Method and apparatus for replica placement based on dynamic network topology

Abstract

The application provides a replica placement method and device based on a dynamic network topology, which discretizes a time axis into time slots and non-overlapping time windows, constructs a dynamic topology and a time-varying request mathematical model, and completes replica planning in three stages: in the first stage, an initial replica set is constructed with coverage priority to achieve full request reachability, ensuring full coverage of services under network fluctuations; in the second stage, replicas are supplemented based on marginal delay gain greediness to reduce the window-weighted total access delay; in the third stage, an epsilon relaxation constraint is introduced to reverse and simplify redundant replicas, and replica synchronization overhead is controlled; the application solves the problems of insufficient reachability of existing edge replica placement, uncoordinated optimization of multiple targets, and excessive online calculation overhead, and takes into account the three indicators of reachability, delay and replica quantity, the greedy iteration architecture has controllable calculation amount, and is suitable for real-time deployment of cloud edge edge cluster dynamic topology.

Description

Technical Field

[0001] This invention relates to the field of computer edge distributed storage technology, and in particular to a method and apparatus for placing replicas based on dynamic network topology. Background Technology

[0002] With the widespread adoption of edge cluster collaborative computing, the scale of remote sensing data at the edge has surged. Edge-edge collaborative distributed file systems have become the core foundation for data storage and sharing, and the placement of replicas directly determines the reliability of edge data access and transmission efficiency. Existing edge replica placement strategies have a core flaw: they cannot simultaneously guarantee the reachability of all requests within a time window, low access latency, and low replica maintenance overhead under dynamic network topologies, and are difficult to adapt to edge scenarios with link jitter, node movement, and network partitioning.

[0003] The four traditional methods—static topology, hotspot centers, predictive learning, and online greedy algorithms—all have shortcomings: static and hotspot strategies cannot adapt to time-varying topologies and are prone to service gaps; predictive algorithms suffer from error propagation and high edge deployment costs; conventional greedy algorithms only optimize latency and lack a controllable replica simplification mechanism for latency escalation, making it impossible to systematically balance reachability, latency, and the number of replicas, and the overhead of online iterative computation and replica synchronization is difficult to manage, thus failing to meet the real-time operation requirements of edge services SLA. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a three-dimensional horizontal wind field inversion method and apparatus based on multi-source radar data. It can solve the problems of insufficient reachability of edge replica placement, inability to coordinate optimization of multiple targets, and excessive online computing overhead. It takes into account the three indicators of reachability, latency, and number of replicas. The greedy iterative architecture has controllable computational load and is adapted to the real-time deployment of dynamic topology of cloud-edge-device edge clusters.

[0005] To achieve the above objectives, the following technical solution is adopted:

[0006] This invention provides a method for replica placement based on dynamic network topology, comprising the following steps:

[0007] Step 1, construct the spatiotemporal system model of the edge cluster: Discretize the time axis into continuous time slots, divide the time decision windows into non-overlapping ones, construct a dynamic topology map for each time slot, and collect data on file request frequency, remaining storage capacity, and link bandwidth of each node within the window;

[0008] Step 2, construct the initial feasible replica set with full coverage priority: filter candidate nodes whose storage capacity of the window's full time slots meets the file storage requirements, iteratively select candidate nodes with the largest weighted coverage request scale and the lowest average access latency to add to the replica set, until all requests within the window can be served by replica nodes, and obtain the initial replica set that satisfies the full request reachability constraint;

[0009] Step 3, greedy replacement of replicas with marginal latency gain: Iterate through the unselected candidate nodes, calculate the window-weighted total latency reduction brought by adding the node, select the node with the largest latency gain in each round to add to the replica set and update the service mapping and latency matrix, and stop replacement when there is no gain;

[0010] Step 4, relax the constraints and reduce redundant replicas: Calculate the window-weighted total latency, and try to remove single replica nodes one by one. Match the original service requests of the node with replacement replicas. If all requests have replacement replicas and the total latency after deletion does not exceed (1+ϵ) times the original total latency, then remove the replica node. Iterate until there are no nodes to reduce, and output the final replica placement scheme, service mapping matrix and latency matrix. Use the time window as the granularity to uniformly characterize dynamic topology and time-varying requests. First, ensure that all requests are reachable, then optimize access latency, and finally reduce the number of replicas in a controllable manner through the relaxation of ϵ. This simultaneously solves the problems of unreachability of dynamic edge network requests, inability to coordinate optimization of multiple objectives, and excessive replica maintenance overhead.

[0011] Preferably, in step 1, the constructed dynamic topology graph is denoted as...

[0012] ;

[0013] In the time slot The network state is assumed to remain constant; V(t) represents the time slot. A set of communicable storage nodes For time slots Available link set This is the set of remaining storage capacity for the nodes. Define the set of available bandwidth for the link; define a binary variable. , representing whether node i stores a copy of file f, a binary variable. The file request of node i is served by node j, and the copy placement variable remains constant within the same time window.

[0014] Preferably, in step 1, the latency model includes single-hop transmission latency and propagation latency. The end-to-end latency of node i served by replica node j is the sum of the latency of all links along the shortest path between the two nodes. The window-weighted total latency is the sum of the latency of all requests accessed, weighted by the frequency of each node's time slot request.

[0015] Preferably, the rule for selecting candidate nodes in step 2 is as follows: select edge devices whose remaining storage capacity is greater than the target file size in all time slots of the current window as candidate nodes. , As the starting point of the window, The time window length, This represents the relative time slot offset within the window.

[0016] Preferably, in step 2, the evaluation criteria for selecting nodes include weighted coverage size. Weighted average delay Weighted coverage size The total number of covered but uncovered requests for node j is accumulated based on request frequency. When multiple candidate nodes have the same weighted coverage size, the weighted average latency is selected. Smaller nodes are added to the replica set.

[0017] Preferably, in step 3, the delay reduction of node j is calculated as follows: ;

[0018] Each round of selection The largest node is added to the result set, and the service mapping of the improved request is updated. With optimal delay ;

[0019] Accumulate all nodes For requests to reduce latency, the original optimal latency and node latency are calculated using request frequency as the weight. The sum of the differences in service latency. The larger the value, the more significant the latency reduction effect of adding this copy.

[0020] Preferably, in step 4, the frequency-weighted total delay within the window is calculated first. Then iteratively scan the current replica set. For any replica node Construct its service request set ;

[0021] Delete replica node First, obtain the replica node. The entire request set for a single service; if the request set is empty, delete the node directly. If a request has no reachable alternative replica, then the node is retained. .

[0022] Preferably, in step 4, the deletion cost is calculated. ϵ is a preset latency easing threshold, which is applied only if the total latency increment after removing a replica node satisfies the condition. At that time, the replica node deletion operation is performed, where To remove a node The resulting increase in total latency.

[0023] This invention provides a replica placement apparatus based on dynamic network topology, used to execute the aforementioned replica placement method based on dynamic network topology, comprising:

[0024] The modeling module is used to discretize time slots and divide time windows to build dynamic topology models and request latency models.

[0025] The coverage construction module is used to filter candidate nodes and iteratively generate a feasible solution with full coverage initial replicas;

[0026] The latency optimization module is used to calculate the marginal latency gain of each candidate node and greedily add replicas to reduce the total access latency.

[0027] The replica reduction module is used to reduce redundant replicas in reverse based on the ε relaxation constraint, and outputs the final replica placement and service mapping results.

[0028] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described copy placement method based on dynamic network topology.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] 1. Guaranteeing full request reachability in dynamic networks, significantly reducing service failure rates. This invention innovatively introduces a time window and time slot discrete modeling mechanism. In the first stage of the algorithm, it prioritizes constructing replica feasible solutions covering all requests within the window, forcibly satisfying the minimum number of replicas for each connected component. This fundamentally solves the service gaps, request unreachability, and frequent retries that occur in traditional static topology and hotspot replica strategies under scenarios of link fluctuations, node disconnections, and network partitioning. It significantly improves the reliability of data access in edge clusters and is suitable for topology-time-varying scenarios such as mobile edge devices and wireless remote sensing.

[0031] 2. Achieve a coordinated and balanced optimization of reachability, latency, and replica quantity, adapting to various service SLAs. Existing algorithms only optimize average latency or hit rate, failing to consider multiple constraints. This invention adopts a three-stage layered optimization architecture of "coverage construction - latency gain supplementation - ε relaxation simplification." It first ensures basic service capabilities, then reduces access latency, and finally, controls the relaxation of latency indicators through a configurable ε relaxation threshold to reduce the number of replicas. It can flexibly balance replica storage and synchronization overhead according to the service's latency tolerance, possessing interpretable and adjustable multi-objective trade-off capabilities, adapting to edge services with different latency and storage cost requirements.

[0032] 3. The algorithm computation and replica maintenance overhead are controllable, enabling lightweight deployment at the edge. This invention employs a lightweight solution logic of greedy iteration and local updates throughout the process, eliminating the need for complex deep learning model training and inference, and avoiding the frequent replica migration problem caused by prediction error propagation. During the replica simplification phase, only the request service relationship corresponding to the deleted node is locally remapped, without the need for global recalculation of topology and request matching. The computational load on the control plane and data plane is low, requiring minimal edge computing power. It can run continuously online in a windowed manner, making it easy to implement and suitable for resource-constrained edge device clusters.

[0033] 4. Comprehensive mathematical constraints, compatible with multi-dimensional hardware limitations of edge devices. The modeling phase uniformly incorporates multiple physical constraints such as node storage capacity, link bandwidth, time slot connectivity, and minimum replica count. The entire process of replica selection, addition, and deletion automatically verifies storage limits and node reachability, preventing unreasonable solutions such as replica placement exceeding device storage or service allocation across disconnected links. The output replica placement results perfectly match the hardware resource limitations of the edge cluster, making the solution highly feasible.

[0034] 5. Service mapping and replica placement are jointly output, providing one-stop support for edge distributed storage scheduling. The algorithm synchronously outputs the optimal replica node set, window-weighted latency matrix, and request-replica service mapping matrix. It not only provides the replica deployment location but also directly provides the optimal service node allocation strategy for all requests in each time slot. No additional secondary matching calculation is required. It can be directly connected to the edge distributed file system scheduling module, simplifying the upper-layer storage business scheduling logic and reducing system integration and development costs. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0036] Figure 1 This is a flowchart illustrating a method for placing replicas based on dynamic network topology provided by the present invention.

[0037] Figure 2 This is a schematic diagram of time window-based decision-making provided by the present invention.

[0038] Figure 3 This is a schematic diagram of the spatiotemporal adaptive copy placement algorithm with ε relaxation provided by the present invention.

[0039] Figure 4 This is a schematic diagram of a replica placement device based on dynamic network topology provided by the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0041] The following is combined with Figure 1The present invention describes a method for placing replicas based on dynamic network topology, comprising the following steps:

[0042] Step 1, construct the spatiotemporal system model of the edge cluster: Discretize the time axis into continuous time slots, divide the time decision windows into non-overlapping ones, construct a dynamic topology map for each time slot, and collect data on file request frequency, remaining storage capacity, and link bandwidth of each node within the window;

[0043] Step 2, construct the initial feasible replica set with full coverage priority: filter candidate nodes whose storage capacity of the window's full time slots meets the file storage requirements, iteratively select candidate nodes with the largest weighted coverage request scale and the lowest average access latency to add to the replica set, until all requests within the window can be served by replica nodes, and obtain the initial replica set that satisfies the full request reachability constraint;

[0044] Step 3, greedy replacement of replicas with marginal latency gain: Iterate through the unselected candidate nodes, calculate the window-weighted total latency reduction brought by adding the node, select the node with the largest latency gain in each round to add to the replica set and update the service mapping and latency matrix, and stop replacement when there is no gain;

[0045] Step 4, relax the constraints and reduce redundant replicas: Calculate the window-weighted total latency, and try to remove single replica nodes one by one. Match the original service requests of the node with replacement replicas. If all requests have replacement replicas and the total latency after deletion does not exceed (1+ϵ) times the original total latency, then remove the replica node. Iterate until there are no nodes to reduce, and output the final replica placement scheme, service mapping matrix and latency matrix. Use the time window as the granularity to uniformly characterize dynamic topology and time-varying requests. First, ensure that all requests are reachable, then optimize access latency, and finally reduce the number of replicas in a controllable manner through the relaxation of ϵ. This simultaneously solves the problems of unreachability of dynamic edge network requests, inability to coordinate optimization of multiple objectives, and excessive replica maintenance overhead.

[0046] Consider a One track, each track One edge device, total A cluster consisting of 10 devices is denoted as 10 ... Each edge device has computing, storage, and transmission capabilities. The storage capacity of the edge devices is... , indicating the first The remaining storage capacity of the edge device; transmission capacity is , representing a node and nodes The data transmission bandwidth between them. Without loss of generality, this invention divides the time axis into time slots to capture the dynamics of topology and file requests in the time dimension, such as... Figure 2 As shown, the timeline is discretized into a set of time slots. .

[0047] In step 1, the constructed dynamic topology graph is denoted as...

[0048] ;

[0049] In the time slot The network state is assumed to remain constant; V(t) represents the time slot. A set of communicable storage nodes For time slots Available link set This is the set of remaining storage capacity for the nodes. This represents the set of available bandwidth for the link.

[0050] Specifically, node set Indicates time slot An edge set is a collection of devices that can participate in communication and storage services. Indicates time slot The set of available links within the network. If links... Then the node is considered and Direct communication is possible within this time slot, and its available bandwidth is denoted as . Further define the link bandwidth set The set of remaining storage capacity of a node is defined as follows: ,in Represents a node In the time slot Available storage capacity.

[0051] This invention uses target files The file size is [size missing], which is the subject of the study. Define nodes at the same time. The collection of files stored is Further defined in time slots Files already placed The node set is .

[0052] In the time slot Internally, the cluster processes files. The request set is denoted as ,in Represents a node In the time slot Internal documents The request frequency is used to characterize the non-uniformity of requests in space and time. This is further defined in time slots. The set of node IDs that initiated the request is .

[0053] To ensure successful processing of requests in each time slot, replica placement needs to be dynamically adjusted according to the cluster topology. Simultaneously, requests exhibit spatiotemporal heterogeneity, with different edge nodes accessing different files at significantly different times slots. If data source nodes directly transmit to requesting nodes, it easily leads to duplicate forwarding and additional communication overhead. Conversely, by combining topology changes with pre-positioning replicas in appropriate locations based on request distribution, the number of forwardings can be significantly reduced and access efficiency improved. Therefore, this invention employs a time window mechanism to determine the information span and update granularity upon which decisions are based: the time axis is discretized into time slots, and windows are set up in continuous, non-overlapping sequences. Divide the window; calculate the copy position at the starting point of each window based on the topology and request information within that window, keep the copy placement unchanged within the window, and readjust it in the next window.

[0054] Based on the distributed storage mechanism, it is assumed that each node stores at most one copy of the same file. In step 1, a binary variable is defined. , representing whether node i stores a copy of file f, a binary variable. The file request from node i is served by node j, and the replica placement variable remains constant within the same time window. In step 1, the latency model includes single-hop transmission latency and propagation latency. The end-to-end latency of node i being served by replica node j is the sum of the latency of all links along the shortest path between the two nodes. The window-weighted total latency is the sum of the latency of all requests accessed, weighted by the request frequency of each node's time slot.

[0055] In this embodiment Indicates time slot node Store files? A copy, Indicates time slot Middle node For the file Is the request from the node? The service uses a time window mechanism, with the window starting point... The placement decision is calculated only once and remains unchanged within the window, i.e., for any ,have Considering that dynamic topology may form multiple connected components, to avoid insufficient replicas leading to unserviceable requests in some connected components due to a lack of replicas, a lower bound constraint on the number of replicas within a window is introduced. .

[0056] In the delay model, the single-hop link delay consists of transmission delay and propagation delay: , Thus, the link delay is obtained. If requesting node In the time slot By service node The processing time is the sum of the on-link delays of the shortest path: The weighted total latency within the window is... ,in For nodes In the time slot For the file Request frequency / weight.

[0057] This invention simultaneously minimizes the number of replicas and the weighted total latency within a single window, establishing a multi-objective 0-1 programming model. :

[0058]

[0059]

[0060] (1)

[0061] (2)

[0062] (3)

[0063] (4)

[0064] (5)

[0065] (6)

[0066] (7)

[0067] The problem contains seven constraints. Equation 1 guarantees the uniqueness of the allocation, meaning that each request will be served by a unique node. Equation 2 guarantees that the service can only be assigned to a node that has placed a replica. Equation 3 guarantees that the service is assigned to a node that is reachable in the current time slot. Equation 4 guarantees that the node that places the replica has sufficient storage space. Equation 5 guarantees that there is a path between the requesting node and at least one node that places a replica. Equation 6 guarantees a lower limit on the number of replicas. Equation 7 specifies the range of values ​​for the decision variables.

[0068] like Figure 3As shown, this invention designs a spatiotemporally adaptive replica placement algorithm with ϵ relaxation to determine replica placement nodes and request service mappings within a specified time window, taking into account both topology and request distribution. The algorithm has two main ideas: First, considering that changes in link visibility and random failures in the cluster topology can cause replicas to be unreachable in some time slots, serviceability takes higher priority than service efficiency. Therefore, the algorithm first constructs an initial node set with the goal of covering all requests within the window, favoring nodes with stable reachability to more requests within the time window while also considering the average access cost of requests. Second, considering that the optimization problem has two objectives—minimizing service latency and minimizing the number of replicas—this algorithm optimizes from these two perspectives sequentially, using ϵ to characterize the relaxation boundary of latency floating during the stage of minimizing the number of replicas. After the above steps, the algorithm outputs the replica placement node set, the optimal service latency matrix, and the service mapping matrix.

[0069] The first phase involves coverage construction based on feasibility. The goal is to ensure that all requests within a window can be served by a replica node, while also considering service latency, under dynamic topology and reachability constraints. First, variables are initialized: Build request collection Mark each request pair as "uncovered" and set the current optimal reachability latency to be unknown; initialize the replica set. Simultaneously using the service mapping matrix With the optimal time delay matrix Record the optimal service node and its latency for each time slot request.

[0070] In step 2, the rule for selecting candidate nodes is: select edge devices whose remaining storage capacity is greater than the target file size in all time slots of the current window as candidate nodes. , As the starting point of the window, The time window length, This represents the relative time slot offset within the window.

[0071] In step 2, the evaluation criteria for selecting nodes include weighted coverage size. Weighted average delay Weighted coverage size The total number of covered but uncovered requests for node j is accumulated based on request frequency. When multiple candidate nodes have the same weighted coverage size, the weighted average latency is selected. Smaller nodes are added to the replica set.

[0072] Phase 1 constructs feasible solutions with coverage priority, and the placement of these solutions may be more biased towards reachability than minimum latency. Therefore, in Phase 2, the algorithm, based on the feasible solutions from Phase 1, introduces new replicas to further reduce the total service latency within the window without compromising serviceability. It employs a greedy strategy of "maximizing marginal gain": for each unselected candidate node... Calculate the latency reduction brought about by its addition as a new replica.

[0073] In step 3, the delay reduction of node j is calculated as follows: ;

[0074] Each round of selection The largest node is added to the result set, and the service mapping of the improved request is updated. With optimal delay ;

[0075] Accumulate all nodes For requests to reduce latency, the original optimal latency and node latency are calculated using request frequency as the weight. The sum of the differences in service latency. The larger the value, the more significant the latency reduction effect of adding this copy.

[0076] Phase two involves adding replicas to reduce latency, which may introduce redundant nodes. Therefore, in phase three, the algorithm performs a "reverse deletion" simplification of the current replica set and uses ϵ to relax constraints and mitigate performance degradation. In step 4, the frequency-weighted total latency within the window is first calculated. Then iteratively scan the current replica set. For any replica node Construct its service request set .

[0077] Delete replica node First, obtain the replica node. The entire request set for a single service; if the request set is empty, delete the node directly. If a request has no reachable alternative replica, then the node is retained. .

[0078] In step 4, the deletion cost is calculated. ϵ is a preset latency easing threshold, which is applied only if the total latency increment after removing a replica node satisfies the condition. At that time, the replica node deletion operation is performed, where To remove a node The resulting increase in total latency.

[0079] like Figure 4As shown, the present invention also provides a method and apparatus for replica placement based on dynamic network topology, used to execute the above-described method for replica placement based on dynamic network topology, comprising:

[0080] The modeling module is used to discretize time slots and divide time windows to build dynamic topology models and request latency models.

[0081] The coverage construction module is used to filter candidate nodes and iteratively generate a feasible solution with full coverage initial replicas;

[0082] The latency optimization module is used to calculate the marginal latency gain of each candidate node and greedily add replicas to reduce the total access latency.

[0083] The replica reduction module is used to reduce redundant replicas in reverse based on the ε relaxation constraint, and outputs the final replica placement and service mapping results.

[0084] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described copy placement method based on dynamic network topology.

[0085] This technology makes joint decisions on replica placement and service mapping at the spatiotemporal granularity of time windows, explicitly guaranteeing reachability coverage for requests within the window, thereby significantly reducing request failures and waiting caused by topology fluctuations. Furthermore, through a multi-stage, multi-objective trade-off mechanism, this technology achieves coordinated optimization of low latency and low replica count while ensuring serviceability. This invention employs a greedy algorithm and local update approach, making it suitable for online windowed operation in cloud-edge-device environments, with controllable computational and maintenance overhead.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for replica placement based on dynamic network topology, characterized in that, Includes the following steps: Step 1, construct the spatiotemporal system model of the edge cluster: Discretize the time axis into continuous time slots, divide the time decision windows into non-overlapping ones, construct a dynamic topology map for each time slot, and collect data on file request frequency, remaining storage capacity, and link bandwidth of each node within the window; Step 2, construct the initial feasible replica set with full coverage priority: filter candidate nodes whose storage capacity of the window's full time slots meets the file storage requirements, iteratively select candidate nodes with the largest weighted coverage request scale and the lowest average access latency to add to the replica set, until all requests within the window can be served by replica nodes, and obtain the initial replica set that satisfies the full request reachability constraint; Step 3, greedy replacement of replicas with marginal latency gain: Iterate through the unselected candidate nodes, calculate the window-weighted total latency reduction brought by adding the node, select the node with the largest latency gain in each round to add to the replica set and update the service mapping and latency matrix, and stop replacement when there is no gain; Step 4, relax constraints and reduce redundant replicas: calculate the window-weighted total delay, attempt to remove single replica nodes in turn, and match replacement replicas for the original service requests of the node; If all requests have alternative replicas and the total latency after deletion does not exceed (1+ϵ) times the original total latency, then remove the replica node, iterate until there are no more simplified nodes, and output the final replica placement scheme, service mapping matrix and latency matrix; By uniformly characterizing dynamic topology and time-varying requests at the time window granularity, we first ensure the reachability of all requests, then optimize access latency, and finally reduce the number of replicas in a controllable manner through ε relaxation, thus simultaneously solving the problems of unreachable dynamic edge network requests, inability to coordinate optimization of multiple objectives, and excessive replica maintenance overhead.

2. The replica placement method based on dynamic network topology according to claim 1, characterized in that, In step 1, the constructed dynamic topology graph is denoted as... ； In the time slot The network state is assumed to remain constant; V(t) represents the time slot. A set of communicable storage nodes For time slots Available link set This is the set of remaining storage capacity for the nodes. Define the set of available bandwidth for the link; define a binary variable. , representing whether node i stores a copy of file f, a binary variable. The file request of node i is served by node j, and the copy placement variable remains constant within the same time window.

3. The replica placement method based on dynamic network topology according to claim 2, characterized in that, In step 1, the latency model includes single-hop transmission latency and propagation latency. The end-to-end latency of node i served by replica node j is the sum of the latency of all links along the shortest path between the two nodes. The window-weighted total latency is the sum of the latency of all requests accessed, weighted by the frequency of each node's time slot request.

4. The replica placement method based on dynamic network topology according to claim 3, characterized in that, Step 2's rule for selecting candidate nodes is: select edge devices whose remaining storage capacity is greater than the target file size in all time slots of the current window as candidate nodes. , As the starting point of the window, The time window length, This represents the relative time slot offset within the window.

5. The replica placement method based on dynamic network topology according to claim 4, characterized in that, In step 2, the evaluation criteria for selecting nodes include weighted coverage size. Weighted average delay Weighted coverage size The total number of covered but uncovered requests for node j is accumulated based on request frequency. When multiple candidate nodes have the same weighted coverage size, the weighted average latency is selected. Smaller nodes are added to the replica set.

6. The replica placement method based on dynamic network topology according to claim 5, characterized in that, In step 3, the delay reduction of node j is calculated as follows: ; Each round of selection The largest node is added to the result set, and the service mapping of the improved request is updated. With optimal delay ; Accumulate all nodes For requests to reduce latency, the original optimal latency and node latency are calculated using request frequency as the weight. The sum of the differences in service latency. The larger the value, the more significant the latency reduction effect of adding this copy.

7. The replica placement method based on dynamic network topology according to claim 6, characterized in that, In step 4, the frequency-weighted total delay within the window is calculated first. Then iteratively scan the current replica set. For any replica node Construct its service request set ; Delete replica node First, obtain the replica node. The entire request set for a single service; if the request set is empty, delete the node directly. ; If a request has no reachable alternative replica, then the node is retained. .

8. The replica placement method based on dynamic network topology according to claim 7, characterized in that, In step 4, the deletion cost is calculated. ϵ is a preset latency easing threshold, which is applied only if the total latency increment after removing a replica node satisfies the condition. At that time, the replica node deletion operation is performed, where To remove a node The resulting increase in total latency.

9. A replica placement device based on dynamic network topology, characterized in that, The method for performing the copy placement method based on dynamic network topology as described in any one of claims 1-8 includes: The modeling module is used to discretize time slots and divide time windows to build dynamic topology models and request latency models. The coverage construction module is used to filter candidate nodes and iteratively generate a feasible solution with full coverage initial replicas; The latency optimization module is used to calculate the marginal latency gain of each candidate node and greedily add replicas to reduce the total access latency. The replica reduction module is used to reduce redundant replicas in reverse based on the ε relaxation constraint, and outputs the final replica placement and service mapping results.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the copy placement method based on dynamic network topology as described in any one of claims 1 to 8.

Images