival rates to achieve CiD. Subsequently, analytical rate adjustment rules are employed to achieve "one-step" rate adjustment, precisely adjusting the transmission rate to the desired value. Traditionally, latency-based CC directly infers the provided information from congestion information for rate adjustment. However, although latency signal measurements have become more precise and rate adjustment rules have been improved, latency signals may still fail to stably and equivalently reflect the true congestion situation under certain conditions; these situations can be summarized as "latency ambiguity conditions" (DAC). In these scenarios, latency-based CC performance may still degrade due to incorrect congestion information inference. For example, underutilized links can make latency and its gradient "appear to be zero," thus masking changes in actual available bandwidth; large queues can cause significant lag in latency signals, leading to delayed perception of congestion changes; when queues experience sudden or severe fluctuations, change points may be "skipped," distorting gradient estimation and causing biases in inferences about the number of packets in transit, aggregation arrival rate, or congestion intensity, or even absurd values ​​(such as negative aggregation arrival rates). When CC uses these "incorrect congestion information inferences" as reliable inputs to drive speed regulation, it can easily lead to performance degradation problems such as insufficient bandwidth utilization, incorrect acceleration or deceleration direction or magnitude, convergence point drift, and queue oscillation. Conversely, if it is too conservative and only uses the inference results as directional cues, it may waste the rich information that the delay signal could provide, making the response to environmental changes slower. Summary of the Invention

[0004] To address the technical problem that existing technologies may suffer from inaccurate congestion information inference under conditions of time delay ambiguity, leading to misleading congestion control (CC) and ultimately performance degradation, this invention proposes a congestion control method, device, and medium based on the reliability sensitivity of congestion information inference, which offers high throughput, low latency, and efficient congestion management.

[0005] To achieve the above-mentioned technical objectives, the technical solution provided by this invention is as follows:

[0006] In a first aspect, the present invention provides a congestion control method sensitive to the reliability of congestion information inference, comprising the following steps:

[0007] Step 1: Measure the round-trip time (RTT) upon receiving the acknowledgment (ACK) message. Then, calculate the queuing delay and the delay gradient representing the trend of queuing delay changes based on the RTT. Finally, infer congestion information based on the weighted sum of the queuing delay and the delay gradient: when the weighted sum is positive, the congestion information is inferred as congestion; otherwise, it is non-congestion.

[0008] Step 2: Based on queuing delay, delay gradient, and congestion information, the system state is divided into trusted and untrusted regions.

[0009] The trusted region is the region where the impact of time delay ambiguity is tolerable, including the trusted speed-up sub-region and the trusted speed-down sub-region. In the trusted speed-up sub-region, the queuing delay is higher than the expected value, the time delay gradient is negative, and the congestion information is inferred as non-congestion. In the trusted speed-down sub-region, the queuing delay is higher than the expected value, the time delay gradient is negative, but the congestion information is inferred as congestion.

[0010] The untrusted region is the area where the impact of latency ambiguity is intolerable, including the untrusted short queue speed-up sub-region, the untrusted short queue restricted speed-up sub-region, the untrusted short queue deceleration sub-region, and the untrusted long queue deceleration sub-region. In the untrusted short queue speed-up sub-region, the queuing latency is lower than the expected value, the latency gradient is negative, and the congestion information is inferred as non-congestion. In the untrusted short queue restricted speed-up sub-region, the queuing latency is lower than the expected value, the latency gradient is positive, but the congestion information is inferred as non-congestion. In the untrusted short queue deceleration sub-region, the queuing latency is lower than the expected value, the latency gradient is positive, and the congestion information is inferred as congestion. In the untrusted long queue deceleration sub-region, the queuing latency is higher than the expected value, the latency gradient is positive, and the congestion information is inferred as congestion.

[0011] Step 3: When the system is in a trusted region, the analytical rate adjustment rule is used to calculate the transmission rate adjustment amount; when the system is in an untrusted region, the probing rate adjustment rule is used to calculate the transmission rate adjustment amount.

[0012] The analytical rate adjustment rule includes: in the trusted speed-up sub-region, the transmission rate to be increased is calculated based on the queuing delay and delay gradient; in the trusted speed-down sub-region, the transmission rate is reduced by multiplying the current rate by an adaptive coefficient that will automatically decrease as congestion eases, and a fairness factor is introduced as a speed adjustment switch; and in both sub-regions, the ratio of the ACK arrival interval to the minimum round-trip time is used as a smoothing factor to smooth the rate adjustment.

[0013] The probing rate adjustment rules include: increasing the transmission rate in untrusted short queue speed-up sub-regions and untrusted short queue speed-limited speed-up sub-regions, and rapidly probing idle bandwidth in exponentially increasing steps; simultaneously, the speed-up coefficient of flows with higher transmission rates is smaller, and the growth stops when the rate exceeds half of the link estimate; decreasing the transmission rate in untrusted short queue speed-down sub-regions and untrusted long queue speed-down sub-regions, and the rate deviates more from the expected value as the current queuing delay increases; and using the ratio of ACK arrival interval to minimum round-trip delay as a smoothing factor to smooth rate adjustment in all four sub-regions, and introducing a fairness factor as a speed adjustment switch.

[0014] Step four: Update the sending rate based on the results of step three, and control the message sending rhythm and the amount of message data sent according to the updated sending rate.

[0015] Preferably, in step one, the queuing delay for:

[0016] ;

[0017] in This represents the round-trip time measured when the sender receives the k-th ACK; It is the minimum RTT value currently recorded; the queuing delay gradient is:

[0018] ;

[0019] in It is an exponentially weighted parameter. It is the time interval between the two most recent ACKs, calculated as follows:

[0020] ;

[0021] in It is the timestamp record of the arrival time of the k-th ACK. This represents the timestamp record of the (k-1)th ACK arrival time.

[0022] Preferably, in step one, the congestion information inference is obtained by a congestion criterion. The calculation formula is:

[0023] ;

[0024] in This is the expected queuing delay; according to The calculation formula, when and hour, 0; then when When extremely small, maintain The expected value of 0 corresponds to The value should be a very large positive value, if at this time The actual value leads to the calculated If the value is a large positive value, the congestion information is inferred to be non-congested; while when... When it is large, maintain The expected value of 0 corresponds to The value should be a very large negative value, if at this time The actual value leads to the calculated If the value is a large negative value, then the congestion information is inferred to be congestion.

[0025] Preferably, step one further includes calculating an approximate congestion criterion using the following piecewise function. to replace To reduce computational overhead:

[0026] ;

[0027] in, It is calculated through displacement operations.

[0028] Preferably, in step two, the system state region is divided into the following six regions:

[0029] 1) For the reliable growth sub-region, by the curve , and Define; 2) For the reliable deceleration sub-region, by curve , and divergence boundary Definition; 3) For the untrusted short queue deceleration sub-region, by the curve , and divergence boundary Definition; 4) For the untrusted long queue deceleration sub-region, via curve , and divergence boundary Define; 5) For untrusted short queue growth sub-regions, through curves , and Define; 6) For unreliable short queues with limited growth sub-regions, the curve is used. and 0 is defined.

[0030] Preferably, in step three, the expression for the analytical speed regulation rule is:

[0031] ;

[0032] in, This indicates that within the reliable growth sub-region, This indicates that within the trusted deceleration sub-region, This represents the rate adjustment amount, where M is the data packet size. For the desired delay, As a smoothing factor, Indicates the ACK interval. To adaptively update the coefficients, , It is a fairness factor, calculated as follows:

[0033] ;

[0034] in It is an independent reference rate for each stream, calculated as follows:

[0035] ;

[0036] in This indicates the line speed of the network interface card (NIC).

[0037] Preferably, in step four, the expression for the probe-based speed regulation rule is:

[0038] ;

[0039] in This indicates that within the untrusted short queue growth sub-region and the untrusted short queue restricted growth sub-region, This indicates that within the untrusted short queue deceleration sub-region and the untrusted long queue deceleration sub-region, , All are adaptive update coefficients; S is used to control the magnitude range of rate adjustment. When the system leaves region A, the value of S will be adjusted according to the most recently recorded... Update the data using the average of the five transmission rates closest to the current time; within the region Inside, And it needs to be multiplied by To ensure fairness when fully loaded; in the region Inside, ; It is an adaptive acceleration factor, and the update formula is: ,in This indicates the number of consecutive RTTs when the link is in a low utilization state.

[0040] In a second aspect, the present invention provides an electronic device, comprising:

[0041] One or more processors;

[0042] Storage device for storing one or more programs;

[0043] The one or more processors execute the one or more programs to implement the aforementioned method.

[0044] Thirdly, the present invention provides a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method.

[0045] The technical advantage of this invention lies in that, before using congestion information to infer information, it first roughly classifies the credibility of the congestion information inference based on the system state. Based on the different credible and untrustworthy regions where the current system state is located, the results of the congestion information inference are used selectively. This allows for efficient rate adjustment by fully utilizing rich congestion information, while suppressing inefficient rate adjustment caused by untrustworthy congestion information inference. Therefore, this invention not only efficiently addresses congestion but also guides the system state towards or maintains a credible region during rate adjustment, thereby continuously improving the credibility of subsequent congestion information inference and achieving comprehensive optimization of congestion control performance. Experimental results show that the congestion control method proposed in this invention has high throughput and low latency, requires no modification to the switch, and is an economical and effective method suitable for data center networks. Attached Figure Description

[0046] Figure 1 This is a diagram illustrating the overall framework of the congestion control method (CCC) based on congestion information inference, which is sensitive to congestion control.

[0047] Figure 2 This diagram illustrates the fairness and efficiency comparison between CCC and DCQCN, a mainstream congestion control method in data centers, in a real-world environment. (a) shows the evolution of rates for different methods, and (b) shows the fairness and efficiency of different methods.

[0048] Figure 3 This diagram illustrates the speed comparison between CCC and mainstream congestion control methods in converging to a fair state.

[0049] Figure 4 This diagram illustrates the differences in Flow Completion Time (FCT) between CCC and other mainstream latency-based congestion control methods under typical data center network workloads. (a) shows a comparison of Facebook's Hadoop workload, and (b) shows a comparison of Microsoft's Web Search workload. Detailed Implementation

[0050] Figure 1This is the overall framework diagram of CCC, a congestion control method based on congestion information inference with sensitivity. The sender obtains the delay from the received ACK and calculates the congestion information inference result, while coarsely classifying the reliability of the inference based on the current system state. When the system is in the reliable region, the negative impact of delay ambiguity is considered tolerable, and analytical speed regulation rules are used to efficiently utilize the current congestion information to guide speed regulation, while ensuring that the system state is maintained in the reliable region as much as possible. When the system is in the unreliable region, probing speed regulation rules use heuristic rules based on the AIMD pattern to perform coarse-grained directional speed regulation and push the system state towards the reliable region. By differentiating the use of congestion information inference under different reliability regions, this method can ensure efficient speed regulation within the reliable region and suppress erroneous speed regulation caused by incorrect congestion information inference, thereby alleviating the bandwidth shortage and queue oscillation problems commonly encountered by delay-based congestion control in dynamic environments. Its operation includes the following steps:

[0051] Step 1: Apply the congestion criterion to the time delay and the time delay gradient to obtain congestion information inference. The congestion criterion is defined as follows:

[0052]

[0053] in This is the expected queuing delay, which is empirically set to 5 microseconds. The queuing delay is calculated based on the k-th ACK message. It is its changing trend (i.e., the time delay gradient). Using the above formula, congestion information can be inferred based on the current time delay signal and the time delay gradient. =0 can be used as a transitional target for analysis rate adjustment, with its desired target point being... This corresponds to the system state when link utilization is sufficient and queuing time is short. Based on the stated criterion, when queuing delay... When it is extremely small, 0 corresponds to The value will increase significantly. At this point, the sending rate can be increased to a larger value to quickly utilize available bandwidth. Conversely, when queuing delay... When the value is large, the transmission rate can be reduced to a smaller value to suppress network congestion.

[0054] The queuing delay in the above formula is defined as follows:

[0055]

[0056] in This represents the round-trip time measured when the sender receives the k-th ACK. It is the minimum known RTT value currently recorded by the method.

[0057] The queuing delay gradient is defined as follows:

[0058]

[0059] in It is the exponential weighting parameter, empirically set to 0.2. It is the time interval between the two most recent ACKs, calculated as follows:

[0060]

[0061] Where t is the timestamp record of the arrival time of the kth ACK.

[0062] In this embodiment, the following piecewise function can be considered for calculating the approximate congestion criterion in the hardware implementation. to replace To reduce expenses:

[0063]

[0064] in, It is calculated through displacement operations.

[0065] Step 2: Based on the congestion criterion and the queuing delay and its changing trend, map the current system state to a preset region. The system state is divided into regions as follows:

[0066]

[0067] in The region is composed of curves , and Define; The region is composed of curves , and divergence boundary Definition. Among them... The region is composed of curves , and divergence boundary ; The area passes through the curve , and divergence boundary Define; The area passes through the curve , and Certainly; The area passes through the curve and 0. Definition. Type A regions refer to reliable regions, where the queuing status and its changing trends reflected by delay samples are generally consistent with the actual system state, and the impact of delay ambiguity is tolerable. Type H regions refer to unreliable regions, where the correspondence between delay signals and congestion states is uncertain or distorted (delay ambiguity), and the impact of delay ambiguity is intolerable. In this case, congestion information inferred from delay signals should not be used as accurate congestion intensity. Region A can be further divided into... (Confidential speed-up sub-region) and (Trusted deceleration sub-region). district and If the congestion criterion is relatively small, meaning the corresponding queue length and its variation are small, the congestion criterion indicates that the resources are not yet fully utilized or are within a range where the rate can be safely increased. In other words, there is still room for the transmission rate to be increased, and the rate should be increased. The congestion criteria described in section H indicate an increase in queuing delay or a congestion trend, suggesting that the queue is too high and although it is shrinking, the reduction is insufficient and has not yet effectively returned to the desired state; therefore, speed should be reduced. Section H can be divided into four sub-sections. (Untrusted short queue deceleration sub-region) and The (untrusted long queue deceleration sub-region) corresponds to the cases of drastic queue fluctuations and large queue size, respectively. In these two regions, the information represented by the time delay may be distorted. (Untrusted short queue speed-up sub-region) and The (Untrusted Short Queue Limited Growth Sub-region) is characterized by low queue latency and slow queue growth, facing underutilization of link resources, which may distort latency information. Due to the small queue size, queue fluctuations are often significant. and It is also susceptible to the distortion of time delay information caused by drastic fluctuations in the queue.

[0068] Step 3: When the system state falls within the trusted zone, the analytical rate adjustment rule is used to calculate the transmission rate adjustment. The calculation method for the analytical rate adjustment rule is as follows:

[0069]

[0070] in The rate adjustment increment is M, where M is the data packet size. To the expected queuing delay, , All coefficients are adaptively updated. The regional, analytical speed regulation rule fully utilizes the results inferred from congestion information to quickly converge the system state to the target point. In the region, the system will spontaneously move to the left towards the curve due to inertia. .when Approaching hour, The rate adjustment step size within the region will gradually approach zero. Therefore, system inertia will dominate, avoiding overshoot and unexpected departure from the time-delay ambiguity tolerance region (region A). Smoothing factor This factor is used to help mitigate the impact of sudden drops in latency caused by long queues. Specifically, this factor is the ACK interval. With minimum round-trip delay The ratio is limited to (0,1], thus keeping the cumulative speed adjustment amplitude stable per unit time: when ACK is dense, it leads to When the size decreases, the step size of a single update is reduced accordingly to avoid overreaction; when ACKs are lost or sparse... When the size increases, the step size of each update is appropriately increased to compensate for missing updates. Through this "time normalization" step size adjustment, the algorithm can effectively suppress speed jitter and erroneous speed adjustment caused by factors such as uneven arrival of ACKs and delayed signal lag.

[0071] In the method described, as In addition, set an independent reference rate for each stream. To accelerate fairness convergence:

[0072]

[0073] in This indicates the line speed of the network interface card (NIC). Traffic sharing the same bottleneck link has the same... Values, for each flow rate They can be considered equal. When When the current sending rate is too low, it indicates that the current sending rate is not competitive with other traffic; conversely, it indicates the same. The direction of the rate adjustment cannot be determined; it is solely determined by... Decide.

[0074] Step four: When the system state falls into the untrusted region, the transmission rate adjustment is calculated using a probe-based rate adjustment rule. The calculation method for the probe-based rate adjustment rule is as follows:

[0075]

[0076] in , , All are adaptive update coefficients. S indicates the magnitude range of the rate adjustment. When the system leaves region A, the value of S will be adjusted according to the recorded data. At that time, the average of the five most recent transmission rates is updated. It is ensured that the S-value for each flow will fairly converge to C / N, where C is the measured link bandwidth and N is the number of flows. In the region... Inside, and multiply by a factor To increase fairness; in the region Inside, ,because The point in the middle will be spontaneously pushed to due to the system's inertia. .in The ratio is used to mitigate the risk of insufficient link utilization. Therefore, data streams with higher transmission rates will gradually increase their rates, and only when the transmission rate exceeds 0... The boosting process stops temporarily to wait for other data streams. within the area, Therefore, its rate decreases as congestion intensity increases to quickly clear the bottleneck queue. AIMD-style rate adjustment keeps the system state aligned with... The path runs smoothly. The update formula is ,in This indicates the number of consecutive RTTs when the link is in a low-utilization state. If exist If the bandwidth utilization is close to 0 within a 5% error range, it is considered a low utilization state. This exponentially increasing rate design can significantly improve bandwidth utilization efficiency. It is the fairness factor, and its calculation method is as follows:

[0077]

[0078] when When a higher rate is suggested, the method will stop processing data at rates exceeding that rate. The transmission rate is adjusted accordingly. Meanwhile, the sending rate is lower than... The traffic can still be accelerated. Conversely, when When it is recommended to reduce the rate, the opposite is true.

[0079] Step 5: Update the sending rate based on the adjusted sending rate, and control the message sending rhythm and the amount of data in transit according to the updated sending rate. The updated sending rate is defined as:

[0080]

[0081] in This indicates the past transmission rate. Indicates the future transmission rate.

[0082] Figure 2This paper demonstrates a comparison of fairness and efficiency between CCC and DCQCN, a mainstream congestion control method in current data centers, in a real-world environment. To address unpredictable external interference, the queuing delay of the target node is set to 20µs. This is illustrated when multiple senders simultaneously transmit long streams of data to the receiver's 40Gbps port (e.g.,...). Figure 2 As shown in (a), the CCC protocol can quickly achieve fair bandwidth allocation while maintaining full link load. In contrast, although the DCQCN protocol can achieve full link load, it failed to achieve fairness convergence during testing. In other words, CCC achieves a better balance between fairness and efficiency. This conclusion is based on... Figure 2 (b) The fairness-efficiency trajectories of CCC and DCQCN are verified—the darker the color, the later the time. CCC can quickly converge to the intersection of the fairness line and the efficiency line in only a few RTTs (round-trip times), which shows that it effectively achieves fair allocation between streams while ensuring data transmission efficiency.

[0083] Figure 3 The figure demonstrates the convergence speed of CCC and mainstream congestion control methods to fairness. As shown, the CCC method converges to a fair state the fastest. Specifically, the CCC method can quickly achieve fairness and continuously approach the efficiency line along the fairness line. This method benefits from its rate adjustment rule, which simultaneously references and incorporates congestion criteria. and independence The congestion metric is [not specified]. In contrast, the HPCC method converges to a fair state more slowly, as described in the literature.

[0084] Figure 4This study demonstrates the differences in FCT between CCC and other mainstream latency-based congestion control methods under typical workloads in data center networks. The experiments employed a large-scale simulation using a Fat tree network topology consisting of 250 hosts. CCC and other congestion control methods were tested using real-world workloads with FB_Hadoop and Web Search loads at 30%. Dashed lines represent INT-based congestion control methods and their variants. In summary, CCC's FCT is comparable to HPCC and PowerTCP in terms of 95% tail FCT latency for short flows (<10KB), except for a higher average (16.4%). In contrast, CCC's 95% tail FCT latency for short flows is 41.0%–43.6% lower than DCQCN and timely FCT latency on FB_Hadoop, and 9.4%–12.5% ​​lower on Web search. Furthermore, another congestion control method employing latency-based analytical rate control rules, θ-PowerTCP, has a similar 95% tail FCT latency to HPCC and PowerTCP, but performs poorly in FCT latency for long flows (>1MB). This is because θ-PowerTCP still performs analytical rate adjustment even when congestion information is inferred to be unreliable. CCC achieves 31.9%–47.4% and 34.7%–48.8% lower average FCT latency for long flows on FB_Hadoop and Web search, respectively, compared to θ-PowerTCP. Furthermore, while Swift performs exceptionally well on Web search, it exhibits higher FCT for short flows on FB_Hadoop.

[0085] According to embodiments of the present invention, the present invention also provides an electronic device and a computer-readable medium.

[0086] Electronic devices include:

[0087] One or more processors;

[0088] Storage device for storing one or more programs.

[0089] When the one or more programs are executed by the one or more processors, the one or more processors implement a congestion control method that is sensitive to the reliability of congestion information inference.

[0090] In practical use, users can interact with servers, which are also electronic devices, via a network to receive or send messages. Terminal devices are generally various electronic devices equipped with a display and used through a human-computer interface, including but not limited to smartphones, tablets, laptops, and desktop computers. Various specific application software can be installed on these terminal devices as needed, including but not limited to web browsers, instant messaging software, social media platforms, and shopping apps.

[0091] A server is a network service provider that offers various services. The congestion control method provided in this embodiment is generally executed by the server. In practical applications, terminal devices can also directly execute the congestion control method under necessary conditions.

[0092] Similarly, the computer-readable medium of the present invention stores a computer program thereon, which, when executed by a processor, implements the congestion control method of congestion information inference reliability sensitive according to embodiments of the present invention.

Claims

1. A congestion control method sensitive to the reliability of congestion information inference, characterized in that, Includes the following steps: Step 1: Measure the round-trip time (RTT) upon receiving the acknowledgment (ACK) message, then calculate the queuing delay and the delay gradient representing the trend of queuing delay changes based on the RTT. Finally, infer congestion information based on the weighted sum of the queuing delay and the delay gradient: when the weighted sum is positive, the congestion information is inferred as congestion. Otherwise, it is not congested; Step 2: Based on queuing delay, delay gradient, and congestion information, the system state is divided into trusted and untrusted regions. The trusted region is the region where the effects of time delay ambiguity are tolerable, including the trusted acceleration sub-region and the trusted deceleration sub-region; In the reliable speed-up sub-region, the queuing delay is higher than the expected value, the delay gradient is negative, and the congestion information is inferred to be non-congested; In the trusted deceleration sub-region, the queuing delay is higher than the expected value, the delay gradient is negative, but the congestion information is inferred to be congestion; The untrusted region is the region where the time delay ambiguity effect is intolerable, including the untrusted short queue speed-up sub-region, the untrusted short queue restricted speed-up sub-region, the untrusted short queue speed-down sub-region, and the untrusted long queue speed-down sub-region; In the untrusted short queue speed-up sub-region, the queuing delay is lower than the expected value, the delay gradient is negative, and the congestion information is inferred as non-congestion; in the untrusted short queue restricted speed-up sub-region, the queuing delay is lower than the expected value, the delay gradient is positive, but the congestion information is inferred as non-congestion; in the untrusted short queue speed-down sub-region, the queuing delay is lower than the expected value, the delay gradient is positive, and the congestion information is inferred as congestion; in the untrusted long queue speed-down sub-region, the queuing delay is higher than the expected value, the delay gradient is positive, and the congestion information is inferred as congestion. Step 3: When the system is in a trusted region, the transmission rate adjustment amount is calculated using analytical rate adjustment rules; When the system is in an untrusted region, a probe-based rate adjustment rule is used to calculate the transmission rate adjustment amount. The analytical rate adjustment rule includes: in the trusted speed-up sub-region, the transmission rate to be increased is calculated based on the queuing delay and delay gradient; in the trusted speed-down sub-region, the transmission rate is reduced by multiplying the current rate by an adaptive coefficient that will automatically decrease as congestion eases, and a fairness factor is introduced as a speed adjustment switch; and in both sub-regions, the ratio of the ACK arrival interval to the minimum round-trip time is used as a smoothing factor to smooth the rate adjustment. The probing rate adjustment rules include: increasing the transmission rate in untrusted short queue speed-up sub-regions and untrusted short queue speed-limited speed-up sub-regions, and rapidly probing idle bandwidth in exponentially increasing steps; simultaneously, the speed-up coefficient of flows with higher transmission rates is smaller, and the growth stops when the rate exceeds half of the link estimate; decreasing the transmission rate in untrusted short queue speed-down sub-regions and untrusted long queue speed-down sub-regions, and the rate deviates more from the expected value as the current queuing delay increases; and using the ratio of ACK arrival interval to minimum round-trip delay as a smoothing factor to smooth rate adjustment in all four sub-regions, and introducing a fairness factor as a speed adjustment switch. Step four: Update the sending rate based on the results of step three, and control the message sending rhythm and the amount of message data sent according to the updated sending rate.

2. The method according to claim 1, characterized in that, In step one, the queuing delay for: ； in This represents the round-trip time measured when the sender receives the k-th ACK; It is the minimum RTT value currently recorded; the queuing delay gradient is: ； in It is an exponentially weighted parameter. It is the time interval between the two most recent ACKs, calculated as follows: ； in It is the timestamp record of the arrival time of the k-th ACK. This represents the timestamp record of the (k-1)th ACK arrival time.

3. The method according to claim 2, characterized in that, In step one, the congestion information inference is obtained from the congestion criterion. The calculation formula is: ； in This is the expected queuing delay; according to The calculation formula, when and hour, 0; then when When extremely small, maintain The expected value of 0 corresponds to The value should be a very large positive value, if at this time The actual value leads to the calculated If the value is a large positive value, the congestion information is inferred to be non-congested; while when... When it is large, maintain The expected value of 0 corresponds to The value should be a very large negative value, if at this time The actual value leads to the calculated If the value is a large negative value, then the congestion information is inferred to be congestion.

4. The method according to claim 2, characterized in that, Step one further includes calculating an approximate congestion criterion using the following piecewise function. to replace To reduce computational overhead: ； in, It is calculated through displacement operations.

5. The method according to claim 3, characterized in that, In step two, the system state region is divided into the following six regions: 1) For the reliable growth sub-region, by the curve , and Define; 2) For the reliable deceleration sub-region, by curve , and divergence boundary Definition; 3) For the untrusted short queue deceleration sub-region, by the curve , and divergence boundary Definition; 4) For the untrusted long queue deceleration sub-region, via curve , and divergence boundary Define; 5) For untrusted short queue growth sub-regions, through curves , and Define; 6) For unreliable short queues with limited growth sub-regions, the curve is used. and 0 is defined.

6. The method according to claim 5, characterized in that, In step three, the expression for the analytical speed regulation rule is: ； in, This indicates that within the reliable growth sub-region, This indicates that within the trusted deceleration sub-region, This represents the rate adjustment amount, where M is the data packet size. For the desired delay, As a smoothing factor, Indicates the ACK interval. To adaptively update the coefficients, , It is a fairness factor, calculated as follows: ； in It is an independent reference rate for each stream, calculated as follows: ； in This indicates the line speed of the network interface card (NIC).

7. The method according to claim 6, characterized in that, In step four, the expression for the probe-based speed regulation rule is: ； in This indicates that within the untrusted short queue growth sub-region and the untrusted short queue restricted growth sub-region, This indicates that within the untrusted short queue deceleration sub-region and the untrusted long queue deceleration sub-region, , All are adaptive update coefficients; S is used to control the magnitude range of rate adjustment. When the system leaves region A, the value of S will be adjusted according to the most recently recorded... Update the data using the average of the five transmission rates closest to the current time; within the region Inside, And it needs to be multiplied by To ensure fairness when fully loaded; in the region Inside, ; It is an adaptive acceleration factor, and the update formula is: ,in This indicates the number of consecutive RTTs when the link is in a low utilization state.

8. An electronic device, comprising: One or more processors; Storage device for storing one or more programs; The feature is that the one or more processors execute the one or more programs to implement the method as described in any one of claims 1-7.

9. A computer-readable medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-7.

Images