Efficient fault-tolerant broadcasting method for dragonfly network

Abstract

The invention provides an efficient fault-tolerant broadcasting method for a dragonfly network. The method comprises the following steps: selecting any node as a root node r; dividing the construction process of the vertex independent spanning tree into two situations according to whether a first sub-node connected with a root node r in the spanning tree is an intra-group node or an out-group node; respectively generating n + h-1 vertex independent spanning trees which take r as a root and cover all vertexes in the network according to a preset construction process of the corresponding situation; and broadcasting data by using the spanning trees. According to the method, based on the topological characteristics of'in-group full interconnection and inter-group regular connection 'of the dragonfly network, unified VISTs construction logic is designed in combination with the types (in-group nodes and out-group nodes) of first-hop connection nodes of root nodes, a construction algorithm does not need to be designed for different values of the dragonfly network, and efficient broadcasting and safe distribution of data can be achieved.

Description

An efficient fault-tolerant broadcasting method for dragonfly networks Technical Field

[0001] This invention relates to the field of high-performance computing interconnect networks, and more specifically to an efficient fault-tolerant broadcasting method for dragonfly networks. Background Technology

[0002] High-performance computing (HPC) systems support complex applications such as scientific computing, engineering simulation, and artificial intelligence training through massively parallel processing capabilities. Their overall performance is largely constrained by the communication efficiency, reliability, and scalability of the underlying interconnect network. The interconnect network not only needs to support point-to-point communication but also needs to efficiently perform collective communication operations such as broadcasting, reduction, and synchronization. Among these, broadcasting is the most fundamental and frequent communication mode, directly impacting system startup, task scheduling, and the efficiency of parallel algorithm execution.

[0003] As the system scale expands from tens of thousands of nodes to hundreds of thousands or even millions of nodes, traditional low-port density, regular topologies (such as two-dimensional / three-dimensional meshes, fat trees, etc.) have gradually exposed problems such as complex wiring, high cost, and large diameter. To this end, the academic and industrial communities have proposed a variety of new interconnection network structures. Among them, the Dragonfly Network has become one of the most representative interconnection topologies in the current ultra-large-scale HPC system due to its comprehensive advantages of "low diameter, high bandwidth utilization, and high scalability". The Dragonfly Network adopts a hierarchical design idea: the switches within the group are fully interconnected to ensure local communication efficiency, and the groups are interconnected through a controlled number of global links. In theory, communication between any two nodes can be completed with a maximum of only three hops. The Dragonfly Network has become a research hotspot in recent years. Its main properties and some related research results are as follows: (1) Dragonfly Network It is an (n+h-1)-regular graph; (2) have Each node and Strip edge; (3) diameter (4) Any two nodes and There are (n + h - 1) disjoint paths; (5) for Any two distinct nodes and The DFDP algorithm can construct (n + h - 1) disjoint paths between them, where the maximum length of these paths is 7, and the time complexity of the algorithm is O(n). (6) When hour, There exists a Hamiltonian cycle in it, when hour, It is Hamiltonian connected; (7) There is one and only one edge between any two groups; (8) Dragonfly networks can be constructed using the CISTODN algorithm. of The set of edges of 16 completely independent spanning trees has a time complexity of O(log n). , for The number of vertices.

[0004] In dragonfly networks, broadcast communication typically relies on tree structures or hop-by-hop propagation mechanisms based on routing. However, as network size and concurrency increase, broadcast methods using single paths or single spanning trees gradually become insufficient to meet performance and reliability requirements: on the one hand, broadcast load is concentrated on a limited number of links, easily forming bottlenecks; on the other hand, the failure of any link or node can lead to broadcast failure or a significant increase in latency. Therefore, how to design an efficient and fault-tolerant broadcast mechanism in dragonfly networks has become an important research topic in the fields of interconnection networks and parallel communication.

[0005] In graph theory, Vertex Independent Spanning Trees (VISTs) serve as an important fault-tolerant structure, providing multiple non-overlapping communication paths between any pair of nodes in a network. Theoretical studies show that if multiple VISTs exist in a network, even if some nodes fail, communication can still be completed through paths in the remaining spanning trees, thus significantly improving the network's fault tolerance. However, for dragonfly networks, there is currently a lack of mature and feasible technical solutions for systematically constructing multiple VISTs. Therefore, it is necessary to propose a method for constructing VISTs for dragonfly networks to overcome the shortcomings of existing technologies. Summary of the Invention

[0006] The objective of this invention is achieved through the following technical solutions.

[0007] This invention relates to Dragonfly Networks Construction schemes for VISTs on other networks. Currently, VIST construction strategies on other networks lack specificity for Dragonfly networks. The existing methods for constructing VISTs cannot be directly reused in dragonfly networks due to their lack of adaptability and insufficient consideration of the hierarchical topology characteristics of dragonfly networks, which are characterized by "full interconnection within groups and regular connections between groups". Therefore, the complete construction logic needs to be redesigned.

[0008] Specifically, this invention provides an efficient fault-tolerant broadcasting method for dragonfly networks, comprising: an efficient fault-tolerant broadcasting method for dragonfly networks, wherein the dragonfly network is D(n, h), where n is the number of nodes in each group, and h is the number of global links for each node, characterized by comprising the following steps: selecting any node in the dragonfly network as the root node r; dividing the construction process of the vertex independent spanning tree into a first case and a second case according to the type of the first child node connected by the root node r in the spanning tree; wherein, in the first case, the first child node connected by the root node r is a node within the same group as it; in the second case, the first child node connected by the root node r is a node outside the same group as it; generating n+h-1 vertex independent spanning trees with the root node r as the root and covering all vertices in the dragonfly network according to the preset construction process corresponding to the first case or the second case respectively; and using the generated n+h-1 vertex independent spanning trees for data broadcasting.

[0009] The advantages of this invention are: This invention is based on Dragonfly Network. Based on the topological characteristics of "full interconnection within groups and rule-based connections between groups," and combined with the type of first-hop connection node of the root node (intra-group node, inter-group node), a unified VISTs construction logic is designed, eliminating the need for different approaches for different Dragonfly networks. , The algorithm for designing repeated values ​​can achieve efficient broadcasting and secure distribution of data. Attached Figure Description

[0010] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: FIG1 is a schematic diagram of the construction process of the first scenario of the invention (root node connecting nodes within a group).

[0011] Figure 2 is a schematic diagram of the construction process of the second scenario of the present invention (root node connects to external nodes of the group).

[0012] Figure 3 is a topological structure diagram of D(3, 2) in a specific embodiment of the present invention.

[0013] Figure 4 is a schematic diagram of the four independent spanning trees T0, T1, T2 and T3 constructed in D(3, 2) shown in Figure 3.

[0014] Figure 5 is a comparison of the transmission failure rate experimental results of the present invention when n=6 and h=3.

[0015] Figure 6 is a comparison of the transmission failure rate experimental results of the present invention when n=7 and h=3.

[0016] Figure 7 is a comparison of the transmission failure rate experimental results of the present invention when n=8 and h=3.

[0017] Figure 8 is a comparison of the transmission failure rate experimental results of the present invention when n=9 and h=3. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0019] The purpose of this invention is to provide a dragonfly network. A method for constructing an independent spanning tree of vertices. A dragonfly network is constructed using the VIST-dragonfly algorithm. The vertices on the tree are used to generate independent spanning trees, enabling parallel transmission of data through multiple paths, thus improving transmission efficiency and fault tolerance.

[0020] According to the definition of an independent spanning tree, to maximize the number of constructible spanning trees, the root node of each independent spanning tree... The degree of the root node must be 1, meaning the root node is connected to only one child node. Therefore, each spanning tree uniquely corresponds to a distinct adjacent edge to the root node. Furthermore, due to the strict symmetry at the vertex level of the dragonfly network, without loss of generality, any node in the network can be chosen as the root node. .

[0021] Set the root node The group contains Each node is configured with [number] nodes. A global link. Based on the type of the first child node connected to the root node in the spanning tree, this invention divides the construction process of VISTs into two cases: the root node connecting to nodes within the group and the root node connecting to nodes outside the group. Corresponding construction processes are designed for each case, ultimately generating a global link. The construction process for the two cases is described in detail below: a vertex independent spanning tree with the root as the root and covering all vertices of the network.

[0022] Scenario 1: The root node connects to nodes within the same group. (This refers to the root node in the spanning tree.) Connect to its group A node within At that time, the spanning tree is formed by edges ( The starting edge is expanded. Because the group uses a fully interconnected topology, the nodes... In the group Inside and outside The rest All nodes are directly adjacent. During construction, the edges are first... Add them sequentially to the spanning tree to achieve grouping. Redundant coverage for all nodes within the system.

[0023] Subsequently, the node pass Each global link is respectively with 1 target group (denoted as Establish unique connections. Let the endpoints of these global links on the target group side be respectively... During the construction process, the edges are first... Add to the spanning tree. Then, with each As its group Local expansion nodes. Because each group has a fully interconnected topology, the nodes... It is directly adjacent to all other nodes in the group, therefore... By adding them sequentially to the spanning tree, the group can be completed. All nodes are covered.

[0024] In the group containing the root node and its direct connections After the target group, there are still One group was not visited. Unvisited groups were then grouped into the target group set in numerical order. Distributed evenly and periodically in a fixed order; specifically, the first The unvisited groups are numbered sequentially in a cyclical fashion. The target group is responsible for making connections. Taking advantage of the property in a dragonfly network that there is one and only one global link between any two groups, each group uniquely determines a cross-group link for each relationship and adds this link to the spanning tree. This periodic allocation ensures that each target group corresponds to... This allows for access to all remaining groups at the group level by identifying different unaccessed groups.

[0025] For each group accessed via a cross-group link, the endpoint of that link within the group is used as a local expansion node. Utilizing the fully interconnected topology within the group, the edges between this node and the remaining nodes in the group are added to the spanning tree, achieving coverage of all vertices within the group. This ultimately results in a spanning tree... A vertex-independent spanning tree is constructed with the root as the root and covering all vertices. The construction process is shown in Figure 1.

[0026] Scenario 2: Root node connects to nodes outside the group when the root node in the spanning tree is... Connected to nodes in another group via a global link. At that time, the spanning tree is formed by edges ( Let ) be the starting edge. Let node... The group I belong to is Because each group in a dragonfly network uses a fully interconnected topology, nodes... In the group Inside and the rest All nodes are directly adjacent. During construction, the edges are first... Add them sequentially to the spanning tree to achieve grouping. Redundant coverage for all nodes within the system.

[0027] After completion After expanding within a group, utilize the group Nodes in the middle outside The global links corresponding to each node will further extend the spanning tree to other nodes. Groups. Based on the inter-group connection structure of the dragonfly network, group numbering can be considered as modular. The arrangement is cyclic, and the global connections between groups are made with a fixed step size. The intervals are symmetrically distributed around the current group. Therefore, with... Directly connected groups are located in the following positions in terms of numbering: (mold In a sense, it forms a shape that unfolds symmetrically on both sides. There are two distinct groups.

[0028] Group In Each node is connected to the above-mentioned node via its unique global link. Each group is a set of nodes. For each newly joined group, the endpoint of the cross-group link within that group is used as a local extension node, and the coverage of all nodes within the group is completed using the fully interconnected topology within the group. Thus, together with... Within this, the spanning tree has been covered. Groups.

[0029] This was then expanded at the group level. Based on the already covered groups, establish inter-group connections between each group and all unvisited groups. In other words, each unvisited group is connected to this... Each of the covered groups will have one connection.

[0030] According to the regularized inter-group connection structure of the Dragonfly Network, there is at most one global link between any two groups. Therefore, once the connection relationships between groups are determined, the corresponding cross-group links are uniquely determined. Each group extends to all unvisited groups, and each node in an unvisited group is directly connected to the spanning tree through its global link.

[0031] Therefore, after this round of inter-group connections, all nodes in the network have been incorporated into the spanning tree. The final result is a tree with... A vertex-independent spanning tree is constructed with the root as the root and covering all vertices. The construction process is shown in Figure 2.

[0032] by For example, below is The topology is shown in Figure 3.

[0033] exist A total of four independent spanning trees were constructed. Two of them belong to case one (the root node connects to nodes within the group); the other two belong to case two (the root node connects to nodes outside the group). The following section will explain this in conjunction with the diagram. With four spanning trees As shown in Figure 4, the specific construction process for the two cases is explained respectively.

[0034] The two spanning trees correspond to case one, with node 1 as the root node. their group .

[0035] structure Root node Connect node 2 within the group, starting with edge (1,2). First, add the edge (2,0) between node 2 and the remaining node 0 within the group. This completes the connection. Full coverage; then node 2 passed A global link connects to the target group (Endpoint 18) and (Endpoint 15), after adding edges (2,18) and (2,15), complete the expansion with 18 and 15 as local expansion nodes. , Within the group, cover; then, sort the remaining 4 unvisited groups by... Rules assigned , After accessing through cross-group links, intra-group expansion is completed, ultimately resulting in... .

[0036] structure Root node Connect node 0 within the group, starting with edge (1,0), and complete the process in the same manner. Full coverage, target group expansion, and access to remaining groups are achieved. .

[0037] For the second scenario involving two spanning trees, node 1 is still the root node. .

[0038] structure Root node Connect to external node 13 (the group it belongs to) via a global link. Starting with edge (1,13), first add the edges (13,12) and (13,14) between node 13 and the other nodes 12 and 14 in the group, and then complete the process. Full coverage; then utilize Except for 13 The global links of nodes (12, 14), with step size Symmetrically extend to two other distinct groups, and after completing intra-group coverage, use the already covered data. Based on each group, connections are established with all unvisited groups to complete the full vertex coverage. .

[0039] structure Root node Connect to external node 10 (in its own group) Following the same process of full coverage within a group, symmetrical group expansion, and inter-group connectivity across the entire network, the following results were obtained: .

[0040] Through the construction of the above two scenarios, in The result yielded four independent spanning trees rooted at node 1, which perfectly matches the given condition. The theoretical number of trees, and the paths between any two trees have no shared vertices except the root node.

[0041] For any This invention proposes in The above constructs a structure with any node as the root node. A unified construction method for VISTs, thereby obtaining the distance between two points. A single-vertex independent path was used. To ensure the statistical validity of the experimental results, an adaptive stepped fault point generation strategy was adopted, generating gradient fault points with a step size of 10% and an upper limit of 70% of the total number of non-root nodes. Fault vertices were randomly selected from non-root nodes, and each fault scale was repeated 1000 times with random sampling. One-to-one communication was simulated in the Dragonfly network under vertex fault conditions. During communication, the vertex-independent paths in VISTs were used as communication channels. The number of failures in 1000 communication attempts was counted, and the ratio of the number of failures to the total number of communication attempts was calculated as the Transmission Failure Rate (TFR). Two comparison schemes were set up: TFR1 represents the transmission failure rate using a single spanning tree path, and TFR represents the transmission failure rate using a single spanning tree path. The transmission failure rate of VISTs multipath communication was measured, and then the fault tolerance performance of the Dragonfly Network was evaluated using TFR when communicating with VISTs.

[0042] This experiment was conducted in Under these circumstances, the number of cross-group connections in a Dragonfly network is fixed. The corresponding total number of nodes are 114, 154, 200, and 252, respectively. The experimental results are shown in Figures 5, 6, 7, and 8. Based on the experimental results, the following conclusions can be drawn: (1) The TFR increases with the number of fault points, and the TFR of the single-channel communication method using a single spanning tree is significantly higher than that of the single-channel communication method using a single spanning tree. Multi-channel communication method for VISTs.

[0043] (2) In use When multiple VISTs communicate via multiple channels, even if there are [various issues] in the path... Even if a path fails, communication can still be successful; and as the network size increases, the number of independent vertex channels available increases, and the number of faulty vertices that the network can accommodate under the same TFR also increases.

[0044] With the same number of faulty vertices, TFR varies with network size. The TFR decreases sharply as the network size increases because the blocking effect of the faulty vertex on the independent path of the global vertex is diluted, and the probability of all communication paths failing is greatly reduced, thus causing a significant decrease in TFR.

[0045] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An efficient fault-tolerant broadcasting method for dragonfly networks, wherein the dragonfly network is D(n,h), where n is the number of nodes in each group and h is the number of global links for each node, characterized in that, Includes the following steps: Select any node in the dragonfly network as the root node r; according to the type of the first child node connected to the root node r in the spanning tree, the construction process of the vertex independent spanning tree is divided into a first case and a second case; wherein, in the first case, the first child node connected to the root node r is a node within the same group as the root node; in the second case, the first child node connected to the root node r is a node outside the same group as the root node. Following the preset construction process corresponding to the first scenario or the second scenario respectively, generate n+h-1 independent vertex spanning trees with the root node r as the root and covering all vertices in the dragonfly network; use the generated n+h-1 independent vertex spanning trees to perform data broadcasting.

2. The method according to claim 1, characterized in that, The preset construction process for the first scenario includes: Step S11: Add the edge (r,v) within the group G0 containing the root node r to the spanning tree, where v is a node within the group; Step S12: Add the edges between node v and all other nodes in group G0 except r and v to the spanning tree, completing the coverage of group G0; Step S13: Determine the h target groups G1, G2, ..., Gh connected to node v through h global links, and the endpoint vi in ​​each target group, and add the edge (v,vi) to the spanning tree; Step S14: For each target group Gi, use endpoint vi as a local expansion node. Step S15: According to the preset allocation rules, the remaining gh-1 unvisited groups are allocated to the h target groups, and each target group is connected to the allocated unvisited group through a unique global link, and the corresponding cross-group link is added to the spanning tree; Step S16: For each unvisited group accessed through a cross-group link, the endpoint of the cross-group link in the unvisited group is used as the local extension node, and the edge between the endpoint and all other nodes in the group is added to the spanning tree, thus completing the coverage of all vertices.

3. The method according to claim 2, characterized in that, The preset allocation rule in step S15 is as follows: according to the numbering order of the unvisited groups, they are periodically and evenly allocated to the h target groups, wherein the kth unvisited group is connected by the target group numbered (kmodh).

4. The method according to claim 1, characterized in that, The preset construction process for the second scenario includes: Step S21: Add the edge (r,v) of the node v outside the group connected to the root node r through the global link to the spanning tree, where the group v belongs to is G1; Step S22: Add the edge between node v and all other nodes in group G1 to the spanning tree to complete the coverage of group G1; Step S23: Use the global links of the n-1 nodes in group G1 other than v to connect to another n-1 distinct groups respectively, and add the corresponding cross-group links to the spanning tree; Step S24: For each newly accessed group, take the endpoint of the cross-group link in the group as the local expansion node, and add the edge between the endpoint and all other nodes in the group to the spanning tree; Step S25: Based on all currently covered groups, make each covered group establish a unique connection with all unvisited groups through its unused global links, add the corresponding cross-group links to the spanning tree, and complete the coverage of all vertices.

5. The method according to claim 4, characterized in that, In step S23, the n-1 distinct groups directly connected to group G1 have their group numbers symmetrically distributed with a step size h in the sense of modulus g.

6. The method according to any one of claims 1 to 5, characterized in that, Among the n+h-1 independent spanning trees generated, no two spanning trees share any vertices except for the root node r.

Images