Load balancing method and apparatus across data center networks

Abstract

This invention provides a load balancing method and apparatus for cross-data center networks. The method is executed by a switching node in a first data center and includes: determining the type of a flow based on the differential service code point (DSC) markers carried in data packets of a data flow received from a source node or other switching nodes; if the flow type is determined to be a cross-data center flow, then employing a state-aware dynamic binding mechanism to select a corresponding link for the flow according to the source node of the flow, so as to forward the flow to a first interconnect node or forward it to the first interconnect node in the order of data packets through other switching nodes in the first data center, and then forward it to a second interconnect node in the second data center through the first interconnect node, so that the second interconnect node can adopt a port virtualization mechanism and a state-aware dynamic binding mechanism, and introduce a conflict arbitration strategy driven by a network topology constraint table to select a corresponding path for the received flow according to the destination node of the received flow, so as to forward the received flow to the destination node.

Description

Technical Field

[0001] This invention relates to the field of data center load balancing technology, and more particularly to load balancing methods and apparatus across data center networks. Background Technology

[0002] RDMA (Remote Direct Memory Access) technology, with its advantages of kernel bypass and protocol stack hardware offloading, can achieve extremely high throughput performance and has become a common paradigm in modern data center networks. With the rapid evolution of cloud storage and large-scale AI model training, a single data center can no longer independently support the distributed training needs of trillion-parameter models in terms of computing scale and storage capacity. Cross-regional distributed deployment has become an inevitable trend. In practical applications, for example, Google needs to complete exabyte-scale data migrations between dozens of sites daily, and the pre-training of ultra-large-scale models such as Gemini 1 Ultra and GPT-4.5 also widely adopts cross-data center parallel training modes. This evolution from single-data center to cross-data center architecture places higher demands on the transmission efficiency and load balancing strategies of RDMA networks under complex topologies.

[0003] The shift from traditional intra-datacenter networks to cross-datacenter networks presents three fundamental challenges to load balancing design. First, data flow patterns are significantly heterogeneous. Compared to the short-message-dominated data flow within a data center, cross-datacenter networks exhibit significantly longer data flows from distributed storage and AI (Artificial Intelligence) training. This means that performance degradation is more severe in the event of load imbalance. Second, topology evolution expands the path decision space. Traditional intra-datacenter networks often employ symmetrical two- or three-layer Clos architectures, resulting in relatively simple load balancing decision paths. Cross-datacenter networks introduce DCI (Data Center Interconnect) switches, significantly increasing end-to-end path diversity while also introducing more complex routing constraints at the receiving end. Finally, heterogeneous propagation delays lead to control loop mismatch. Round-trip latency within a data center is typically in the microsecond range, while cross-datacenter propagation latency can reach milliseconds. This heterogeneous propagation time makes it difficult for traditional load balancing strategies to achieve effective and rapid responses in cross-datacenter environments. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a load balancing method and apparatus across data center networks to eliminate or improve one or more defects existing in the prior art.

[0005] One aspect of the present invention provides a load balancing method across a data center network, the method being performed by a switching node in a first data center, comprising the following steps: The type of the data stream is determined based on the differential service code point markings carried in the data packets of the received data stream from the source node or other switching nodes; If the data flow is determined to be a cross-data center data flow, a state-aware dynamic binding mechanism is adopted. Based on the source node of the data flow, a corresponding link is selected for the data flow to forward the data flow to the first interconnect node via other switching nodes in the first data center in the order of data packets, or forwarded to the first interconnect node, and then forwarded to the second interconnect node in the second data center via the first interconnect node. This enables the second interconnect node to adopt a port virtualization mechanism and a state-aware dynamic binding mechanism, and introduces a conflict arbitration strategy driven by a network topology constraint table to select a corresponding path for the received data flow based on the destination node of the received data flow to forward the received data flow to the destination node.

[0006] In some embodiments of the present invention, the method further includes: If the data flow is determined to be an internal data flow within the data center, a greedy path selection strategy based on local queue state is adopted to select a corresponding link for each data packet of the data flow, so as to forward each data packet to other switching nodes of the first data center.

[0007] In some embodiments of the present invention, the queue state maintained at each port of each switching node in the data center includes the queue length and the cross-data center packet count; a greedy path selection strategy based on local queue state is adopted to select a corresponding link for each packet of the data stream, including: For each data packet in the data stream, the port with the shortest queue length and the corresponding link in the switching node are selected preferentially.

[0008] In some embodiments of the present invention, the queue state maintained at each port of each switching node in the data center includes the queue length and the cross-data center packet count; a state-aware dynamic binding mechanism is adopted to select the appropriate link for the data flow according to the source node of the data flow, including: Select the same port and corresponding link of the switching node for data streams originating from the same source node; When the cross-data center packet count of a port of a switching node remains zero for a preset time threshold, the port and the corresponding link are selected for the data flow. When all ports of the switching node are busy, a port and corresponding link are randomly selected for the data flow.

[0009] In some embodiments of the present invention, if the round-trip latency of the data stream is small, the corresponding differential service code point is marked as an intra-data center data stream; if the round-trip latency of the data stream is large, the corresponding differential service code point is marked as an inter-data center data stream.

[0010] Another aspect of the present invention provides a load balancing method across a data center network, the method being executed by a second interconnection node in a second data center, comprising the following steps: The system receives a data stream from a first interconnect node in a first data center. The data stream is obtained by a switching node in the first data center determining that the type of the data stream is a cross-data center data stream based on the differential service code point markers carried in the data packets of the data streams received from the source node or other switching nodes. Then, a state-aware dynamic binding mechanism is used to select the appropriate link for the data stream according to the source node of the data stream, so as to forward the data stream to the first interconnect node or forward it to the first interconnect node in the order of the data packets through other switching nodes in the first data center. It adopts a port virtualization mechanism and a state-aware dynamic binding mechanism, and introduces a conflict arbitration strategy driven by a network topology constraint table. It selects the appropriate path for the received data stream according to the destination node of the received data stream, so as to forward the received data stream to the destination node.

[0011] In some embodiments of the present invention, a port virtualization mechanism and a state-aware dynamic binding mechanism are employed, and a conflict arbitration strategy driven by a network topology constraint table is introduced to select the appropriate path for the received data stream based on the destination node of the received data stream, including: Based on the ratio between the link bandwidth between the second interconnect node and the switching node connected to the second interconnect node and the network card speed of the source node, each port of the second interconnect node is divided into multiple logical virtual ports. Select the same port of the second interconnect node for data streams received that are destined for the same destination node; When the cross-data center packet count of the port of the second interconnection node remains zero for a preset time threshold, the port is selected for the received data stream; When all ports of the second interconnect node are busy, a port is randomly selected for the received data stream; The location relationship of the destination node of the received data stream is determined by querying the network topology constraint table. A conflict arbitration strategy is then adopted to select the corresponding logical virtual port and corresponding path for the received data stream based on the location relationship of the destination node and the network topology constraint table. The network topology constraint table includes each node of the second data center and the links between each node.

[0012] In some embodiments of the present invention, a conflict arbitration strategy is employed to select appropriate logical virtual ports and paths for the received data stream based on the location relationships of the destination nodes and the network topology constraint table, including: When the destination node of the received data stream is located in the same Pod domain, different logical virtual ports are selected for the received data stream, and non-conflicting paths are selected according to the network topology constraint table. When the destination nodes of the received data streams are located in the same rack, logical virtual ports with non-intersecting paths are selected for the received data streams, and non-intersecting paths are selected according to the network topology constraint table.

[0013] Another aspect of the present invention provides a load balancing device across a data center network, the device comprising: a computer device including a processor and a memory, the memory storing computer instructions, the processor executing the computer instructions stored in the memory, and the device implementing the steps of the aforementioned method when the computer instructions are executed by the processor.

[0014] Another aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned method.

[0015] Another aspect of the present invention provides a computer program product including computer instructions that, when executed by a processor, implement the steps of the aforementioned method.

[0016] The load balancing method and apparatus for cross-data center networks of the present invention can identify the data flow type in the cross-data center network through a data flow classification method. When the data flow is identified as a cross-data center data flow, load balancing of cross-data center data flow scheduling can be achieved, thereby improving link utilization, reducing out-of-order packet overhead, and shortening flow completion time, thereby improving transmission performance and network resource utilization efficiency.

[0017] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the description and drawings.

[0018] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, are not intended to limit the scope of the invention. The components in the drawings are not drawn to scale but are merely illustrative of the principles of the invention. For ease of illustration and description of certain parts of the invention, corresponding portions in the drawings may be enlarged, i.e., may appear larger relative to other components in an exemplary device actually manufactured according to the invention. In the drawings: Figure 1 This is a schematic diagram of the overall workflow of a load balancing method across a data center network according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating a load balancing method across a data center network according to an embodiment of the present invention. Figure 3 This is a flowchart illustrating a load balancing method across a data center network according to another embodiment of the present invention. Figure 4 This is a flowchart illustrating a load balancing method across a data center network according to another embodiment of the present invention. Figure 5 This is a schematic diagram of a cross-data center network topology in one embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0021] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0022] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0023] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0024] In the following description, embodiments of the invention will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0025] To address the challenges posed by data flow characteristics, complex network topologies, and long propagation delays in cross-data center networks, this invention aims to improve link utilization, reduce out-of-order packet overhead, and shorten flow completion time in complex heterogeneous cross-data center network environments, thereby enhancing overall transmission performance and system stability. A load balancing method and apparatus for cross-data center networks are proposed. In achieving these objectives, this invention primarily needs to address the following two challenges: The first challenge lies in how to accurately identify and hierarchically schedule cross-data center data flows and intra-data center data flows, which is a crucial prerequisite for building a hierarchical load balancing mechanism. Since these two types of data flows differ significantly in path length, latency characteristics, and bandwidth constraints, the lack of an effective differentiation mechanism and a unified scheduling strategy for both types of data flows can easily lead to long-distance data flows consuming critical cross-domain link resources, or short-distance data flows experiencing unnecessary path detours in local networks, thus affecting overall transmission efficiency. Therefore, designing a low-overhead and accurate data flow category identification method in cross-data center networks and executing hierarchical scheduling accordingly is one of the core technical challenges in achieving efficient load balancing. The second challenge lies in how to achieve efficient and balanced utilization of multi-path bandwidth resources while ensuring transmission orderliness and stability. In cross-data center network scenarios, fine-grained scheduling will directly lead to uncontrollable deviations in the arrival order of cross-path data packets, thereby causing large-scale out-of-order problems. On the other hand, it is difficult to fully utilize multi-path bandwidth resources if coarse-grained or static scheduling strategies are adopted. Therefore, how to achieve an effective balance between out-of-order control and bandwidth utilization, and how to perform adaptive path selection based on network conditions, is a key issue affecting system performance and the second core technical challenge in achieving efficient load balancing.

[0026] To address the aforementioned challenges, the overall workflow of this invention is as follows: Figure 1As shown, firstly, a data flow awareness mechanism based on features such as RTT (Round-Trip Time) is designed to dynamically mark streaming data packets according to these features, thereby distinguishing between cross-data center data flows and intra-data center data flows. Secondly, within the data center, differentiated path selection is performed for different types of data flows based on local queue status and network congestion. Specifically, for the forwarding process of cross-data center data flows within the data center, a state-aware dynamic binding mechanism is adopted to maintain path stability while avoiding out-of-order issues caused by frequent path switching, thus reducing out-of-order overhead. Furthermore, at the cross-data center level, forwarding decisions (path selection and forwarding control) for cross-data center data flows are made by combining network topology constraints and path characteristics, thereby improving overall network performance while ensuring transmission stability and ultimately improving overall network resource utilization efficiency. For intra-data center data flows, a greedy path selection strategy based on local queue status is adopted to prioritize the use of low-latency, low-congestion paths, achieving efficient bandwidth utilization.

[0027] Figure 2 This is a flowchart of a load balancing method across a data center network according to an embodiment of the present invention, as follows: Figure 2 As shown, this method is executed by the switching node of the first data center and includes the following steps: Step S210: Determine the type of the data stream based on the differential service code point marker carried by the data packets of the received data stream from the source node or other switching nodes.

[0028] Current data centers exhibit a hybrid flow of AI-driven data flows and traditional business flows. The prerequisite for differentiated load balancing in this invention lies in the accurate distinction between data flows within a data center and those across data centers. Because intermediate switching devices in the network (such as switches, acting as switching nodes) lack complete contextual information, typically only end hosts and cross-data center switching devices (i.e., DCI switching devices, acting as interconnect nodes within data centers) possess data flow classification capabilities. Although cross-data center switching devices can identify cross-data center data flows through routing tables, because these devices are deployed at the network edge, the path selection within the data center has already been completed by the time the data flow arrives at that location. Relying on this device for data flow classification would require notifying upstream switching devices through complex reverse signaling mechanisms, which not only incurs significant control plane overhead but also introduces substantial latency.

[0029] In contrast, the source or sender (host or server, as the source node) naturally possesses the initial state information of the data stream to be transmitted and can effectively distinguish different types of data streams by utilizing the significant differences in characteristics such as RTT between cross-data center data streams and intra-data center data streams. Therefore, this invention decentralizes the data stream classification function to the sender, which determines whether the stream is an intra-data center data stream or a cross-data center data stream based on characteristics such as the round-trip time (RTT). Then, it uses the DSCP (Differentiated Services Code Point) field in the IP (Internet Protocol) header of the data packet to explicitly mark the data stream type, enabling all forwarding devices (switching nodes) and load balancing decision nodes (any switching node) along the path to identify the data stream type. Specifically, if the RTT of the data stream is small, the corresponding DSCP is marked as an intra-data center data stream; if the RTT of the data stream is large, the corresponding DSCP is marked as a cross-data center data stream. Based on this source-end labeling mechanism, the load balancing decision nodes within the data center can execute differentiated scheduling strategies without maintaining additional states, thereby achieving an efficient, lightweight, and commercially compatible data flow classification method while avoiding the introduction of additional network overhead.

[0030] Step S220: If the data flow is determined to be a cross-data center data flow, a state-aware dynamic binding mechanism is adopted. The corresponding link is selected for the data flow according to the source node of the data flow, so that the data flow is forwarded to the first interconnect node through other switching nodes of the first data center in the order of data packets, or forwarded to the first interconnect node, and then forwarded to the second interconnect node of the second data center through the first interconnect node. This enables the second interconnect node to adopt a port virtualization mechanism and a state-aware dynamic binding mechanism, and introduces a conflict arbitration strategy driven by a network topology constraint table. The corresponding path is selected for the received data flow according to the destination node of the received data flow, so that the received data flow is forwarded to the destination node.

[0031] The queue status metrics maintained at each port of each switching device acting as a switching node in a data center include queue length and cross-data center packet count, which can be used to support differentiated path forwarding processing for different types of data streams.

[0032] By introducing a state-aware dynamic binding mechanism and combining it with a network topology constraint table for path conflict avoidance, the out-of-order overhead of cross-data center data streams is effectively reduced and the utilization rate of cross-domain links is improved while ensuring the sequential transmission of cross-data center data streams, thereby improving the overall performance in long-distance transmission scenarios.

[0033] In some embodiments, a state-aware dynamic binding mechanism is employed to select a corresponding link for the data stream based on the source node of the data stream, including: Select the same port and corresponding link of the switching node for data streams originating from the same source node; When the cross-data center packet count of a port of a switching node remains zero for a preset time threshold, the port and the corresponding link are selected for the data flow. When all ports of the switching node are busy, a port and corresponding link are randomly selected for the data flow.

[0034] For cross-datacenter data flows, a state-aware dynamic binding mechanism is introduced. By classifying flows at the sender level, data flows from the same sender are consistently forwarded to the same egress port to avoid out-of-order packet transmission during long-distance cross-datacenter transmission. Under this mechanism, a port is considered idle and used for new flow allocation only when the cross-datacenter queue packet count on a certain port remains zero for a preset time threshold. When all ports are busy, random port selection is used as a fallback strategy to alleviate head-of-queue congestion while maintaining the orderly transmission of cross-datacenter data flows as much as possible.

[0035] To balance the high throughput requirements of data flow within a data center with the strict requirements of transmission order for data flow across data centers, this invention designs a scheduling logic decoupling strategy for different types of data flow in the load balancing module within the data center. This strategy includes steps S220 and S230.

[0036] In some embodiments, such as Figure 3 As shown, the method also includes the following steps: Step S230: If it is determined that the data flow is an internal data flow of the data center, a greedy path selection strategy based on the local queue state is adopted to select a corresponding link for each data packet of the data flow, so as to forward each data packet to other switching nodes of the first data center.

[0037] In some embodiments, a greedy path selection strategy based on local queue state is employed to select a corresponding link for each data packet of the data stream, including: For each data packet in the data stream, the port with the shortest queue length and the corresponding link in the switching node are selected preferentially.

[0038] For each data packet in the data flow within the data center, a path selection mechanism based on a greedy strategy is adopted. The port with the smallest queue length (if there are both data flows within the data center and data flows across the data center in the queue, the queue length is the sum of the number of data packets of the two types of data flows) is selected for forwarding. This forwards the data packet to the intermediate switching node within the data center or the switching node connected to its destination node, thereby achieving fine-grained data packet distribution and maximizing link utilization. This allows the data flow to make priority use of idle bandwidth resources, while avoiding congested paths that are occupied by data flows across the data center.

[0039] This invention employs a greedy path selection strategy based on queue length within the data center to dynamically select low-congestion paths for data flows, thereby achieving efficient utilization of multi-path bandwidth and avoiding uneven link utilization caused by excessively coarse flow-level scheduling granularity.

[0040] By introducing a hierarchical load balancing mechanism based on data flow classification, the scheduling of data flows within a data center and data flows across data centers is decoupled. This breaks the limitations of traditional unified scheduling strategies, which struggle to balance throughput and orderliness, enabling differentiated optimization and load balancing of different types of data flows, thereby improving overall network resource utilization efficiency. Furthermore, through this decoupling mechanism, efficient collaborative scheduling of different types of data flows within the same data center can be achieved.

[0041] Figure 4 This is a flowchart of a load balancing method across data center networks according to another embodiment of the present invention, such as... Figure 4 As shown, this method is executed by the second interconnect node of the second data center and includes the following steps: Step S410: Receive a data stream from a first interconnect node in a first data center. The data stream is obtained by a switching node in the first data center determining that the type of the data stream is a cross-data center data stream based on the differential service code point markers carried in the data packets of the data streams received from the source node or other switching nodes. Then, a state-aware dynamic binding mechanism is used to select the corresponding link for the data stream according to the source node of the data stream, so that the data stream is forwarded to the first interconnect node or forwarded to the first interconnect node in the order of the data packets through other switching nodes in the first data center. Step S420: A port virtualization mechanism and a state-aware dynamic binding mechanism are adopted, and a conflict arbitration strategy driven by a network topology constraint table is introduced. The appropriate path is selected for the received data stream according to the destination node of the received data stream, so as to forward the received data stream to the destination node.

[0042] In some embodiments, a port virtualization mechanism and a state-aware dynamic binding mechanism are employed, and a conflict arbitration strategy driven by a network topology constraint table is introduced to select the appropriate path for the received data stream based on the destination node of the received data stream, including: Based on the ratio between the link bandwidth between the second interconnect node and the switching node connected to the second interconnect node and the network card speed of the source node, each port of the second interconnect node is divided into multiple logical virtual ports. Select the same port of the second interconnect node for data streams received that are destined for the same destination node; When the cross-data center packet count of the port of the second interconnection node remains zero for a preset time threshold, the port is selected for the received data stream; When all ports of the second interconnect node are busy, a port is randomly selected for the received data stream; The location relationship of the destination node of the received data stream is determined by querying the network topology constraint table. A conflict arbitration strategy is then adopted to select the corresponding logical virtual port and corresponding path for the received data stream based on the location relationship of the destination node and the network topology constraint table. The network topology constraint table includes each node of the second data center and the links between each node.

[0043] In some embodiments, a conflict arbitration strategy is employed to select appropriate logical virtual ports and paths for the received data stream based on the location relationships of the destination nodes and the network topology constraint table, including: When the destination node of the received data stream is located in the same Pod domain, different logical virtual ports are selected for the received data stream, and non-conflicting paths are selected according to the network topology constraint table. When the destination nodes of the received data streams are located in the same rack, logical virtual ports with non-intersecting paths are selected for the received data streams, and non-intersecting paths are selected according to the network topology constraint table.

[0044] DCI switches are network devices specifically designed for data center interconnection scenarios. Their core function is to establish high-speed, low-latency, and highly reliable interconnection channels between geographically dispersed data centers, supporting critical services such as cross-domain computing power collaboration, data synchronization, and load balancing. DCI switches feature ultra-high bandwidth and high-density ports, supporting high-speed interfaces such as 400G / 200G / 100G to meet the extreme bandwidth requirements of AI training and large-scale model collaboration. Taking a data center network with a fat-tree topology as an example... Figure 5 As shown, the link bandwidth between the DCI switch and the core layer switch in the cross-data center link. Significantly higher than the speed of the host network card This can lead to bandwidth mismatch and resource waste. To address this issue, this invention introduces a port virtualization mechanism in the cross-data center load balancing module (DCI node). Specifically, each physical DCI egress port of the DCI switch is virtualized according to a ratio. It is divided into multiple logical virtual ports to enable finer-grained path selection and bandwidth allocation, reducing the waste of bandwidth resources.

[0045] Based on this, to alleviate the potential pseudo-congestion problem caused by network topology constraints, such as Figure 5 To address network path congestion in data center B, this invention categorizes received data streams by granularity, using either the receiving end or the destination end (as the destination node). Data streams destined for the same receiving end are consistently forwarded to the same egress port. Furthermore, the dynamic binding mechanism within the data center's load balancing module is extended by introducing a conflict arbitration strategy driven by a network topology constraint table. This strategy queries the network topology constraint table and allocates paths based on the location relationships of the destination ends of the received data streams and the network topology in the table. Figure 5 As shown in Pod2, when the destination of a newly received data stream is in the same Pod domain as the destination of a data stream with an already allocated path, a different logical virtual port (logical virtual egress port) is assigned to the newly received data stream compared to the data stream with the already allocated path; furthermore, as... Figure 5 As shown in Rack2, when the destination of the received data stream is located within the same rack (a rack is a switch that directly connects the source or destination node), a logical virtual port group with disjoint paths is selected for forwarding, thereby reducing link contention and congestion risks. The above table-query-based decision logic can effectively avoid path conflicts.

[0046] In summary, the load balancing method for cross-data center networks provided by this invention provides a decoupled hierarchical load balancing scheme. By decoupling the scheduling logic of data flows within a data center from those across data centers, different routing strategies are adopted for different data flows. This enables differentiated path selection and resource allocation for different types of data flows. Under the premise of ensuring the orderliness and stability of transmission, the utilization rate of multi-path bandwidth resources can be improved and the flow completion time can be reduced, thereby meeting the high-performance RDMA transmission requirements under complex cross-data center network topologies.

[0047] Corresponding to the above method, embodiments of the present invention also provide a load balancing device across a data center network. The device includes a computer device, which includes a processor and a memory. The memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device implements the steps of the aforementioned method.

[0048] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned method. The computer-readable storage medium may be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, floppy disk, hard disk, removable storage disk, CD-ROM, or any other form of storage medium known in the art.

[0049] This invention also provides a computer program product, including computer instructions that, when executed by a processor, implement the steps of the aforementioned method.

[0050] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0051] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0052] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0053] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A load balancing method across data center networks, characterized in that, The method is executed by a switching node in the first data center, and the method includes: The type of the data stream is determined based on the differential service code point markings carried in the data packets of the received data stream from the source node or other switching nodes; If the data flow is determined to be a cross-data center data flow, a state-aware dynamic binding mechanism is adopted. Based on the source node of the data flow, a corresponding link is selected for the data flow to forward the data flow to the first interconnect node via other switching nodes in the first data center in the order of data packets, or forwarded to the first interconnect node, and then forwarded to the second interconnect node in the second data center via the first interconnect node. This enables the second interconnect node to adopt a port virtualization mechanism and a state-aware dynamic binding mechanism, and introduces a conflict arbitration strategy driven by a network topology constraint table to select a corresponding path for the received data flow based on the destination node of the received data flow to forward the received data flow to the destination node.

2. The method according to claim 1, characterized in that, The method further includes: If the data flow is determined to be an internal data flow within the data center, a greedy path selection strategy based on local queue state is adopted to select a corresponding link for each data packet of the data flow, so as to forward each data packet to other switching nodes of the first data center.

3. The method according to claim 2, characterized in that, The queue status maintained at each port of each switching node in the data center includes the queue length and the cross-data center packet count; A greedy path selection strategy based on local queue state is adopted to select a corresponding link for each data packet in the data stream, including: For each data packet in the data stream, the port with the shortest queue length and the corresponding link in the switching node are selected preferentially.

4. The method according to claim 1, characterized in that, The queue status maintained at each port of each switching node in the data center includes the queue length and the cross-data center packet count; A state-aware dynamic binding mechanism is adopted to select the appropriate link for the data stream based on the source node of the data stream, including: Select the same port and corresponding link of the switching node for data streams originating from the same source node; When the cross-data center packet count of a port of a switching node remains zero for a preset time threshold, the port and the corresponding link are selected for the data flow. When all ports of the switching node are busy, a port and corresponding link are randomly selected for the data flow.

5. The method according to any one of claims 1 to 4, characterized in that, If the round-trip latency of the data stream is small, the corresponding differential service code point is marked as an intra-data center data stream; if the round-trip latency of the data stream is large, the corresponding differential service code point is marked as an inter-data center data stream.

6. A load balancing method across data center networks, characterized in that, The method is executed by a second interconnection node in the second data center, and the method includes: The system receives a data stream from a first interconnect node in a first data center. The data stream is obtained by a switching node in the first data center determining that the type of the data stream is a cross-data center data stream based on the differential service code point markers carried in the data packets of the data streams received from the source node or other switching nodes. Then, a state-aware dynamic binding mechanism is used to select the appropriate link for the data stream according to the source node of the data stream, so as to forward the data stream to the first interconnect node or forward it to the first interconnect node in the order of the data packets through other switching nodes in the first data center. It adopts a port virtualization mechanism and a state-aware dynamic binding mechanism, and introduces a conflict arbitration strategy driven by a network topology constraint table. It selects the appropriate path for the received data stream according to the destination node of the received data stream, so as to forward the received data stream to the destination node.

7. The method according to claim 6, characterized in that, It employs a port virtualization mechanism and a state-aware dynamic binding mechanism, and introduces a conflict arbitration strategy driven by a network topology constraint table. Based on the destination node of the received data stream, it selects the appropriate path for the received data stream, including: Based on the ratio between the link bandwidth between the second interconnect node and the switching node connected to the second interconnect node and the network card speed of the source node, each port of the second interconnect node is divided into multiple logical virtual ports. Select the same port of the second interconnect node for data streams received that are destined for the same destination node; When the cross-data center packet count of the port of the second interconnection node remains zero for a preset time threshold, the port is selected for the received data stream; When all ports of the second interconnect node are busy, a port is randomly selected for the received data stream; The location relationship of the destination node of the received data stream is determined by querying the network topology constraint table. A conflict arbitration strategy is then adopted to select the corresponding logical virtual port and corresponding path for the received data stream based on the location relationship of the destination node and the network topology constraint table. The network topology constraint table includes each node of the second data center and the links between each node.

8. The method according to claim 7, characterized in that, A conflict arbitration strategy is adopted to select the appropriate logical virtual port and path for the received data stream based on the location relationship of the destination node and the network topology constraint table, including: When the destination node of the received data stream is located in the same Pod domain, different logical virtual ports are selected for the received data stream, and non-conflicting paths are selected according to the network topology constraint table. When the destination nodes of the received data streams are located in the same rack, logical virtual ports with non-intersecting paths are selected for the received data streams, and non-intersecting paths are selected according to the network topology constraint table.

9. A load balancing device for a cross-data center network, comprising a processor, a memory, and computer instructions stored in the memory, characterized in that, The processor is configured to execute the computer instructions, and when the computer instructions are executed, the device implements the steps of the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method as described in any one of claims 1 to 8.

Images