Data center mixed flow exchange buffer management method oriented to high efficiency

Abstract

A hybrid traffic exchange buffer management method oriented to a high-efficiency data center adopts a modular design, and comprises a switch monitoring module used for periodically recording indexes such as the service rate, the arrival rate and the queue length of a data packet; the calculation module calculates a loss rate based on a queuing model, constructs a utility function in combination with the throughput and the delay gradient, and quantifies the performance value allocated to the current buffer area; and the control module iteratively optimizes the admission control parameter through a micro experiment, alternately explores the influence of the increase and decrease of the admission parameter on the utility value, selects a parameter updating queue threshold value of the maximum utility value, and allows the two types of flow to freely seize resources if the occupancy rate of the buffer area does not reach a set threshold value, and continuously and cooperatively updates the admission parameter through the control module. According to the method, the mixed flow performance can be effectively optimized, the fair bandwidth sharing of the two types of flows can still be ensured even in a buffer scrambling scene with a higher lossy flow ratio and a lower lossless flow ratio, and the balance of high throughput and low delay is realized.

Description

Technical Field

[0001] This invention belongs to the field of data center networks, specifically relating to a method for managing hybrid traffic exchange buffers in data centers for high-efficiency applications. Background Technology

[0002] In early data center networks, traditional applications relied on packet loss-based transmission mechanisms using TCP traffic and its related extensions. However, with the exponential growth of data center data volumes, the demand for high bandwidth and low latency has become increasingly urgent. This need is alleviated by RoCE (RDMA over converged Ethernet), a technology based on lossless Ethernet aggregation, which employs a switch-configured PFC mechanism. When the ingress port detects that the buffer occupancy rate exceeds the PFC threshold, it stops packet transmission on the upstream port by using a pause frame. Mixed traffic is very common in data center networks. In practice, traditional web-based applications utilize high-speed storage backends, mixing end-user TCP requests with RDMA storage requests to achieve data read and write operations. RDMA device management and software-defined control planes typically rely on the TCP protocol, while RoCE is used for data communication. AI / ML applications use RoCE to interconnect GPUs and CPUs, but may still employ TCP-based storage solutions.

[0003] In mixed traffic scenarios (lossy traffic using TCP and lossless traffic using RDMA), lossy traffic is far more greedy for buffer resources than lossless traffic. These issues can severely impact the quality of service (QoS) of high-performance computing and storage networks, potentially leading to server crashes and user data loss, causing incalculable damage. Specifically, this manifests in two main challenges: First, lossy traffic consumes significantly more buffer resources. Due to the significant differences between the two traffic types, the admission control threshold parameter α of the ingress buffer pool is typically set to infinity to accommodate lossy traffic. Conversely, the configuration of lossless traffic requires limiting the number of packets entering the ingress buffer pool.

[0004] The round-trip time for lossy traffic is almost 15 times that of lossless traffic, causing the lossy traffic queue to over-consume buffer resources and frequently trigger traffic priority control. This not only degrades the performance of lossless traffic but also leads to queue head congestion, routing deadlock, and PFC pause frame storms. Secondly, lossless traffic buffering is more difficult. When a switch receives burst traffic exceeding a threshold, lossy traffic will prioritize discarding non-burst packets because the burst traffic has higher priority. However, a key insight is that when a switch triggers a PFC pause frame due to a burst of packets reaching the threshold, even if the burst traffic has higher priority, the upstream switch will immediately stop all traffic transmission upon receiving the pause frame. Last but not least, the buffer allocation method is not performance-oriented.

[0005] To simplify management, buffer allocations may exceed actual memory requirements. Allocating a larger buffer for a traffic class may slightly improve the performance of that class at the expense of other traffic classes. This behavior can significantly impact overall traffic performance, potentially leading to longer tail latency for some traffic types and undermining fairness when multiple traffic classes coexist. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a data center hybrid traffic switching buffer management method for high efficiency. This method fully utilizes limited hardware resources, dynamically allocating switch buffer resources for both lossy and lossless traffic with the goal of high efficiency. This reduces the preemptive nature of lossy traffic on buffer space, provides predictable traffic performance, dynamism, isolation, and the convenience of rapid iteration after deployment, better meets the diverse and complex data center network traffic patterns, and ensures that lossy and lossless traffic can fairly share link bandwidth.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solution: A method for managing hybrid traffic exchange buffers in high-efficiency data centers includes the following steps: Step 1: The monitoring module sets the monitoring period. Whenever the monitoring interval is reached, the switch records the key indicators of the queue and passes them to the next module. Step 2: After receiving the aggregated metrics, the calculation module determines the specific values ​​of each reward and penalty item based on the established utility function and calculates the utility value. Step 3: The control module consists of two cycles, each consisting of two sets of monitoring cycles and a calculation control cycle; the impact of increasing or decreasing the admission control parameters on the utility value is explored alternately, and the parameter that maximizes the utility value is selected for updating. Step 4: The control module receives the updated cycle value from the calculation module and sends the modified cycle value back to the monitoring module to adjust the monitoring cycle for the new stage.

[0008] This invention discloses a high-performance data center hybrid traffic exchange buffer management method. It combines an online learning-like approach to continuously optimize admission control parameters, aiming to solve the problem that when lossy and lossless traffic share a switch buffer, lossy traffic excessively preempts resources due to longer round-trip latency, leading to frequent pause frames, increased tail latency, and unbalanced buffer resource allocation for lossless traffic. The method employs a modular design, including a switch monitoring module for periodically recording metrics such as packet service rate, arrival rate, and queue length; a calculation module that calculates the loss rate based on a queuing model and constructs a utility function based on throughput and latency gradients to quantify the performance value of the current buffer allocation; and a control module that iteratively optimizes admission control parameters through micro-experiments, alternately exploring the impact of increasing or decreasing admission parameters on the utility value, selecting the parameter that maximizes the utility value to update the queue threshold. If the buffer occupancy rate does not reach the set threshold, both types of traffic are allowed to freely preempt resources, while the control module continuously updates the admission parameters collaboratively. This invention effectively optimizes hybrid traffic performance, ensuring fair bandwidth sharing between the two types of traffic even in buffer contention scenarios where lossy traffic accounts for a higher proportion and lossless traffic accounts for a lower proportion, achieving a balance between high throughput and low latency.

[0009] The beneficial effects of this invention: This invention proposes a high-efficiency data center hybrid traffic exchange buffer management method, aiming to solve the problem that lossy traffic in a data center network will greedily preempt the shared buffer resources of lossless traffic. Specifically, it has the following advantages: 1. This invention implements a real-time monitoring method for mixed traffic bandwidth utilization. In a data center switch, a monitoring module periodically collects core indicators such as service rate and arrival rate of TCP lossy traffic and RDMA lossless traffic. Combined with EWMA smoothing and utility function quantification evaluation, it can accurately monitor the bandwidth utilization of both types of traffic in real time, enabling timely detection of unfair behavior such as lossy traffic excessively preempting resources. This method provides an efficient traffic monitoring and management mechanism, significantly improving the buffer utilization efficiency and network transmission stability of mixed-traffic switches. 2. The hybrid traffic utility function proposed in this invention, combining shared buffer allocation and packet loss rate, possesses significant advantages in multi-dimensional performance optimization and precise control. This function deeply couples the core resource configuration parameter of shared buffer allocation with packet loss rate. On the one hand, it overcomes the limitations of traditional functions that only focus on resource allocation or traffic loss. On the other hand, the differentiated definition of packet loss rate precisely matches the transmission characteristics of the two types of traffic, enabling the utility function to specifically capture the core performance bottlenecks of different traffic types, thus providing a precise basis for subsequent parameter optimization. It effectively provides a unified standard for addressing the key problem of the difficulty in coordinating the optimization of resource allocation and performance loss in hybrid traffic scenarios.

[0010] 3. This invention proposes a dynamic occupancy threshold mechanism for the switch's shared buffer. By iteratively optimizing the admission control parameter α using a method similar to online learning, the threshold criteria for lossy / lossless queues are dynamically adjusted. This ensures that when the shared buffer occupancy is low, lossy traffic is allowed to dynamically seize buffer resources based on actual load demand, without being restricted by a fixed ratio. When the occupancy increases and unfair allocation occurs, resource balancing is achieved through threshold constraints and rate reduction. This solution manages buffer resource allocation between the two types of traffic more flexibly and efficiently, avoiding resource waste under low load and preventing allocation imbalance under high load, significantly improving network resource utilization and overall network performance. Attached Figure Description Figure 1 A flowchart of a method for managing hybrid traffic exchange buffers in a data center for high-efficiency applications, according to some embodiments of the present invention, is shown.

[0012] Figure 2 A switch internal architecture diagram is shown, which is applicable to a data center hybrid traffic switching buffer management method for high-efficiency applications according to some embodiments of the present invention.

[0013] Figure 3 A full-process architecture diagram of a hybrid traffic exchange buffer management method for high-efficiency data centers, based on some embodiments of the present invention, is shown. Figure 4 The diagram shows the 99th percentile flow completion time for TCP and RDMA traffic. Detailed Implementation

[0014] The embodiments of the present invention will be disclosed below with reference to the drawings. For clarity, many practical details will be described in the following description. However, it should be understood that these practical details are not intended to limit the invention. That is, in some embodiments of the invention, these practical details are not essential.

[0015] like Figure 1 As shown, a method for managing hybrid traffic exchange buffers in data centers for high-efficiency applications includes the following steps: Step 1: The monitoring module sets the monitoring period. Whenever the monitoring interval is reached, the switch records the key indicators of the queue and passes them to the next module. Step 2: After receiving the aggregated metrics, the calculation module determines the specific values ​​of each reward and penalty item based on the established utility function and calculates the utility value. Step 3: The control module consists of two cycles, each consisting of two sets of monitoring cycles and a calculation control cycle; the impact of increasing or decreasing the admission control parameters on the utility value is explored alternately, and the parameter that maximizes the utility value is selected for updating. Step 4: The control module receives the updated cycle value from the calculation module and sends the modified cycle value back to the monitoring module to adjust the monitoring cycle for the new stage.

[0016] like Figure 2 As shown, an internal architecture of a switch for managing hybrid traffic switching buffers in high-efficiency data centers specifically includes: The monitoring module is implemented based on an FPGA hardware architecture. It performs high-frequency, low-latency data acquisition and temporary storage operations through hardware logic circuits. Specifically, during initialization, a PCIe 4.0 data interface is configured to establish hardware-level connections with the eight ports of the port module. Simultaneously, the port's outgoing packet counter, incoming packet counter, and queue status register are associated to ensure direct reading of hardware register data, reducing software layer latency. Secondly, a 100MHz clock module is configured within the FPGA, generating a 50kHz trigger signal through frequency division logic. Each trigger initiates a parallel data acquisition process, simultaneously reading the service rate of the eight port queues. Arrival rate Number of queues N and queue length ; A 1KB temporary buffer is allocated in the SRAM built into the FPGA. An 8-partition FIFO logic design is adopted, with each partition of 128B corresponding to the index storage of one port. After each acquisition, the cached data of the 8 partitions is packaged into JSON format and transmitted to the kernel of the computing module via the PCIe 4.0 bus. At the same time, the buffer is cleared to prepare for the next round of acquisition.

[0017] If the buffer is about to overflow, the overflow protection mechanism will be automatically triggered, discarding the latest data collected and marking the "buffer overflow" event in the FPGA log register to ensure the integrity of the stored data.

[0018] In data center networks, each port of a switch maintains eight different priority queues. Within the switching buffer, the MMU maintains two types of queues: inbound queues and outbound queues. These queues are also called counters. When a packet arrives at the switch, it must pass inbound and outbound admission control before entering the buffer. After a packet enters the buffer, the MMU maps it to an inbound counter based on the source port s and priority p, and to an outbound counter based on the destination port i and priority p. Each packet is stored only once in the buffer. Because lost and lossless traffic are assigned to different priority queues, these queues are classified as: inbound lossy queue, inbound lossless queue, outbound lossy queue, and outbound lossless queue. Each queue in the switching buffer is allocated a fixed buffer capacity to ensure basic forwarding functionality. To improve buffer utilization, shared buffers can dynamically allocate resources to specific queues; this sharing mechanism is called a buffer pool. In actual industrial production, buffer pools are typically configured as inbound pools, outbound lossy pools, outbound lossless pools, and clearance pools. Lossless and lossy traffic need to share the same inbound pool because partitioning the buffer would severely impact utilization. If the buffer space occupied by a certain type of traffic is not used by other types, it will lead to a serious waste of resources and reduce the ability to absorb burst traffic. The clearing pool is specifically designed to serve lossless traffic. Its original purpose was to accommodate data packets in transit after the lossless queue is paused by the flow control mechanism.

[0019] To determine the relationship between switch buffer allocation and throughput, queuing theory and network traffic models are used to analyze the characteristics of the switch buffers described above. This modeling is based on steady-state conditions to obtain the loss rate.

[0020] Among them, for lossless flow and lossy traffic The queue occupancy status is as follows:

[0021]

[0022] The representation is detrimental. It represents no loss. The size of the outlet or inlet pool. For admission control parameters, N is the number of exit queues using the buffer.

[0023] The throughput of each queue versus admission control parameters under both lossy and lossless traffic conditions. The relationship between them can be derived as follows:

[0024] Delay Indicates the time period Average packet queuing delay within the queue, this parameter is determined by the queue length. The number of data packets N in the queue and the service rate The decision is made jointly, and the calculation formula is as follows:

[0025] Ultimately, the utility function maintained by each queue of each exchange can be obtained. This function helps the exchange make trade-offs when allocating buffers to ensure that each queue achieves its maximum utility value:

[0026] Using throughput As a reward metric, packet loss rate will also be considered. and average packet delay gradient As a penalty factor; where the coefficient and All are constants; coefficients Set a tolerance threshold for random packet loss. Latency coefficient. The theoretical maximum flow rate corresponding to a specific bottleneck under balanced conditions without flow expansion is the sum of the total flow transmission rate and the bottleneck capacity.

[0027] The computing module primarily handles tasks such as monitoring data parsing, key indicator derivation, utility value calculation, and data interaction. When data flows from the monitoring module to the computing module, the specific implementation is as follows: During the module initialization phase, the PCIe 4.0 slave interface is configured through the device tree to establish a data receiving link with the monitoring module's FPGA. Simultaneously, the pre-compiled M / M / 1 / K queuing model algorithm library and EWMA smoothing functions are loaded. Upon receiving data from the monitoring module, the data parsing thread is started to extract the raw indicator values ​​from each port and adjust the service rate. Arrival rate Perform EWMA smoothing and derive the utility value of the current admission control parameters based on the M / M / 1 / K model, then transfer the result to the control module.

[0028] The control module contains two cycles, each cycle containing two elements labeled T. e This phase is used to explore the impact of access control parameters on system stability. Each T... e The phase consists of a monitoring period T m and calculate the control period T c Composition. Before each formal parameter update, the control module executes a two-phase iterative process. The first cycle serves as a pairwise exploration, studying the changes in the admission control parameters. Increase and reduce The impact of timeliness value, including the increase Marked as ;reduce Marked as ;here, This indicates increasing or decreasing the number of consecutive iterations in the same direction, aiming to accelerate parameter processing. The convergence; after the control module receives the utility values ​​from two exploratory tests, it will combine the utility values ​​corresponding to the two different adjustment schemes. and Compare them; if you find In subsequent testing cycles, the module will undergo in-depth analysis. and The impact on utility value, or and Based on this new set of comparative data, the module ultimately adopted... Within two cycles Update the value; otherwise, the system retains it. As the updated value.

[0029] During the monitoring phase, the control module also receives T from the computing module. c Value, update the parameter, and change the T value. c Send back to reset monitoring period T m This continuous feedback loop ensures that the system remains adaptive and responsive, thereby promoting rapid and efficient parameter convergence.

[0030] This invention proposes a switch buffer management method suitable for mixed traffic in data center networks, which divides responsibilities between mixed-traffic switches and non-mixed-traffic switches: like Figure 2As shown, four pool types are defined in data center switches: ingress pool, egress lossy pool, egress lossless pool, and clearance pool. Each port of a switch in a data center network maintains eight different priority queues. Within the switching buffer, the MMU maintains two types of queues: ingress queues and degress queues. These queues are also called counters. When a packet arrives at the switch, it must pass ingress and degress admission control before entering the buffer. After a packet enters the buffer, the MMU maps it to the ingress counter based on the source port s and priority p, and to the degress counter based on the destination port i and priority p. Each packet is stored only once in the buffer. Because dropped and non-dropped traffic are assigned to different priority queues, these queues are classified as: ingress lossy queue, ingress lossless queue, egress lossy queue, and egress lossless queue. Each queue in the switching buffer is allocated a fixed buffer capacity to ensure basic forwarding functionality. To improve buffer utilization, shared buffers can dynamically allocate resources to specific queues; this sharing mechanism is called a buffer pool. Lossless and lossy traffic need to share the same inbound pool because partitioning the buffer would severely impact utilization. If the buffer space occupied by one type of traffic is not used by other types, it will lead to a serious waste of resources and reduce the ability to absorb burst traffic. The clear pool is dedicated to serving lossless traffic, and its original design purpose is to accommodate data packets in transit after the lossless queue is paused by the flow control mechanism.

[0031] The complete process architecture diagram of this invention is as follows: Figure 3 As shown, the present invention provides a method for managing hybrid traffic exchange buffers in data centers for high-efficiency applications, which specifically includes the following steps: Step 1: The monitoring module sets the monitoring period. Whenever the monitoring interval is reached, the switch records the key indicators of the queue and passes them to the next module. To effectively manage and optimize queue performance, we divide the timeline into segments of duration T. e The time block, called the micro-experiment interval, typically ranges from 15.2 to 20.7 seconds in real-world scenarios, with best practices for this interval being between 15.2 and 20.7 seconds. T e The choice of value needs to be determined based on the specific application scenario and the required monitoring accuracy. Since queue latency is relatively short compared to the round-trip time, a shorter T value is used. e This value is sufficient to meet the accuracy requirements of the micro-experiment. During the initialization phase, a hardware-level connection is established between the PCIe 4.0 data interface and the eight ports of the port module. Simultaneously, the port's outgoing packet counter, incoming packet counter, and queue status register are associated to reduce software layer latency. Furthermore, a 100MHz clock module inside the FPGA is configured, generating a 50kHz trigger signal through frequency division logic. Each trigger initiates a parallel data acquisition process, synchronously reading the service rate of the eight port queues. Arrival rate Number of queues N and queue length Meanwhile, the switch calculates the service rate by integrating monitoring data. and arrival rate The data is then batch-packaged into JSON format and sent to the computing module via the PCIe 4.0 bus. At the same time, the cache area is cleared to prepare for the next round of collection. During this period, the data packets are constantly flowing in and out of the queue, and the entire transmission process will not be interrupted while waiting for the performance test results.

[0032] Step 2: After receiving the aggregated metrics, the calculation module determines the specific values ​​of each reward and penalty item based on the established utility function and calculates the utility value. The calculation module receives the service rate µ from the monitoring module. t Arrival rate λ t After obtaining other relevant performance metrics, the core task is to calculate the utility value used to evaluate buffer allocation. The specific calculation process is as follows: first, determine the specific values ​​of each reward and penalty item based on a predetermined utility function; then, weightedly combine these values. The utility function is based on throughput. As a reward metric, packet loss rate will also be considered. and average packet delay gradient As a penalty factor, and the coefficients in the function and All are constants; coefficients Used to set the tolerance threshold for random packet loss. ( The utility function represents the theoretical maximum flow rate of a specific bottleneck under equilibrium conditions without flow expansion; it is the sum of the total traffic transmission rate and the bottleneck capacity. This utility function predicts packet loss rate by establishing a queuing model, proactively providing feedback before actual packet loss occurs to reduce the risk of triggering flow control mechanisms. The delay d involved in the function... t Refers to the time period T e The average packet queuing delay within the queue is determined by the queue length. The number of data packets N in the queue and the service rate The decision is made jointly. The calculated utility value of the admission control is then passed to the control module for subsequent iterations.

[0033] Step 3: The control module consists of two cycles, each consisting of two sets of monitoring cycles and a calculation control cycle. The impact of increasing or decreasing the admission control parameters on the utility value is explored alternately, and the parameter that maximizes the utility value is selected for updating. The control module consists of two cycles, each consisting of two sets of monitoring cycles (T). m ) and calculation control cycle (T)c It consists of ) and each cycle further contains two labeled T e The (micro-experimental interval) phase is used to explore the impact of the admission control parameters on system stability. Before each formal update of the admission control parameters, the control module will first execute a two-stage iterative process to alternately explore the admission control parameters. The impact of increases or decreases on utility values; the first period serves as a paired exploration to study the admission control parameters separately. Increase (marked as) ) and reduce (marked as) Changes in utility value (here) This represents the number of consecutive iterations (increasing or decreasing) in the same direction, with the aim of accelerating parameter processing. (convergence), when the control module receives the utility values ​​corresponding to these two exploratory tests. and Then, the two will be compared, and if any discrepancies appear... (Or the opposite), then further analysis will be conducted in subsequent testing cycles. and (or correspondingly adjusted to) and The impact on utility value, and then based on the new utility value comparison results (such as... ), select Within two cycles Update the value; retain if the condition is not met. Finally, the update is completed by selecting the parameter that maximizes the utility value through the above process.

[0034] Step 4: The control module receives the updated cycle value from the calculation module and sends the modified cycle value back to the monitoring module to adjust the monitoring cycle for the new stage. The control module will first receive the update cycle value from the calculation module, which is the calculation control cycle T. c After adjusting and modifying the periodic value, the modified periodic value is sent back to the monitoring module. Upon receiving the modified periodic value, the monitoring module adjusts the monitoring period T for the new phase based on it. m Through this continuous feedback loop, the system is ensured to maintain its adaptability and responsiveness, thereby promoting the rapid and efficient convergence of key parameters such as admission control parameters, and providing a stable basis for the dynamic optimization and allocation of subsequent buffer resources.

[0035] Figure 4The 99th percentile flow completion times (FCTs) for TCP and RDMA traffic are displayed. The method (HBM) demonstrates a significant FCT latency advantage across different traffic categories. The DT scheduling scheme exhibits a clear FCT latency growth trend with increasing load. Under high load, HBM can reduce the 99th percentile FCT latency of RDMA traffic by approximately 55%-75%. With increasing TCP load, TCP traffic shows no performance degradation. Compared to the DT scheme, while ABM can alleviate RDMA FCT latency degradation, its effect is relatively limited because it cannot distinguish between lost and normal traffic. HBM still achieves an FCT latency optimization rate of nearly 50%. Compared to L2BM and Reverie, HBM shows a slower FCT latency growth with increasing load, demonstrating superior performance. However, HBM has more stable performance characteristics, maintaining a consistent FCT latency under different loads, with a peak of approximately 11.5. Compared to the other two schemes, HBM has a shorter FCT latency, with an average optimization rate of approximately 3.6%-5.0%. This invention allows different types of traffic to occupy buffer space fairly, eliminating the greed of lossy traffic for buffer space and ensuring the fairness of the coexistence of lossy and lossless traffic. This method not only improves the utilization efficiency of network resources such as switch buffers but also ensures fairness among different traffic types and overall network performance optimization, making it suitable for the needs of modern complex network environments.

[0036] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A method for managing hybrid traffic exchange buffers in data centers for high-efficiency applications, characterized in that: Includes the following steps: Step 1: The monitoring module sets the monitoring period. Whenever the monitoring interval is reached, the switch records the key indicators in the queue and then passes them to the next module. Step 2: After receiving the aggregated metrics, the calculation module determines the specific values ​​of each reward and penalty item based on the established utility function and calculates the utility value. Step 3: The control module consists of two cycles, each consisting of two sets of monitoring cycles and a calculation control cycle; the impact of increasing or decreasing the admission control parameters on the utility value is explored alternately, and the parameter that maximizes the utility value is selected for updating. Step 4: The control module receives the updated cycle value from the calculation module and sends the modified cycle value back to the monitoring module to adjust the monitoring cycle for the new stage.

2. The data center hybrid traffic exchange buffer management method according to claim 1, characterized in that: In step 1, to effectively manage and optimize queue performance, the time axis is divided into durations T. e The time block is called the micro-experiment interval; T e Values ​​are set between 15.2 and 20.

7. Within the range.

3. The data center hybrid traffic exchange buffer management method according to claim 2, characterized in that: During the monitoring process, the module calculates the service rate by integrating the monitoring data. and arrival rate During this period, data packets continuously flow in and out of the queue; Calculated service rate and arrival rate The value will be transferred to the calculation module for in-depth analysis.

4. The data center hybrid traffic exchange buffer management method according to claim 3, characterized in that: The calculation module receives the service rate from the monitoring module. Arrival rate Data on other performance metrics are used to calculate the utility value for evaluating buffer allocation; firstly, the specific values ​​of each reward and penalty item are determined according to the established utility function, and then the values ​​are weighted and combined; the utility function is as follows: Using throughput As a reward metric, packet loss rate will also be considered. and average packet delay gradient As a penalty factor; where the coefficient and All are constants; coefficients Set a tolerance threshold for random packet loss. Latency coefficient. The theoretical maximum flow rate corresponding to a specific bottleneck under balanced conditions without flow expansion is the sum of the total flow transmission rate and the bottleneck capacity.

5. The data center hybrid traffic exchange buffer management method according to claim 4, characterized in that: By establishing a queuing model to predict packet loss rate, proactive feedback can be provided before actual packet loss occurs, thereby reducing the risk of triggering flow control mechanisms. Among the delays Indicates the time period Average packet queuing delay within the queue, this parameter is determined by the queue length. The number of data packets N in the queue and the service rate A joint decision.

6. The data center hybrid traffic exchange buffer management method according to claim 5, characterized in that: In step 3: the control module comprises two cycles, each cycle containing two markers. This phase is used to explore the impact of admission control parameters on system stability; each The phase consists of a monitoring cycle and calculation control cycle Composition: Before each formal parameter update, the control module executes a two-stage iterative process.

7. The data center hybrid traffic exchange buffer management method according to claim 6, characterized in that: The first cycle serves as a paired exploration to study the influence of admission control parameters. Increase and reduce The effect of timeliness value, including the increase Marked as ;reduce Marked as ;here, This indicates increasing or decreasing the number of consecutive iterations in the same direction, aiming to accelerate parameter processing. The convergence; after the control module receives the utility values ​​from two exploratory tests, it will combine the utility values ​​corresponding to the two different adjustment schemes. and Compare them; if you find In subsequent testing cycles, the module will undergo in-depth analysis. and The effect on utility value, or and Based on this new set of comparative data, the module ultimately adopted... Within two cycles Update the value; otherwise, the system retains it. As the updated value.

8. The data center hybrid traffic exchange buffer management method according to claim 1, characterized in that: During the monitoring phase, the control module also receives data from the computing module. Value, update the parameter, and change the value. Send back to reset the monitoring period. ; This continuous feedback loop ensures that the system remains adaptive and responsive, thereby promoting rapid and efficient convergence of parameters.

Images