A low-delay communication scheduling system for a computing power shelter

Abstract

This invention discloses a low-latency communication scheduling system for a computing power shelter, comprising a data parsing module, a delay conversion module, an offline pre-scheduling module, and an adaptive fine-tuning module. The data parsing module automatically extracts physical pipeline lengths, equipment coordinates, and topology information from the BIM model to generate a structured dataset. The delay conversion module combines physical transmission time and queuing processing time, introducing a dynamic media calibration factor and a congestion-aware weight adjustment mechanism to generate a composite delay weight matrix. The offline pre-scheduling module pre-calculates the globally optimal path based on multi-constraint QoS requirements and compiles it into a static forwarding table containing an MPLS label stack, which is then fixed to the switch. The adaptive fine-tuning module achieves rapid network readiness through a startup unit and detects topology deviations on-site through an incremental update unit, performing differential recalculation and atomic updates only on locally affected areas. This invention achieves low-latency communication assurance for the computing power shelter network throughout its entire lifecycle, from design and production to operation.

Description

Technical Field

[0001] This invention relates to a low-latency communication scheduling system for computing power container, belonging to the field of low-latency communication technology. Background Technology

[0002] A computing power container is a modular data center integrating computing, storage, and networking IT infrastructure. It is typically used in edge computing, emergency command, and field operations, characterized by rapid deployment, flexibility, and strong environmental adaptability. With the explosive growth in computing power demand, communication latency between computing nodes within the container has become a key bottleneck restricting overall performance. This is especially true in scenarios such as distributed training and real-time inference, where even microsecond-level latency fluctuations can impact user experience. Therefore, achieving efficient, stable, and low-latency communication within the computing power container's internal network has become a crucial issue that urgently needs to be addressed in this field.

[0003] Currently, network deployment in computing power shelters primarily employs a combination of on-site configuration and dynamic routing protocols. After the shelter is assembled, network engineers complete the cable connections on-site according to the design drawings. They then manually configure the switches via command line or network management platform, or enable dynamic routing protocols such as OSPF and IS-IS, which automatically discover neighbors, calculate routes, and achieve network-wide convergence. Some advanced solutions introduce SDN controllers, centrally distributing flow tables to achieve a degree of automated operation and maintenance. Additionally, some solutions attempt to plan the network topology during the 3D simulation design phase, but a significant gap remains between the design and actual deployment.

[0004] Existing technical solutions have drawbacks, including significant network startup latency. After the modular shelter is powered on, the dynamic routing protocol needs to undergo a series of processes such as neighbor discovery, link-state database synchronization, and SPF calculation. For a modular shelter network containing dozens of switches, the convergence time often reaches tens of seconds or even minutes, failing to meet the plug-and-play and ultra-fast readiness requirements of the computing power modular shelter. Furthermore, it has poor adaptability to on-site construction deviations. During actual cabling, cable lengths, connection relationships, and equipment identification may differ from the BIM design model. These deviations can cause the pre-designed network configuration to fail. While the dynamic routing protocol can adaptively discover new topologies, it triggers network-wide route recalculation, consuming significant computing resources and potentially causing service interruptions and latency jitter.

[0005] Existing technologies artificially separate the network design phase from the operational phase: physical topology information (such as actual cable length and precise device location) in the design phase cannot be transmitted to the routing decisions during runtime, causing path selection to rely solely on abstract metrics such as hop count, making it difficult to achieve the lowest physical latency. At the same time, the general design of dynamic routing protocols aims to cope with unknown topology changes, and their convergence mechanism is effective across the entire network, making it impossible to perform fine-grained processing for local changes. Therefore, in scenarios like makeshift hospitals where there is relatively high determinism but a small amount of deviation, it is like "using a sledgehammer to crack a nut," wasting resources and affecting stability. Summary of the Invention

[0006] The technical problem this invention aims to solve is that traditional computing power container network deployment relies on on-site manual configuration or dynamic routing protocol convergence, which suffers from large startup delays, poor adaptability to on-site construction deviations, high routing recalculation overhead, and the inability to inherit physical topology information from the design stage.

[0007] To solve the above-mentioned technical problems, this invention proposes a low-latency communication scheduling system for computing power container, comprising a data parsing module, a latency conversion module, an offline pre-scheduling module, and an adaptive fine-tuning module; The data parsing module is used to interface with the BIM model of the 3D simulation design software. Based on the layer markers, it automatically identifies and extracts the spatial path coordinate sequence of physical pipelines and the 3D spatial absolute coordinates of network devices in the weak current / network communication layer, and generates a structured dataset containing the actual physical length matrix of cables L, the equipment location information table D, and the design topology diagram G0. The delay conversion module is used to determine the cable's actual physical length matrix L, combined with the reference signal propagation rate U of the cable medium. ij base Calculate the physical transmission time T ij prop Simultaneously obtain the queuing time T of the switch. ij queue The propagation rate of the reference signal is corrected by introducing a dynamic medium calibration factor γij generated from the measured round-trip time (RTT) based on the feedback of the field lightweight detection protocol, and then according to formula W. ij =α(t)·T ij prop +β(t)·T ij queue Calculate the composite delay weights to form the composite delay weight matrix W; where the weight coefficients α(t) and β(t) are dynamically adjusted according to the real-time monitored network congestion status; The offline pre-scheduling module is used during the factory assembly stage to run a multi-constraint shortest path optimization algorithm in an offline environment based on a composite delay weight matrix W and a structured dataset. It pre-calculates the lowest-latency communication path between all computing nodes that satisfies QoS constraints regarding bandwidth and packet loss rate, and compiles the path into a static underlying forwarding table containing an MPLS label stack pre-encapsulation mechanism, resulting in a pre-scheduled routing table R. pre It is stored in the switch's non-volatile memory; where MPLS stands for Multiprotocol Label Switching. The adaptive fine-tuning module includes a startup unit and an incremental update unit; The startup unit is used to intercept the dynamic routing protocol startup process after the modular shelter is powered on, forcing the switch to directly load and apply the hardened pre-scheduled routing table R from non-volatile memory. pre To the hardware forwarding table, achieve millisecond-level ultra-fast network readiness based on physical design; Incremental update units are used to construct the actual physical topology map G in real time during field operation using a lightweight probe protocol. actual ; G actual By comparing the design topology diagram G0 layer by layer, missing links, newly added links, deviations in link attributes, and changes in device identifiers are detected, and the set of affected local nodes V is determined. affected ; with V affected The local subgraph G is formed by the nodes of the subgraph and its adjacent nodes. sub Only the composite delay weight is recalculated for the local subgraph and the shortest path algorithm is run to generate a local route update patch ΔR. The patch is then distributed using an atomic update mechanism that first writes to the backup table and then switches to the backup table at once, thus maintaining the pre-scheduled path of the unchanged nodes and extremely low latency communication.

[0008] The data parsing module is specifically used to read BIM model files through API or parsing tools and filter out layers marked as low-voltage / network communication. Identify the geometric entities of the communication cable, extract the spatial path coordinate sequence and calculate the actual physical length Lij, and output the cable length matrix L; Identify network device entities, extract their three-dimensional absolute coordinates Pi and pre-assigned network identifier information, and output a device information list Di; Establish a preliminary physical topology connection diagram G0 based on the device ports connected at both ends of the cable; The output is a structured dataset containing a cable length matrix L, a device information table D, and an initial topology graph G0.

[0009] As mentioned above, when the delay conversion module generates the composite delay weight matrix, it obtains the reference signal propagation rate Uijbase based on the cable medium parameter library, and combines it with the measured round-trip time RTT fed back by the lightweight physical link detection on site to generate a dynamic medium calibration factor γij, and updates Uij with the actual rate. Calculate the physical transmission time: Tijprop = Lij / Uij; Obtain the switch queue processing time Tijqueue; According to formula W ij =α(t)·T ij prop +β(t)·T ij queue Calculate the composite delay weights, where α(t) and β(t) are adjustable weight coefficients that are dynamically adjusted based on the real-time network congestion status. For non-directly connected node pairs, the weights are set to infinity, ultimately forming the composite delay weight matrix W.

[0010] The dynamic adjustment of the weighting coefficients α(t) and β(t) mentioned above is based on the monitoring of port utilization, queue depth, and packet loss rate. When the utilization rate of any port exceeds 70% for three consecutive samplings, or the queue depth exceeds 80% or the packet loss rate is greater than 0.1%, it is determined to be a link congestion. The system gradually increases the weight of β(t) and decreases the weight of α(t). After the congestion is eliminated, the weight coefficients are smoothly restored to the default value within a preset time.

[0011] As mentioned above, when calculating the lowest latency communication path, the offline pre-scheduling module, based on multi-constraint QoS requirements, marks the available bandwidth B according to the port bandwidth in the device information table. ij Based on the required bandwidth (breq) and maximum tolerable packet loss rate (lreq) specified by the service flow, link filtering is first performed to eliminate links that do not meet the bandwidth and packet loss rate requirements, forming a feasible subgraph. The Floyd-Warshall algorithm based on composite delay weight is then run on the feasible subgraph to obtain the lowest delay path that meets the QoS constraints. For critical business flows, a primary and backup path separation strategy is adopted: first calculate the primary path, then temporarily set the weight of all links on the primary path to infinity, recalculate the backup path, and ensure that the primary and backup paths do not intersect at the link layer or node layer; if they cannot be completely separated, the separation is relaxed to the SRLG level of shared risk link group.

[0012] The pre-scheduled routing table uses an MPLS label stack pre-encapsulation mechanism, specifically as follows: Assign a globally unique label to each destination node in the entire network, and forward each hop on the path according to the label; When generating the pre-scheduled routing table, the maximum label stack depth of the switch is obtained according to the switch model. For devices with limited depth, a label aggregation strategy is adopted to segment multi-hop paths. A single-layer label is used within each segment, and segments are connected through IP forwarding or VLAN conversion. For devices that do not support MPLS, the alternative is to enable VLANQinQ to simulate a tag stack using dual VLAN tags, or to use IP-in-IP tunnels for encapsulation and forwarding.

[0013] The startup unit is specifically used to intercept the automatic startup process of traditional dynamic routing protocols after the switch is powered on; directly read the pre-scheduled routing table Rpre from the non-volatile memory and load it into the hardware forwarding table of the forwarding engine; when the physical link status changes to Up, the corresponding forwarding table entry takes effect immediately, achieving millisecond-level network readiness; and send a network readiness signal to the management platform.

[0014] A low-latency communication scheduling method for a computing power container includes the following steps: S1. By connecting to the 3D simulation design software through the data parsing module, the pipeline routing data and network equipment information marked as weak current / network communication layers are automatically extracted to generate a structured dataset; S2. The delay conversion module calculates the physical transmission time based on the actual physical length of the cable and the propagation rate of the medium signal, and then weights and combines it with the queuing processing time of the switch to generate a composite delay weight matrix. S3. The shortest path optimization algorithm is run based on the composite delay weight matrix through the offline pre-scheduling module. The lowest latency communication path between all computing power nodes is pre-calculated, compiled into a static underlying forwarding table, and the pre-scheduled routing table is obtained. It is then solidified into the non-volatile memory of the switch during the factory assembly process. S4. After the modular shelter is powered on, the startup unit of the adaptive fine-tuning module forces the switch to directly load and apply the pre-scheduled routing table stored in the memory, thereby achieving ultra-fast network readiness. S5. The incremental update unit of the adaptive fine-tuning module performs lightweight physical link detection on site. When a deviation from the design model is found, differential route recalculation is initiated only for the affected local nodes, generating local route update patches and distributing them to maintain extremely low latency communication for unchanged nodes.

[0015] The incremental update unit in step S5 described above specifically performs the following sub-steps: S51. Start the LLDP or BFD lightweight probe protocol in the background to collect neighbor information at second intervals and build the actual physical topology graph Gactual. S52. Compare the actual topology Gactual with the designed topology G0, detect missing links, added links, link attribute deviations and device identifier changes, and determine the affected local node set Vaffected. S53. Using Vaffected and its adjacent nodes to form a local subgraph Gsub, recalculate the composite delay weights within the subgraph, run Dijkstra's algorithm to update the optimal path, and generate a local route update patch ΔR, where the path from the subgraph boundary node to the external node remains unchanged from the original pre-scheduling table. S54. The ΔR is sent to the affected switches through the management channel, and an atomic update mechanism is adopted to write to the backup table first and then switch once to ensure that the update process does not cause packet loss or routing black holes. S55. Record the detected deviation information and report it to the operation and maintenance center for design optimization.

[0016] The offline pre-scheduling module generating the pre-scheduling routing table in step S3, as described above, specifically includes the following sub-steps: S31: Input the composite delay weight matrix W and the device information table D. Based on the available bandwidth Bij marked by the port bandwidth in the device information table, and combined with the required bandwidth breq and the maximum tolerable packet loss rate lreq specified by the service flow, filter the links and remove links that do not meet the bandwidth and packet loss rate requirements to form a feasible subgraph. S32: Run the Floyd-Warshall algorithm based on composite delay weights on the feasible subgraph to obtain the lowest delay path that satisfies QoS constraints; S33: For critical business flows, adopt a primary and backup path separation strategy: first calculate the primary path, then temporarily set the weight of all links on the primary path to infinity, recalculate the backup path, and ensure that the primary and backup paths do not intersect at the link layer or node layer; if they cannot be completely separated, relax the separation to the SRLG level of shared risk link group. S34: Convert each shortest path into a static forwarding table entry on the switch, use the MPLS label stack pre-encapsulation mechanism to assign a globally unique label to each destination node in the entire network, and select the label encapsulation format or alternative scheme according to the switch model compatibility. S35: Group the generated forwarding table entries by device to form a pre-scheduled routing table Rpre, which is then stored in the switch's non-volatile memory during the factory assembly process.

[0017] This invention has positive effects: (1) The present invention introduces a composite delay weight matrix through a delay conversion module, which comprehensively considers physical transmission time and queuing processing time, and utilizes dynamic medium calibration factor and congestion-aware weight adjustment mechanism to make path selection closer to the real physical delay and real-time network status, significantly improving the accuracy and adaptability of low-latency communication.

[0018] (2) The present invention uses the incremental update unit of the adaptive fine-tuning module to perform differential recalculation and atomic update only on the local area of ​​the on-site topology deviation, which greatly reduces the on-site computing overhead and routing oscillation risk. While ensuring network stability, it effectively absorbs construction deviation and maintains the extremely low latency communication of the core computing nodes.

[0019] (3) The present invention pre-calculates and solidifies the global optimal path in the factory stage through the offline pre-scheduling module, and achieves millisecond-level network readiness in conjunction with the startup unit, completely getting rid of the dependence on the convergence of the on-site dynamic routing protocol, and greatly shortening the time from powering on the container to service availability. Attached Figure Description

[0020] The invention will now be further described with reference to the accompanying drawings.

[0021] Figure 1 This is a schematic diagram of the modules of the present invention; Figure 2 This is a schematic diagram of the module operation process of the present invention. Detailed Implementation

[0022] Example 1 See Figure 1 and Figure 2 This embodiment 1 includes a data parsing module, a delay conversion module, an offline pre-scheduling module, and an adaptive fine-tuning module; The data parsing module is used to interface with 3D simulation design software to automatically extract pipeline routing data marked as "weak current / network communication" layer, including the actual physical wiring length of communication cables inside the cabin and when connecting across cabins. At the same time, it extracts the 3D spatial absolute coordinates of network devices inside the cabin and the network identification information pre-assigned during the design, providing basic data for subsequent delay calculations.

[0023] The delay conversion module is used to calculate the physical transmission time between any two network nodes based on the extracted actual physical wiring length of the cable and the inherent signal propagation rate of the cable medium. It then weights and combines this physical transmission time with the theoretical queuing processing time of the switch node to generate a composite delay weight matrix oriented towards the lowest actual physical delay, replacing the traditional evaluation standard based on hop count.

[0024] When generating composite delay weights, the weight coefficients α(t) and β(t) are adjusted based on real-time monitoring of network congestion status. Specifically, the system periodically collects the following metrics for each switch port via SNMP or Telemetry: port utilization (usually calculated as a 5-second average utilization), output queue depth (the ratio of the current queue length to the maximum queue length), and packet loss rate (the number of packets lost per unit time). When the utilization of any port exceeds the 70% threshold for three consecutive samples, or the queue depth exceeds 80%, or the packet loss rate is greater than 0.1%, the link is determined to be in a congested state. At this time, the system automatically increases the weight of β(t) (e.g., gradually increasing it from the default 0.3 to 0.7 in steps of 0.1), while decreasing the weight of α(t), making path selection more inclined towards links with lower queuing delays, thus achieving congestion avoidance. After congestion is eliminated, the weight coefficients smoothly recover to their default values ​​within 5 minutes. This dynamic adjustment mechanism is decoupled from changes in the overall network topology, affecting only the weight values ​​of local links and avoiding routing oscillations.

[0025] The default values ​​for α and β can be preset according to the business scenario. For latency-sensitive control services (such as industrial automation), α=0.8 and β=0.2 are preset, emphasizing the dominance of physical transmission delay; for high-bandwidth data services (such as video transmission), α=0.4 and β=0.6 are preset, focusing more on queuing processing time to avoid congestion. During the initialization phase, the system automatically assigns initial weight coefficients according to the service flow type marked at design time, and dynamically adjusts them according to the above rules during operation.

[0026] The offline pre-scheduling module is used to run the shortest path optimization logic based on the generated composite delay weight matrix, calculate in advance the global communication path with the lowest latency between all computing power nodes in the container, and compile these paths into a static underlying forwarding table, i.e., a pre-scheduling routing table, for the factory assembly process to solidify into the non-volatile memory of the switch.

[0027] When calculating the global path, in addition to composite delay weights, multiple QoS requirements such as bandwidth and packet loss rate must be met. The specific implementation is as follows: First, based on the port bandwidth capabilities recorded in the device information table D, the available bandwidth B_ij is labeled for each link. For service flows (such as critical control flows and video flows), the required bandwidth b_req and the maximum tolerable packet loss rate l_req have been specified during the design phase. Before running the shortest path algorithm, link filtering is performed: links with available bandwidth less than b_req and links with historical packet loss rates (obtained from the device information table or preset based on empirical values) exceeding l_req are removed. The remaining links form a feasible subgraph. Then, the Floyd-Warshall algorithm based on composite delay weights is run on the feasible subgraph to obtain the lowest delay path that meets QoS constraints. For critical service flows, a primary / backup path separation strategy is adopted: first, the primary path P_primary is calculated; then, the weights of all links on the primary path are temporarily set to infinity, and the backup path P_backup is recalculated to ensure that the primary and backup paths do not intersect at the link layer (or node layer), avoiding a single failure from affecting both primary and backup simultaneously. If a completely disjoint backup path cannot be found, the constraints are relaxed to the Shared Risk Link Group (SRLG) level separation, meaning that primary and backup paths are prevented from passing through the same physical pipe or the same switch board. The calculated primary and backup path sets are stored together in the pre-scheduled routing table.

[0028] The Shared Risk Link Group (SRLG) is mainly used to ensure that the primary and backup paths will not be interrupted simultaneously due to the same point of failure (such as the same fiber optic cable, the same physical conduit, or the same switch board) when calculating the primary and backup paths for critical business flows.

[0029] To achieve extremely low forwarding latency, MPLS label stack pre-encapsulation is used.

[0030] MPLS is primarily used for traditional IP route lookup, which is based on the longest match lookup of the destination IP address. This process is slow at the software level and cannot flexibly specify a complete, calculated, physically optimal path.

[0031] Label allocation follows a per-hop label strategy based on the destination node: each destination node in the entire network is assigned a globally unique label (e.g., a 20-bit label space, supporting up to approximately 1 million nodes, far exceeding the number of nodes inside the shelter), and each hop on the path is forwarded based on the label. Label space management adopts a centralized controller approach, with unified allocation by a central orchestrator during the offline pre-scheduling phase to ensure no conflicts. For situations where some older switches may not support MPLS or have limited label stack depth (e.g., only supporting two label layers), two alternative solutions are designed: 1. VLAN tag replacement: For devices that do not support MPLS, use 802.1Q VLAN ID (12 bits) for path identification, and implement a tag stack function through multi-level VLAN nesting (the device needs to support QinQ).

[0032] 2. IP Tunnel Encapsulation: If the device only supports IP forwarding, GRE or IP-in-IP tunneling is used to encapsulate the original packet before forwarding. The tunnel endpoint is the next hop in the path. During pre-scheduling route table generation, the system automatically selects the optimal encapsulation method based on the switch model and writes the corresponding configuration into the table entry. Before programming, the system automatically matches a compatible encapsulation format using the model field in the device information table.

[0033] The adaptive fine-tuning module includes a startup unit and an incremental update unit. The startup unit is used to force the underlying network hardware to skip the neighbor discovery and network topology convergence waiting period of traditional dynamic routing protocols after the mobile cabin is in place and powered on, and directly load and apply the pre-scheduled routing table fixed in memory to achieve millisecond-level ultra-fast network readiness. The incremental update unit is used to start lightweight physical link detection in the system background to obtain the actual connectivity status on site. If a deviation is found from the BIM design model, differential route recalculation is initiated only for the local network nodes that have the deviation, generating local route update patches and distributing them to ensure that the core computing power nodes that have not been changed maintain their original extremely low latency communication state.

[0034] The automatic extraction in the above data parsing module is as follows: Input the original model files from BIM and other 3D simulation design software, such as .ifc and .rvt formats, including architectural, structural, mechanical and electrical information.

[0035] First, the model file is read using API and parsing tools. All layers are traversed, and specific layers marked "weak current / network communication" are selected to exclude irrelevant professional data.

[0036] Next, in the filtered layers, the geometric entities representing communication cables are identified, their spatial path coordinate sequences are extracted, and the actual physical length L along the path is calculated. ij The output is a cable length matrix L, where the elements L... ij This represents the length of the cable between node i and node j. If there is no direct physical connection between the nodes, it is marked as infinity.

[0037] Then, identify network device entities such as switches, servers, and routers in the layer and extract their three-dimensional absolute coordinates P. i = (x i y i , z iIt reads the network identification information pre-assigned during the design phase, including: IP address and subnet mask, MAC address, device name, port number, and VLAN segmentation information, and outputs the network identification information as a device information list D. i ={ID i P i MAC i IP i VLAN i , ...}.

[0038] Next, based on the device ports connected at both ends of the cable, a preliminary physical topology diagram G0=(V,E0) is established, where vertex V represents all network device nodes, edge E0 represents direct physical connections, and attribute L is attached. ij .

[0039] Finally, a structured dataset is output, which includes a cable length matrix L, a device information table D, and an initial topology graph G0.

[0040] The specific details of the composite delay weight matrix are as follows: First, input a structured dataset and a cable media parameter library, which includes a pre-stored table of media types and signal propagation rates. Based on the cable type designation in the design drawings, obtain the corresponding reference signal propagation rate U from the media parameter library. ij base Based on this, and combined with the measured round-trip time (RTT) data from the detection feedback of medium-lightweight physical links, the actual propagation rate of the cables in the field is deduced, and a dynamic medium calibration factor γ is generated. ij If the deviation between the actual rate and the reference rate exceeds a preset threshold, then the U value will be updated based on the actual rate. ij If no actual measured data is available or the deviation is within the allowable range, the reference rate will be used to ensure that the calculation of physical transmission time can truly reflect the aging state of the cable and the differences in construction process, and to avoid delay estimation errors caused by the discrepancy between theoretical and actual values.

[0041] For any pair of nodes (i,j) with a direct physical connection, calculate the physical transmission time T. ij prop =L ij / U ij .

[0042] Based on the switch device model and performance parameters, obtain the theoretical queuing time T. ij queue Queue processing time is related to device processing capacity, port speed, and expected load, and is estimated based on the port forwarding latency provided by the device manufacturer. If precise data is unavailable, a constant t is set.q This represents a typical queuing delay.

[0043] The physical transmission time and queuing processing time are weighted together to generate a composite delay weight w. ij For directly connected node pairs, the weight calculation formula is: W ij =α(t)⋅T ij prop +β(t)⋅T ij queue α(t) and β(t) are adjustable weighting coefficients that are dynamically adjusted according to the application scenario and the real-time network congestion status. When the overall network load is low, the proportion of α is appropriately increased; when local links become congested, the system automatically increases the weight of β, enabling the weight matrix to dynamically avoid congested links and achieve dual optimization of load balancing and low latency. For non-directly connected node pairs, the initial weights are set to infinity. The weights of all directly connected nodes are used to construct a weight matrix W of size n×n, where n is the number of nodes.

[0044] Finally, the composite delay weight matrix W is output.

[0045] The pre-scheduled routing table is as follows: Input the composite delay weight matrix W and the device information table D from the structured dataset.

[0046] Based on the number of nodes within the shelter and business requirements, a multi-constraint path optimization algorithm is selected. Using the input composite delay weight matrix W, and combining device performance parameters (such as port bandwidth) and business flow type, an improved Bellman-Ford-Moore algorithm is employed to ensure the lowest possible latency while simultaneously meeting QoS requirements such as bandwidth reservation and packet loss rate limits. For critical business flows, two paths, primary and backup, are automatically calculated to form a primary / backup path set.

[0047] Given a weight matrix W, run the Floyd-Warshall algorithm to iteratively update the minimum cumulative weight between any pair of nodes and record the corresponding next-hop information. This yields the minimum composite delay and the complete path between all pairs of nodes.

[0048] The Floyd-Warshall algorithm is a dynamic programming algorithm for finding the shortest path from all sources. It is suitable for dense graphs with a limited number of nodes and is used to calculate the lowest latency path between all pairs of computing power nodes based on the composite delay weight matrix W. Specifically, it involves inputting the set of nodes and the composite delay weight matrix W, then initializing and constructing the distance matrix and the path matrix, iteratively updating the distance between all node pairs (i,j) for each intermediate node, and finally obtaining the final distance matrix and path matrix after considering all intermediate nodes. This final distance matrix represents the lowest composite delay from node i to node j. The complete path node sequence can be obtained by recursively tracing the matrix.

[0049] When converting each shortest path into a static forwarding table entry on the switch, an MPLS label stack pre-encapsulation mechanism is introduced. For a path from node i to node j, a fixed label value is pre-assigned to each hop on the path, and the label stack is encapsulated in the forwarding table entry. For example, for the path i→a→b→j, on node i, the outgoing port is encapsulated with a label. a On node a, based on the label a Swap to Label b It also forwards the data, further reducing node processing latency.

[0050] Considering that different switch vendors may support varying MPLS label stack depths (commonly 2-3 layers), the central orchestrator queries the switch model field in the device information table to obtain its maximum label stack depth during pre-scheduled routing table generation. For devices with strict depth limitations, a label aggregation strategy is automatically adopted: multi-hop paths are segmented, each segment uses a single-layer label, and segments are connected via IP forwarding or VLAN translation. If the device does not support MPLS at all, an alternative solution is enabled: 1. VLAN QinQ: Uses a dual-layer VLAN tag to simulate a tag stack. The outer VLAN identifies the path, and the inner VLAN identifies the destination node. The device must support 802.1ad.

[0051] 2. IP-in-IP Tunnel: The original packet is encapsulated at the source node, with the destination address being the next-hop node. Upon arrival, the packet is decapsulated and forwarded. During the programming phase, the optimal encapsulation format is automatically selected based on the device model, and the corresponding configuration is generated.

[0052] Group these entries by device to generate a pre-scheduled routing table R. pre ={R pre switch1 R pre switch2 Each entry includes the destination node identifier, tag, next hop, outgoing port, and VLAN.

[0053] Then, on the equipment assembly line, the MES system retrieves the currently assembled modular shelter number and corresponding design data. Simultaneously, it retrieves the 3D model of the shelter and the simulation verification report of the pre-scheduled routing table from the digital twin platform. Before programming, the switch serial number is automatically compared with the predefined device ID in the model to ensure a one-to-one correspondence between the physical device and the digital twin. If the match is successful, the R... pre The version number, burning time, and operator information are recorded in the blockchain evidence storage system.

[0054] Perform the burning process as follows: 1. Connect the switch to the programming tool via a debugging interface, such as the Console or Ethernet port.

[0055] 2. The programming tool automatically identifies the switch model and sets the forwarding table entry R belonging to that switch. pre switch Convert to a configuration format supported by this model.

[0056] 3. Write to the switch's non-volatile memory, employing a dual-backup redundant storage strategy and adding a cyclic redundancy check code.

[0057] 4. Perform verification, read the written content and compare it with the original data to ensure that there are no errors. If the data in the main storage area is damaged, the startup unit will automatically restore from the backup area.

[0058] Finally, the switch with the pre-scheduled routing table is shipped with the modular housing.

[0059] The specific startup unit in the adaptive fine-tuning module is as follows: First, power on the switch to perform a hardware self-test, and then boot the operating system.

[0060] During the protocol stack initialization phase, the startup unit intercepts the automatic startup process of traditional dynamic routing protocols to prevent them from performing neighbor discovery and link state flooding. This can be achieved by modifying the startup script or disabling related services at the system service layer.

[0061] The system reads the pre-scheduled routing table R from non-volatile memory. pre This data is directly loaded into the hardware forwarding table of the forwarding engine. The physical link status is checked (Link Up), and once the port is up, the corresponding forwarding table entry takes effect. There's no need to wait for routing protocol convergence; the entire network enters a communicable state within milliseconds after the port stabilizes.

[0062] The switch sends a "network ready" signal to the management platform, indicating that it has started up as designed and the network has entered the operating state based on the pre-scheduled routing table.

[0063] The incremental update unit in the adaptive fine-tuning module is as follows: In the field link deviation detection and local adaptive update, the actual physical topology is input, topology G0 is designed, and pre-scheduled routing table R is prepared. pre , Composite Delay Weight Generation Logic.

[0064] 1. First, perform physical link detection. The background process initiates either the Link Layer Discovery Protocol (LLDP) or a lightweight probe protocol like Bidirectional Forwarding Detection (BFD) to periodically collect neighbor information and construct the actual physical topology graph G. actual The detection frequency is set to the second level to balance load and real-time performance.

[0065] The detection frequency is dynamically adjusted based on network size and equipment performance. For a typical modular deployment (≤100 nodes), the LLDP message sending interval is set to 5 seconds, and the BFD detection interval is set to 1 second, ensuring second-level perception of topology changes while minimizing CPU load impact (tested with 100 switches, single detection CPU utilization is <1%). Conditions triggering incremental updates include: 1. Missing Link: A link that exists in the design topology but is not actually found is considered missing if it does not appear for 3 consecutive probe cycles; 2. Add Link: If a link that does not exist in the design topology appears in the actual topology, it will be triggered immediately; 3. Link attribute deviation: The RTT measured by BFD is compared with the propagation time T_prop converted from the design length. If the deviation between RTT / 2 and T_prop exceeds 20% or the absolute value exceeds 50μs (take the larger value), it is judged as an abnormal delay and a recalculation is triggered. 4. Device Identifier Change: If the device identifier (such as system name, port description) carried by LLDP does not match the design, a recalculation is triggered.

[0066] Simultaneously monitor port status, error count, etc., to determine link quality. 2. Then perform deviation detection: The actual topology G actual Compare with the designed topology G0 and identify the differences: The missing link is one that is designed to exist but is not actually connected; it may be due to a port being down or not being wired. The newly added link was not designed but is actually connected, which may be due to a construction error; Changes in link attributes, such as abnormal cable length causing delay deviations, are addressed by estimating the actual length through round-trip time (RTT) and comparing it with the design length. If the difference exceeds a threshold, it is considered a deviation.

[0067] At the same time, check whether the device identification, such as IP and MAC, is consistent with the design. Any changes are also considered deviations.

[0068] Determine the set of affected local nodes V affected K-hop limits the range of nodes directly connected to the deviation link and the surrounding area that may be affected by the routing.

[0069] 3. Next, perform differential route recalculation: Extracting local subgraphs: using V affected The subgraph G is formed by the subgraph and its adjacent nodes. sub This includes all links connecting these nodes, as well as links that have not changed. For node pairs within the subgraph, the composite delay weights of S2 are recalculated to generate a local weight matrix Wsub. The weight calculation formula remains the same: W ij =α(t)⋅T ij prop +β(t)⋅T ij queue ; Run the shortest path algorithm Dijkstra on the subgraph, starting from each affected node, and update the optimal path between nodes in the subgraph. During computation, the paths from subgraph boundary nodes (i.e., nodes connected to external nodes) to external nodes remain unchanged from the original pre-scheduled table as constraints.

[0070] Generate a local route update patch ΔR: This only includes forwarding table entries that need to be modified due to topology changes, including newly added, deleted, and modified entries.

[0071] 4. Patch distribution and atomic updates: The ΔR is sent to the affected switches via the management channel.

[0072] An atomic update mechanism is employed, which first writes to the backup table and then performs a one-time switch to ensure that no packet loss or routing black holes occur during the update process. After the update is complete, the network returns to normal operation.

[0073] 5. Finally, provide data feedback: Detected deviations are recorded and reported to the operation and maintenance center for subsequent design optimization. If the deviation involves design flaws, construction rectification suggestions can be generated. Throughout the process, the network maintains low-latency communication matching the actual situation, and the update scope is minimized. Although this solution highly relies on the accuracy of the BIM model, construction deviations can be effectively absorbed through on-site detection and partial recalculation of incremental update units. For equipment or links not modeled in the BIM model, the incremental update unit will automatically include them in the partial recalculation after discovering new links, generating corresponding routing table entries and distributing them, achieving the capability of "design not covered, on-site adaptive".

[0074] Obviously, the above embodiments are merely examples to clearly illustrate the embodiments of the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively list all embodiments here. However, these obvious variations or modifications derived from the spirit of the present invention are still within the protection scope of the present invention.

Claims

1. A low-latency communication scheduling system for a computing power container, characterized in that, It includes a data parsing module, a delayed conversion module, an offline pre-scheduling module, and an adaptive fine-tuning module; The data parsing module is used to interface with the BIM model of the 3D simulation design software. Based on the layer markers, it automatically identifies and extracts the spatial path coordinate sequence of physical pipelines and the 3D spatial absolute coordinates of network devices in the weak current / network communication layer, and generates a structured dataset containing the actual physical length matrix L of cables, the device location information table D, and the design topology diagram G0. The delay conversion module is used to determine the cable's actual physical length matrix L, combined with the cable medium's reference signal propagation rate U. ij base Calculate the physical transmission time T ij prop Simultaneously obtain the queuing time T of the switch. ij queue ; By introducing a dynamic medium calibration factor γ generated based on the measured round-trip time (RTT) fed back by a field lightweight detection protocol. ij The propagation rate of the reference signal is corrected according to formula W. ij =α(t)·T ij prop +β(t)·T ij queue Calculate the composite delay weights to form the composite delay weight matrix W; where the weight coefficients α(t) and β(t) are dynamically adjusted according to the real-time monitored network congestion status; The offline pre-scheduling module is used during the factory assembly stage to run a multi-constraint shortest path optimization algorithm in an offline environment based on the composite delay weight matrix W and the structured dataset. It pre-calculates the lowest-latency communication path between all computing nodes that satisfies QoS constraints regarding bandwidth and packet loss rate, and compiles the path into a static underlying forwarding table containing an MPLS label stack pre-encapsulation mechanism, thus obtaining a pre-scheduled routing table R. pre It is stored in the switch's non-volatile memory; where MPLS stands for Multiprotocol Label Switching. The adaptive fine-tuning module includes a startup unit and an incremental update unit; The startup unit is used to intercept the dynamic routing protocol startup process after the modular shelter is powered on, forcing the switch to directly load and apply the hardened pre-scheduled routing table R from the non-volatile memory. pre To the hardware forwarding table, achieve millisecond-level ultra-fast network readiness based on physical design; The incremental update unit is used to construct the actual physical topology map G in real time during field operation using a lightweight detection protocol. actual ; G actual By comparing the design topology diagram G0 layer by layer, missing links, newly added links, deviations in link attributes, and changes in device identifiers are detected, and the set of affected local nodes V is determined. affected ; with V affected The local subgraph G is formed by the nodes of the subgraph and its adjacent nodes. sub Only the composite delay weight is recalculated for the local subgraph and the shortest path algorithm is run to generate a local route update patch ΔR. The patch is then distributed using an atomic update mechanism that first writes to the backup table and then switches to the backup table at once, thus maintaining the pre-scheduled path of the unchanged nodes and extremely low latency communication.

2. The low-latency communication scheduling system for computing power container according to claim 1, characterized in that, The data parsing module is specifically used to read BIM model files through API or parsing tools and filter out layers marked as low-voltage / network communication. Identify the geometric entities of the communication cable, extract the spatial path coordinate sequence, and calculate the actual physical length L. ij Output cable length matrix L; Identify network device entities, extract their three-dimensional absolute coordinates Pi and pre-assigned network identifier information, and output a device information list D. i ; Establish a preliminary physical topology connection diagram G0 based on the device ports connected at both ends of the cable; The output is a structured dataset containing a cable length matrix L, a device information table D, and an initial topology graph G0.

3. The low-latency communication scheduling system for computing power container according to claim 2, characterized in that, When the delay conversion module generates the composite delay weight matrix, it includes obtaining the reference signal propagation rate U based on the cable medium parameter library. ij base Combined with the measured round-trip time (RTT) from the on-site lightweight physical link detection feedback, a dynamic medium calibration factor γ is generated. ij Update U at the actual rate ij ; Calculate the physical transmission time T ij prop =L ij / U ij ; Obtain the switch queuing time T ij queue ; According to formula W ij =α(t)·T ij prop +β(t)·T ij queue Calculate the composite delay weights, where α(t) and β(t) are adjustable weight coefficients that are dynamically adjusted based on the real-time network congestion status. For non-directly connected node pairs, the weights are set to infinity, ultimately forming the composite delay weight matrix W.

4. The low-latency communication scheduling system for computing power container according to claim 3, characterized in that, The dynamic adjustment of the weighting coefficients α(t) and β(t) is based on the monitoring of port utilization, queue depth, and packet loss rate. When the utilization rate of any port exceeds 70% for three consecutive samplings, or the queue depth exceeds 80% or the packet loss rate is greater than 0.1%, it is determined to be a link congestion. The system gradually increases the weight of β(t) and decreases the weight of α(t). After the congestion is eliminated, the weight coefficients are smoothly restored to the default value within a preset time.

5. The low-latency communication scheduling system for computing power container according to claim 4, characterized in that, When calculating the lowest latency communication path, the offline pre-scheduling module, based on multi-constraint QoS requirements, uses the available bandwidth Bij marked in the port bandwidth of the device information table, combined with the required bandwidth breq and the maximum tolerable packet loss rate lreq specified by the service flow, to first perform link filtering, eliminating links that do not meet the bandwidth and packet loss rate requirements, forming a feasible subgraph; then, it runs the Floyd-Warshall algorithm based on composite delay weights on the feasible subgraph to obtain the lowest latency path that meets the QoS constraints. For critical business flows, a primary and backup path separation strategy is adopted: first calculate the primary path, then temporarily set the weight of all links on the primary path to infinity, recalculate the backup path, and ensure that the primary and backup paths do not intersect at the link layer or node layer; if they cannot be completely separated, the separation is relaxed to the SRLG level of shared risk link group.

6. The low-latency communication scheduling system for computing power container according to claim 5, characterized in that, The pre-scheduled routing table uses an MPLS label stack pre-encapsulation mechanism, specifically: Assign a globally unique label to each destination node in the entire network, and forward each hop on the path according to the label; When generating the pre-scheduled routing table, the maximum label stack depth of the switch is obtained according to the switch model. For devices with limited depth, a label aggregation strategy is adopted to segment multi-hop paths. A single-layer label is used within each segment, and segments are connected through IP forwarding or VLAN conversion. For devices that do not support MPLS, the alternative is to enable VLANQinQ to simulate a tag stack using dual VLAN tags, or to use IP-in-IP tunnels for encapsulation and forwarding.

7. The low-latency communication scheduling system for computing power container according to claim 6, characterized in that, The startup unit is specifically used to intercept the automatic startup process of traditional dynamic routing protocols after the switch is powered on; and directly read the pre-scheduled routing table R from the non-volatile memory. pre It is loaded into the hardware forwarding table of the forwarding engine; when the physical link state changes to U p At that time, the corresponding forwarding entries take effect immediately, achieving millisecond-level network readiness; a network readiness signal is sent to the management platform.

8. A low-latency communication scheduling method for a computing power container in the system as described in claim 1, characterized in that, Includes the following steps: S1. By connecting to the 3D simulation design software through the data parsing module, the pipeline routing data and network equipment information marked as weak current / network communication layers are automatically extracted to generate a structured dataset; S2. The delay conversion module calculates the physical transmission time based on the actual physical length of the cable and the propagation rate of the medium signal, and then weights and combines it with the queuing processing time of the switch to generate a composite delay weight matrix. S3. The shortest path optimization algorithm is run based on the composite delay weight matrix through the offline pre-scheduling module. The lowest latency communication path between all computing power nodes is pre-calculated, compiled into a static underlying forwarding table, and the pre-scheduled routing table is obtained. It is then solidified into the non-volatile memory of the switch during the factory assembly process. S4. After the modular shelter is powered on, the startup unit of the adaptive fine-tuning module forces the switch to directly load and apply the pre-scheduled routing table stored in the memory, thereby achieving ultra-fast network readiness. S5. The incremental update unit of the adaptive fine-tuning module performs lightweight physical link detection on site. When a deviation from the design model is found, differential route recalculation is initiated only for the affected local nodes, generating local route update patches and distributing them to maintain extremely low latency communication for unchanged nodes.

9. The low-latency communication scheduling method for computing power container according to claim 8, characterized in that, The incremental update unit in step S5 specifically performs the following sub-steps: S51. Start the LLDP or BFD lightweight probe protocol in the background to collect neighbor information at second intervals and build the actual physical topology graph Gactual. S52. Compare the actual topology Gactual with the designed topology G0, detect missing links, added links, link attribute deviations and device identifier changes, and determine the affected local node set Vaffected. S53. Using Vaffected and its adjacent nodes to form a local subgraph Gsub, recalculate the composite delay weights within the subgraph, run Dijkstra's algorithm to update the optimal path, and generate a local route update patch ΔR, where the path from the subgraph boundary node to the external node remains unchanged from the original pre-scheduling table. S54. The ΔR is sent to the affected switches through the management channel, and an atomic update mechanism is adopted to write to the backup table first and then switch once to ensure that the update process does not cause packet loss or routing black holes. S55. Record the detected deviation information and report it to the operation and maintenance center for design optimization.

10. The low-latency communication scheduling method for computing power container according to claim 8, characterized in that, The offline pre-scheduling module generating the pre-scheduling routing table in step S3 specifically includes the following sub-steps: S31: Input the composite delay weight matrix W and the device information table D. Based on the available bandwidth Bij marked by the port bandwidth in the device information table, and combined with the required bandwidth breq and the maximum tolerable packet loss rate lreq specified by the service flow, filter the links and remove links that do not meet the bandwidth and packet loss rate requirements to form a feasible subgraph. S32: Run the Floyd-Warshall algorithm based on composite delay weights on the feasible subgraph to obtain the lowest delay path that satisfies QoS constraints; S33: For critical business flows, adopt a primary and backup path separation strategy: first calculate the primary path, then temporarily set the weight of all links on the primary path to infinity, recalculate the backup path, and ensure that the primary and backup paths do not intersect at the link layer or node layer; if they cannot be completely separated, relax the separation to the SRLG level of shared risk link group. S34: Convert each shortest path into a static forwarding table entry on the switch, use the MPLS label stack pre-encapsulation mechanism to assign a globally unique label to each destination node in the entire network, and select the label encapsulation format or alternative scheme according to the switch model compatibility. S35: Group the generated forwarding table entries by device to form the pre-scheduled routing table R. pre It is used to be permanently stored in the non-volatile memory of the switch during the final assembly process in the factory.

Images