Multiprocessor system fault diagnosis method based on directed tree structure

Abstract

The invention belongs to the field of multiprocessor systems, and discloses a multiprocessor system fault diagnosis method based on a directed tree structure. A fault of a processor in a multiprocessor system is inevitable, and how to quickly and accurately find out the fault is a key problem for keeping normal operation and stable performance of the system. According to the method, a directed tree structure is provided, a fault diagnosis method is provided based on the directed tree structure and according to a classic MM comparison diagnosis model, and whether the processor breaks down or not can be quickly and accurately diagnosed through the method. The structure is simple, can be easily embedded into a multiprocessor system, is low in algorithm complexity, and can quickly and accurately complete fault diagnosis of the processor.

Description

Technical Field

[0001] The invention belongs to the field of multi-processors, and in particular relates to a multi-processor system fault diagnosis method based on a directed tree structure. Background Art

[0002] Multiprocessor system is the core technology for high performance computing and is widely used in weather forecasting, oil exploration, biological information processing, etc. The reliability and fault tolerance of multiprocessor system are the key to ensure its normal operation and stable performance.

[0003] In high-speed multi-processor systems, processor failure is inevitable. When one or more processors fail, it may cause the system to crash, making it unable to operate normally or produce erroneous calculation results. Therefore, timely and effective detection of faulty processors (or components) in the system, and then repairing or replacing them to ensure the normal operation of the system has always been one of the research hotspots of multi-processor systems.

[0004] In a multiprocessor system, the process of detecting all faulty processors is called system diagnosis. If all processors in a system G can be accurately diagnosed as faulty when the number of faulty processors does not exceed t, then the system is called t-diagnosable. The maximum value of t that makes the system G t-diagnosable is called the diagnostic degree of G, denoted by t(G).

[0005] In the field of system diagnosis research, researchers have designed various models to identify faulty processors, such as the Maeng, Malek (MM) model (comparison model) and the Preparata, Metze, Chien (PMC) model. Maeng and Malek proposed the MM model (J.Maeng and M.Malek.A comparison connection assignment for self-diagnosis of multiprocessor systems.In Proceedings of 11th International Symposium on Fault-Tolerant Computing,pages 173–175,1981.), which sends the same task from the processor to two different neighbors and then compares their responses for diagnosis; by summarizing all comparisons, the system can determine the status of each processor. Preparata et al. proposed the PMC model (F.Preparata,G.Metze,andR.Chien.On the connection assignment problem of diagosis systems.IEEETransactions Electronic Computers,EC-16(6):848–854,1967.), which effectively completes the diagnosis of faults by interconnecting and testing adjacent processors.

[0006] In some cases, we may only focus on whether a specific processor (or component) in a multiprocessor system is faulty. Hsu and Tan proposed a new diagnosis definition (G.-H.Hsu and JJ-M.Tan.A local diagnosiability measure for multiprocessor systems.IEEE Transactions on Parallel and Distributed Systems, 18(5):598–607, 2007.), called local diagnosis, which means only diagnosing the status of a specific processor. The global diagnosis strategy requires collecting a large number of test results between processors in the system, which is highly complex. In comparison, the local diagnosis strategy has higher application value.

[0007] The topology of a multiprocessor system is usually represented as an undirected graph, reflecting the corresponding communication structure, where connections allow bidirectional data transfer. However, in practical implementations, bidirectional communication between two processors is usually implemented using two independent unidirectional channels, each channel handling data transfer in the opposite direction. In order to manage a large number of vertices, reduce hardware complexity, and address the challenges of manufacturing large-scale integrated circuits, researchers introduced the concept of directed networks. These networks, such as directed hypercubes and directed star graphs, provide an efficient way to organize communication by using unidirectional links. This approach simplifies the system architecture and enhances scalability.

[0008] From the basic definition of system diagnosis such as PMC model and comparison model, the diagnosis of one vertex on other vertices can be regarded as directed edges. Based on the comparison model, Lin et al. constructed a tree structure WB(u; t) and the corresponding local diagnosis algorithm LDA (C.-K.Lin, Y.-H.Teng, JJ-M.Tan, and G.-H.Hsu. Local diagnosis algorithms for multiprocessor systems under the comparison diagnosis model. IEEE Transactions on Reliability, 62(4):800–810, 2013.), through which the state of a vertex can be determined. However, this method is not applicable to directed graphs because each directed edge does not always have a corresponding opposite directed edge. The tree structure and LDA algorithm cannot be applied to directed topological structures such as directed hypercubes and directed star graphs. Summary of the invention

[0009] Aiming at the deficiencies of the existing diagnosis technology, the present invention proposes a multi-processor system fault diagnosis method based on a directed tree structure to solve the fault diagnosis problem in the multi-processor system.

[0010] The present invention discloses a multi-processor system fault diagnosis method based on a directed tree structure. According to the interconnection structure of the multi-processor system, a specific directed tree structure is constructed with the processor to be diagnosed as the root. A classic comparative diagnosis model is used to perform diagnostic tests in the directed tree structure, and the test results are collected. Then, the state of whether the processor is faulty or not is obtained by an algorithm. The specific steps are as follows:

[0011] Step 1, constructing a specific directed tree structure with the processor to be diagnosed as the root according to the interconnection structure of the multi-processor system;

[0012] Step 2, using the classic comparative diagnostic model, perform diagnostic tests in a directed tree structure and collect test results;

[0013] Step 3, executing the algorithm and deriving the state of whether the processor is faulty or not from the test results;

[0014] Step 4: Perform fault diagnosis on all processors in the multi-processor system in the same way to complete the system-level fault diagnosis of the multi-processor system and find all faulty processors.

[0015] Furthermore, in step 1, based on graph theory knowledge, the interconnection structure of the multi-processor system is represented by a graph G = (V, E), where the vertices in the graph correspond to the processors in the multi-processor system, and the edges correspond to the links between the processors. V and E are the vertex set and edge set in the graph, respectively, representing the set of processors in the multi-processor system and the set of links between the processors.

[0016] The symbol u represents the root of the directed tree structure, which represents the processor to be detected. p and q represent the number of Type T and Type C subgraphs in the directed tree structure, respectively. is the vertex in the tree, E1 and E2 are two edge sets in the tree; a specific directed tree structure DT(u,p,q) is given, where:

[0017]

[0018] E(DT(u,p,q))=E1∪E2

[0019]

[0020] The directed tree structure is embedded in a multi-processor system. Based on the structure, a diagnostic test is performed to determine whether the processor corresponding to vertex u is faulty or not.

[0021] Furthermore, in step 2, a comparative diagnostic model, namely the MM model, is used;

[0022] In the MM model, a test processor w sends the same test task to its two adjacent tested processors u and v at the same time, and then compares the test results fed back by them to identify the faulty processor; this test result is called a symptom, denoted by σ; σ w (u, v) represents the comparison result of the processor w on the output of the tested point pair {u, v}. If the output results are inconsistent, then σ w (u,v)=1, otherwise σ w (u,v)=0; if σ w (u,v)=1, which means at least one of u, v, w is faulty; if σ w (u,v)=0 and w is known to be fault-free, then both u and v are fault-free. If w is known to be faulty, the test result is unreliable.

[0023] In the directed tree structure DT(u,p,q), is the test point, and the test results will be obtained:

[0024]

[0025] Furthermore, in step 3, the diagnostic result is obtained by analyzing the test result; the directed tree structure contains two substructures of the state, Type T and Type C, and the comparison result is statistically analyzed by two variables g(u,i) and h(u,j), and the formula is as follows:

[0026]

[0027] Let F be the set of fault vertices in the directed tree structure, representing the set of fault processors distributed in the structure; let For 1≤i≤p,

[0028]

[0029] The derivation is proved as follows:

[0030] When g(u,i)=0, we get or if if thereby,

[0031] When g(u,i)=1 and u∈F, we get By u∈F and have to Depend on and have to thereby,

[0032] When g(u,i)=-1 and When At this time, if because have to thereby, if because and have to And then by have to thereby,

[0033] remember For 1≤j≤q,

[0034]

[0035] The derivation is proved as follows:

[0036] When h(u,j)=0, we get or if if thereby,

[0037] When h(u,j)=1 and u∈F, we get By u∈F and have to Depend on and have to thereby,

[0038] When h(u,i)=-1 and When At this time, if because have to thereby, if because and have to And then by have to thereby,

[0039] Now assume that p+q≥1, and the number of fault points in the directed tree structure is at most p+q;

[0040] Sum the values ​​of g(u,i) and h(u,i). If Vertex u is non-faulty, otherwise vertex u is faulty; its corollary proof is as follows:

[0041] Note that g0={i|g(u,i)=0}, g1={i|g(u,i)=1}, g2={i|g(u,i)=-1}, h0={j|h(u,j)=0}, h1={j|h(u,j)=1}, h2={j|h(u,j)=-1}.

[0042] Apparently remember From formula (1) and (2), we can get: g = |g1| - |g2|, h = |h1| -

[0043] |h2|, g+h=|g1|+|h1|-|g2|-|h2|; From formula (3) (4), when i∈g0, |F∩P i |≥1, when j∈h0, |F∩Q j |≥1; when g+h≥0, |g1|+|h1|≥|g2|+|h2|; if u is faulty, then when i∈g1, |F∩P i |≥2, when j∈h1, |F∩Q j |≥2; hence, |F|≥|g0|+2|g1|+|h0|+2|h1|+1≥|g0|+|g1|+|g2|+|h0|+|h1|+|h2|+1=p+q+1; based on the assumption |F|≤p+q, a contradiction is obtained, so u is non-faulty;

[0044] When g+h<0, |g1|+|h1|<|g2|+|h2|; if u is non-faulty, then when i∈g2, |F∩P i |≥2, when j∈h2, |F∩Q j |≥2; hence, |F|≥|g0|+2|g2|+|h0|+2|h2|>|g0|+|g1|+|g2|+|h0|+|h1|+|h2|=p+q; suppose |F|≤p+q, we get a contradiction, and thus u is faulty.

[0045] Furthermore, step 2 and step 3 collect the test results and obtain the diagnosis results. The algorithm framework is as follows:

[0046]

[0047]

[0048] The algorithm flow is as follows:

[0049] Input vertex u, the number of substructures Type T p, the number of substructures Type C q, the test results g(u,i) of p Type T structures, 1≤i≤p, and the test results h(u,j) of q Type C structures, 1≤j≤q; for 1≤i≤p, sum g(u,i) and record it as g; for 1≤j≤q, sum h(u,j) and record it as h; determine whether h+g≥0 holds. If so, output 0, indicating that vertex u is not faulty. Otherwise, output 1, indicating that vertex u is faulty.

[0050] The beneficial effects achieved by the present invention are:

[0051] (1) The present invention provides a local diagnosis structure D(u,p,q) and two test functions. On this basis, a local diagnosis algorithm for directed graphs is proposed. Under the comparative diagnosis model, when the number of fault vertices does not exceed p+q, the time complexity of the algorithm for diagnosing u is O(p+q). According to the algorithm, a directed graph with a local diagnosis structure D(u,p,q) is locally p+q diagnosable for vertex u.

[0052] (2) The structure and algorithm are designed based on directed graphs. For undirected graphs, each edge represents a bidirectional link between processors, which can be represented as two directed edges in different directions. The same method can be used to embed directed tree structures, collect test data, and call algorithms to complete system diagnosis. Therefore, the structure and algorithm can be well implemented in both directed and undirected graphs, and are more suitable for fault diagnosis of multi-processor systems.

[0053] (3) A large number of experiments show that when the fault probability of each vertex in the structure is 50%, the accuracy of the algorithm in diagnosing faulty elements is still above 70%. The directed tree structure constructed by this method, around the vertex to be diagnosed, combines the characteristics of the comparative diagnosis model, selects the minimum number of vertices and directed edges, and avoids the influence of more information; secondly, a good algorithm is constructed, and only necessary test results are collected to input the algorithm to avoid interference from other test results. This ensures the accuracy and efficiency of the diagnosis results. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 Schematic diagram of the Type T structure and diagnosis results in an embodiment of the present invention.

[0055] Figure 2 Schematic diagram of the Type C structure and diagnostic results in an embodiment of the present invention.

[0056] Figure 3 Schematic diagram of the directed tree structure D(u, p, q) in an embodiment of the present invention. DETAILED DESCRIPTION

[0057] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings.

[0058] This embodiment is based on a directed tree structure and a classic comparative diagnosis model (MM model) to propose a fault diagnosis method, by which the fault of the processor can be quickly and accurately diagnosed. The structure is simple, easy to embed into a multi-processor system, the algorithm complexity is low, and the fault diagnosis of the processor can be completed quickly and accurately.

[0059] According to graph theory, the topology of a multi-processor system is represented by a graph G. The vertex set V(G) of the graph represents the set of processors, and the edge set E(G) of the graph represents the set of links between processors. If the links between processors are bidirectional, the corresponding graph G is an undirected graph, and the corresponding edges are undirected edges. In an undirected graph, two vertices u and v are connected by an edge, denoted as {u, v}, that is, {u, v}∈E(G), and u is called an adjacent point of v. The degree d of vertex u is G (u) = |{{u, v}|{u, v}∈E(G)}|, which is the number of vertices adjacent to vertex u. If the link between processors is unidirectional, the corresponding graph G is a directed graph, and the corresponding edge is a directed edge. (u, v)∈E(G) represents a directed edge from u to v. For u, (u, v) is an outgoing edge, and v is u's outgoing neighbor. For v, (u, v) is an incoming edge, and u is v's incoming neighbor. For vertex u, its outgoing neighbor set is The set of incoming adjacent nodes is remember

[0060] In comparison-based diagnostic methods, the tester (also called comparator) acts as a third processor that provides certain inputs to the two processors and compares their output results. This comparison produces the same or different results between the outputs of the two processors. w (u,v) represents the comparison result between processor u and processor v, where:

[0061]

[0062] If σ w (u,v)=1, which means at least one of u,v,w is faulty. w (u,v)=0 and w is known to be fault-free, then both u and v are fault-free. If w is known to be faulty, the test result is unreliable and the status of u and v cannot be determined.

[0063] First, two structures Type T and Type C are given, such as Figure 1 and 2 As shown, all possible test results obtained in these two structures are listed and analyzed, as shown in Tables 1 and 2 below.

[0064] Table 1 All possible test results under Type T structure

[0065]

[0066] Table 2 All possible test results under Type C structure

[0067]

[0068] Combining structures Type T and Type C constructs a directed tree structure DT(u,p,q), such as Figure 3 shown.

[0069] based on Figure 3 The directed tree structure of is used to collect the test results and draw conclusions on whether vertex u is faulty or not through the algorithm shown in the following table.

[0070] Table 3 Local diagnosis algorithm flow of directed graph

[0071]

[0072] The detailed steps of the algorithm are:

[0073] (1) Input vertex u, the number of substructures of Type T p, the number of substructures of Type C q, and the test result D in substructure Type T T , test result D in substructure Type C T ;

[0074] (2) Classify and count the test results of the p Type T structures collected, that is, for 1≤i≤p, if p0 value increases by 1, if p1 value increases by 1;

[0075] (3) Classify and count the test results of the q Type C structures collected, that is, for 1≤j≤q, if The value of q0 is increased by 1. If The value of q1 increases by 1;

[0076] (4) Compare the size of p0+q0 and p1+q1. If p0+q0≥p1+q1, output 0, indicating that vertex u is not faulty. Otherwise, output 1, indicating that vertex u is faulty.

[0077] In the diagnosis structure, when the number of faulty elements is not greater than p+q, the state of u can be accurately diagnosed by the algorithm. When the number of faulty elements is greater than p+q, all the adjacent points of vertex u are faulty. At this time, the state of u cannot be obtained. In practical applications, the probability of this happening is almost 0. The simulation experiment once again verifies that the algorithm performs well when there are more faulty elements.

[0078] In the simulation experiment, assuming that the probability of a faulty element lying is 50%, for different values ​​of p and q, each vertex randomly generates a faulty or not state when the fault probability is r, and runs the diagnosis algorithm 1000 times. The diagnostic accuracy is the number of correct diagnoses / 1000. The following table lists the experimental results. It can be seen from the experiment that when the fault probability r = 0.2, the diagnostic accuracy is almost 100%, because in this case the number of faulty elements is close to p + q. Even if there are more faults, the diagnostic effect of the algorithm is good. Table 4 also shows that when the fault probability of each vertex in the structure is 50%, the diagnostic accuracy is still above 70%.

[0079] Table 4 Diagnostic accuracy of simulation experiments

[0080]

[0081]

[0082] The above description is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiment. Any equivalent modifications or changes made by ordinary technicians in this field based on the contents disclosed by the present invention should be included in the protection scope recorded in the claims.

Claims

1. A multiprocessor system fault diagnosis method based on a directed tree structure, characterized in that: In a multi-processor system, adjacent processors and links are selected around the processor to be diagnosed to form a directed tree structure. Comparative test diagnosis is performed in the directed tree structure and test results are collected. The test results are input into the algorithm to determine whether the processor is faulty or not. Then, the same method is used to diagnose the status of other processors to complete system-level fault diagnosis. The specific steps include: Step 1, constructing a directed tree structure with the processor to be diagnosed as the root according to the interconnection structure of the multi-processor system; Step 2, using the Maeng and Malek comparative diagnostic model, perform diagnostic tests in a directed tree structure and collect test results; Step 3, deriving the state of whether the processor to be diagnosed is faulty or not through an algorithm based on the test results; Step 4, using the method of step 3 to perform fault diagnosis on all processors in the multi-processor system in turn, complete system-level fault diagnosis of the multi-processor system, and find all faulty processors.

2. The multiprocessor system fault diagnosis method based on a directed tree structure according to claim 1, characterized in that: In step 1, the topological structure of the multi-processor system is represented by a graph G = (V, E), where the vertices in the graph correspond to the processors in the multi-processor system, and the edges correspond to the links between the processors. V and E are the set of vertices and the set of edges in the graph, respectively, representing the set of processors in the multi-processor system and the set of links between the processors. Depending on whether the links between the processors in the multi-processor system are bidirectional or unidirectional, the corresponding graphs are undirected graphs and directed graphs, respectively.

3. The multiprocessor system fault diagnosis method based on a directed tree structure according to claim 2 is characterized in that: If the links between processors are bidirectional, the corresponding graph G is an undirected graph, and the corresponding edges are undirected edges. In an undirected graph, if two vertices u and v are connected by an edge, denoted as {u, v}, that is, {u, v}∈E(G), u is called an adjacent point of v. The degree of vertex u is d G (u) = |{{u, v}|{u, v}∈E(G)}|, i.e., the number of vertices adjacent to vertex u; If the link between processors is unidirectional, the corresponding graph G is a directed graph, and the corresponding edge is a directed edge; (u, v) ∈ E (G) represents a directed edge from u to v. For u, (u, v) is an outgoing edge, and v is u's outgoing neighbor. For v, (u, v) is an incoming edge, and u is v's incoming neighbor. For vertex u, its outgoing neighbor set is The number of vertices in the set is u, and the out-degree is denoted by The set of incoming adjacent nodes is The number of vertices in the set is u, denoted by Right now 4. The multiprocessor system fault diagnosis method based on a directed tree structure according to claim 3 is characterized in that: A directed tree structure is embedded in the interconnection structure of a multiprocessor system, that is, a subgraph of a graph G is taken, and the subgraph is a specific directed tree structure; firstly, two graph structures Type T and Type C are given as two forms of subgraphs in the directed tree structure; The graph structure Type T contains 7 vertices and 6 directed edges, V(TypeT) = {u, y1, y2, y3, y4, y5, y6}, E(TypeT) = {(y1,u), (y1,y2), (y3,y2), (y3,y4), (y5,y1), (y5,y6)}; the graph structure Type C contains 6 vertices and 6 directed edges, V(TypeC) = {u, z1, z2, z3, z4, z5}, E(TypeC) = {(z1,u), (z1,z2), (z3,z2), (z3,z4), (z5,z1), (z5,z3)}.

5. The multiprocessor system fault diagnosis method based on a directed tree structure according to claim 4 is characterized in that: For a directed tree structure, the symbol u is used to represent the root of the directed tree, which represents the processor to be tested; p and q represent the number of Type T and Type C substructures in the directed tree structure, that is, the number of subgraphs of these two forms in the directed tree; the set and The symbol in represents all vertices in the tree except u; E1 and E2 are two edge sets in the tree, which contain edges in the Type T structure and edges in the Type C structure respectively; the specific description of the directed tree structure DT(u,p,q) is as follows: E(DT(u,p,q))=E1∪E2 The directed tree is embedded in a multi-processor system. Based on this structure, diagnostic tests are performed to determine whether the processor corresponding to vertex u is faulty or not.

6. The multiprocessor system fault diagnosis method based on a directed tree structure according to claim 5, characterized in that: In step 2, MM was used to compare diagnostic models; In the MM comparison diagnosis model, a vertex w representing a test processor sends the same test task to its two adjacent vertices u and v representing the tested processors at the same time, and then compares the test results fed back by them to identify the faulty processor; this test result is called a symptom, denoted by σ; σ w (u, v) represents the comparison result of the output of vertex w to the tested point pair {u, v}. If the output results are inconsistent, then σ w (u,v)=1, otherwise σ w (u,v)=0; if σ w (u,v)=1, which means at least one of u, v, w is faulty; if σ w (u,v)=0 and w is known to be fault-free, then both u and v are fault-free. If w is known to be faulty, the test result is unreliable. In the directed tree structure DT(u,p,q), is the test point, and the test results will be obtained:

7. The method for diagnosing multiprocessor system faults based on a directed tree structure according to claim 6, characterized in that: In step 3, the diagnostic results are obtained by analyzing the test results. The directed tree structure contains two substructures, Type T and Type C. The test results are statistically analyzed by two variables g(u,i) and h(u,j). The formula is as follows: Let F be the set of fault vertices in the directed tree structure, representing the set of fault processors distributed in the structure; let the vertex set For 1≤i≤p, remember For 1≤j≤q, Now assume that p+q≥1, and the number of fault points in the directed tree structure is at most p+q; Sum the values ​​of g(u,i) and h(u,i). If Vertex u is non-faulty, otherwise vertex u is faulty.

8. The multiprocessor system fault diagnosis method based on a directed tree structure according to claim 7 is characterized in that: In step 3, the process of collecting test results to derive a diagnosis result is as follows: Input vertex u, the number of Type T substructures p, the number of Type C substructures q, the test results g(u,i) of p Type T structures, 1≤i≤p, and the test results h(u,j) of q Type C structures, 1≤j≤q; For 1≤i≤p, sum g(u,i), denoted as g; For 1≤j≤q, sum h(u,j), denoted as h; Determine whether h+g≥0 is true. If so, output 0, indicating that vertex u is not faulty. Otherwise, output 1, indicating that vertex u is faulty.

Images