A data routing and queue scheduling method and system for asymmetric information exchange networks

By using a shared memory switching structure and a two-level routing mapping mechanism, the routing complexity and bandwidth imbalance problems in asymmetric information exchange networks are solved, achieving non-blocking forwarding and preventing data loss, thereby improving data throughput.

CN123074997APending Publication Date: 2026-11-10CHINA ELECTRONIC TECH GRP CORP NO 38 RES INST
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611414194.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-11
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Asymmetric information exchange networks suffer from complex routing, unbalanced bandwidth, and the risk of data loss, especially when the number of input ports and output ports are not equal and their bandwidths are not equal, which existing technologies cannot effectively solve.

Method used

A shared memory switching architecture is adopted, and a two-level routing mapping mechanism is used for data frame buffer management. The first-level routing mapping maps the destination port number of the input port to the queue number, and matrix operations are used to realize the mapping from the queue to the output port. Shared memory is used for data caching, and the number of queues and scheduling algorithm are optimized by combining the sequential access characteristics of NVMe hard drives.

Benefits of technology

It enables non-blocking forwarding in asymmetric networks, simplifies routing rules, prevents data loss, and improves data throughput by smoothing out bandwidth imbalances through large-capacity caching.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123074997A_ABST
    Figure CN123074997A_ABST
Patent Text Reader

Abstract

This invention discloses a data routing and queue scheduling method and system for asymmetric information exchange networks. The method employs a shared memory switching architecture and manages data frames through output queuing. It includes: a first-level routing mapping: mapping the destination port number of each input port received data frame to a queue number, using a many-to-one mapping; writing the data frame into the corresponding logical queue address space in the shared memory according to the queue number; a second-level routing mapping: converting a queue selection row vector of length M into an output port selection row vector of length L through matrix operations, replacing the summation operation in the matrix operations with an OR operation, where M is the number of queues and L is the number of output ports; reading the data frame from the shared memory and distributing it to the corresponding output port according to the output port selection row vector. This invention solves problems such as input / output bandwidth mismatch, complex routing rules, and disk cache performance optimization.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of network communication technology, specifically to a data routing and queue scheduling method and system for asymmetric information exchange networks. Background Technology

[0002] If an asymmetric information exchange network exists, meaning the number of input ports is not equal to the number of output ports, and the total bandwidth of the input ports is not equal to the total bandwidth of the output ports, then a large-capacity buffer is needed to smooth out the bandwidth imbalance between input and output ports and prevent data loss. Appropriate data queuing and scheduling algorithms also need to be selected to ensure non-blocking forwarding of data frames. Using a direct mapping method from input ports to output ports would lead to complex routing rules. Summary of the Invention

[0003] The technical problem to be solved by this invention is the complexity of routing, bandwidth imbalance, and risk of data loss in asymmetric information exchange networks (where the number of ports and bandwidth are not equal).

[0004] The present invention solves the above-mentioned technical problems through the following technical means: A data routing and queue scheduling method for an asymmetric information exchange network, wherein the number of input ports and output ports of the information exchange network are unequal, a shared memory switching structure is adopted, and data frames are buffered and managed in an output queuing manner, including: First-level routing mapping: Maps the destination port number of the data frame received by each input port to the queue number, which is a many-to-one mapping; writes the data frame into the corresponding logical queue address space in the shared memory according to the queue number; Second-level routing mapping: The queue selection row vector of length M is converted into the output port selection row vector of length L through matrix operations. The summation operation in the matrix operation is replaced by the OR operation, where M is the number of queues and L is the number of output ports. Select the row vector based on the output port, read the data frame from the shared memory and distribute it to the corresponding output port.

[0005] Furthermore, the second-level routing mapping uses a mapping matrix M_CVT to achieve forward mapping. The size of the mapping matrix M_CVT is M×L, and the forward mapping calculation formula is: K1 = Q1 × M_CVT Where Q1 is the queue selection row vector, K1 is the output port selection row vector, each row of the mapping matrix M_CVT contains at most one non-zero element, and the summation operation in matrix multiplication is replaced by an OR operation.

[0006] Furthermore, it also includes reverse mapping: mapping the output port enable state row vector K2 to the queue output enable state row vector Q2, calculated as follows: Q2=K2×

[0007] in The transpose of the mapping matrix M_CVT is of size L×M. The summation operation in matrix multiplication is replaced by an OR operation. K2[m]=1 indicates that the m-th output port is allowed to output, and Q2[n]=1 indicates that the n-th queue is allowed to output.

[0008] Furthermore, the mapping rules for the second-level routing are pre-built in the software, and the mapping rules are not lost after the shared memory is powered off, so there is no need to rewrite them.

[0009] Furthermore, the rules for the first-level routing mapping are fixed and preset; for illegal destination port numbers, they are mapped to the exception queue.

[0010] Furthermore, it also includes a data multiplexing step: the data frame requests of each input port are detected by polling, and the cache write control module is interacted with through the request handshake interface. After the handshake is successful, the data frame is read from the front-end cache of the input port and pushed to the shared memory; the module that performs the data multiplexing does not have its own cache.

[0011] Furthermore, the shared storage is implemented using NVMe hard drives; by controlling the total number of queues, the read and write operations of the NVMe hard drives are mainly sequential access, thereby improving data throughput.

[0012] This invention also provides a data routing and queue scheduling system for an asymmetric information exchange network, wherein the number of input ports and the number of output ports in the information exchange network are unequal, including: Shared memory, allowing read and write operations on all input and output ports; The data multiplexing module is used to poll the data frames of each input port, perform the first-level routing mapping, map the destination port number of the data frame to the queue number, map it to a many-to-one mapping, and push the data frame carrying the queue number to the cache write control module. The cache write control module is used to write data frames into the corresponding logical queue address space in the shared memory according to the queue number; The cache read control module is used to read data frames from the shared memory in logical queue order; The data demultiplexing module is used to perform the second-level routing mapping. It converts the queue selection row vector of length M into the output port selection row vector of length L through matrix operations. The summation operation in the matrix operation is replaced with an OR operation, and the data frame is distributed to the corresponding output port according to the output port selection row vector.

[0013] Furthermore, the data tapping module includes: The vector mapping unit stores the mapping matrix M_CVT and performs forward and reverse mapping through matrix operations. Forward mapping converts the queue selection row vector Q1 into the output port selection row vector K1, and reverse mapping converts the output port enabled state row vector K2 into the queue output enabled state row vector Q2. The output port gating unit controls the gating of each output port protocol conversion module according to the row vector selected by the output port, and distributes the data frames read from the shared memory to the corresponding output ports.

[0014] Furthermore, the data multiplexing module is not equipped with a local cache; the data multiplexing module has request handshake interfaces with the front-end input port and the cache write control module respectively. After both handshakes are successful, the data frame is read from the front-end cache and pushed to the cache write control module.

[0015] Furthermore, the shared storage is an NVMe hard drive; the cache write control module and the cache read control module primarily access the NVMe hard drive in a sequential read / write manner.

[0016] Furthermore, the mapping matrix used for the second-level routing mapping in the data demultiplexing module is pre-frozen in the software. The mapping matrix is ​​not lost after the shared memory is powered off and does not need to be rewritten.

[0017] The advantages of this invention are: This invention employs a shared memory architecture, where all ports read from and write to the same large-capacity cache (such as an NVMe hard drive). This large capacity helps to mitigate bandwidth imbalances.

[0018] Single-level output queue: Each queue / multiple queues in the cache corresponds to one output port; routing is achieved through two levels of logical mapping, rather than two levels of physical caching. Specifically, the first level (input port mapping): maps the destination port number to the queue number; multiple input ports can be mapped to the same queue, simplifying routing rules. The second level (output port mapping): maps the queue number to the output port number through matrix operations (forward / reverse mapping), supporting flexible configuration of "one output port corresponding to multiple queues".

[0019] The two-level routing mapping mechanism is as follows: First, the queue number is determined based on the destination port number; then, the output port number is determined based on the queue number, and vector conversion is achieved through matrix operations. This invention does not limit the specific scheduling algorithm; its focus is on reducing the number of queues to adapt to the sequential read / write characteristics of NVMe. The scheduling process uses matrix mapping to select the output port. This invention does not emphasize service differentiation; its core objective is to achieve non-blocking forwarding, prevent data loss, and simplify routing. Optimizing the queue number and scheduling method to address the characteristic of NVMe hard drives being "fast for sequential read / write and slow for random read / write." Attached Figure Description

[0020] Figure 1 This is a schematic diagram illustrating the queuing principle of the method in this embodiment of the invention; Figure 2 This is a block diagram illustrating the principle of the data multiplexing module in an embodiment of the present invention. Figure 3 This is a block diagram illustrating the principle of the data tapping module in an embodiment of the present invention. Figure 4 This is a schematic diagram illustrating the principle of the 6-input to 4-output vector mapping in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] The steps of a data routing and queue scheduling method for an asymmetric information exchange network are as follows: I. Selection of Switching Network Structure like Figure 1 As shown, the switching structure selected in this embodiment is a shared memory structure. In the shared memory structure, all input and output ports can perform read and write operations on the shared mass memory. Data from each input port is written into the shared mass memory. The data storage addresses are stored in the form of logical queues, and the logical queues are scheduled. According to the queue scheduling results, the data frames are read from the shared mass memory to each output port.

[0023] The data multiplexing module completes the polling of the input port and the management of data frames, the buffer write control module moves the data frames to the shared mass storage, the buffer read control module reads the data frames from the shared mass storage, and the data demultiplexing module completes the queue scheduling, moving the data read from the shared mass storage to the corresponding output port.

[0024] Therefore, the throughput of a shared large-capacity memory architecture depends on the size of the shared cache region and the bandwidth of cache writes and reads.

[0025] II. Queuing Method Selection Based on the location of the buffer in the switching structure, the queuing methods of the switching structure can be divided into input queuing, output queuing, and combined input and output queuing.

[0026] The switching structure selected in this system is a shared memory structure. The most commonly used queuing method for shared memory structure is output queuing, and the shared memory structure can implement simple scheduling algorithms with higher cache bandwidth.

[0027] When using the output queuing method, each queue / multiple queues in the buffer corresponds to one output port. When a data frame is written to the buffer, it is written to the corresponding output queue according to the destination port. At the same time as writing, the data frames at the head of each output queue are read sequentially to form an output frame stream, which is then distributed to each output port via the data demultiplexing module.

[0028] A queue corresponds to a segment of address space in shared memory, and queue management is essentially address space management of shared memory. The more queues there are, the more complex the address space management of large-capacity memory becomes.

[0029] In practice, address spaces of varying sizes are allocated based on the data bandwidth of each queue, allowing for scalability of the queue size. This system uses NVMe hard drives to implement a large-capacity data cache. While hard drives offer maximum data throughput during sequential read / write operations, their read / write bandwidth is minimal during random access. Therefore, queue scheduling should aim to simplify the number of data queues and optimize hard drive read / write performance.

[0030] III. Data Multiplexing and First-Level Routing Mapping 1) Data multiplexing The data multiplexing module receives data frames from each input port and pushes them to a shared mass storage device via the buffer write control module. The data multiplexing module completes the first-level routing mapping, comprising queue number generation and multiplexed data selection. Its principle block diagram is shown below. Figure 2 As shown.

[0031] The data multiplexing module multiplexes the data from multiple input ports into a single data stream, and performs multiplexing processing on the multiple input ports according to the arbitrator data selection signal written into the back-end control logic.

[0032] The data multiplexing module itself does not have a cache and has no ability to store data. The data read must be pushed to the next stage. Therefore, in addition to having a request handshake interface with the front stage, the data multiplexing module also has a request handshake interface with the state machine of the write control module in the next stage. After the handshake is successful, the data multiplexing module reads the data from the cache of the front stage.

[0033] In addition to performing data combining, the data multiplexing module also performs the first-level routing mapping, mapping the destination port number of each input port data to a queue number. This mapping is many-to-one, meaning that data from multiple input ports may be mapped to the same queue.

[0034] 2) First-level route mapping The first-level routing mapping maps the destination port number to the queue number. The queue number is used for cache queue and address space management. The table below shows an example of the first-level routing mapping.

[0035] Table 1. Destination Port Number to Queue Number Mapping

[0036] The table above is an example of a mapping rule. The queue number can be obtained by combining the local station number, the destination station number, and the output port number. In actual use, the mapping rule can be flexibly set according to the actual situation.

[0037] The rules for the first-level routing mapping are fixed and preset. The destination port number is converted into a queue number according to the preset rules. These rules can be embedded in the software and do not require additional annotation.

[0038] IV. Data Splitting and Second-Level Routing Mapping 1) Data tapping The data demultiplexing module distributes the data frames read from the shared large-capacity buffer to the output port. The data demultiplexing module also completes the second-level routing mapping, namely the output port mapping, using the mapping matrix M_CVT to map the queue number to the output port number.

[0039] Data frames read from the shared large-capacity buffer are simultaneously connected to all output port protocol conversion modules. The selection of each protocol conversion module is controlled by the output port selection vector. Data despinning consists of two parts: vector mapping and output port selection. The principle block diagram is as follows: Figure 3 As shown. 2) Second-level route mapping After a data frame is read from the buffer, a second-level routing mapping is performed, mapping the data's queue number to the output port number and pushing it to the corresponding output port. The second-level routing mapping refers to the mapping from queue number to output port number; there is a fixed correspondence between output ports and physical output ports, and physical output ports are bound to hardware ports.

[0040] Table 2. Example of queue number to output port mapping

[0041] In the engineering implementation, the second-level routing mapping is achieved through vector transformation. The transformation rules can be calculated as transformation matrices and embedded in the software without additional annotation. This way, even if a large-capacity cache loses power, the transformation matrix will not be lost.

[0042] In implementation, the queue selection vector needs to be mapped to the output port selection vector, or the output port state vector needs to be mapped to the queue state vector. The first mapping is called the forward mapping and the second mapping is called the reverse mapping. Vector transformation is achieved through matrix operations.

[0043] For example, in this system, if there are 6 queues, let row vector Q1 represent the data queue selection vector, with a length of M (M=6), and Q1[n]=1 indicates selecting the nth queue for output; and if there are 4 output ports, let row vector K1 represent the output port's output enable state, with a length of L (L=4), and K1[m]=1 indicates selecting the mth output port for output. The forward mapping maps vector Q1 to vector K1, and its calculation method is as follows:

[0044] M_CVT is an M×L matrix used to map "selecting a queue" to "selecting an output port". Matrix multiplication is performed using an OR operation, and each row of the matrix can contain at most one 1, indicating that a queue will only output from one output port (there is no data broadcasting requirement in this system). The M_CVT matrix has a size of 18×14. This indicates that the nth queue outputs at the mth output port, and the M_CVT matrix is ​​based on the actual task input.

[0045] M_CVT is called the mapping matrix. The principle of mapping from a 6×1 vector to a 1×4 vector is as follows: Figure 4 As shown, Figure 4 The red circle indicates the position where the element of the mapping matrix is ​​equal to 1.

[0046] Correspondingly, the state vector of the output port needs to be converted into the queue output state vector, that is, a 1×4 vector needs to be converted into a 1×6 vector. This conversion is called reverse mapping, which is also achieved through matrix operations.

[0047] Let row vector K2 represent the output enable state of the output port. This vector has a length of L (L=4), and its elements are Boolean values. K2[m]=1 indicates that the m-th output port is enabled. Let row vector Q2 represent the output enable state of the data queue. This vector has a length of M (M=6), and Q2[n]=1 indicates that the n-th queue is enabled. The conversion from vector K2 to Q2 is calculated as follows:

[0048] in It is the transpose of the M_CVT matrix with size L×M, which maps the state of "output port is allowed to output" to "data from a queue is allowed to output". When multiplying matrices, the summation operation is an OR operation.

[0049] The core technologies of this embodiment include: (1) Two-level routing mapping mechanism (core innovation) Level 1 (Input Port Mapping): Maps the destination port number to a queue number. Multiple input ports can be mapped to the same queue, simplifying routing rules.

[0050] Level 2 (Output Port Mapping): Maps queue numbers to output port numbers through matrix operations (forward / reverse mapping), supporting flexible configuration of "one output port corresponding to multiple queues".

[0051] Two-level mapping decouples input and output routing, avoiding the complexity of direct mapping.

[0052] (2) Optimization and adaptation of shared memory structure and output queuing method For asymmetric networks (where the number of input / output ports and bandwidth are unequal), a shared large-capacity storage device (such as an NVMe hard drive) is selected as a cache to mitigate bandwidth imbalances.

[0053] The output queuing (OQ) method is adopted, with each queue corresponding to an output port, which simplifies the scheduling logic.

[0054] In particular, considering that the sequential read / write performance of NVMe hard drives is much higher than that of random access, the number of queues can be simplified to optimize hard drive read / write efficiency.

[0055] (3) Vectorized queue scheduling based on matrix operations The transformation between the queue selection vector (Q1) and the output port enable vector (K2) is solidified into a mapping matrix (M_CVT) and its transpose matrix multiplication (the summation operation is actually an "OR operation").

[0056] Mapping rules can be pre-built into the software, are not lost when power is off, require no additional annotation, and are simple and reliable to implement in engineering.

[0057] (4) Systematic solutions for asymmetric networks It integrates the complete process of data multiplexing, cache write control, shared storage management, cache read control, and data demultiplexing, and is specifically designed to solve problems such as input / output bandwidth mismatch, complex routing rules, and hard disk cache performance optimization.

[0058] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A data routing and queue scheduling method for an asymmetric information exchange network, wherein the number of input ports and the number of output ports in the information exchange network are unequal, characterized in that, A shared memory switching architecture is used, with output queuing for data frame buffering management, including: First-level routing mapping: Maps the destination port number of the data frame received by each input port to the queue number, which is a many-to-one mapping; writes the data frame into the corresponding logical queue address space in the shared memory according to the queue number; Second-level routing mapping: The queue selection row vector of length M is converted into the output port selection row vector of length L through matrix operations. The summation operation in the matrix operation is replaced by the OR operation, where M is the number of queues and L is the number of output ports. Select the row vector based on the output port, read the data frame from the shared memory and distribute it to the corresponding output port.

2. The method according to claim 1, characterized in that, The second-level routing mapping uses a mapping matrix M_CVT to achieve forward mapping. The size of the mapping matrix M_CVT is M×L, and the forward mapping calculation formula is: K1 = Q1 × M_CVT Where Q1 is the queue selection row vector, K1 is the output port selection row vector, each row of the mapping matrix M_CVT contains at most one non-zero element, and the summation operation in matrix multiplication is replaced by an OR operation.

3. The method according to claim 2, characterized in that, It also includes reverse mapping: mapping the output port enable state row vector K2 to the queue output enable state row vector Q2, calculated as follows: Q2=K2× in The transpose of the mapping matrix M_CVT is of size L×M. The summation operation in matrix multiplication is replaced by an OR operation. K2[m]=1 indicates that the m-th output port is allowed to output, and Q2[n]=1 indicates that the n-th queue is allowed to output.

4. The method according to claim 1, characterized in that, The mapping rules for the second-level routing are pre-built in the software. The mapping rules are not lost when the shared storage loses power and do not need to be rewritten.

5. The method according to claim 1, characterized in that, The rules for first-level routing mapping are fixed and preset; for illegal destination port numbers, they are mapped to the exception queue.

6. The method according to claim 1, characterized in that, It also includes a data multiplexing step: the data frame request of each input port is detected by polling, the request handshake interface is used to interact with the cache write control module, and after the handshake is successful, the data frame is read from the front-end cache of the input port and pushed to the shared memory; The module that performs data reconnection does not have its own cache.

7. The method according to claim 1, characterized in that, The shared storage is implemented using NVMe hard drives; by controlling the total number of queues, the read and write operations of the NVMe hard drives are mainly sequential access, thereby improving data throughput.

8. A data routing and queue scheduling system for an asymmetric information exchange network, wherein the number of input ports and the number of output ports of the information exchange network are unequal, characterized in that, include: Shared memory, allowing read and write operations on all input and output ports; The data multiplexing module is used to poll the data frames of each input port, perform the first-level routing mapping, map the destination port number of the data frame to the queue number, map it to a many-to-one mapping, and push the data frame carrying the queue number to the cache write control module. The cache write control module is used to write data frames into the corresponding logical queue address space in the shared memory according to the queue number; The cache read control module is used to read data frames from the shared memory in logical queue order; The data demultiplexing module is used to perform the second-level routing mapping. It converts the queue selection row vector of length M into the output port selection row vector of length L through matrix operations. The summation operation in the matrix operation is replaced with an OR operation, and the data frame is distributed to the corresponding output port according to the output port selection row vector.

9. The system according to claim 8, characterized in that, The data tapping module includes: The vector mapping unit stores the mapping matrix M_CVT and performs forward and reverse mapping through matrix operations. Forward mapping converts the queue selection row vector Q1 into the output port selection row vector K1, and reverse mapping converts the output port enabled state row vector K2 into the queue output enabled state row vector Q2. The output port gating unit controls the gating of each output port protocol conversion module according to the row vector selected by the output port, and distributes the data frames read from the shared memory to the corresponding output ports.

10. The system according to claim 8, characterized in that, The data multiplexing module is not equipped with a local cache; the data multiplexing module has a request handshake interface with the front-end input port and the cache write control module respectively. After both handshakes are successful, the data frame is read from the front-end cache and pushed to the cache write control module.

11. The system according to claim 8, characterized in that, The shared storage is an NVMe hard drive; the cache write control module and the cache read control module access the NVMe hard drive in a sequential read / write manner.

12. The system according to claim 8, characterized in that, The mapping matrix used for the second-level path in the data tapping module is pre-built in the software.