Dynamic traffic scheduling method based on inaccurate information in data center network

By designing the initial queue allocation and fine-grained priority degradation mechanism in the data center network, the traffic size prediction model is used to allocate priority and degradation thresholds to traffic, the problem of small flow being blocked by medium and large flow being solved, and the rapid transmission of small flows and overall performance optimization is achieved.

CN120528877APending Publication Date: 2025-08-22CENT SOUTH UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202510780522.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-06-12
Publication Date
2025-08-22

AI Technical Summary

Technical Problem

The existing data center network traffic scheduling schemes are difficult to effectively utilize inaccurate information in scenarios with unknown traffic size, resulting in small flows being blocked by medium and large flows, increasing latency, and high deployment complexity or insufficient performance.

Method used

By training the traffic size prediction model offline, differentiate the initial priority and degradation threshold for different traffic based on the predicted traffic size, design the initial queue allocation and fine-grained priority degradation mechanism, optimize the opportunity for small streams and reduce the blockage of small streams by large streams.

Benefits of technology

The completion time of small streams is significantly optimized, the average and tail completion time of small streams is reduced, and the overall performance and resource utilization efficiency of the network are improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120528877A_ABST
    Figure CN120528877A_ABST
Patent Text Reader

Abstract

The invention discloses a dynamic traffic scheduling method based on inaccurate information in a data center network, which comprises the following steps of: in response to arrival of a current data packet, judging whether the traffic to which the current data packet belongs is new traffic, if so, predicting the size of the traffic, distributing an initial queue for the traffic according to an interval in which the size of the traffic is located, calculating a degradation threshold value of the flow in each queue; otherwise, taking the priority of the queue distributed by the flow as the priority of the current data packet, and executing the forwarding of the current data packet according to the priority; judging whether the number of sent data packets of the flow in the currently distributed queue reaches a degradation threshold value of the flow in the currently distributed queue or not, and if so, carrying out degradation operation on the flow; otherwise, the queue distributed by the flow is kept unchanged. According to the application, the blockage of medium and large flows to small flows can be greatly reduced, and the completion time of the small flows is remarkably optimized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of data center networks, and in particular relates to a dynamic traffic scheduling method based on imprecise information in a data center network. Background Art

[0002] Many data center networks (DCNs) support a variety of applications, such as web search, key-value storage, and distributed machine learning training. The traffic generated exhibits a heavy-tailed distribution, with a large number of small, latency-sensitive flows. Traffic scheduling utilizes queues to prioritize small flows over larger flows, effectively optimizing their completion time and significantly improving the performance of latency-sensitive traffic.

[0003] Existing traffic scheduling schemes aim to provide ultra-low latency for applications and can be categorized as precise traffic-aware scheduling and imprecise traffic-aware scheduling. Precise traffic-aware scheduling schemes rely on precise, pre-known traffic information to approximate minimum remaining processing time scheduling, but their deployability is poor. For example, pFabric implements the minimum remaining processing time policy through an infinite number of priority levels, but requires fully customized switches, resulting in extremely complex deployment. EPN simulates infinite priority levels with only two queues, but requires deploying a priority controller and mapper on the switch and host, respectively. pHost and Homa also schedule traffic based on a limited number of queues, but require restructuring the network stack or relying on programmable switches. Imprecise traffic scheduling schemes, on the other hand, do not require information about traffic size and adjust priorities solely based on the number of bytes sent. PIAS assigns the highest priority to all new traffic and gradually reduces the priority of current packets as the data volume increases. This approach is adaptable to the multiple queues of commodity switches and traditional TCP / IP protocol stacks. However, its static degradation thresholds make it difficult to adapt to dynamic network traffic, and its "indiscriminate scheduling" mechanism can cause high queues to be occupied by medium and large flows for extended periods, severely blocking small flows. Recent research has used historical traffic trajectories to infer traffic flow sizes, employing heuristic buffer-aware methods and machine learning algorithms to achieve imprecise traffic flow size predictions, providing strong support for scheduling scheme design. For example, QCLIMB proposes a "climb first, then descend" scheduling strategy, which differentiates between large and small flows based on an imprecise lower bound of the prediction. However, this strategy still suffers from insufficient traffic differentiation and can introduce out-of-order behavior, leading to increased latency.

[0004] Therefore, in scenarios where the traffic size is unknown, how to obtain and effectively utilize imprecise traffic information to prioritize small flows over medium and large flows to minimize their completion time is a topic with research value. Summary of the Invention

[0005] To address the trade-off between deployability and optimized performance in existing traffic scheduling solutions, this paper proposes a dynamic traffic scheduling method for data center networks based on imprecise information. By assigning differentiated initial priorities and unique degradation thresholds to predicted flows of varying sizes, this method achieves fine-grained degradation control, significantly reducing the impact of medium and large flows on small flows and significantly optimizing small flow completion times.

[0006] In a first aspect, the present application provides a method for dynamic traffic scheduling based on imprecise information in a data center network, comprising:

[0007] Step 1: In response to the arrival of the current data packet, obtain the flow to which the current data packet belongs; determine whether the first flow is new flow, if so, predict the size of the first flow, and go to step 2; otherwise, go to step 3; wherein the first flow is the flow to which the current data packet belongs;

[0008] Step 2: Allocate an initial queue to the first flow based on the size interval of the first flow, and calculate a degradation threshold for the first flow in each queue; the degradation threshold for a flow in a queue represents the number of packets that can be transmitted by the flow in the queue;

[0009] Step 3: Using the priority of the queue allocated to the first flow as the priority of the current data packet, and forwarding the current data packet accordingly;

[0010] Step 4: Obtain the number of data packets sent by the first traffic in the currently assigned queue and the degradation threshold of the first traffic in the currently assigned queue, and determine whether it is necessary to downgrade the first traffic; if the number of data packets sent by the first traffic in the currently assigned queue reaches the degradation threshold of the first traffic in the currently assigned queue, downgrade the first traffic; otherwise, the queue assigned to the first traffic remains unchanged; wherein, downgrading the first traffic means allocating the first traffic to a queue with a lower transmission priority.

[0011] According to the first aspect, in one possible implementation, before step 1, the method further includes: offline training a traffic size prediction model based on an actual traffic data set in the data center network, so that the trained traffic size prediction model has the function of predicting the size of the traffic based on the network characteristics of the traffic and the real-time network status;

[0012] Predicting the size of the first flow in step 1 includes: inputting the network characteristics and real-time network status of the first flow into a trained flow size prediction model, and using the trained flow size prediction model to predict the size of the first flow.

[0013] This step uses the traffic size prediction model to predict traffic size through offline training, providing a basis for subsequent fine-grained priority scheduling.

[0014] According to the first aspect, in a possible implementation method, the offline training of the traffic size prediction model enables the trained traffic size prediction model to have the function of predicting the size of the traffic based on the network characteristics of the traffic and the real-time network status, including: offline training of the traffic size prediction model, enables the trained traffic size prediction model to have the function of predicting the size of the traffic based on the network characteristics of the traffic and the real-time network status and calculating the confidence.

[0015] In the step 1, using the trained flow size prediction model to predict the size of the first flow includes: using the trained flow size prediction model to predict the size of the first flow and calculating the confidence.

[0016] In other words, this solution uses the traffic size prediction model to predict the traffic size, providing a basis for subsequent fine-grained priority scheduling, and also outputs the confidence level to correct the prediction value deviation. , and combined with the confidence α to perform error correction to determine the final predicted flow size, where The closer it is to 1, the closer the final predicted first flow is to the actual predicted first flow. .

[0017] According to the first aspect, in a possible implementation manner, the traffic size prediction model is a random forest model.

[0018] According to the first aspect, in a possible implementation, in step 2, the size of the first flow is the size of the corrected flow , the specific calculation formula is:

[0019]

[0020] Where, is the size of the first flow predicted, is the confidence level, represents the scaling factor, which limits and The maximum difference between them can be combined with confidence to perform error correction.

[0021] This step dynamically corrects the predicted first flow based on the confidence α , and based on the scaling factor Adaptively scale the prediction results to overcome the deviation of the prediction model caused by data noise, burst traffic, etc., thereby reducing resource allocation errors caused by prediction deviation.

[0022] According to the first aspect, in one possible implementation, the priority of the queue may include in is the highest priority, is the lowest priority, is the number of priorities, that is, the number of queues; The priority of the queue is .

[0023] According to the first aspect, in a possible implementation, the flow rate is Demotion threshold in queues The calculation method is:

[0024]

[0025] Where, Represents the number of packets that can be transmitted in the initial queue. Represents the traffic degradation threshold adjustment coefficient, represents the priority of the initial queue to which traffic is assigned, where .

[0026] According to the first aspect, in a possible implementation, the flow degradation threshold adjustment coefficient for:

[0027]

[0028] Where, is the experience value, k and It is inversely proportional to the traffic degradation threshold, which is used to increase the degradation threshold of small flows and ensure that small flows degrade more slowly.

[0029] Furthermore, the following method is used to determine The value of:

[0030] Keeping other conditions unchanged, the workload was tested with a step size of 50. The value of 0-500 corresponds to the flow completion time (FCT) in different situations. Experiments show that when When set to 50-100, the performance of this application algorithm is optimal. The value of is preferably 50-100. The value of is from 0 to 500, and the FCT curve shows a trend of first decreasing and then increasing: When the value of k is too small, it will decrease, which will cause the small flow to be frequently degraded and unable to complete the transmission in the high queue. When the value is too large, although small flows can be transmitted through a single degradation, large flows will transmit too much data in the high queue, thus affecting the overall performance.

[0031] According to the first aspect, in one possible implementation, the number of packets that can be transmitted in the initial queue is for:

[0032]

[0033] Where, It is an empirical value used for adjustment to avoid large flows taking up too much space in the initial queue. and Represents the two dividing points for dividing the traffic size range. S limits the minimum number of packets that each flow needs to transmit in the initial priority queue to avoid transmission tailing due to degradation operations when a small flow has a small number of packets remaining.

[0034] Further, The specific steps for determining S are as follows:

[0035] In exploring the parameters In the experiment on the effect of traffic scheduling, when the traffic size is known and fixed The test was carried out under the conditions of It is positively correlated with the number of bytes transmitted by the large flow in the initial queue. Specifically, The larger the value, the more serious the obstruction of the large flow to the small flow, and the average FCT of the small flow increases accordingly; when When it is 0, the average / tail FCT of small flows is optimal, but it will cause starvation of large flows and cause TCP timeout retransmission, which will reduce application performance. Set to 0.001, this value can be dynamically adjusted based on the expected traffic distribution of the data center. The smaller the value, the better it is for improving small flow performance.

[0036] Regarding the parameter S, it is set to 10KB based on the traffic distribution characteristics in the data center network. The function of this parameter is to prevent a small flow from being downgraded to a lower queue even when there are only a few packets left, thereby reducing the adverse effects of the small flow transmission process. At the same time, when calculating the number of packets a that can be transmitted in the initial queue for each flow, S can ensure that when When the packet size is ≤10KB, the entire traffic is forwarded in the queue with the highest priority, effectively reducing tail delay.

[0037] In a second aspect, the present application provides an electronic device, comprising: a memory and a processor;

[0038] The memory is used to store computer programs;

[0039] The processor is configured to call the computer program to execute the method described above.

[0040] In a third aspect, the present application provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed on an electronic device, the electronic device implements the method described above.

[0041] In a fourth aspect, the present application provides a computer program product, including a computer program, which, when executed on an electronic device, enables the electronic device to implement the method described above.

[0042] The specific implementation methods of the second to fourth aspects of this application can refer to the implementation method of the first aspect above, and will not be repeated here.

[0043] Beneficial effects:

[0044] The present invention proposes a dynamic traffic scheduling method based on imprecise information in a data center network, and designs an initial queue allocation and fine-grained priority degradation mechanism. In the initial queue allocation module, the initial queue is classified and allocated according to the predicted traffic size. The smaller the traffic size, the higher the priority of the allocated initial queue, thereby increasing the chance of small flows being transmitted first, thereby reducing the average and tail FCT of small flows; in the fine-grained priority degradation module, differentiated degradation thresholds are set according to the traffic size, so that large flows are downgraded to low queues faster as the transmission progresses, reducing the blockage of small flows. The present invention can dynamically adjust different levels of priority for flows of different sizes with the help of predicted traffic size information, reduce the blockage of smaller flows by larger flows, optimize the completion time of small flows, and significantly optimize the flow completion time. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0046] Figure 1 This is a flow chart of a dynamic traffic scheduling method based on imprecise information in a data center network provided by an embodiment of the present invention.

[0047] Figure 2This is a performance comparison of the embodiment of the present invention with the QCLIMB and PIAS solutions in terms of flow completion time for small and medium flows and large flow throughput under the K-Means workload; (a), (b), (c), and (d) are schematic diagrams of the changes in the average FCT of small flows, the tail FCT of small flows, the average FCT of medium flows, and the large flow throughput under different load conditions (Load, ranging from 0.3 to 0.9).

[0048] Figure 3 This is a performance comparison of the embodiments of the present invention with the QCLIMB and PIAS solutions in terms of FCT of small and medium flows and large flow throughput under the PageRank workload; (a), (b), (c), and (d) are schematic diagrams of the changes in the average FCT of small flows, the tail FCT of small flows, the average FCT of medium flows, and the large flow throughput under different load conditions (Load, ranging from 0.3 to 0.9).

[0049] Figure 4 This is a performance comparison of the FCT of small and medium flows and the throughput of large flows provided by an embodiment of the present invention with the QCLIMB and PIAS solutions under the SGD workload; (a), (b), (c), and (d) are schematic diagrams of the changes in the average FCT of small flows, the tail FCT of small flows, the average FCT of medium flows, and the throughput of large flows under different load conditions (Load, ranging from 0.3 to 0.9). DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be described in detail below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other implementation methods obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. An embodiment of the present invention provides a dynamic traffic scheduling method based on imprecise information in a data center network. The method of the present application can be divided into an initial queue allocation stage, a fine-grained priority degradation stage and a packet marking stage.

[0051] The initial queue allocation phase is as follows:

[0052] After receiving data traffic from the application side, the initial queue allocation module performs traffic classification based on the traffic size predicted by the traffic size prediction model. For example, the specific classification criteria are: 0-100KB is a small flow, 100KB-10MB is a medium flow, and greater than 10MB is a large flow. Based on the above classification results, initial queues are allocated for traffic of different sizes. Generally speaking, the smaller the traffic size, the higher the priority of the initial queue assigned. This is because small flows are more sensitive to delays and need to be processed preferentially to ensure their transmission efficiency and timeliness; conversely, large flows are assigned to relatively low-priority initial queues.

[0053] The fine-grained priority degradation stages are as follows:

[0054] This stage is achieved through the formula Calculate the degradation threshold corresponding to each queue. Here 、 、 Parameters such as [number of traffic flows] need to be optimized and set, and their values ​​will affect the dynamic adjustment of traffic priority during transmission. As data traffic continues to be transmitted, fine-grained priority adjustments are made to the traffic based on the results of the formula calculation. Specifically, even if a certain traffic is given a higher priority in the initial stage, if it meets the degradation conditions set by the formula and parameters during transmission, its priority will be appropriately lowered. This mechanism aims to achieve balanced utilization of network resources, prevent high-priority traffic from occupying too many resources for a long time, provide reasonable transmission opportunities for other traffic, and thus ensure the orderly scheduling of the entire network traffic.

[0055] The details of the group marking process are as follows:

[0056] This step tags the current packet's priority based on the flow it belongs to and the priority of the queue it's assigned to. For example, the packet's priority can be embedded in the packet header. The switch reads the priority tag to identify the priority and performs packet forwarding and scheduling operations accordingly. Packets with high-priority tags are transmitted first, thus achieving efficient network traffic scheduling, ensuring that the network needs of various applications are differentiated and meeting them, thereby improving overall network performance.

[0057] The embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0058] Example 1:

[0059] The present application embodiment provides a dynamic traffic scheduling method based on imprecise information in a data center network (hereinafter referred to as Sadra). The workflow diagram is as follows: Figure 1 As shown, the specific steps include:

[0060] Step 1: In response to the arrival of the current data packet, obtain the flow to which the current data packet belongs; determine whether the first flow is new flow, if so, predict the size of the first flow, and go to step 2; otherwise, go to step 3; wherein the first flow is the flow to which the current data packet belongs;

[0061] Specifically, the flow to which the current packet belongs (set as the first flow) is obtained to determine whether the flow is new. If not, the process jumps directly to the packet marking phase and fine-grained priority degradation phase (i.e., steps 3 and 4). If it is determined to be new, the initialization process is initiated, i.e., the size of the first flow is estimated based on the network characteristics of the first flow and the real-time network status, and the process then jumps to step 2.

[0062] In some embodiments, before step 1, the method further includes: offline training a traffic size prediction model based on an actual traffic data set in the data center network, so that the trained traffic size prediction model has the function of predicting the size of the traffic based on the network characteristics of the traffic and the real-time network status;

[0063] Predicting the size of the first flow in step 1 includes: inputting the network characteristics and real-time network status of the first flow into a trained flow size prediction model, and using the trained flow size prediction model to predict the size of the first flow.

[0064] This step uses the traffic size prediction model to predict traffic size through offline training, providing a basis for subsequent fine-grained priority scheduling.

[0065] In some embodiments, the offline training of the traffic size prediction model enables the trained traffic size prediction model to have the function of predicting the size of the traffic based on the network characteristics of the traffic and the real-time network status, including: offline training of the traffic size prediction model enables the trained traffic size prediction model to have the function of predicting the size of the traffic based on the network characteristics of the traffic and the real-time network status and calculating the confidence.

[0066] In the step 1, using the trained flow size prediction model to predict the size of the first flow includes: using the trained flow size prediction model to predict the size of the first flow and calculating the confidence.

[0067] In some embodiments, the traffic size prediction model is a random forest model.

[0068] In some embodiments, the network characteristics of the traffic and the real-time network status include the start time of traffic generation, the interval with the end time of the previous traffic, and the status information of the data packet five-tuple.

[0069] Step 2: Allocate an initial queue to the first flow based on the size interval of the first flow, and calculate a degradation threshold for the first flow in each queue; the degradation threshold for a flow in a queue represents the number of packets that can be transmitted by the flow in the queue;

[0070] Illustratively, the traffic size ranges include 0-100KB, 100KB-10MB, and 10MB-∞, corresponding to three types of traffic: small flow, medium flow, and large flow, respectively.

[0071] Different initial queues and priority degradation rates are assigned to different types of traffic. The initial queues assigned to small, medium, and large flows can be the highest priority, second highest priority, and third priority queues, respectively, to prioritize transmission of smaller flows.

[0072] In some embodiments, the size of the first flow is the size of the corrected flow , the specific calculation formula is:

[0073]

[0074] Where, is the size of the first flow predicted, is the confidence level, represents the scaling factor, which limits and The maximum difference between them can be combined with confidence to perform error correction.

[0075] In some embodiments, multiple comparative experiments show that The best effect is achieved when the value is around 0.9.

[0076] In some embodiments, the priority of the queue may include in is the highest priority, is the lowest priority, is the number of priorities, that is, the number of queues; The priority of the queue is .

[0077] In some embodiments, the flow rate is Demotion threshold in queues The calculation method is:

[0078]

[0079] Where, Represents the number of packets that can be transmitted in the initial queue. Represents the traffic degradation threshold adjustment coefficient, represents the priority of the initial queue to which traffic is assigned, where .

[0080] According to the first aspect, in a possible implementation, the flow degradation threshold adjustment coefficient for:

[0081]

[0082] Where, is the experience value, k and It is inversely proportional to the traffic degradation threshold, which is used to increase the degradation threshold of small flows and ensure that small flows degrade more slowly.

[0083] In some embodiments, the following method is used to determine The value of:

[0084] Keeping other conditions unchanged, the workload was tested with a step size of 50. The value of 0-500 corresponds to the flow completion time (FlowCompletionTime, FCT) in different situations. Experiments show that when When set to 50-100, the performance of this algorithm is optimal. The value of is from 0 to 500, and the FCT curve shows a trend of first decreasing and then increasing: When the value of k is too small, it will decrease, which will cause the small flow to be frequently degraded and unable to complete the transmission in the high queue. When the value is too large, although small flows can be transmitted through a single degradation, large flows will transmit too much data in the high queue, thus affecting the overall performance.

[0085] In some embodiments, the number of packets that can be transmitted in the initial queue for:

[0086]

[0087] Where, It is an empirical value used for adjustment to avoid large flows taking up too much space in the initial queue. and Represents the two dividing points for dividing the traffic size range. S limits the minimum number of packets that each flow needs to transmit in the initial priority queue to avoid transmission tailing due to degradation operations when a small flow has a small number of packets remaining.

[0088] Further, The specific steps for determining S are as follows:

[0089] In exploring the parameters In the experiment on the effect of traffic scheduling, when the traffic size is known and fixed The test was carried out under the conditions of It is positively correlated with the number of bytes transmitted by the large flow in the initial queue. Specifically, The larger the value, the more serious the obstruction of the large flow to the small flow, and the average FCT of the small flow increases accordingly; when When it is 0, the average / tail FCT of small flows is optimal, but it will cause starvation of large flows and cause TCP timeout retransmission, which will reduce application performance. Considering performance and fairness comprehensively, The value of is preferably 0.001, which can be dynamically adjusted according to the expected traffic distribution of the data center. Under the premise of ensuring that there is no risk of starvation for large flows, The smaller the value, the better it is for improving small flow performance.

[0090] Regarding the parameter S, it is preferably set to 10KB based on the traffic distribution characteristics in the data center network. The function of this parameter is to prevent a small flow from being downgraded to a lower queue even when there are only a few packets left, thereby reducing the adverse effects of the small flow transmission process. At the same time, when calculating the number of packets a that can be transmitted in the initial queue for each flow, S can ensure that when When the packet size is ≤10KB, the entire traffic is forwarded in the queue with the highest priority, effectively reducing tail delay.

[0091] In this step, the number B of data packets sent by the first flow in the currently allocated queue is initialized to 0.

[0092] Step 3: Using the priority of the queue allocated to the first flow as the priority of the current data packet, and forwarding the current data packet accordingly;

[0093] In this step, the priority of the current data packet is determined based on the parameter information corresponding to the first flow and embedded into the packet header. At the same time, the number of data packets B sent by the first flow in the currently allocated queue is updated;

[0094] Step 4: Obtain the number of data packets sent by the first traffic in the currently assigned queue and the degradation threshold of the first traffic in the currently assigned queue, and determine whether it is necessary to downgrade the first traffic; if the number of data packets sent by the first traffic in the currently assigned queue reaches the degradation threshold of the first traffic in the currently assigned queue, downgrade the first traffic; otherwise, the queue assigned to the first traffic remains unchanged; wherein, downgrading the first traffic means allocating the first traffic to a queue with a lower transmission priority.

[0095] In this step, a determination is made as to whether the number of packets B sent by the first traffic flow in the currently assigned queue has reached the demotion threshold for the first traffic flow in the currently assigned queue. If so, the first traffic flow is assigned to a queue with a lower priority level (increased by one). If not, the queue assigned to the first traffic flow remains unchanged, i.e., the first traffic flow maintains its current priority. After the determination is complete, processing of subsequent packets continues.

[0096] It should be understood that when the system detects that the number of packets sent by a particular flow in its currently assigned queue has reached its corresponding demotion threshold in the current queue, it triggers a priority update mechanism. Specifically, if the newly arrived packet is the last packet that can be forwarded by that flow in the current queue, the priority of all subsequent packets of that flow will be adjusted to a lower level. By gradually reducing the priority of a flow, the system can achieve fine-grained control over the priority of all packets.

[0097] It should also be noted that the method provided in this embodiment requires initialization before implementing fine-grained scheduling of input traffic. This initialization involves offline training of the prediction model. Specifically, the process involves training a high-precision traffic volume prediction model based on a real-world traffic dataset from the data center network. Compared to a single decision tree model, this model effectively reduces prediction error through an ensemble learning mechanism and quantifies the confidence level of the prediction results by calculating the variance between each decision tree.

[0098] The above embodiment provides a dynamic traffic scheduling method based on imprecise information in a data center network, which optimizes the flow completion time by designing the initial queue allocation and fine-grained priority degradation mechanism. In the initial queue allocation mechanism, the system classifies the traffic based on the predicted traffic size and assigns an initial priority, following the allocation principle of "the smaller the size of the predicted first traffic, the higher the initial priority", thereby increasing the priority transmission opportunities for small flows and reducing their average and tail FCT; in the fine-grained priority degradation mechanism, differentiated degradation thresholds are set according to the size of the predicted first traffic, so that large flows are downgraded to low queues faster during transmission, reducing the obstruction of small flows. This method provides priority protection for flows of different sizes with the help of predicted traffic size information, thereby achieving refined scheduling and performance optimization of traffic.

[0099] Example 2:

[0100] This embodiment provides an electronic device, including: a memory and a processor;

[0101] The memory is used to store computer programs;

[0102] The processor is configured to call the computer program to execute the method described in the first embodiment.

[0103] Example 3:

[0104] This embodiment provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed on an electronic device, the electronic device implements the method according to the first embodiment.

[0105] Example 4:

[0106] This embodiment provides a computer program product, including a computer program. When the computer program is run on an electronic device, the electronic device implements the method described in the first embodiment.

[0107] The specific implementation methods of a system, electronic device, computer-readable storage medium, and computer program product provided in the embodiments of the present application can refer to the specific embodiments of the above-mentioned method and will not be repeated here.

[0108] Obviously, those skilled in the art should understand that the above-mentioned units or steps of the present application can be implemented using a general-purpose computing device. They can be concentrated on a single computing device or distributed across a network composed of multiple computing devices. Alternatively, they can be implemented using program codes executable by the computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps can be made into a single integrated circuit module for implementation. Thus, the present application is not limited to any specific combination of hardware and software.

[0109] The performance of the solution of the present invention is tested in conjunction with an experimental simulation to verify the feasibility and advantages of the solution of the present invention.

[0110] Experimental Simulation: This study conducted performance evaluation based on the NS3.31 network simulation platform, employing a leaf-spine topology consisting of four core switches and nine edge switches. Each edge switch connected to 16 servers via 40Gbps links and to four core switches via 100Gbps links. End hosts used DCTCP as the transport protocol. To validate the effectiveness of the proposed Sadra algorithm, this study selected two representative traffic scheduling schemes for comparative experiments: QCLIMB (which supports scheduling based on imprecise traffic information) and PIAS (scheduling based on the number of bytes sent without prior traffic size assumptions). The experiments employed three classic heavy-tailed distributed workloads: K-Means, PageRank, and SGD. Their traffic characteristics closely align with those of real-world data centers (i.e., a majority of small flows, with a few large flows dominating the bandwidth). Network load was controlled by varying the time intervals between traffic arrivals.

[0111] Figures 2 to 4Subfigures (a) and (b) show how the average and tail FCT of small flows vary with load for different schemes under various workloads. Experimental data shows that under all tested workloads, the Sadra scheme significantly outperforms the PIAS scheme in both the average and tail FCT of small flows. Specifically, compared to PIAS, Sadra reduces the average and tail FCT of small flows by 61% / 69%, 53% / 71%, and 55% / 89% in the K-Means, PageRank, and SGD workloads, respectively. In the PIAS scheme, small flows are easily severely blocked by large flows in high queues, resulting in limited performance. Compared to the QCLIMB scheme, Sadra also outperforms the QCLIMB scheme in terms of the average and tail FCT of small flows. In the K-Means, PageRank, and SGD workloads, it reduces the average and tail FCT of small flows by 25% / 9%, 19% / 10%, and 23% / 38%, respectively. It should be noted that in some scenarios, the performance of Sadra is close to that of QCLIMB. This is because some small flows have completed transmission during the queue climbing phase and have not triggered the complete priority adjustment mechanism of QCLIMB.

[0112] Figure 2-Figure 4 The sub-graph (c) shows the trend of the average FCT of the mid-stream under different workloads. The experimental results show that the Sadra scheme is significantly better than the PIAS and QCLIMB schemes in terms of mid-stream scheduling performance. Figure 3 Taking subgraph (c) of Figure 2 as an example, under the PageRank workload, Sadra achieved an 8%-21% and 12%-16% performance improvement in average FCT for midstream flows compared to PIAS and QCLIMB, respectively. This result is as expected, as the synergy between the Initial Priority Assignment (IPA) and Fine-Grained Priority Demotion (FGD) mechanisms enables midstream flows to be transmitted before large flows. Furthermore, since small flows are transmitted and completed quickly, when all small flows are prioritized over midstream flows, the queuing delay of midstream flows does not increase significantly compared to when they are blocked by large flows.

[0113] Figure 2-Figure 4Subgraph (d) shows the average throughput performance of large flows under different experimental settings. The following conclusions can be drawn from data analysis: Among various traffic scheduling schemes, the difference in the average throughput of large flows is negligible. It is worth noting that the Sadra scheme performs better than the PIAS scheme in the transmission performance of large flows in some scenarios. This phenomenon shows that the negative impact of the priority scheduling of small flows on the transmission of large flows is not absolute. The reason is that the network bandwidth resources occupied by small flows are significantly lower than those of large flows. Taking the PageRank workload with a network load of 0.5 as an example, according to Figure 3 The experimental data of subgraph (d) shows that compared with the PIAS solution, the Sadra solution improves the average throughput of large flows by 1.2%.

[0114] In summary, the present invention provides a dynamic traffic scheduling method based on imprecise information in a data center network, which has better performance, can provide fine-grained priority control for traffic of different sizes, and optimize flow completion time.

Claims

1. A dynamic traffic scheduling method based on imprecise information in a data center network, characterized in that: include: Step 1: In response to the arrival of the current data packet, obtain the flow to which the current data packet belongs; determine whether the first flow is new flow, if so, predict the size of the first flow, and go to step 2; otherwise, go to step 3; wherein the first flow is the flow to which the current data packet belongs; Step 2: Allocate an initial queue to the first flow based on the size interval of the first flow, and calculate a degradation threshold for the first flow in each queue; the degradation threshold for a flow in a queue represents the number of packets that can be transmitted by the flow in the queue; Step 3: Using the priority of the queue allocated to the first flow as the priority of the current data packet, and forwarding the current data packet accordingly; Step 4: Obtain the number of data packets sent by the first traffic in the currently assigned queue and the degradation threshold of the first traffic in the currently assigned queue, and determine whether it is necessary to downgrade the first traffic; if the number of data packets sent by the first traffic in the currently assigned queue reaches the degradation threshold of the first traffic in the currently assigned queue, downgrade the first traffic; otherwise, the queue assigned to the first traffic remains unchanged; wherein, downgrading the first traffic means allocating the first traffic to a queue with a lower transmission priority.

2. The method according to claim 1, characterized in that Before step 1, the method further includes: offline training a traffic size prediction model based on an actual traffic data set in the data center network, so that the trained traffic size prediction model has the function of predicting the size of the traffic based on the network characteristics of the traffic and the real-time network status; Predicting the size of the first flow in step 1 includes: inputting the network characteristics and real-time network status of the first flow into a trained flow size prediction model, and using the trained flow size prediction model to predict the size of the first flow.

3. The method according to claim 2, characterized in that The offline training of the traffic size prediction model enables the trained traffic size prediction model to have the function of predicting the size of traffic based on the network characteristics of the traffic and the real-time network status, including: Offline training of the traffic size prediction model enables the trained traffic size prediction model to have the function of predicting the traffic size and calculating the confidence level based on the network characteristics of the traffic and the real-time network status; In step 1, using the trained flow size prediction model to predict the size of the first flow includes: The trained flow size prediction model is used to predict the size of the first flow and calculate the confidence level.

4. The method according to claim 3, characterized in that In step 2, the size of the first flow is the size of the corrected flow , the specific calculation formula is: ; Where, is the size of the first flow predicted, is the confidence level, represents the scaling factor, which limits and The maximum difference between them can be combined with confidence to perform error correction.

5. The method according to claim 1, wherein Traffic in the Demotion threshold in queue The calculation method is: ; Where, Represents the number of packets that can be transmitted in the initial queue. Represents the traffic degradation threshold adjustment coefficient, represents the priority of the initial queue to which traffic is assigned, where , is the priority number.

6. The fine-grained priority scheduling method according to claim 5, characterized in that: Traffic degradation threshold adjustment coefficient for: ; Where, It's experience points.

7. The fine-grained priority scheduling method according to claim 5, characterized in that: The number of packets that can be transmitted in the initial queue for: ; Where, It is an empirical value used for adjustment to avoid large flows taking up too much space in the initial queue. and Represents the two dividing points for dividing the traffic size range. S limits the minimum number of packets that each flow needs to transmit in the initial priority queue to avoid transmission tailing due to degradation operations when a small flow has a small number of packets remaining.

8. An electronic device, characterized in that: include: memory and processor; The memory is used to store computer programs; The processor is configured to call the computer program to execute the method according to any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed on an electronic device, the electronic device implements the method according to any one of claims 1 to 7.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed on an electronic device, the electronic device implements the method according to any one of claims 1 to 7.