A method and apparatus for data center network congestion control based on residence time

Abstract

The application relates to the technical field of computer network communication, and discloses a data center network congestion control method and device based on residence time. The method comprises the following steps: maintaining a ghost queue on the switch side, the ghost queue being a virtual byte counter and not actually buffering data packets; the ghost queue increasing the count according to the length of the data packets entering the physical queue of the switch and decreasing at a set emptying rate; obtaining the residence time of the data packets in the physical queue; when the occupancy rate of the ghost queue is higher than a high-order space threshold value and the smooth value of the residence time is higher than a high-order time threshold value, increasing a continuous confirmation counter; when the continuous confirmation counter reaches a preset number of times, determining that the current state is a sustained congestion state; and performing an explicit congestion notification marking operation in the sustained congestion state. The application avoids short flow misjudgment and control oscillation, and significantly improves the overall throughput and flow completion time performance of the data center network.

Description

Technical Field

[0001] This invention relates to the field of computer network communication technology, and specifically to a data center network congestion control method and apparatus based on dwell time. Background Technology

[0002] Modern data center networks typically handle two distinct types of traffic simultaneously. One type is intra-domain traffic, whose round-trip time (RTT) is usually in the microsecond range, and whose completion time is extremely sensitive to queuing latency. The other type is cross-data center or cross-domain traffic, whose RTT can reach the millisecond range, significantly increasing the bandwidth-delay product (BDP). When these two types of traffic coexist in the same network link and switch queue, there is a significant mismatch between congestion signal feedback latency, queue backlog evolution speed, and buffer absorption capacity.

[0003] Commercial data center switches typically employ a shallow buffer design with limited physical buffer capacity. For cross-domain long round-trip latency traffic, by the time congestion signals based on queue levels are fed back to the sender, the switch's physical buffer may already be severely overloaded or even overflowing. For intra-domain short flows, premature or excessively strong congestion marking can cause the sender to unnecessarily slow down, thereby lengthening the completion time of the short flow tail.

[0004] Existing solutions such as Data Center Transmission Control Protocol (DCTCP) and Data Center Quantitative Congestion Notification (DCQCN) primarily trigger Explicit Congestion Notification (ECN) flags based on physical queue levels or hard thresholds. While these solutions are simple to implement, the thresholds are typically fixed, making it difficult to simultaneously adapt to mixed traffic environments where round-trip latency differs by tens to hundreds of times or even more.

[0005] Uno and other solutions based on ghost queues or virtual queues introduce early warning capabilities. These solutions maintain virtual queues with an emptying rate (also known as drainage rate) lower than the physical line rate, allowing the virtual queues to reflect potential congestion trends before the physical queues. However, existing ghost queue solutions typically still rely primarily on the virtual queue level at a given moment to make congestion decisions, lacking independent verification of congestion persistence. Short-term bursts of traffic can cause ghost queues to spike instantaneously. Even if the physical queues can absorb and empty the congestion within a short time, these solutions may still trigger premature explicit congestion notification flags, resulting in false positives and decreased link utilization.

[0006] In existing solutions, ghost queue capacity, emptying rate, and explicit congestion notification threshold are mostly static parameters, or they only switch between a few discrete states. When the load, round-trip time, and traffic distribution change dynamically, the switching between discrete states is prone to oscillations around the threshold, and it is also difficult to continuously adjust the control strength according to the severity of congestion.

[0007] Therefore, the main technical problems of existing technologies include: lack of effective technical means to distinguish between short-term bursts and continuous congestion; difficulty in adapting fixed threshold or discrete state control to dynamic traffic environments; and the tendency to mislabel, control oscillations, and reduced link utilization in mixed round-trip delay scenarios. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a data center network congestion control method and apparatus based on dwell time.

[0009] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: In a first aspect, the present invention provides a data center network congestion control method based on dwell time, comprising: A ghost queue is maintained on the switch side. The ghost queue is a virtual byte counter and does not actually cache data packets. The ghost queue increments its count based on the length of the data packets entering the switch's physical queue and decrements at a set emptying rate. Get the dwell time of the data packet in the physical queue; When the ghost queue occupancy rate is higher than the space high-order threshold and the smoothed value of the dwell time is higher than the time high-order threshold, the continuous acknowledgment counter is increased; when the continuous acknowledgment counter reaches a preset number of times, the current state is determined to be a continuous congestion state. In a state of persistent congestion, an explicit congestion notification marking operation is performed.

[0010] In one embodiment, a hysteresis state switching step is also included: when the exit condition of the persistent congestion state is met, the persistent congestion state is switched to the normal state; the exit condition is: the ghost queue occupancy rate is lower than the spatial low threshold, or the smoothed value of the dwell time is lower than the time low threshold; wherein, the spatial low threshold is less than the spatial high threshold, and the time low threshold is less than the time high threshold.

[0011] In one embodiment, a PID adaptive control step is also included: Construct a composite error; the composite error includes: ghost queue occupancy rate deviation, physical queue occupancy rate deviation, and emptying rate gap; The composite error is input into the PID controller to obtain the control output; Adjust the effective capacity, emptying rate, and explicit congestion notification flag threshold of the ghost queue based on the control output.

[0012] In one embodiment, the composite error is expressed as: ; in, This is a smoothing value for ghost queue occupancy. Ghost queue target occupancy rate, This is due to the deviation in ghost queue occupancy rate; This is a smoothed value for physical queue occupancy. Physical queue target occupancy rate, This refers to the deviation in physical queue occupancy. This is the normalized venting rate gap smoothing value; and These are the weighting coefficients.

[0013] In one embodiment, adjusting the effective capacity, emptying rate, and explicit congestion notification tagging threshold of the ghost queue based on the control output specifically includes: The control output includes capacity control quantity. , Exhaust control quantity and threshold control quantity ; According to capacity control amount Update the ghost queue's effective capacity factor Effective capacity of ghost queues ,in, Basic capacity; According to the discharge control volume Update the ghost queue emptying rate factor This is used to adjust the emptying rate of the ghost queue; Based on threshold control quantity Update the explicit congestion notification labeling threshold or labeling probability range; The capacity control quantity, emptying control quantity, and threshold control quantity are respectively smoothed and limited. The capacity factor, emptying rate factor, and explicit congestion notification flag threshold are respectively set with minimum and maximum value constraints.

[0014] In one embodiment, the method further includes: maintaining a continuous congestion timer, which increments the continuous congestion timer when the ghost queue occupancy change rate is positive, the emptying rate gap is positive, and the smoothing value of the physical queue or ghost queue occupancy exceeds its respective set threshold. When the emptying rate gap disappears and both the physical queue and the ghost queue are at low occupancy, reduce the continuous congestion timer or reset it to zero. If neither the conditions for increasing the continuous congestion timer nor the conditions for decreasing or clearing the continuous congestion timer are met, the continuous congestion timer is decreased in a semi-decay manner. When the continuous congestion timer exceeds the set threshold, increase the PID controller gain or the scaling factor of the control output.

[0015] In one embodiment, performing the explicit congestion notification marking operation specifically includes: When the number of bytes occupied by the ghost queue is lower than the low marking threshold, no explicit congestion notification is performed; when the number of bytes occupied by the ghost queue is higher than the high marking threshold, explicit congestion notification is performed; when the number of bytes occupied by the ghost queue is between the low marking threshold and the high marking threshold, explicit congestion notification is performed with a probability that increases linearly with the number of bytes occupied by the ghost queue.

[0016] In a second aspect, the present invention provides a control apparatus for implementing the data center network congestion control method based on dwell time according to any embodiment of the first aspect, comprising: The ghost queue maintenance module is used to maintain the ghost queue on the switch side. The ghost queue is a virtual byte counter that does not actually cache data packets. The ghost queue increments its count based on the length of the data packets entering the switch's physical queue and decrements its count at a set emptying rate. The dwell time calculation module is used to obtain the dwell time of data packets in the physical queue; The continuous congestion gating module is used to increase the continuous acknowledgment counter when the ghost queue occupancy rate is higher than the spatial high-order threshold and the smoothed value of the dwell time is higher than the time high-order threshold; when the continuous acknowledgment counter reaches a preset number of times, the current state is determined to be a continuous congestion state. The explicit congestion notification marking execution module is used to perform explicit congestion notification marking operations in a persistent congestion state.

[0017] In one embodiment, a hysteresis state switching module is further included, which is used to switch the continuous congestion state to a normal state when the exit condition of the continuous congestion state is met; the exit condition is: the ghost queue occupancy rate is lower than the spatial low threshold, or the smoothed value of the dwell time is lower than the time low threshold; wherein, the spatial low threshold is less than the spatial high threshold, and the time low threshold is less than the time high threshold.

[0018] In one embodiment, a PID adaptive control module is also included, for: Construct a composite error; the composite error includes: ghost queue occupancy rate deviation, physical queue occupancy rate deviation, and emptying rate gap; The composite error is input into the PID controller to obtain the control output; Adjust the effective capacity, emptying rate, and explicit congestion notification flag threshold of the ghost queue based on the control output.

[0019] Compared with the prior art, the beneficial technical effects of the present invention are: This invention distinguishes between short-term bursts and persistent congestion. It combines the spatial signal of ghost queue occupancy with the temporal signal of packet dwell time. Persistent congestion is only confirmed when both signals are consistently high, thereby reducing misjudgments of short-term bursts that can be quickly absorbed by the physical cache.

[0020] Reduce control oscillations. This invention uses a hysteretic state machine with a high threshold for entry and a low threshold for exit, along with a continuous confirmation mechanism, to avoid oscillations caused by repeated triggering of a single threshold near the critical state.

[0021] This invention achieves parameter self-adaptation. It employs a composite error and PID continuous controller to adaptively adjust the ghost queue capacity, emptying rate, and ECN marking threshold, thereby improving adaptability to different RTT, load, and traffic distributions.

[0022] Improve congestion assessment accuracy. This invention utilizes predictive virtual queue backlog, actual physical queue occupancy, and emptying rate gaps to enable the controller to assess congestion severity from multiple dimensions.

[0023] It is feasible to deploy the switch locally. The signals required by this invention mainly come from the local queue status of the switch and the inbound and outbound timestamps of data packets, without relying on additional telemetry between switches; Exponentially weighted moving average (EWMA) smoothing, counter and PID updates can all be implemented in O(1) state.

[0024] The experimental results are quantifiable. In a packet-level simulation experiment, under 40% network load and shallow buffer conditions, the P99 FCT (99th percentile flow completion time) of short flow within the domain, the P99 FCT of the entire flow, and the P99 FCT of long flow across the domain were all reduced compared to the baseline scheme. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall architecture of the present invention.

[0026] Figure 2 This is a schematic diagram of the continuous congestion gating state machine of the present invention.

[0027] Figure 3 This is a schematic diagram of the closed-loop PID continuous control of the present invention.

[0028] Figure 4 This is a schematic diagram of the ECN probability labeling function of the present invention.

[0029] Figure 5 This is a flowchart of the method of the present invention. Detailed Implementation

[0030] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.

[0031] like Figure 5 As shown, a data center network congestion control method based on dwell time in this invention includes the following steps: S1, maintain a ghost queue on the switch side. The ghost queue is a virtual byte counter that does not actually cache data packets. The ghost queue increments its count based on the length of the data packets entering the switch's physical queue and decrements at a set emptying rate. S2, Get the dwell time of the data packet in the physical queue; S3, when the ghost queue occupancy rate is higher than the space high-order threshold and the smoothed value of the dwell time is higher than the time high-order threshold, the continuous acknowledgment counter is increased; when the continuous acknowledgment counter reaches a preset number of times, the current state is determined to be a continuous congestion state. S4, in the state of continuous congestion, performs an explicit congestion notification marking operation.

[0032] This invention maintains a phantom queue on the switch side, using the phantom queue occupancy rate as an early warning signal in the spatial dimension; simultaneously, it records the packet enqueue time and dequeue time, calculates the packet dwell time in the physical queue, and uses the dwell time as a congestion verification signal in the time dimension. In this invention, the phantom queue refers to a virtual byte counter maintained on the switch side. It does not actually buffer packets, but increases according to the number of bytes enqueued and decreases at a set emptying rate, used to anticipate potential congestion trends.

[0033] Based on this, the present invention constructs a two-stage continuous sensing gating mechanism of "early warning-confirmation": when the ghost queue occupancy rate increases, only a congestion warning is triggered; only when the ghost queue occupancy rate and the smoothed value of the dwell time are both at a high level, and this condition is met continuously for a preset number of times, is it determined to be a continuous congestion state. Thus, the present invention can distinguish between the instantaneous virtual queue amplification effect and the real continuous queuing.

[0034] Furthermore, this invention introduces a hysteresis state switching mechanism. Entering the persistent congestion state uses a higher threshold and continuous acknowledgment requirements, while exiting the persistent congestion state uses a lower threshold and release acknowledgment requirements, thereby avoiding frequent round-trip switching near the critical threshold.

[0035] Furthermore, this invention constructs a PID controller based on continuous congestion judgment. The PID controller generates a composite error based on ghost queue occupancy rate deviation, physical queue occupancy rate deviation, and emptying rate gap, and outputs a continuous control quantity to adaptively adjust the effective capacity of the ghost queue, the ghost queue emptying rate, and the explicit congestion notification (ECN) flag threshold. Compared to discrete state transitions, this invention can continuously adjust the control strength according to the degree of congestion. The emptying rate gap g refers to the value obtained by normalizing the normalized difference between the packet inbound rate and the actual outbound rate (or the set emptying rate) within the current sampling window. An emptying rate gap greater than 0 indicates that the queue backlog continues to worsen, while an emptying rate gap less than or equal to 0 indicates that the queue backlog is stabilizing or is receding.

[0036] The switch-side component of this invention may include: a basic signal acquisition module, a ghost queue maintenance module, a dwell time calculation module, a signal smoothing module, a continuous congestion gating module, a hysteresis state machine module, a PID adaptive control module, and an ECN tag execution module. The transmitting end can employ existing ECN response mechanisms to adjust the transmission rate based on the received ECN tags; the specific rate adjustment algorithm at the transmitting end is not a necessary limitation of this invention.

[0037] Figure 1 This is a schematic diagram of the overall system architecture of the present invention, illustrating the relationship between basic signal acquisition on the switch side, ghost queue maintenance, dwell time calculation, continuous congestion gating, PID adaptive control, and ECN marking execution. The present invention will be described in detail below in several parts.

[0038] 1. Ghost queue maintenance and basic signal acquisition.

[0039] On the output port side of the switch, a ghost queue is maintained. This ghost queue can be implemented as a virtual byte counter, without actually buffering data packets. When a data packet enters the physical queue, the ghost queue counter increases by the number of bytes corresponding to the length of the data packet; the ghost queue decreases over time according to a set emptying rate.

[0040] In one embodiment, the ghost queue emptying rate is lower than the physical link line rate. For example, the base emptying rate can be set to 80% to 98% of the physical link line rate, preferably about 90%. By using an emptying mechanism that is lower than the physical line rate, the ghost queue can reflect potential congestion trends earlier than the physical queue.

[0041] Record or read the timestamp of each data packet entering the physical queue. and the timestamp of leaving the physical queue And calculate the dwell time of the data packet. : .

[0042] Dwell time reflects the actual duration a data packet stays in the switch's buffer. When short bursts are quickly absorbed and emptied by the physical buffer, the ghost queue may spike momentarily, but the packet dwell time remains low. When real, sustained congestion occurs, data packets need to wait continuously, and the dwell time increases accordingly.

[0043] 2. Signal smoothing and derived quantity calculation.

[0044] To reduce the jitter caused by a single data packet or a single sample, the switch uses an exponentially weighted moving average (EWMA) to smooth signals such as ghost queue occupancy, physical queue occupancy, dwell time, enqueue rate, dequeue rate, and emptying rate gap.

[0045] In one embodiment, for any instantaneous quantity smoothing value You can update it as follows: ; in, This is a smoothing factor, which can be set according to the switch update cycle and the target response speed. For example, The value can be between 0.01 and 0.2, with a preferred value of approximately 0.05. Exponentially weighted moving average (EWMA) smoothing only requires storing one historical state for each variable, with an O(1) storage complexity, making it suitable for local implementation on a switch.

[0046] In one embodiment, the ghost queue occupancy rate p is defined as the number of bytes currently occupied by the ghost queue. With the current effective ghost queue capacity The ratio; define the physical queue occupancy rate r as the number of bytes currently occupied by the physical queue. With physical queue capacity The ratio. The enqueue rate can also be calculated based on the number of enqueued bytes and the number of dequeued bytes within the sampling window. Departure rate And the normalized venting rate gap.

[0047] 3. Continuous congestion gating mechanism based on dwell time perception.

[0048] The persistent congestion gating mechanism includes an early warning phase and a confirmation phase.

[0049] Early warning phase: Real-time monitoring of ghost queue occupancy rate p or its smoothed value. .when Above the spatial high threshold At that time, it was believed that there was a potential congestion trend in the virtual space dimension, triggering the early warning condition.

[0050] Diagnostic stage: Further determination of the smoothing value of the length of stay Is it higher than the high-order threshold of time? Only if the spatial dimension condition Conditions related to the time dimension Congestion is considered persistent only when both conditions are met simultaneously.

[0051] To improve robustness, a continuous acknowledgment counter (cnt) is set. When the above two-dimensional conditions are met for N consecutive sampling periods, the current state is set to a persistent congestion state; if either condition is not met, the acknowledgment counter is reset or reduced. N can be set according to the scenario, for example, it can be 2 to 8, preferably 3 to 5.

[0052] The technical effect of this mechanism is that a high ghost queue occupancy rate only indicates a backlog trend in the virtual space dimension, but this backlog may originate from short-term bursts; a continuously increasing dwell time indicates that the actual waiting time of data packets in the physical queue is increasing. The joint judgment of these two factors can achieve cross-verification of spatial and temporal signals, thereby reducing the false labeling of short-term bursts.

[0053] Figure 2 This can be a schematic diagram of a continuous congestion gating state machine, showing the normal state, warning conditions, continuous confirmation, continuous congestion state, and hysteresis exit conditions.

[0054] 4. Hysteresis state switching mechanism.

[0055] A hysteresis mechanism is used for state switching between the normal state and the persistent congestion state. Entering the persistent congestion state uses a combination of high-order thresholds, such as... and It requires N consecutive confirmations; exiting the persistent congestion state uses a combination of low-order thresholds, for example... or It requires M consecutive release confirmations.

[0056] in, Less than , Less than M can be the same as N or less than N, so that normal control can be restored more quickly when congestion is relieved. By using an asymmetric design of entry and exit thresholds, a buffer is formed at the state boundary to avoid repeated switching due to small fluctuations in queue level or dwell time.

[0057] In normal conditions, sensitive early warnings can be maintained to ensure compatibility with existing ghost queue schemes and protect short-flow completion time; in persistent congestion conditions, stronger early ECN marking and adaptive control based on ghost queues are enabled to suppress real persistent congestion.

[0058] 5. PID continuous controller with compound error.

[0059] To avoid insufficient adaptability caused by fixed parameters and discrete state switching, this invention further employs a PID continuous controller to dynamically adjust the relevant parameters of the ghost queue.

[0060] In one embodiment, the input to the PID controller is the composite error e of the three signals: ; in, This is a smoothing value for ghost queue occupancy. Ghost queue target occupancy rate; This is a smoothed value for physical queue occupancy. Physical queue target occupancy rate; This is the normalized venting rate gap smoothing value; and These are the weighting coefficients for the deviation in physical queue occupancy and the gap in emptying rate, respectively.

[0061] The aforementioned composite error means that: ghost queue occupancy deviation reflects predictive virtual backlog, physical queue occupancy deviation reflects actual cache pressure, and emptying rate gap reflects whether the enqueue rate consistently exceeds the dequeue or emptying capacity. The combination of these three factors can more accurately characterize the severity of congestion.

[0062] The controller maintains proportional, integral, and derivative terms. The integral term can be set with upper and lower limits to avoid integral saturation, and the derivative term reflects the trend of congestion changes. In one embodiment, the control output u can be expressed as: ; The ghost queue capacity factor, ghost queue emptying rate factor, and ECN marker threshold offset can be adjusted based on this continuous control output or based on multiple control outputs, respectively. These represent the proportional gain, integral gain, and derivative gain of the PID controller, respectively. This represents the integral accumulation term of the composite error e over time.

[0063] Figure 3 This can be a schematic diagram of a closed-loop PID continuous control, showing the feedback relationship between the three-signal composite error, controller output and capacity factor, evacuation rate factor, and ECN threshold.

[0064] 6. Three adaptive control outputs.

[0065] First, capacity factor adjustment. Based on the capacity control amount... Update the ghost queue's effective capacity factor The effective ghost queue capacity can be expressed as: ,in Based on the basic capacity. When the number of entries increases, the ghost queue occupancy rate decreases for the same number of entries, and the ECN triggering time is delayed; When the threshold is reduced, the ghost queue is more likely to reach the marking threshold, and ECN is triggered earlier.

[0066] Second, venting rate factor adjustment. Based on the venting control quantity... Update ghost queue emptying rate factor . Increasing the value indicates that the ghost queue is emptied faster. A decrease indicates that the ghost queue emptying becomes slower. In one embodiment, A feedforward term related to the evacuation rate gap can also be added to improve the control response speed when the inbound rate is consistently faster than the outbound rate.

[0067] Third, ECN labeling threshold adjustment. Based on the threshold control amount... Update the ECN labeling threshold or labeling probability range. When the composite error indicates increased congestion, lower the labeling threshold to trigger the ECN earlier; when the composite error indicates that congestion has eased, raise or restore the labeling threshold to reduce over-labeling.

[0068] The three control variables mentioned above can be smoothed and limited using exponentially weighted moving average (EWMA) to prevent instability caused by excessively large single control outputs. Minimum and maximum values ​​can be set for the capacity factor, venting rate factor, and ECN threshold to ensure that the control parameters remain within deployable ranges.

[0069] 7. Continuous congestion timer and gain scheduling.

[0070] In one embodiment, a continuous congestion timer Tc is also maintained. When the queue occupancy change rate is positive, the emptying rate gap is positive, and the physical queue or ghost queue is at a high level, Tc is increased; when the emptying rate gap disappears and the queue occupancy is at a low level, Tc is reset to zero or decreased; in the transition state, Tc can be decreased in a half-decay manner.

[0071] When Tc exceeds the set threshold, the PID controller gain or control output scaling factor is increased to enable the controller to intervene more effectively during real and continuous congestion; when Tc does not exceed the set threshold, the baseline gain is maintained to avoid overreacting to transient fluctuations.

[0072] 8. Explicit congestion notification probability marking is implemented.

[0073] In a state of persistent congestion, explicit congestion notification probability marking is performed based on an adaptive threshold. Let the low marking threshold be... The high-labeling threshold is ,in Greater than .

[0074] When the ghost queue occupies fewer bytes (Qp) than When the Qp is higher than 1, no ECN is marked on the data packet; when the Qp is higher than 1, no ECN is marked on the data packet. When, the data packet is marked with ECN; when Qp is located and When the time interval is between, the values ​​are labeled according to a probability that increases linearly with Qp.

[0075] The probabilistic marking method described above allows the marking behavior to change smoothly with the level of congestion, avoiding the abrupt changes caused by hard threshold marking. After receiving the ECN mark, the sending end can adjust the transmission rate using existing end-side congestion control mechanisms.

[0076] Figure 4 This is a schematic diagram of the ECN probability marking function, showing the process where ghost queues occupying less than the low threshold theta_min are not marked, those occupying more than the high threshold theta_max are all marked, and those occupying the probability marking area between the two thresholds are marked with a linear increase in probability.

[0077] The present invention will be further described below with reference to embodiments. It should be understood that the following embodiments are used to explain the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Any equivalent substitutions or adjustments made to the threshold, control parameters, smoothing coefficients, update cycles, or state machine conditions under the concept of the present invention may fall within the scope of implementation of the present invention.

[0078] Example 1: Adaptive Ghost Queue Control in Hybrid Round-Trip (RTT) Data Center Networks: Step 1: The switch maintains physical queues and ghost queues on the outgoing port side. When a data packet enters the physical queue, the ghost queue counter increments by the corresponding number of bytes; the ghost queue decreases over time according to the emptying rate determined by the emptying rate factor.

[0079] Step two: The switch records the enqueue timestamp and dequeue timestamp for each data packet and calculates the dwell time. Within each control cycle, it reads the number of bytes occupied by the physical queue, the number of bytes occupied by the ghost queue, the number of enqueue bytes, and the number of dequeue bytes, and calculates the occupancy rate, speed, and emptying rate gap.

[0080] Step 3: The switch performs EWMA smoothing on the above signals to obtain smoothed values ​​for ghost queue occupancy, physical queue occupancy, dwell time, and emptying rate gap.

[0081] Step four: The switch performs a continuous congestion gating judgment. When the smoothed value of the ghost queue occupancy rate exceeds the spatial high-order threshold and the smoothed value of the residence time exceeds the time high-order threshold, the acknowledgment counter is incremented; when the above conditions are met N times consecutively, the switch enters a continuous congestion state.

[0082] Step 5: The switch runs the PID controller. The PID controller constructs a composite error based on the ghost queue occupancy rate deviation, physical queue occupancy rate deviation, and emptying rate gap, outputs the capacity factor, emptying rate factor, and ECN threshold offset, and smooths and limits the three control parameters.

[0083] Step 6: In the continuous congestion state, the switch marks the packets according to the adaptive ECN threshold and probability marking function; in the normal state, it can maintain the basic ghost queue marking logic or the existing ECN marking logic.

[0084] Step 7: The sender performs rate adjustment based on the received ECN flags. The sender can use DCTCP, DCQCN, BBR algorithms combined with ECN or other congestion response mechanisms that support ECN.

[0085] Example 2: Quick confirmation configuration in a low RTT scenario: In data center intranets with shorter RTT and faster control feedback, the number of consecutive acknowledgments N required to enter a persistent congestion state can be set to a smaller value, such as 2 or 3, while the number of exit acknowledgments M can be set to 1 or 2. This configuration can improve the speed of persistent congestion identification while maintaining short burst filtering capabilities.

[0086] Example 3: Early warning configuration in a shallow-buffered switch scenario: In switching devices with shallower physical buffers, the spatial high-order threshold or the temporal high-order threshold can be appropriately reduced to allow for earlier initiation of the continuous congestion acknowledgment process. Simultaneously, the lower limit of the ECN tagging threshold can be set more conservatively to prevent the physical queue from being rapidly filled under long RTT feedback delays.

[0087] Example 4, Configuration: In an experimental verification, evaluation can be conducted using a packet-level network simulator such as HTSIM. The topology can employ two 8-port fat-tree interconnected data centers, with a single-link bandwidth of 100Gbps, an intra-domain RTT of 14μs, and an inter-domain RTT of 2ms. Intra-domain traffic can use a short-flow distribution similar to Google Web Search, while cross-domain traffic can use a long-flow distribution similar to Alibaba WAN, with a load ratio of 4:1.

[0088] Under 40% network load and shallow buffer conditions, metrics such as intra-domain short-flow P99 FCT, full-flow P99 FCT, and cross-domain long-flow P99 FCT can be recorded and compared with baseline solutions such as MPRDMA+BBR, Gemini, and Uno. The above experimental results are used to demonstrate the technical effectiveness of this invention in distinguishing between short-term bursts and continuous congestion, mitigating fixed threshold mislabeling, and improving control stability in hybrid RTT scenarios.

[0089] This invention introduces packet dwell time as a time-dimensional verification signal based on the ghost queue spatial signal. Through continuous acknowledgment, hysteresis switching, and adaptive control, the switch can distinguish between short-term bursts and real continuous congestion, and dynamically adjust ECN marking behavior accordingly. The invention also uses a joint judgment of ghost queue occupancy rate and smoothed packet dwell time value to avoid misjudging short-term bursts as continuous congestion based solely on the instantaneous level of the ghost queue.

[0090] This invention reduces repeated state switching near critical thresholds and lowers control oscillations by using continuous acknowledgment and a hysteresis state machine; it improves the accuracy of congestion severity assessment by integrating ghost occupancy rate, physical occupancy rate, and emptying rate gap, and enables continuous adjustment of control strength; it dynamically adjusts ghost queue capacity factor, emptying rate factor, and ECN tag threshold to adapt to different RTT, load, and traffic distributions, avoiding mismatch of fixed parameters; and it uses timestamps, EWMA, counters, and local queue information, without relying on additional telemetry between switches, making it easy to deploy in shallow-buffered switch scenarios.

[0091] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0092] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple steps or stages, which are not necessarily completed at the same time, but may be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0093] In one embodiment, the present invention also provides an apparatus. The apparatus may be a system (including a distributed system), software (application), module, component, server, client, etc., that uses the methods described in the embodiments of this specification, combined with necessary hardware implementation. Since the implementation schemes and methods for solving the problem by the apparatus are similar, the implementation of the specific apparatus in the embodiments of this specification can be referred to the implementation of the foregoing methods, and repeated details will not be described again. Although the apparatus described in the embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0094] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0095] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0096] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A data center network congestion control method based on residence time, characterized in that, include: A ghost queue is maintained on the switch side. The ghost queue is a virtual byte counter and does not actually cache data packets. The ghost queue increments its count based on the length of the data packets entering the switch's physical queue and decrements at a set emptying rate. Get the dwell time of the data packet in the physical queue; When the ghost queue occupancy rate is higher than the space high-order threshold and the smoothed value of the dwell time is higher than the time high-order threshold, the continuous acknowledgment counter is increased; when the continuous acknowledgment counter reaches a preset number of times, the current state is determined to be a continuous congestion state. In a state of persistent congestion, an explicit congestion notification marking operation is performed.

2. The data center network congestion control method based on dwell time according to claim 1, characterized in that, It also includes a hysteresis state switching step: when the exit condition of the continuous congestion state is met, the continuous congestion state is switched to the normal state; the exit condition is: the ghost queue occupancy rate is lower than the spatial low threshold, or the smoothed value of the dwell time is lower than the time low threshold; wherein, the spatial low threshold is less than the spatial high threshold, and the time low threshold is less than the time high threshold.

3. The data center network congestion control method based on dwell time according to claim 1, characterized in that, It also includes PID adaptive control steps: Construct a composite error; the composite error includes: ghost queue occupancy rate deviation, physical queue occupancy rate deviation, and emptying rate gap; The composite error is input into the PID controller to obtain the control output; Adjust the effective capacity, emptying rate, and explicit congestion notification flag threshold of the ghost queue based on the control output.

4. The data center network congestion control method based on dwell time according to claim 3, characterized in that, The composite error is expressed as: ； in, This is a smoothing value for ghost queue occupancy. Ghost queue target occupancy rate, This is due to the deviation in ghost queue occupancy rate; This is a smoothed value for physical queue occupancy. Physical queue target occupancy rate, This refers to the deviation in physical queue occupancy. This is the normalized venting rate gap smoothing value; and These are the weighting coefficients.

5. The data center network congestion control method based on dwell time according to claim 3, characterized in that, The adjustment of the effective capacity, emptying rate, and explicit congestion notification tagging threshold of the ghost queue based on the control output specifically includes: The control output includes capacity control quantity. , Exhaust control quantity and threshold control quantity ; According to capacity control amount Update the ghost queue's effective capacity factor Effective capacity of ghost queues ,in, Basic capacity; According to the discharge control volume Update the ghost queue emptying rate factor This is used to adjust the emptying rate of the ghost queue; Based on threshold control quantity Update the explicit congestion notification labeling threshold or labeling probability range; The capacity control quantity, emptying control quantity, and threshold control quantity are respectively smoothed and limited. The capacity factor, emptying rate factor, and explicit congestion notification flag threshold are respectively set with minimum and maximum value constraints.

6. The data center network congestion control method based on dwell time according to claim 3, characterized in that, Also includes: Maintain a persistent congestion timer. When the ghost queue occupancy change rate is positive, the emptying rate gap is positive, and the smoothing value of the physical queue or ghost queue occupancy exceeds their respective set thresholds, increase the persistent congestion timer. When the emptying rate gap disappears and both the physical queue and the ghost queue are at low occupancy, reduce the continuous congestion timer or reset it to zero. If neither the conditions for increasing the continuous congestion timer nor the conditions for decreasing or clearing the continuous congestion timer are met, the continuous congestion timer is decreased in a semi-decay manner. When the continuous congestion timer exceeds the set threshold, increase the PID controller gain or the scaling factor of the control output.

7. The data center network congestion control method based on dwell time according to claim 1, characterized in that, The execution of the explicit congestion notification marking operation specifically includes: When the number of bytes occupied by the ghost queue is lower than the low marking threshold, no explicit congestion notification is performed; when the number of bytes occupied by the ghost queue is higher than the high marking threshold, explicit congestion notification is performed; when the number of bytes occupied by the ghost queue is between the low marking threshold and the high marking threshold, explicit congestion notification is performed with a probability that increases linearly with the number of bytes occupied by the ghost queue.

8. A control apparatus for implementing the data center network congestion control method based on dwell time according to any one of claims 1 to 7, characterized in that, include: The ghost queue maintenance module is used to maintain the ghost queue on the switch side. The ghost queue is a virtual byte counter that does not actually cache data packets. The ghost queue increments its count based on the length of the data packets entering the switch's physical queue and decrements its count at a set emptying rate. The dwell time calculation module is used to obtain the dwell time of data packets in the physical queue; The continuous congestion gating module is used to increase the continuous acknowledgment counter when the ghost queue occupancy rate is higher than the spatial high-order threshold and the smoothed value of the dwell time is higher than the time high-order threshold; when the continuous acknowledgment counter reaches a preset number of times, the current state is determined to be a continuous congestion state. The explicit congestion notification marking execution module is used to perform explicit congestion notification marking operations in a persistent congestion state.

9. The control device according to claim 8, characterized in that, It also includes a hysteresis state switching module, used to switch the continuous congestion state to a normal state when the exit conditions of the continuous congestion state are met; the exit conditions are: the ghost queue occupancy rate is lower than the spatial low threshold, or the smoothed value of the dwell time is lower than the time low threshold; wherein, the spatial low threshold is less than the spatial high threshold, and the time low threshold is less than the time high threshold.

10. The control device according to claim 8, characterized in that, It also includes a PID adaptive control module, used for: Construct a composite error; the composite error includes: ghost queue occupancy rate deviation, physical queue occupancy rate deviation, and emptying rate gap; The composite error is input into the PID controller to obtain the control output; Adjust the effective capacity, emptying rate, and explicit congestion notification flag threshold of the ghost queue based on the control output.

Images