Dynamic graph processing method based on 1T2R memristor array

By mapping the adjacency matrix of the dynamic graph to the 1T2R memristor array and using the programming operation of the memristor, the time and area overhead problems of traditional hardware when processing the dynamic graph are solved, and efficient dynamic graph processing and storage are achieved, which significantly Improved computing speed and storage density.

CN119007775AActive Publication Date: 2024-11-22PEKING UNIV
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202411006856.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-25
Publication Date
2024-11-22
Estimated Expiration
2044-07-25

AI Technical Summary

Technical Problem

When the traditional von Neumann architecture computers process dynamic graph data, the data is frequently transferred due to separation from memory, which cannot meet the needs of efficient information processing. The existing hardware is difficult to flexibly adjust when processing dynamic graphs, resulting in time and area. Large overhead.

Method used

Using a dynamic graph processing scheme based on 1T2R memristor array, the adjacency matrix of the dynamic graph is mapped to the 1T2R memristor array, and the addition and deletion of nodes and edges are realized through the programming operation of the memristor, using high-impedance state and low The resistive state indicates the existence or absence of nodes and edges, reducing time and area overhead.

Benefits of technology

It effectively reduces the time complexity and area overhead of dynamic graph processing, improves calculation speed and storage density, and realizes O(1) time complexity in terms of nodes and edges, which significantly improves the efficiency of dynamic graph processing.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119007775A_ABST
    Figure CN119007775A_ABST
Patent Text Reader

Abstract

The invention discloses a dynamic graph processing method based on a 1T2R memristor array, and belongs to the technical field of in-memory calculation and graph calculation. An adjacent matrix of a graph is stored on a memristor array composed of 1T2R units, the state of nodes is stored in the 1T2R unit at the (i, i) position, and the information of edges is stored in the 1T2R unit at the (i, j) position and represents the connection relation between the node i and the node j; the operation on nodes and edges of the graph is mapped into the operation on corresponding units in the 1T2R memristor array, and dynamic graph storage and dynamic graph processing are efficiently carried out. By means of the array operation mode, the time complexity of the dynamic graph processing process is effectively reduced, node increasing, node deleting, edge increasing and edge deleting can be achieved under the O (1) time complexity, the calculation speed of dynamic graph processing is effectively improved, and higher storage density is achieved on dynamic graph data storage.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of memristor in-memory computing and graph computing, and specifically relates to a dynamic graph computing solution based on a memristor in-memory computing architecture. Background Art

[0002] With the rapid development of the Internet of Things and artificial intelligence technologies, human society's demand for computing and information processing has grown rapidly. The deep neural networks and large models that have emerged in the past decade have a high demand for information processing computing. At present, computers based on the traditional von Neumann architecture have a design that separates computing and storage, and their computing efficiency is restricted by the frequent movement of data, and cannot adapt to the needs of new computing for efficient information processing. Graph processing is a computing method widely used in applications such as causal inference, molecular prediction, and social networks. The data used in graph processing has strong irregularity and places extremely high demands on the efficiency of the processor. Therefore, graph processing urgently needs efficient dedicated hardware to accelerate it.

[0003] A dynamic graph is a special kind of graph that can change over time. At a specific time point, a dynamic graph has the characteristics of a typical graph; at different time points, the graph structure of a dynamic graph will change. Between different time points, the operations performed on a dynamic graph include adding nodes, deleting nodes, adding edges, and deleting edges.

[0004] Dynamic graphs can also be used to represent information such as social networks and material structures. Using a graph to represent a social network at a certain moment, the structure of the social network often evolves over time, and the storage and calculation of such problems become more complicated. The hardware used to process dynamic graph information needs to have a certain degree of flexibility, and be able to adjust to the changes of the dynamic graph at different times with a small hardware overhead, while also having the efficiency of a graph processor.

[0005] Emerging in-memory computing and new memory technologies are a solution that can alleviate the bottleneck of the von Neumann architecture. By placing storage and computing in the same location, the hardware overhead of data access can be reduced. In-memory computing based on memristor arrays can achieve a storage-computing integrated computing architecture with a higher storage density. The memristor itself can store information non-volatilely, which can avoid repeated data loading and further reduce memory access overhead. In-memory computing technology based on memristor arrays is expected to provide an effective solution to the dynamic graph problem. Summary of the invention

[0006] The purpose of the present invention is to provide a circuit architecture and a dynamic graph computing solution for processing dynamic graph data based on a memristor array, so as to achieve high parallel hardware acceleration of dynamic graph processing.

[0007] Based on the existing 1T2R memristor cells and the 1T2R memristor array composed of them, the present invention develops a solution for specifically mapping a dynamic graph and the processing operations of the dynamic graph to the 1T2R memristor array and the programming operations of the memristor array, so as to realize dynamic graph processing based on the 1T2R memristor array. This solution maps the adjacency matrix A of the dynamic graph to the 1T2R memristor array, maps the operations on the nodes of the graph to the operations on the (i, i) cells in the 1T2R memristor array, and maps the operations on the edges of the graph to the operations on the (i, j) cells in the 1T2R memristor array, where i ≠ j.

[0008] As Figure 1 shown in (a) therein, the 1T2R memristor cell is composed of one transistor and two memristors; one 1T2R memristor cell is connected to four lines in the array, namely BLL, WL, BLR, and SL; among them, BLL and WL are parallel, BLR and SL are parallel, and BLL and WL are perpendicular to BLR and SL. In the cell structure, the two memristors are respectively named L-RRAM and R-RRAM, and the two memristors are commonly connected to the drain terminal of the transistor; the other end of L-RRAM is connected to BLL; the other end of R-RRAM is connected to BLR; the gate of the transistor is connected to WL; the source of the transistor is connected to SL.

[0009] As Figure 1 shown in (b) therein, a memristor array based on the 1T2R cell structure contains M × M (M rows and M columns) 1T2R memristor cells; the array has a total of M WLs, M BLLs, M BLRs, and M SLs; the cells in one row share the same BLL and the same WL, and the cells in one column share the same BLR and the same SL.

[0010] On the 1T2R memristor array, a graph is stored using the representation method of the adjacency matrix of the graph. For a graph with M nodes, the adjacency matrix A of the graph is stored as a square matrix of M × M to store the entire graph. In the square matrix, the element A at the coordinates (i, j) ij stores the connection relationship between node i and node j (i.e., the edge between i and j), where i and j are integers, 0 ≤ i < M, 0 ≤ j < M. The element A at the coordinates (i, i) ii stores the information of node i itself. For the actual deployment of the graph, it is not required that A ij = A ji .

[0011] A dynamic graph is defined as a graph whose properties of nodes or edges change over time. If an adjacency matrix A is used to store a dynamic graph, then the adjacency matrix A will change over time. In conventional processors or traditional 1T1R memristor arrays, processing dynamic graphs requires a large time overhead and area cost. The dynamic graph processing scheme based on the 1T2R memristor array proposed in the present invention can effectively reduce the time overhead and area cost of dynamic graph calculations. The dynamic graph processing scheme based on the 1T2R memristor array is as follows:

[0012] 1) Graph adjacency matrix mapping: For a graph with M nodes, there can be directed connections between its nodes, that is, edges between nodes. The adjacency matrix A of the graph is an M×M square matrix. In the square matrix, the element A with coordinates (i, j) is ij Stores the connection relationship between node i and node j (i.e., the edge between i and j). ij The size can be -1, 0, +1. ij When the size of is 0, it means that there is no connection from node i to node j. The method proposed in the present invention maps the adjacency matrix A to a 1T2R memristor array of size M×M. In the memristor array, the memristor unit with position coordinates (i, i) stores the element A in the adjacency matrix A. ii , i.e., the node state; the memristor unit with position coordinates (i, j) stores the element A in the adjacency matrix ij , that is, the state of the edge.

[0013] 2) Node addition: If the i-th node is added, the cell with the position coordinate (i, i) in the 1T2R memristor array is operated, and both memristors in the cell are set to the HRS (high resistance) state. This state indicates that the i-th node has been added. During the data readout process, the cell at the (i, i) position will not significantly increase the current at the BLR end. The current is detected at the BLR end, and no abnormal current is detected.

[0014] 3) Node deletion: If the i-th node is deleted, the unit with the position coordinate (i, i) in the 1T2R memristor array is operated, and both memristors in the unit are set to the LRS (low resistance) state. This state indicates that the i-th node has been deleted. During the data readout process, the unit at the (i, i) position will significantly increase the current at the BLR end because it has two low resistance states. When the current is detected at the BLR end, an abnormal current will be detected. By the abnormal current in the i-th column, the outside world can determine that the i-th node has been deleted.

[0015] 4) Edge addition: If an edge from the i-th node to the j-th node is added, the cell with the position coordinate (i, j) in the 1T2R memristor array is operated according to A ij If Aij = +1, then the two memristors L-RRAM and R-RRAM of the (i, j) unit are set to HRS and LRS states respectively, that is, the unit is used to represent +1. ij =-1, the two memristors L-RRAM and R-RRAM of the (i, j) unit are set to LRS and HRS states respectively, that is, the unit is used to represent -1.

[0016] 5) Edge deletion: If the edge from the i-th node to the j-th node is deleted, the cell with the position coordinate (i, j) in the 1T2R memristor array is operated. Both memristors of the (i, j) cell are set to the HRS state, that is, the cell is used to represent 0.

[0017] 6) Data readout: In the data readout phase, an address i is selected in each clock cycle, and the WL in the 1T2R memristor array i With BLL i is selected (set to high level). All BLRs are grounded, and the analog-to-digital converter (ADC) is used to read the current value passing through it. All SLs are floating, and the analog-to-digital converter (ADC) is used to read the voltage value it has. One clock cycle can read out the M bits of data in the i-th row. Among them, if the i-th BLR reads out a high current, it means that the node i has been deleted, and the i-th row data read out this time is invalid; if the i-th BLR reads out a low current, it means that the node i has been added, and the i-th row data read out this time is valid. By traversing the selected address i from 0 to M-1, all M rows of data of the 1T2R memristor array can be read out, that is, all the data of the adjacency matrix A can be read out. It takes a total of M clock cycles to read out the entire adjacency matrix A.

[0018] In the adjacency matrix mapping of the above processing scheme, for the unit with position coordinate (i, i) in the 1T2R memristor array, if both memristors are HRS, the node is valid, and if both memristors are LRS, the node is invalid. The read current of each BLR line in the 1T2R memristor array can be used to determine whether the corresponding node is valid.

[0019] For the cell with position coordinates (i, j) in the 1T2R memristor array, when the combination of L-RRAM and R-RRAM is (HRS, HRS), it represents the number 0, that is, an edge does not exist; when the combination of L-RRAM and R-RRAM is (HRS, LRS), it represents the number +1, that is, the edge has a value of +1; when the combination of L-RRAM and R-RRAM is (LRS, HRS), it represents the number -1, that is, the edge has a value of -1.

[0020] The edge addition operation is based on A ijThe value of , sets the two memristors of the 1T2R memristor unit with the position coordinates (i, j) from (HRS, HRS) to (LRS, HRS) or (HRS, LRS); performs edge deletion operation, that is, sets the two memristors of the 1T2R memristor unit with the position coordinates (i, j) from (LRS, HRS) or (HRS, LRS) to (HRS, HRS).

[0021] In the above processing scheme, the adjacency matrix of the graph of M nodes is read out in M ​​clock cycles in data reading. The validity of the i-th node is determined according to the current on the i-th BLR line read out. If the i-th BLR line is high current, the i-th node is invalid; if the i-th BLR line is low current, the i-th node is valid.

[0022] The memristor in the 1T2R memristor array described in the present invention is preferably a memristor based on a metal-insulator-insulator-metal structure, or a memristor based on a metal-semiconductor-semiconductor-metal structure, which has multiple stable resistance states, wherein the highest resistance state is a high resistance state HRS, and the lowest resistance state is a low resistance state LRS; the transistor is preferably an N-type metal oxide semiconductor field effect transistor.

[0023] Beneficial effects of the present invention:

[0024] By constructing a dynamic graph processing solution based on a 1T2R memristor array, the time complexity of the dynamic graph processing process can be effectively reduced. As shown in the following table, for a graph with M nodes, the size of its adjacency matrix is ​​M×M. When using a traditional 1T1R memristor array to store dynamic graph data, both adding and deleting nodes require O(M) time complexity, that is, the time consumed is proportional to M. Adding and deleting edges require O(1) time complexity, that is, the time consumed is a constant. When using the 1T2R memristor array proposed in the present invention to store dynamic graph data, adding and deleting nodes only require O(1) time complexity, that is, the time consumed is a constant. Adding and deleting edges requires O(1) time complexity, that is, the time consumed is a constant. In practical applications, the time consumed by operations with O(1) time complexity is less than that of operations with O(M) time complexity, and as M increases, the advantages of operations with O(1) time complexity become more obvious. Therefore, the dynamic graph mapping method based on 1T2R array proposed in the present invention can effectively improve the calculation speed of dynamic graph processing of memristor array.

[0025] Time complexity Traditional 1T1R memristor array 1T2R Memristor Array Adding Nodes O(M) O(1) Deleting a Node O(M) O(1) Add Edge O(1) O(1) Delete Edge O(1) O(1)

[0026] In terms of storage density, considering that the value range of each edge is -1, 0, +1, the traditional 1T1R memristor array needs to use 2 1T1R units to store the value of one edge, while the graph mapping method based on the 1T2R array in the present invention only needs to use 1 1T2R unit to store the value of one edge. Among them, the area of ​​the 1T1R unit and the 1T2R unit are both about the size of 1 transistor. Therefore, the graph mapping method based on the 1T2R array proposed in the present invention can effectively improve the area density of the memristor array for storing graph data.

[0027] In summary, the graph mapping method based on 1T2R array proposed in the present invention has improvements in both dynamic graph processing speed and area density, and can perform dynamic graph calculations more efficiently in real applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Schematic diagram of (a) a 1T2R memristor unit and (b) a 1T2R memristor array composed of 1T2R memristor units used in the present invention.

[0029] Figure 2 This is a graph data mapping scheme based on 1T2R memristor array proposed by the present invention, wherein the cells in the dotted box represent nodes, and the other cells represent edges.

[0030] Figure 3 An example of a specific graph adjacency matrix and its mapping on the 1T2R memristor array is shown, where the cells in the dashed box represent nodes and the other cells represent edges.

[0031] Figure 4 Shows a Figure 3 An example of dynamic graph processing based on CNN and its mapping on the 1T2R memristor array, where the cells in the dashed box represent nodes and the other cells represent edges. DETAILED DESCRIPTION

[0032] The present invention will be further described below through specific embodiments in conjunction with the accompanying drawings.

[0033] The following embodiments are carried out using a memristor array with an actual 1T2R unit structure. The structure of the memristor array is as follows: Figure 1 The operations applied in the key programming steps are given, and an embodiment of the memristor storage-computing integrated architecture is given.

[0034] 1) Graph adjacency matrix mapping: such as Figure 3 As shown, for a node with M nodes (node ​​N0 to node N M-1 ) graph, there can be directed connections between nodes, that is, edges between nodes. The adjacency matrix A of a graph is an M×M square matrix, in which the element A with coordinates (i, i)ii Represents the node, element A with coordinates (i, j) ij Stores the connection relationship between node i and node j (i.e., the edge between i and j). ij The size can be -1, 0, +1. ij When the size is 0, it means that there is no connection from node i to node j. In this example, there is a bidirectional connection between node N0 and node N1, and the connection size is +1. M-1 There is a range from N0 to N M-1 unidirectional connection, the connection size is -1. Node N0 is valid and node N1 is invalid. The method proposed in the present invention maps the adjacency matrix A to a 1T2R memristor array of size M×M. In the memristor array, the memristor unit with position coordinates (i, j) stores A in the adjacency matrix ij In the memristor array, both the (0, 1) and (1, 0) positions represent +1, the information stored in the position (0, 0) indicates that the node N0 is valid, and the information stored in the position (1, 1) indicates that the node N1 is invalid.

[0035] 2) Node addition: Figure 4 Indicates the change of the dynamic graph at the next moment. At this moment, the N1 node is added, and the cell at the (1, 1) position in the memristor array is operated, and both memristors in the cell are set to the HRS (high resistance) state. This state indicates that the N1 node has been added. During the data readout process, a low current is read out on the BLR1 data line, proving that the N1 node is valid.

[0036] 3) Node deletion: Figure 4 Shows the deletion of node N M-1 The process of operating the cell at position (M-1, M-1) in the memristor array sets both memristors in the cell to the LRS (low resistance) state. This state indicates that N M-1 The node has been deleted. During the data readout process, the BLR M-1 The high current is read on the data line, proving that N M-1 The node is invalid.

[0037] 4) Edge increase: Figure 4 Indicates the change of the dynamic graph at the next moment, from node N1 to node N M-1 The edge of the node is operated on the cell at position (1, M-1) in the memristor array, and the left side of the two memristors in the cell is set to HRS and the right side is set to LRS, that is, the cell is used to represent +1. This operation represents the N1 node to N M-1 The node's edge has been added with a connection size of +1.

[0038] 5) Edge deletion: Figure 4The process of deleting the edge from node N0 to node N1 is shown. The cell at position (0, 1) in the memristor array is operated, and the two memristors in the cell are set to HRS, HRS, that is, the cell is used to represent 0. This operation indicates that the edge from node N0 to node N1 has been deleted.

[0039] The above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. A person skilled in the art may modify or make equivalent substitutions for the technical solutions of the present invention without departing from the spirit and scope of the present invention. The protection scope of the present invention shall be subject to the claims.

Claims

1. A method for processing a dynamic graph based on a 1T2R memristor array, which maps the dynamic graph and the processing operation of the dynamic graph to the 1T2R memristor array and the programming operation of the 1T2R memristor array, wherein: The 1T2R memristor array comprises M×M 1T2R memristor units, each 1T2R memristor unit is connected to four lines BLL, WL, BLR and SL, wherein BLL is parallel to WL, BLR is parallel to SL, and the two sets of parallel lines are perpendicular to each other; the 1T2R memristor unit is composed of a transistor and two memristors L-RRAM and R-RRAM, one end of L-RRAM and R-RRAM is commonly connected to the drain end of the transistor, the other end of L-RRAM is connected to BLL, and the other end of R-RRAM is connected to BLR; the gate of the transistor is connected to WL; the source of the transistor is connected to SL; the 1T2R memristor array has a total of M WLs, M BLLs, M BLRs and M SLs; the 1T2R memristor units in a row share the same BLL and the same WL, and the 1T2R memristor units in a column share the same BLR and the same SL; For a dynamic graph with M nodes, the adjacency matrix A of the graph uses an M×M square matrix to store the entire graph, in which the element A with coordinates (i, j) ij Stores the connection relationship between node i and node j, that is, the edge between nodes i and j, and the element A with coordinates (i, i) ii Store the information of node i itself; map the adjacency matrix A of the dynamic graph to the 1T2R memristor array, map the operation of the nodes of the graph to the operation of the unit with coordinates (i, i) in the 1T2R memristor array, and map the operation of the edges of the graph to the operation of the unit with coordinates (i, j) in the 1T2R memristor array, wherein, i and j are integers from 0 to M-1, and i≠j.

2. The dynamic graph processing method according to claim 1, characterized in that: include: 1) Adjacency matrix mapping of the graph: In the 1T2R memristor array, the memristor unit with position coordinates (i, i) stores element A in the adjacency matrix A ii , i.e., the node state. The two memristors in the memristor unit are both HRS, i.e., the high resistance state, which means that the node is valid. The two memristors are both LRS, i.e., the low resistance state, which means that the node is invalid. The memristor unit with the position coordinate (i, j) stores the element A in the adjacency matrix A. ij , that is, the state of the edge, element A ij The size is -1, 0 or +1. ij When the size is 0, it means that there is no connection from node i to node j; in the memristor unit with the position coordinates (i, j), when the two memristors L-RRAM and R-RRAM are in the HRS state, it represents the number 0, when the L-RRAM is in the HRS state and the R-RRAM is in the LRS state, it represents the number +1, and when the L-RRAM is in the LRS state and the R-RRAM is in the HRS state, it represents the number -1; 2) Node addition: If the i-th node is added, the unit with the position coordinate (i, i) in the 1T2R memristor array is operated, and both memristors in the unit are set to the HRS state, indicating that the i-th node has been added; 3) Node deletion: If the i-th node is deleted, the unit with the position coordinate (i, i) in the 1T2R memristor array is operated, and both memristors in the unit are set to the LRS state, indicating that the i-th node has been deleted; 4) Edge addition: If an edge from the i-th node to the j-th node is added, the cell with the position coordinate (i, j) in the 1T2R memristor array is operated according to A ij Program the value of: If A ij = +1, then the two memristors L-RRAM and R-RRAM of the (i, j) unit are set to HRS and LRS states respectively, that is, the unit is used to represent +1; if A ij = -1, then the two memristors L-RRAM and R-RRAM of the (i, j) unit are set to LRS and HRS states respectively, that is, the unit is used to represent -1; 5) Edge deletion: If the edge from the i-th node to the j-th node is deleted, then the position coordinates in the 1T2R memristor array are (i, The unit (i, j) is operated to set both memristors of the unit (i, j) to the HRS state, that is, the unit is used to represent 0.

3. The dynamic graph processing method according to claim 2, characterized in that: In the data readout phase, an address i is selected in each clock cycle, and WL in the 1T2R memristor array i With BLL i is selected and set to a high level; all BLRs are grounded, and the analog-to-digital converter is used to read the current value passing through it; all SLs are floating, and the analog-to-digital converter is used to read the voltage value it has; one clock cycle can read out M bits of data in the i-th row, among which, if the i-th BLR is read as a high current, it means that the node i has been deleted, and the i-th row data read out this time is invalid; if the i-th BLR is read as a low current, it means that the node i has been added, and the i-th row data read out this time is valid; the selected address i is traversed from 0 to M-1, and all M rows of data of the 1T2R memristor array are read out, that is, all the data of the adjacency matrix A is read out.

4. The dynamic graph processing method according to claim 1, characterized in that: The memristor is a memristor based on a metal-insulator-insulator-metal structure, or a memristor based on a metal-semiconductor-semiconductor-metal structure.

5. The dynamic graph processing method according to claim 1, characterized in that: The transistor is an N-type metal oxide semiconductor field effect transistor.

Citation Information

Patent Citations

  • Resistive random access memory unit and preparation method

    CN111312747A

  • In-memory computing accelerator based on 1T2R resistive random access memory array and application thereof

    CN114783486A

  • Read-write reconfigurable memristor storage and calculation integrated system

    CN116741235A

  • High-order dynamic graph data storage method and device, equipment and storage medium

    CN117807277A

  • System, Method, and Computer Program Product for Dynamic Node Classification in Temporal-Based Machine Learning Classification Models

    US20240078416A1