A topology reconstruction method based on traffic awareness in multi-chip interconnection networks

Through the topological reconstruction method based on traffic perception in multi-core particle interconnect networks, the links are dynamically adjusted, and the problems of high latency and high cost between core particle networks are solved, achieving more efficient performance improvement and cost reduction.

CN116260760BActive Publication Date: 2024-11-12ZHEJIANG LAB +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202211613274.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-12-15
Publication Date
2024-11-12
Estimated Expiration
2042-12-15

AI Technical Summary

Technical Problem

In the existing multi-core integration technology, the inter-core interconnect network has high latency, cost-restricted link number, low frequency, and the lack of existing topological reconstruction algorithms, which limits the improvement of network performance.

Method used

Using the topological reconstruction method based on traffic perception, the shortest path of the entire source is calculated through the Floyd-Warshall algorithm, the mapping relationship between the node and the interface router is determined, the traffic is counted, and the cross-switch network can be configured using the depth-first search algorithm to dynamically adjust the link. Reduce latency and power consumption.

Benefits of technology

It realizes the design of reducing average hops and delays in multi-core interconnect networks, reducing chip area and cost, and improving performance. It is suitable for general-purpose processors and dedicated processors.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116260760B_ABST
    Figure CN116260760B_ABST
Patent Text Reader

Abstract

The present invention provides a topology reconstruction method based on traffic perception in a multi-chip interconnection network. The multi-chip interconnection network topology includes an intra-chip network topology and an inter-chip network topology. The method adds a configurable cross-switch network in the inter-chip network and uses a heuristic reconstruction method to configure these configurable cross-switches to provide a directly connected virtual link for frequently communicating routers to reduce the average number of hops and network delay. The present invention reduces the number of chip pins, reduces the average number of hops and delay of the network, thereby reducing the area and cost of the chip and improving its performance. It can be used in the design of general-purpose processors and dedicated processors for multiple applications.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of integrated circuit technology, and in particular to a topology reconstruction method based on traffic perception in a multi-chip interconnection network. Background Art

[0002] Multi-chiplet integration technology is an emerging design paradigm for continuing to increase chip scale in the post-Moore's Law era. Multi-chiplet integration technology is an emerging design paradigm for continuing to increase chip scale in the post-Moore's Law era. Multi-chiplet integration technology refers to the following design paradigm: modular bare chips (i.e., chiplets) are stacked and connected into multi-chiplet modules (Multi-Chiplet Module) using the "building blocks" method and advanced packaging. Because the yield of large-area chips is much lower than that of small chips, multi-chiplet integration technology can improve the yield of chips and reduce costs, allowing manufacturers to manufacture larger-scale chips under the same economic conditions. However, the existing multi-chiplet integration technology has problems such as the delay of inter-chip interconnection is much higher than the intra-chip interconnection, the number of inter-chip interconnection links is restricted by cost, and the frequency of the inter-chiplet network is often lower than the intra-chiplet network.

[0003] Existing chiplet interconnect network topologies mainly include fixed topology, reconfigurable chiplet interconnect topology and traditional reconfigurable on-chip network topology. They have poor performance in multi-chip integrated systems due to the following defects:

[0004] (1) Fixed topologies, such as the Kite family topology, cannot be dynamically adjusted based on the network traffic of the running applications, resulting in increased network latency and power consumption.

[0005] (2) The current reconfigurable core interconnect topology Adapt-NoC only considers the switching between four regular topologies. These four regular topologies are two-dimensional mesh topology (mesh), CMesh, two-dimensional torus topology (torus) and tree topology (tree). In addition, there are a large number of irregular topologies that have not been considered; this limits the possibility of network topology and cannot further reduce the network's latency and power consumption. In addition, Adapt-NoC does not provide a topology reconstruction algorithm. Users can only test the performance of the application under different topologies multiple times and select the topology with the best performance. The lack of a topology reconstruction algorithm limits the scope of application of Adapt-NoC.

[0006] (3) The traditional reconfigurable on-chip network topology neither considers the limitation of the number of chip pins on the number of inter-chip links nor the delay difference between chips. The traditional reconfigurable on-chip network topology is based on a single die design, which will greatly affect the performance in a multi-chip integrated system. Summary of the invention

[0007] The object of the present invention is to provide a topology reconstruction method based on traffic perception in a multi-chip interconnection network in view of the deficiencies of the prior art.

[0008] The object of the present invention is achieved through the following technical solution: A topology reconstruction method based on traffic perception in a multi-chip interconnection network, comprising the following steps:

[0009] (1) Based on the multi-chip interconnection network topology, the Floyd-Warshall algorithm is used to calculate the all-source shortest path of the entire network;

[0010] (2) According to the all-source shortest path obtained in step (1), determine the mapping relationship between the node pair and the interface router pair: suppose there are two nodes u and v, the first interface router is r, and the last interface router is s. If there is the first interface router r and the last interface router s on the shortest path between nodes u and v, then there exists a mapping relationship from the node pair (u, v) to the interface router pair (r, s), and the mapping relationship is recorded as: (u, v)→(r, s);

[0011] (3) Wait for a reconstruction period and count the traffic of each node pair during this period;

[0012] (4) According to the mapping relationship obtained in step (2), the traffic of the node pair obtained in step (3) is added one by one to the interface router pair corresponding to the mapping relationship to obtain the traffic of each interface router pair;

[0013] (5) Arrange all the interface router pairs in descending order according to the traffic of each interface router pair obtained in step (4) and obtain an arrangement order;

[0014] (6) Traverse all the interface router pairs in the order obtained in step (5): use a depth-first search algorithm to find the path of the interface router pair on the configurable crossbar switch network, and then find the reconfigurable link that the interface router pair needs to occupy, and output all the paths of the corresponding reconfigurable link; if the required reconfigurable link is different from the occupied reconfigurable link, then use the interface router pair;

[0015] (7) resetting the configurable crossbar switch network, performing port connection operations on all paths outputted from step (6), and resetting the multi-chip interconnection network topology to a topology that does not include the configurable crossbar switch network, and then calculating a routing table according to the new topology;

[0016] (8) Return to step (3) and perform the next iterative calculation until the network is closed.

[0017] Furthermore, the step (6) includes the following sub-steps:

[0018] (6.1) marking all reconfigurable links as unoccupied, and setting the first interface router pair in the arrangement order as the current interface router pair;

[0019] (6.2) Use depth-first search to find the path of the current interface router pair on the configurable crossbar switch network, and then find the reconfigurable link that the interface router pair needs to occupy, and output all paths of the corresponding reconfigurable link; if there is no reconfigurable link that needs to be occupied, go to step (6.4);

[0020] (6.3) determining whether the reconfigurable link outputted in step (6.2) is the same as the occupied reconfigurable link; if the required reconfigurable link is not the same as the occupied reconfigurable link, applying the interface router pair corresponding to the required reconfigurable link;

[0021] (6.4) If the current interface router pair has a next interface router pair, set the next interface router pair as the current interface router pair and repeat step (6.2), otherwise end this step.

[0022] Furthermore, the step (6.2) includes the following sub-steps:

[0023] (6.2.1) For the current crossbar switch, access the adjacent interface crossbar switch in the horizontal direction. If the adjacent crossbar switch has not been accessed and the reconfigurable link between the adjacent crossbar switch and the current crossbar switch is not occupied, set it as the current crossbar switch and record it in the path; otherwise, proceed to the next step;

[0024] (6.2.2) Visit the adjacent crossbar switch in the vertical direction. If the adjacent crossbar switch has not been visited and the reconfigurable link between the adjacent crossbar switch and the current crossbar switch is not occupied, set it as the current crossbar switch and record it in the path; otherwise, proceed to the next step;

[0025] (6.2.3) Determine whether the current cross-switch is the destination switch: If the current cross-switch is the destination cross-switch, output the path of the current interface router pair and the reconfigurable links it occupies; if the current cross-switch is not the destination cross-switch, mark the current cross-switch as visited and remove it from the path of the current interface router pair, then set the previous cross-switch in the path as the current cross-switch and re-determine whether it is the destination switch until the source cross-switch is marked as visited.

[0026] Furthermore, the operation of accessing the horizontally adjacent interface cross switches in the step (6.2.1) is specifically as follows: if the horizontal coordinate of the current cross switch is smaller than the horizontal coordinate of the destination cross switch, then access the adjacent cross switch on the east side; if the horizontal coordinate of the current cross switch is larger than the horizontal coordinate of the destination cross switch, then access the adjacent cross switch on the west side; until the horizontal coordinate of the current cross switch is equal to the horizontal coordinate of the destination cross switch.

[0027] Furthermore, the operation of accessing the adjacent cross switches in the vertical direction in the step (6.2.2) is specifically as follows: if the vertical coordinate of the current cross switch is smaller than the vertical coordinate of the destination cross switch, then access the adjacent cross switch in the north; if the vertical coordinate of the current cross switch is larger than the vertical coordinate of the destination cross switch, then access the adjacent cross switch in the south; until the vertical coordinate of the current cross switch is equal to the vertical coordinate of the destination cross switch.

[0028] Furthermore, the step (6.2) of using depth-first search to find the path of the current interface router pair on the configurable crossbar switch network is specifically as follows: all interface routers are initialized to an unvisited state, each configurable crossbar switch network is selected in turn, the configurable crossbar switch corresponding to the source router of the current interface router pair is set as the source crossbar switch, the configurable crossbar switch corresponding to the destination router is set as the destination crossbar switch, and the source crossbar switch is set as the current crossbar switch.

[0029] Furthermore, the step (7) includes the following sub-steps:

[0030] (7.1) Reconfigure the configurable crossbar switch network to provide virtual links for the interface router pairs applied in step (6), and perform the following operations on all paths output from step (6):

[0031] Let the path be (v1, v2, ..., v p ), where v is a configurable crossbar switch and p is a positive integer; in the configurable crossbar switch v1, the local port and the port connected to v2 are connected; in the configurable crossbar switch v p In the local port and connection v p-1 ports are connected; in other configurable crossbar switches v i In which, the subscript i∈(1,p), connect v i-1 Ports and connections i+1 The ports are connected; wherein the ports include input ports and output ports;

[0032] (7.2) Reset the multi-chip interconnect network topology to a topology that does not include the configurable crossbar switch network, and add edges to the interface router pairs applied in step (6); then run the Floyd-Warshall algorithm to calculate the routing table of the network and send it to each router.

[0033] Furthermore, the reconfigurable links refer to links between configurable crossbar switches and between a configurable crossbar switch and a router.

[0034] Furthermore, the multi-chip interconnection network topology is implemented on an active interposer or a passive interposer; when the network topology is implemented on an active interposer, the interface router and the configurable cross switch are located on the active interposer, and the communication between the router in the chip and the interface router passes through the micro-bumps and the physical layer transceiver; while the communication between the interface routers and the configurable cross switches does not need to pass through the micro-bumps and the physical layer transceiver; when the network topology is implemented on a passive interposer, the interface router and the configurable cross switch are located in the chip; the communication between the router in the chip and the interface router does not pass through the micro-bumps and the physical layer transceiver, but the communication between the interface routers and the configurable cross switches needs to pass through the micro-bumps and the physical layer transceiver.

[0035] Furthermore, in step (5), all the interface router pairs may be arranged in descending order according to the product of the flow and delay of each interface router pair obtained in step (4).

[0036] Compared with the prior art, the invention has the following beneficial effects: by adding a configurable crossbar network to the inter-chiplet network and configuring these crossbars using a heuristic reconstruction method according to the distribution of network traffic, the invention provides a directly connected virtual link for routers that frequently communicate, reduces the number of chiplet pins, reduces the average number of network hops and delay, thereby reducing the area and cost of the chip and improving its performance. The method is relatively efficient and flexible and can be used in the design of general-purpose processors and dedicated processors for multiple applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A topological diagram of a multi-chip interconnection network of the present invention;

[0038] Figure 2 It is a structural diagram of the multi-chip interconnection network in the present invention when an active intermediate layer is used to implement it;

[0039] Figure 3 It is a structural diagram of the multi-chip interconnection network in the present invention when a passive intermediate layer is used to implement it;

[0040] Figure 4 A micro-architecture diagram of a router according to the present invention;

[0041] Figure 5 FIG. 4 is a logic circuit diagram of a configurable crossbar switch according to the present invention. DETAILED DESCRIPTION

[0042] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0043] The multi-chip interconnection network architecture of the present invention includes a network topology, a router micro-architecture and a configurable crossbar switch structure. The multi-chip interconnection network consists of an intra-chip network and an inter-chip network. When an active interposer is used, Figure 1 As shown in the figure, the topology of the intra-chip network is a 4×4 mesh topology, and every 4 intra-chip network routers are connected to an interface router. The intra-chip network routers can communicate with other intra-chip routers through the interface routers. In the inter-chip network, the interface routers not only form a mesh network, but also connect to n configurable crossbar switch networks; the configurable crossbar switch network is a mesh network. During operation, the configurable crossbar switches will dynamically adjust their connection status according to the distribution of network traffic, provide directly connected virtual links for frequently communicating interface routers, reconstruct the network topology, and thus reduce network latency.

[0044] The location of the interface routers and configurable crossbar switches varies depending on the type of interposer used in the network architecture:

[0045] (1) When the network architecture is implemented on an active interposer, the interface routers and configurable crossbar switches are located on the interposer; e.g. Figure 2 As shown in the figure, there are wires and logic circuits on the active interposer, so the interface router and configurable crossbar switch can be placed on it. The communication between the router in the chip and the interface router must pass through the microbump and the physical layer transceiver (that is, through the chip), while the communication between the interface routers and the configurable crossbar switches does not need to pass through the microbump and the physical layer transceiver; and the interface router and the configurable crossbar switch forward data packets without passing through the microbump and the physical layer transceiver. The communication between the chiplets only needs to pass through the chip twice, that is, starting from the source chiplet → passing through the chiplet → reaching the interposer → passing through the chiplet → reaching the destination chiplet, and the network delay is low.

[0046] (2) When the network architecture is implemented on a passive interposer, the interface routers and configurable crossbar switches are located in the core; e.g. Figure 3 As shown in the figure, there are only wires on the passive interposer, but no logic circuits, so interface routers and configurable crossbar switches cannot be placed. The communication between the router in the chip and the interface router does not pass through the microbumps and physical layer transceivers, while the communication between interface routers and configurable crossbar switches needs to pass through microbumps and physical layer transceivers; and the interface routers and configurable crossbar switches need to pass through microbumps and physical layer transceivers to forward data. Communication between chiplets often needs to pass through the inter-chip multiple times (if a data packet passes through n interface routers, it needs to pass through the inter-chip 2(n-1) times), resulting in high network latency.

[0047] The microarchitecture of the interface router is as follows Figure 4As shown in the figure, the router consists of input ports, virtual channels, routing tables, routing calculation modules, virtual channel arbiters, switch arbiters, multiplexers, cross switches and output ports. The router has input and output ports for reconfigurable cross switches; the routing calculation module is used to calculate the next hop route according to the routing table. In each reconstruction cycle, the system updates the routing tables of all routers. The input ports of the router include input ports connected to adjacent routers and input ports connected to configurable cross switch networks. Among them, all configurable cross switch networks share one input port, and a multiplexer determines the configurable cross switch network currently using the port; each input port has a group of virtual channels, and each virtual channel has a group of input buffers. The virtual channel arbiter and the switch arbiter control the virtual channels and cross switches of the router respectively. There is a set of multiplexers inside the router, and the routing calculation module uses this set of multiplexers to control the output of data packets to adjacent routers or connected configurable cross switches; if the data packet is output to an adjacent router, a cross switch determines which output port the data packet arrives at, and the output port of the router is connected to the adjacent router; if the data packet is output to a configurable cross switch, another cross switch determines which configurable cross switch network the data packet is output to.

[0048] The structure of the configurable crossbar switch is as follows Figure 5 As shown. The configurable crossbar switch consists of multiple multiplexers and AND gates. In each reconstruction cycle, the system updates the connection status of all configurable crossbar switches and determines its connection status through the control port of the configurable crossbar switch. The configurable crossbar switch has five directions, namely local, east, south, west and north; the control port of each direction includes its selection port and enable port; the data port of each direction includes its input port and output port; among them, in this network architecture, the bit width of the configurable crossbar switch data port is the same as the bit width of the router port; in this reconfigurable network, the first core of the first core particle is the network reconstruction controller, which is responsible for executing the topology reconstruction method and sending the configuration calculated by the topology reconstruction method to each router and the configurable crossbar switch.

[0049] The present invention provides a topology reconstruction method based on traffic perception in a multi-chip interconnection network, the method comprising the following steps:

[0050] (1) Based on the topology of the multi-chip interconnection network (excluding the configurable crossbar switch network), the Floyd-Warshall all-source shortest path algorithm is used to calculate the all-source shortest path of the entire network.

[0051] (2) Based on the all-source shortest path obtained in step (1), determine the mapping relationship between node pairs and interface router pairs:

[0052] Traverse all node pairs, assume that there are two nodes u and v, assume that the first interface router is r and the last interface router is s. If there is the first interface router r and the last interface router s on the shortest path between nodes u and v (where r and s can be the same), then there is a mapping relationship from the node pair (u, v) to the interface router pair (r, s), recorded as (u, v) → (r, s). If the shortest path between nodes u and v does not pass through any interface router, then its mapping relationship is recorded as a null value.

[0053] (3) Wait for a reconstruction cycle and count the traffic of each node pair during this period.

[0054] (4) According to the mapping relationship obtained in step (2), the traffic of the node pair obtained in step (3) is added to the interface router pair corresponding to the mapping relationship one by one to obtain the traffic of each interface router pair. If the mapping relationship of the node pair is a null value, its traffic is not added to any interface router pair.

[0055] (5) All the interface router pairs are arranged in descending order according to the traffic of each interface router pair obtained in step (4) to obtain an arrangement order; or all the interface router pairs are arranged in descending order according to the product of the traffic and delay of each interface router pair obtained in step (4), wherein the unit of traffic is flit and the unit of network delay is cycle.

[0056] (6) Traverse all the interface router pairs in the order of arrangement obtained in step (5) to find the interface router pairs that need to establish virtual links; wherein the virtual links refer to links established through the configurable crossbar switch network.

[0057] Step (6) includes the following sub-steps:

[0058] (6.1) Initialization: Mark all reconfigurable links as unoccupied and set the first interface router pair in the arrangement order as the current interface router pair.

[0059] (6.2) Use depth-first search to find the path of the current interface router pair on the configurable crossbar switch network, and then find the reconfigurable link that the interface router pair needs to occupy, and output all paths of the corresponding reconfigurable link; if there is no reconfigurable link that needs to be occupied, go to step (6.4).

[0060] Initialize all interface routers to the unaccessed state. Select each configurable crossbar network in turn, set the configurable crossbar corresponding to the source router of the current interface router pair as the source crossbar, and the configurable crossbar corresponding to the destination router as the destination crossbar; then set the source crossbar as the current crossbar. For the current crossbar, continuously perform the following sub-steps:

[0061] (6.2.1) First check the horizontal adjacent interface cross switches: if the horizontal coordinate of the current cross switch is smaller than the horizontal coordinate of the destination cross switch, then check the adjacent cross switch on the east side; if the horizontal coordinate of the current cross switch is larger than the horizontal coordinate of the destination cross switch, then check the adjacent cross switch on the west side; until the horizontal coordinate of the current cross switch is equal to the horizontal coordinate of the destination cross switch, proceed to the next step; if the adjacent cross switch has not been accessed and the reconfigurable link between it and the current cross switch is not occupied, then set it as the current cross switch and record it in the path, otherwise, proceed to the next step.

[0062] (6.2.2) Check the adjacent cross switches in the vertical direction again: if the vertical coordinate of the current cross switch is smaller than the vertical coordinate of the destination cross switch, check the adjacent cross switch in the north; if the vertical coordinate of the current cross switch is larger than the vertical coordinate of the destination cross switch, check the adjacent cross switch in the south; until the vertical coordinate of the current cross switch is equal to the vertical coordinate of the destination cross switch, execute the next step; if the adjacent cross switch has not been visited and the reconfigurable link between it and the current cross switch is not occupied, set it as the current cross switch and record it in the path, otherwise, execute the next step.

[0063] (6.2.3) Determine whether the current cross-switch is the destination switch: If the current cross-switch is the destination cross-switch, output the path of the current interface router pair and the reconfigurable links it occupies; if the current cross-switch is not the destination cross-switch, mark the current cross-switch as visited and remove it from the path of the current interface router pair, then set the previous cross-switch in the path as the current cross-switch and re-determine whether it is the destination switch until the source cross-switch is marked as visited.

[0064] (6.3) Determine whether the reconfigurable link outputted in step (6.2) is the same as the occupied reconfigurable link. If the required reconfigurable link is not the same as the occupied reconfigurable link, apply the interface router pair corresponding to the required reconfigurable link. If step (6.2) does not output a reconfigurable link, skip this step.

[0065] (6.4) If the current interface router pair has a next interface router pair, set the next interface router pair as the current interface router pair and repeat step (6.2), otherwise end this step.

[0066] (7) Resetting the configurable crossbar switch network, performing port connection operations on all paths outputted from step (6), and resetting the multi-chip interconnection network topology to a topology that does not include the configurable crossbar switch network, and then calculating a routing table based on the new topology.

[0067] Step (7) includes the following sub-steps:

[0068] (7.1) The network reconfiguration controller reconfigures the configurable crossbar switch network to provide virtual links for the interface router pairs applied in step (6). The following operations are performed on all paths output from step (6):

[0069] Let the path be (v1, v2, ..., v p ), where v is a configurable crossbar switch and p is a positive integer. In the configurable crossbar switch v1, the local port and the port connected to v2 are connected; in the configurable crossbar switch v p In the local port and connection v p-1 ports are connected; in other configurable crossbar switches v i In which, the subscript i∈(1,p), connect v i-1 Ports and connections i+1 The ports include input ports and output ports.

[0070] (7.2) Reset the multi-chip interconnect network topology stored in the network reconstruction controller to a topology that does not include the configurable crossbar switch network, and add edges to the interface router pairs that have been applied in step (6). Then run Fl oy d-War s Hall's all-source shortest path algorithm calculates the network's routing table and sends it to each router.

[0071] (8) Return to step (3) and perform the next iterative calculation until the network is closed.

[0072] Those skilled in the art will readily appreciate other embodiments of the present application after considering the description and practicing the contents disclosed herein. The present application is intended to cover any variations, uses or adaptations of the present application, which follow the general principles of the present application and include common knowledge or customary technical means in the art that are not disclosed in the present application. The description and examples are intended to be exemplary only.

[0073] It should be understood that the present application is not limited to the exact construction that has been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof.

Claims

1. A topology reconstruction method based on traffic perception in a multi-chip interconnection network, characterized in that: The steps include: (1) Based on the multi-chip interconnection network topology, the Floyd-Warshall algorithm is used to calculate the all-source shortest path of the entire network; (2) According to the all-source shortest path obtained in step (1), determine the mapping relationship between the node pair and the interface router pair: suppose there are two nodes u and v, the first interface router is r, and the last interface router is s. If there is a first interface router r and a last interface router s on the shortest path between nodes u and v, then there exists a mapping relationship from the node pair (u, v) to the interface router pair (r, s). This mapping relationship is recorded as: (u, v) → (r, s); (3) Wait for a reconstruction period and count the traffic of each node pair during this period; (4) According to the mapping relationship obtained in step (2), the traffic of the node pair obtained in step (3) is added one by one to the interface router pair corresponding to the mapping relationship to obtain the traffic of each interface router pair; (5) Arrange all the interface router pairs in descending order according to the traffic of each interface router pair obtained in step (4) and obtain an arrangement order; (6) Traverse all the interface router pairs in the order obtained in step (5): use a depth-first search algorithm to find the path of the interface router pair on the configurable crossbar switch network, and then find the reconfigurable link that the interface router pair needs to occupy, and output all the paths of the corresponding reconfigurable link; if the required reconfigurable link is different from the occupied reconfigurable link, then use the interface router pair; (6.1) marking all reconfigurable links as unoccupied, and setting the first interface router pair in the arrangement order as the current interface router pair; (6.2) Use depth-first search to find the path of the current interface router pair on the configurable crossbar switch network, and then find the reconfigurable link that the interface router pair needs to occupy, and output all paths of the corresponding reconfigurable link; if there is no reconfigurable link that needs to be occupied, go to step (6.4); (6.2.1) For the current crossbar switch, access the adjacent interface crossbar switch in the horizontal direction. If the adjacent crossbar switch has not been accessed and the reconfigurable link between the adjacent crossbar switch and the current crossbar switch is not occupied, set it as the current crossbar switch and record it in the path; otherwise, proceed to the next step; (6.2.2) Visit the adjacent crossbar switch in the vertical direction. If the adjacent crossbar switch has not been visited and the reconfigurable link between the adjacent crossbar switch and the current crossbar switch is not occupied, set it as the current crossbar switch and record it in the path; otherwise, proceed to the next step; (6.2.3) Determine whether the current crossbar switch is the destination switch: If the current crossbar switch is the destination crossbar switch, output the path of the current interface router pair and the reconfigurable links it occupies; If the current crossbar switch is not the destination crossbar switch, mark the current crossbar switch as visited and remove it from the path of the current interface router pair, then set the previous crossbar switch in the path as the current crossbar switch and re-determine whether it is the destination switch until the source crossbar switch is marked as visited; (6.3) determining whether the reconfigurable link outputted in step (6.2) is the same as the occupied reconfigurable link; if the required reconfigurable link is not the same as the occupied reconfigurable link, applying the interface router pair corresponding to the required reconfigurable link; (6.4) If the current interface router pair has a next interface router pair, set the next interface router pair as the current interface router pair and repeat step (6.2), otherwise end this step; (7) resetting the configurable crossbar switch network, performing port connection operations on all paths outputted from step (6), and resetting the multi-chip interconnection network topology to a topology that does not include the configurable crossbar switch network, and then calculating a routing table according to the new topology; (7.1) Reconfigure the configurable crossbar switch network to provide virtual links for the interface router pairs applied in step (6), and perform the following operations on all paths output from step (6): Let the path be (v1,v2,...,v p ), where v is a configurable crossbar switch and p is a positive integer; in the configurable crossbar switch v1, the local port and the port connected to v2 are connected; in the configurable crossbar switch v p In the local port and connection v p-1 ports are connected; in other configurable crossbar switches v i In which, the subscript i∈(1,p), connect v i-1 Ports and connections i+1 The ports are connected; wherein the ports include input ports and output ports; (7.2) Reset the multi-chip interconnect network topology to a topology that does not include the configurable crossbar switch network, and add edges to the interface router pairs that have been applied in step (6); then run the Floyd-Warshall algorithm to calculate the routing table of the network and send it to each router; (8) Return to step (3) and perform the next iterative calculation until the network is closed.

2. The method for topology reconstruction based on traffic perception in a multi-chip interconnection network according to claim 1, characterized in that: The operation of accessing the horizontally adjacent interface cross switches in the step (6.2.1) is specifically as follows: if the horizontal coordinate of the current cross switch is smaller than the horizontal coordinate of the destination cross switch, then access the adjacent cross switch on the east side; if the horizontal coordinate of the current cross switch is larger than the horizontal coordinate of the destination cross switch, then access the adjacent cross switch on the west side; until the horizontal coordinate of the current cross switch is equal to the horizontal coordinate of the destination cross switch.

3. The method for topology reconstruction based on traffic perception in a multi-chip interconnection network according to claim 1, characterized in that: The operation of accessing the adjacent cross switches in the vertical direction in the step (6.2.2) is specifically as follows: if the vertical coordinate of the current cross switch is smaller than the vertical coordinate of the destination cross switch, then access the adjacent cross switch in the north; if the vertical coordinate of the current cross switch is larger than the vertical coordinate of the destination cross switch, then access the adjacent cross switch in the south; until the vertical coordinate of the current cross switch is equal to the vertical coordinate of the destination cross switch.

4. The method for topology reconstruction based on traffic perception in a multi-chip interconnection network according to claim 1, characterized in that: The step (6.2) of using depth-first search to find the path of the current interface router pair on the configurable crossbar switch network is specifically as follows: all interface routers are initialized to an unvisited state, each configurable crossbar switch network is selected in turn, the configurable crossbar switch corresponding to the source router of the current interface router pair is set as the source crossbar switch, the configurable crossbar switch corresponding to the destination router is set as the destination crossbar switch, and the source crossbar switch is set as the current crossbar switch.

5. The method for topology reconstruction based on traffic perception in a multi-chip interconnection network according to claim 1, characterized in that: The reconfigurable links refer to links between configurable crossbar switches and between configurable crossbar switches and routers.

6. The method for topology reconstruction based on traffic perception in a multi-chip interconnection network according to claim 1, characterized in that: The multi-chip interconnection network topology is implemented on an active interposer or a passive interposer; when the network topology is implemented on an active interposer, the interface router and the configurable cross switch are located on the active interposer, and the communication between the router in the chip and the interface router passes through the micro-bump and the physical layer transceiver; while the communication between the interface routers and the configurable cross switches does not need to pass through the micro-bump and the physical layer transceiver; when the network topology is implemented on a passive interposer, the interface router and the configurable cross switch are located in the chip; the communication between the router in the chip and the interface router does not pass through the micro-bump and the physical layer transceiver, but the communication between the interface routers and the configurable cross switches needs to pass through the micro-bump and the physical layer transceiver.

7. The method for topology reconstruction based on traffic perception in a multi-chip interconnection network according to claim 1, characterized in that: In the step (5), all the interface router pairs may be arranged in descending order according to the product of the flow and delay of each interface router pair obtained in the step (4).