Resource collaboration and time synchronization method and system for large-scale computing power cluster

Abstract

The invention provides a large-scale computing power cluster-oriented resource collaboration and time synchronization method and system, and relates to the technical field of computing power network architecture. Firstly, resource registration and state maintenance are carried out on a computing power node through an authoritative DNS; after a task is received, the system intelligently selects an optimal target computing power node and plans a data transmission path; in order to realize accurate collaboration, the system dynamically creates a clock synchronization domain according to a task, elects a master clock and plans an optimal clock synchronization path; and finally, the system encapsulates and packs the optimal task data transmission path, the main clock timestamp and the optimal clock synchronization path identifier, and distributes the data to related nodes, so that efficient coordination and high-precision time synchronization of resources are realized. According to the method, optimal matching of tasks and resources is achieved, the high efficiency of data transmission is guaranteed, the precision and reliability of time synchronization are guaranteed from a physical link, and finally efficient coordination and high-precision time synchronization of the resources in the computing power cluster are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computing power network architecture, and more particularly to a resource coordination and time synchronization method and system for large-scale computing power clusters. BACKGROUND

[0002] A computing power cluster refers to a centralized, scaled, and cooperatively working supercomputing resource pool formed by interconnecting a large number of computing nodes such as GPU cards through a high-speed network. A large-scale computing power cluster network involves a large amount of resources that are distributed in different nodes and systems. In order to coordinate the large amount of resources of each party, it is necessary to break the resource island, integrate and share the resources, and provide a unified resource view for tasks. However, in the existing heterogeneous computing power cluster network, the types of devices are complex, the scale is large, the network resources and computing power resources are managed in a decentralized manner, the resource registration relies on the Kubernetes platform or private protocols, and there is a lack of unified service registration and scheduling mechanism, making it difficult to realize cross-domain resource retrieval and load balancing.

[0003] In the cloud computing scenario, Segment Routing IPv6 (SRv6) based on the IPv6 forwarding plane is the core protocol of the next-generation network technology, supports flexible traffic scheduling across data centers, and dynamic network orchestration in the edge computing environment. The computing power network emphasizes the deep integration of computing power and network, the network can be programmed according to the business needs, flexibly schedule ubiquitous resources, coordinate the computing power and network resources of the whole network, realize computing power routing, and improve the overall performance and efficiency of the system. However, the existing SRv6 protocol supports path programming, but lacks deep cooperation with the whole network resource management system. The resource scheduling cannot deeply utilize the flexible coding resource capability of SRv6 and Security Identifiers (SID), and the SID configuration is usually static, which is difficult to dynamically respond to resource state changes, resulting in insufficient cooperation between the control and data planes, and the inability to deeply link with the global resource state monitoring and scheduling system, and dynamically adapt to resource changes and meet the needs of low-latency high-computing power scenarios.

[0004] In addition, time synchronization is the core foundation of the computing power cluster network, and its accuracy and real-time performance directly determine the distributed computing efficiency, real-time business reliability, and global resource utilization. The existing Precise Time Protocol (PTP) high-precision synchronization scheme needs to rely on special hardware, and some devices lack hardware support. Moreover, the existing master clock election cannot adapt to dynamic scenarios, and is not integrated into the network control plane or resource scheduling decision as a core element, forming an optimization blind area. SUMMARY

[0005] To address the problem that existing computing power clusters cannot simultaneously achieve resource coordination and time synchronization, this invention proposes a method and system for resource coordination and time synchronization for large-scale computing power clusters, enabling efficient resource coordination and high-precision time synchronization within the computing power cluster.

[0006] To achieve the above-mentioned technical effects, the technical solution of the present invention is as follows: Firstly, this application proposes a resource coordination and time synchronization method for large-scale computing power clusters: S1. Manage and initialize the computing nodes in the computing power cluster, register resources with the authoritative DNS using the computing nodes, and synchronize and maintain the status of the computing nodes. S2. Receive task requirements, divide the task area based on the task requirements, determine the optimal target computing power node within the task area to respond to the task requirements, and plan the optimal task data transmission path. S3. Based on the task area range, the computing power cluster is divided into several corresponding clock synchronization domains. A computing power node in each clock synchronization domain is selected as the master clock node, and its clock status is used as the master clock to generate the master clock timestamp. S4. Starting from the master clock node, plan the optimal clock synchronization path and generate the optimal clock synchronization path identifier; S5. Encapsulate the optimal task data transmission path, master clock timestamp, and optimal clock synchronization path identifier, and use the master clock node to forward them to all other computing power nodes in the clock synchronization domain to achieve resource collaboration and time synchronization.

[0007] In this technical solution, the computing nodes are first registered and their status maintained using an authoritative DNS. Upon receiving a task, the system intelligently selects the optimal target computing node and plans the data transmission path. To achieve precise collaboration, the system dynamically creates a clock synchronization domain based on the task, elects a master clock, and plans the optimal clock synchronization path. Finally, the system encapsulates the optimal task data transmission path, the master clock timestamp, and the optimal clock synchronization path identifier, and distributes them to the relevant nodes, thereby achieving efficient resource collaboration and high-precision time synchronization. This method not only achieves optimal matching between tasks and resources, ensuring efficient data transmission, but also guarantees the accuracy and reliability of time synchronization from the physical link perspective, ultimately realizing efficient resource collaboration and high-precision time synchronization in the computing cluster.

[0008] Preferably, step S1, which involves managing and initializing the computing nodes of the computing power cluster, comprises the following steps: After the computing node starts up, it establishes a secure channel with SDN based on the Secure Shell Protocol and the Transport Layer Security Protocol, and communicates through the Network Configuration Protocol NETCONF. The SDN verifies the identity of the computing power node based on a certificate or a pre-shared key and determines the device type of the computing power node, and records the device address of the computing power node; The SDN uses a network configuration protocol NETCONF to issue a master clock address to the computing power node, and receives time synchronization confirmation information of the computing power node; The SDN issues a DNS server address and a resource registration template to the computing power node, which is used to configure a DNS server resolution rule for the computing power node, and completes the management and initialization of the computing power node.

[0009] Preferably, the step S1 uses the computing power node to register resources with the authoritative DNS, and the process is as follows: The computing power node generates a DNS resource record based on the DNS server address and the resource registration template issued by the SDN, and uses transaction signature TSIG authentication to ensure the legality of the request; the DNS resource record includes a SID-INFO record and an SRV record, the SID-INFO record is used to record the resource state, including the IP address of the computing power node, the device type of the computing power node, the real-time load rate and the geographic location of the computing power node; the SRV record is used to locate the standard record type of the service and points to the service port; The computing power node sends a DNS dynamic update protocol request to the authoritative DNS server, and after the request is passed, uses the DNS dynamic update protocol to register the resource state to the DNS server in real time; The authoritative DNS server verifies whether the IP address of the computing power node sending the DNS dynamic update protocol request is in the pre-authorized list of the SDN, and verifies the validity of the transaction signature TSIG; if both verifications are passed, the registration is successful, if any one of the two verifications is not passed, the registration fails, and the computing power node re-registers according to the exponential backoff strategy until the SDN intervenes to repair.

[0010] Preferably, the state synchronization and maintenance of the computing power node in the step S1 includes constructing a state synchronization and maintenance mechanism of the computing power node to synchronize the resource state of the computing power node; The state synchronization and maintenance mechanism of the computing power node includes an SDN periodic monitoring mechanism and a short-period monitoring mechanism; The process of the SDN periodic monitoring mechanism is as follows: The computing power node periodically reports the load and health state to the SDN, and the SDN judges whether the load of the computing power node exceeds the pre-set threshold; if it exceeds, the SDN triggers the DNS record weight adjustment to reduce the priority of the SRV record; if it does not exceed, the SDN continues to monitor the load and health state of the computing power node; The short period monitoring mechanism comprises: setting a survival period TTL of the SID-INFO record as a short period, in each short period, the computing power node sends at least one DNS UPDATE request to the SDN to refresh the SID-INFO record, otherwise the SDN considers that the computing power node is offline, and deletes the node configuration by using the NETCONF and deletes the SID-INFO record by using the DNS API.

[0011] Preferably, the process of step S2 is: S21. The SDN divides a task area range based on a task requirement, the task requirement comprising a computing power node type and quantity, generates a task area identifier Task-ID and a computing power node member list; S22. It is judged whether the task requirement needs specific computing resources, if yes, the computing power node sends a resource request to the SDN, the SDN sends a resource query request to the local DNS, the local DNS returns a matched candidate resource node list, and S23 is executed; if the task requirement does not need specific computing resources, S23 is directly executed; S23. The optimal computing node is selected to execute the task, a multi-factor weighted algorithm is used to calculate the node selection weight, the node with the highest weight is selected as the target computing node, and a SID sequence explicit path comprising the target computing node IP address and intermediate forwarding nodes is generated; S24. The SDN distributes the SID sequence explicit path comprising the target computing node IP address and intermediate forwarding nodes to the switch of the computing power node, and completes the response of the task requirement.

[0012] Preferably, the process that the SDN divides the computing power cluster into several clock synchronization domains based on the task area range in step S3 is: The SDN takes the computing power node located in the physical topology center in each task area range as the core of the clock synchronization domain, jumps outwards based on the core of the time synchronization domain, until the maximum node jump number set in advance is reached, defines the computing power node at the time when the maximum node jump number set in advance is reached as the boundary of the time synchronization domain, and completes the division of the time synchronization domain.

[0013] Preferably, the selection process of the master clock in step S3 is: The running parameters of all the computing power nodes in the same time synchronization domain are collected as the scoring basis, the running parameters comprising the in-out degree, the throughput, the clock accuracy level and the link delay, and the weight of the scoring basis is initialized based on the network type; Based on the scoring basis, the scores of the computing power nodes are calculated, and the scores of the computing power nodes are updated according to a preset time period based on the network state; The clock state of the highest-scored computing power node is set as the master clock, the highest-scored computing power node is the master clock node, and the rest of the computing power nodes are slave clock nodes, if there are two or more computing power nodes with the same score, then based on the preset score basis priority, a step-by-step comparison is performed until the only optimal computing power node is selected, and the priority of the score basis is in turn: out-degree, clock accuracy level, throughput, and link delay.

[0014] Preferably, the process of generating the optimal clock synchronization path identifier is: A minimum spanning tree is constructed using the Dijkstra algorithm, the root node of the minimum spanning tree is the master clock node, and the minimum spanning tree is used to plan the shortest path from the master clock computing power node to all other nodes in the same time synchronization domain; Based on the shortest path, an optimal clock synchronization path identifier is generated.

[0015] In a second aspect, the application also proposes a resource coordination and time synchronization system for a large-scale computing power cluster, the system comprising: A computing power node initialization unit is configured to manage and initialize the computing power nodes in the computing power cluster, register resources to the authoritative DNS using the computing power nodes, and synchronize and maintain the state of the computing power nodes; A data transmission path planning unit is configured to receive task requirements, divide the task area range based on the task requirements, determine the optimal target computing power node responding to the task requirements in the task area range, and plan the optimal task data transmission path; A master clock timestamp generation unit is configured to divide the computing power cluster into corresponding clock synchronization domains based on the task area range, select a computing power node in each clock synchronization domain as a master clock node, set its clock state as the master clock, and generate a master clock timestamp; A clock synchronization path generation unit is configured to plan an optimal clock synchronization path from the master clock node as the starting point and generate an optimal clock synchronization path identifier; A resource coordination and time calibration unit is configured to encapsulate the optimal task data transmission path, the master clock timestamp, and the optimal clock synchronization path identifier, and forward them to all other computing power nodes in the clock synchronization domain using the master clock node to realize resource coordination and time synchronization.

[0016] In a third aspect, the application also proposes a computer device, which comprises a memory, a processor, and a computer program stored on the memory and executable by the processor.

[0017] Compared with the prior art, the application has the following advantages: The application provides a resource coordination and time synchronization method and system for a large-scale computing power cluster, resource registration and state maintenance of computing power nodes are performed through an authoritative DNS; after receiving a task, an optimal target computing power node is intelligently selected and a data transmission path is planned; in order to realize accurate coordination, a clock synchronization domain is dynamically created according to the task, a master clock is elected and an optimal clock synchronization path is planned; finally, the optimal task data transmission path, the master clock timestamp and the optimal clock synchronization path identifier are packaged and distributed to related nodes, so that efficient resource coordination and high-precision time synchronization are realized. The method not only realizes optimal matching of tasks and resources, guarantees the efficiency of data transmission, but also guarantees the accuracy and reliability of time synchronization from the physical link, and finally realizes efficient resource coordination and high-precision time synchronization in the computing power cluster. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 a flowchart of a resource coordination and time synchronization method for a large-scale computing power cluster proposed in embodiment 1 of the application; Figure 2 a system architecture diagram of the resource coordination and time synchronization for a large-scale computing power cluster proposed in embodiment 1 of the application; Figure 3 a structural diagram of a resource coordination and time synchronization system for a large-scale computing power cluster proposed in embodiment 3 of the application; Figure 4 a structural diagram of a computer device proposed in embodiment 4 of the application. DETAILED DESCRIPTION

[0019] The drawings are only used for illustrative description and cannot be understood as a limitation of the patent; In order to better illustrate the embodiments, some parts of the drawings are omitted, enlarged or reduced, and do not represent the actual size; For those skilled in the art, it is understandable that some well-known descriptions in the drawings can be omitted.

[0020] The technical solutions of the application will be further described below in combination with the drawings and embodiments.

[0021] The positional relationship described in the drawings is only used for illustrative description and cannot be understood as a limitation of the patent; Embodiment 1 Embodiment 1 proposes a resource coordination and time synchronization method for a large-scale computing power cluster, and the flowchart of the method is shown in Figure 1 , including the following steps: S1. The computing power nodes in the computing power cluster are managed and initialized, the computing power nodes are registered to the authoritative DNS, and the state synchronization and maintenance of the computing power nodes are performed; S2. Receiving a task requirement, dividing a task area range based on the task requirement, determining an optimal target computing power node responding to the task requirement in the task area range, and planning an optimal task data transmission path; S3. Dividing the computing power cluster into corresponding clock synchronization domains based on the task area range, selecting a certain computing power node in each clock synchronization domain as a master clock node, taking the clock state of the master clock node as the master clock, and generating a master clock timestamp; S4. Taking the master clock node as a starting point, planning an optimal clock synchronization path, and generating an optimal clock synchronization path identifier; S5. Encapsulating the optimal task data transmission path, the master clock timestamp, and the optimal clock synchronization path identifier, and forwarding them to all other computing power nodes in the clock synchronization domain by using the master clock node, to realize resource coordination and time synchronization.

[0022] In this embodiment, first, the computing power nodes are registered and maintained by the authoritative DNS; after receiving a task, the system intelligently selects an optimal target computing power node and plans a data transmission path; to realize precise coordination, the system dynamically creates a clock synchronization domain according to the task, elects a master clock, and plans an optimal clock synchronization path; finally, the system encapsulates and packages the optimal task data transmission path, the master clock timestamp, and the optimal clock synchronization path identifier, and distributes them to the relevant nodes, thereby realizing efficient coordination of resources and high-precision time synchronization. This method not only realizes the optimal matching of tasks and resources, and guarantees the efficiency of data transmission, but also guarantees the accuracy and reliability of time synchronization from the physical link, and finally realizes efficient coordination of resources and high-precision time synchronization in the computing power cluster.

[0023] Specifically, the architecture implementing the method is divided into a collaborative application layer 1, a controller layer 2, and an infrastructure layer 3; the structural diagram of the architecture is shown in Figure 2 The collaborative application layer 1 includes time service, DNS service, and applications (computing tasks), the controller layer 2 mainly includes an SDN controller, and the infrastructure layer 3 includes switches, routers, servers, virtual machines, and other intelligent computing cluster nodes; the collaborative application layer 1 and the controller layer 2 are extended based on an SDN northbound interface 4, and the controller layer 2 and the infrastructure layer 3 are extended based on an SDN southbound interface 5.

[0024] Embodiment 2 In this embodiment, the process of managing and initializing the computing power nodes of the computing power cluster in step S1 is as follows: After the computing power node is started, a secure shell protocol and a transmission layer security protocol are used to establish a secure channel with the SDN, and a network configuration protocol NETCONF is used for communication; The SDN verifies the identity of the computing power node based on a certificate or a pre-shared key and determines the device type of the computing power node, and records the device address of the computing power node; The SDN uses a network configuration protocol NETCONF to issue a master clock address to the computing power node, and receives time synchronization confirmation information of the computing power node; The SDN issues a DNS server address and a resource registration template to the computing power node, which is used to configure a DNS server resolution rule for the computing power node, and completes the management and initialization of the computing power node.

[0025] In the embodiment, the process of registering a resource to an authoritative DNS by the computing power node in step S1 is as follows: The computing power node generates a DNS resource record based on the DNS server address and the resource registration template issued by the SDN, and uses a transaction signature TSIG authentication to ensure the legality of the request; the DNS resource record includes a SID-INFO record and an SRV record, the SID-INFO record is used to record the resource state, including: the IP address of the computing power node, the device type of the computing power node, the real-time load rate and the geographical position of the computing power node; the SRV record is used to locate the standard record type of the service and points to the service port; The computing power node sends a DNS dynamic update protocol request to the authoritative DNS server, after the request is passed, the resource state is registered to the DNS server in real time by using the DNS dynamic update protocol; The authoritative DNS server verifies whether the IP address of the computing power node sending the DNS dynamic update protocol request is in the pre-authorized list of the SDN, and verifies the validity of the transaction signature TSIG; if both verifications are passed, the registration is successful, if any one of the two verifications is not passed, the computing power node re-registers according to the exponential backoff strategy until the SDN intervenes to repair.

[0026] Specifically, after the computing power node is started, a secure channel can be established with the SDN controller through SSH / TLS, and the NETCONF protocol is used for communication; the SDN controller verifies the identity of the node based on a certificate or a pre-shared key, confirms the device role such as "server node", and synchronizes the device address to the controller.

[0027] Specifically, after the master clock is determined, the SDN controller uses the <edit-config>In the operation, the main clock address is issued to the computing power node, for example: time.example.com; after the node synchronizes the time, the controller returns <rpc-reply>Confirm the configuration takes effect; after taking effect, the SDN controller issues the DNS server address, such as dns.example.com, and the resource registration template, and configures the node DNS resolution rule.

[0028] Specifically, the computing node generates a DNS resource record according to the resource registration template issued by the SDN controller, including: a SID-INFO record: a custom record type containing node IP, computing type, such as GPU-NVIDIA-A100, real-time load rate, geographic location, and other metadata; and an SRV record: pointing to a service port, such as srv6._tcp.example.com 86400 IN SRV 0 5 8080 node1.example.com.

[0029] In the present application, three types of resource records are added in the DNS server, including: SRV-COMPUTE: register computing resource attributes (CPU / GPU model, memory / storage capacity, real-time load); SRV-NETWORK: register network resources (bandwidth margin, latency, topology location); and SRV-MODEL: register AI model parameter version, storage location, and calling interface.

[0030] In the present embodiment, the state synchronization and maintenance of the computing node in step S1 include constructing a state synchronization and maintenance mechanism for the computing node to synchronize the resource state of the computing node; The state synchronization and maintenance mechanism of the computing node includes an SDN periodic monitoring mechanism and a short-period monitoring mechanism; The process of the SDN periodic monitoring mechanism is as follows: The computing node periodically reports the load and health status to the SDN, and the SDN determines whether the load of the computing node exceeds a pre-set threshold; if it does, the SDN triggers DNS record weight adjustment to reduce the priority of the SRV record; if it does not, the SDN continues to monitor the load and health status of the computing node; The short-period monitoring mechanism includes: setting the survival period TTL of the SID-INFO record as a short period, and in each short period, the computing node sends at least one DNS UPDATE request to the SDN to refresh its SID-INFO record, otherwise the SDN considers the computing node offline, and deletes the node configuration using NETCONF and deletes the SID-INFO record using the DNS API.

[0031] Specifically, the computing node sends a DNS UPDATE request to the DNS server through NETCONF <notification>The message periodically reports the load and health status to the SDN controller. The threshold of the computing power node is 90%, and if the node load exceeds the threshold, the SDN controller triggers the DNS record weight adjustment to reduce the priority of the SRV record.

[0032] Specifically, one period of the survival period TTL is 60 seconds.

[0033] Specifically, the state synchronization and maintenance of the computing power node further includes exception handling and security mechanisms, including: The NETCONF configuration rollback mechanism, specifically: if the time synchronization or DNS address fails to be issued, the controller rolls back to the previous version of the configuration, and records an alarm; the DNS registration retry mechanism, specifically: when the node fails to register the resource, it retries according to the exponential backoff strategy until the SDN controller intervenes to repair.

[0034] In the embodiment, the process of step S2 is: S21. The SDN divides the task area range based on the task demand, the task demand including the type and quantity of computing power nodes, generates a task area identifier Task-ID and a computing power node member list; S22. Determine whether the task demand requires specific computing resources, if yes, the computing power node sends a resource request to the SDN, the SDN sends a resource query request to the local DNS, and the local DNS returns a list of matching candidate resource nodes, and executes S23; if the task demand does not require specific computing resources, directly execute S23; S23. Select the optimal computing node to execute the task, calculate the node selection weight using a multi-factor weighting algorithm, select the node with the highest weight as the target computing node, and generate a SID sequence explicit path containing the target computing node IP address and intermediate forwarding nodes; S24. The SDN sends the SID sequence explicit path containing the target computing node IP address and intermediate forwarding nodes to the switch of the computing power node, and completes the response of the task demand.

[0035] Specifically, the division rule of the task area range is: First, the SDN controller calls the DNS global resource view to filter nodes that meet the task conditions; then, based on the k-means clustering algorithm, the nodes are divided into task areas according to their geographical location, resource type, and network delay; finally, the number of nodes in the area is adjusted considering the area size limit, specifically, the number of single-area nodes ≤128 (adjusted according to the task size), and the maximum hop count ≤8 (to avoid time synchronization precision decay).

[0036] Specifically, when the task requirement requires a specific computing resource or the load of the computing node reaches a threshold, the node sends a resource request to the SDN controller through a southbound interface (such as NETCONF or gRPC). The request content includes: resource type (GPU computing power, storage space), demand priority, service QoS requirement (latency, bandwidth), and other metadata; after the SDN receives the resource request, the SDN controller sends a query to the local DNS through a northbound interface (RESTful API), carrying the resource type and service constraint condition (such as "GPU type = Tesla V100, latency ≤ 50 ms"); after sending the resource request, the local DNS returns a list of matching candidate resource nodes, or returns after being queried by the local DNS from the authoritative DNS, including the IP address of each node, resource type, current load rate, and SID encoding template (such as Locator prefix).

[0037] Filtering invalid nodes, specifically, the controller excludes unreachable nodes based on network topology (link state collected by LLDP protocol).

[0038] The multi-factor weighted algorithm weight includes: resource load (such as GPU utilization ≤ 80% for high score), path cost (number of hops of SRv6 explicit path, latency), and service priority (high priority service is forced to select low load node).

[0039] The SID structure design includes: Locator field, used to fix the prefix (such as 2001:db8:1:: / 64) to identify the resource pool domain; Function field, used to dynamically write the last 64 bits of the target resource IP, and embed the resource type identifier (such as 0xA1 represents GPU resource).

[0040] The encoding example of the SID is: if the target resource IP is 2001:db8:2:1::10 and the resource type is GPU, then the SID encoding is 2001:db8:1::A1:2001:db8:2:1::10.

[0041] Specifically, the process of SID sequence explicit path distribution includes: OpenFlow flow table injection, the controller distributes the SRv6 SID encoding and forwarding rule to the switch of the initial or intermediate node through the southbound interface, and specifies the End.X behavior (explicit next hop); BGP-LS synchronization, in the cross-domain scenario, the controller broadcasts the SID path information to the adjacent domain through the BGP-LS protocol, to ensure global route reachability.

[0042] In the distribution process, it also includes bidirectional forwarding detection (BFD): the controller and the node monitor the path connectivity through the BFD protocol, and if the packet loss rate > 1%, the re-routing is triggered.

[0043] In the embodiment, the process of dividing the computing power cluster into several clock synchronization domains based on the task area range in step S3 is as follows: The SDN takes the computing power node located at the center of the physical topology in each task area range as the core of the clock synchronization domain, and jumps outwards based on the core of the time synchronization domain until the maximum node jump number is reached, defines the computing power node at the time when the maximum node jump number is reached as the boundary of the time synchronization domain, and completes the division of the time synchronization domain.

[0044] Specifically, the maximum node jump number is set to 8 hops (referring to the default value of the MSTP protocol) to avoid the decrease of synchronization accuracy and the accumulation of delay caused by too deep jump number.

[0045] Specifically, the range of the time synchronization domain is dynamically adjusted. The SDN controller collects the link state in real time through the southbound interface (such as OpenFlow), predicts the network congestion and topology changes in combination with federated learning, and dynamically adjusts the domain range. When the number of hops in a certain domain exceeds the threshold, splitting is triggered, and a new domain elects a local master clock.

[0046] In the embodiment, the process of selecting the master clock in step S3 is as follows: The running parameters of all computing power nodes in the same time synchronization domain are collected as the scoring basis, including the in-degree, throughput, clock accuracy level and link delay, and the weights of the scoring basis are initialized based on the network type; Based on the scoring basis, the scores of each computing power node are calculated, and the scores of the computing power nodes are updated according to the preset time period based on the network state; The clock state of the computing power node with the highest score is set as the master clock, the computing power node with the highest score is the master clock node, and the remaining computing power nodes are slave clock nodes. If there are two or more computing power nodes with the same score, a hierarchical comparison is performed based on the priority of the preset scoring basis until a unique optimal computing power node is selected. The priority of the scoring basis is in turn: in-degree, clock accuracy level, throughput, and link delay.

[0047] Specifically, the weights of the scoring basis include: in-degree weight , which is used to measure the frequency of being referenced by other nodes and reflects the reliability of the node as a time reference. The expression is:

[0048] Among them, is the in-degree score, is the in-degree weight, , in the industrial internet, ; out-degree weight , for measuring the ability of a node to verify other nodes, embodying its contribution to global time synchronization, the expression is:

[0049] wherein, is the out-degree score, is the out-degree weight, In the industrial internet, ; The comprehensive out-in-degree score , the expression is:

[0050] Throughput weight , for indicating the ability of a node to process PTP synchronization messages, the expression of the throughput score is:

[0051] wherein, is the throughput score, is the throughput weight, In the industrial internet, ; Clock accuracy level score Based on the clock quality level weight and the clock stability weight , the expression is:

[0052] wherein, represents the clock quality level, represents the stability, In the industrial internet, , ; Link delay score Based on the delay weight and the jitter weight , the expression is:

[0053] wherein, represents the actual transmission delay, represents the actual jitter, In the industrial internet, ,

[0054] The comprehensive out-in-degree score, the throughput score, the clock accuracy level score and the link delay score, the expression is:

[0055] wherein, is the weight of the score, which is dynamically adjusted in different application scenarios, for example: when the network topology is stable, the (in-degree weight) is increased; in a high-concurrency scenario, the (throughput weight) is increased; in a clock accuracy sensitive scenario (such as industrial control), the ; Specifically, the running parameter score of the industrial Internet of Things is based on the weights: in-degree weight , out-degree weight , throughput weight , clock quality level weight , clock stability weight , time delay weight , and jitter weight . In a satellite network, the time delay weight is increased to 0.5, and the throughput weight is reduced to 0.1.

[0056] Specifically, the preset time period is one hour.

[0057] In this embodiment, the process of generating the optimal clock synchronization path identifier is: A minimum spanning tree is constructed using the Dijkstra algorithm, the root node of the minimum spanning tree being the master clock node, for planning the shortest path from the master clock computing node to all other nodes in the same time synchronization domain; Based on the shortest path, the optimal clock synchronization path identifier is generated.

[0058] Specifically, the weight of the path planning is the time delay link utilization rate, and a low-load, low-jitter path is preferentially selected.

[0059] Specifically, when the node performs clock synchronization forwarding according to the clock synchronization path identifier, the time delay of the path delay forwarding is also considered, and a Time-Sync TLV field is embedded in the SRv6 extension header, carrying the master clock timestamp and path delay compensation value. The intermediate node parses the TLV and corrects the local clock to realize hop-by-hop synchronization (accuracy ≤ 1 μs); Specifically, after selecting the master clock, further comprising: periodically detecting, collecting throughput and time delay every 1 hour; checking clock accuracy through NETCONF rpc / rpc-reply message interaction every 24 hours; if the master clock triggers the reselection threshold, reselecting the master clock, the reselection threshold includes: absolute threshold: forced reselection when the master clock score is less than 0.7 (full score 1.0); relative threshold: the score of the suboptimal node exceeds 85% of the current master clock, and lasts for 3 monitoring periods (default period 1 hour); fast switching mechanism: when the master clock fails, select a node from the rotation queue to take over.

[0060] Embodiment 3 The embodiment provides a resource coordination and time synchronization system for a large-scale computing power cluster, in the embodiment, the system is used to implement a resource coordination and time synchronization method for a large-scale computing power cluster, and a structural schematic diagram is as shown in Figure 3 The embodiment provides a resource coordination and time synchronization system for a large-scale computing power cluster, in the embodiment, the system is used to implement a resource coordination and time synchronization method for a large-scale computing power cluster, and a structural schematic diagram is as shown in The computing power node initialization unit is configured to manage and initialize the computing power nodes in the computing power cluster, register resources to the authoritative DNS by using the computing power nodes, and synchronize and maintain the states of the computing power nodes. The data transmission path planning unit is configured to receive a task demand, divide a task area range based on the task demand, determine an optimal target computing power node responding to the task demand in the task area range, and plan an optimal task data transmission path. The master clock timestamp generation unit is configured to divide the computing power cluster into corresponding clock synchronization domains based on the task area range, select a certain computing power node in each clock synchronization domain as a master clock node, take the clock state of the master clock node as a master clock, and generate a master clock timestamp. The clock synchronization path generation unit is configured to take the master clock node as a starting point, plan an optimal clock synchronization path, and generate an optimal clock synchronization path identifier. The resource coordination and time calibration unit is configured to encapsulate the optimal task data transmission path, the master clock timestamp and the optimal clock synchronization path identifier, and forward them to all other computing power nodes in the clock synchronization domain by using the master clock node, so as to realize resource coordination and time synchronization.

[0061] Embodiment 4 In the embodiment, a computer device is provided, the computer device comprises a memory 101, a processor 102 and a computer program stored on the memory 101 and capable of being run on the processor, the processor 102 executes the computer program, and a resource coordination and time synchronization method for a large-scale computing power cluster is realized, and a structural schematic diagram of the device is as shown in Figure 4 .

[0062] Obviously, the above embodiments of the present application are only examples for clearly explaining the present application, and are not intended to limit the embodiments of the present application. Based on the above description, other different forms of changes or variations can be made by those skilled in the art. Here, it is not necessary and impossible to exhaust all the embodiments. Any modification, equivalent replacement and improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the claims of the present application.< / notification>

Claims

1. A method for resource coordination and time synchronization for large-scale computing clusters, characterized in that, Includes the following steps: S1. Manage and initialize the computing nodes in the computing power cluster, register resources with the authoritative DNS using the computing nodes, and synchronize and maintain the status of the computing nodes. S2. Receive task requirements, divide the task area based on the task requirements, determine the optimal target computing power node within the task area to respond to the task requirements, and plan the optimal task data transmission path. S3. Based on the task area range, the computing power cluster is divided into several corresponding clock synchronization domains. A computing power node in each clock synchronization domain is selected as the master clock node, and its clock status is used as the master clock to generate the master clock timestamp. S4. Starting from the master clock node, plan the optimal clock synchronization path and generate the optimal clock synchronization path identifier; S5. Encapsulate the optimal task data transmission path, master clock timestamp, and optimal clock synchronization path identifier, and use the master clock node to forward them to all other computing power nodes in the clock synchronization domain to achieve resource collaboration and time synchronization.

2. The resource coordination and time synchronization method for large-scale computing power clusters according to claim 1, characterized in that, Step S1 describes the process of managing and initializing the computing nodes of the computing power cluster. After the computing node starts up, it establishes a secure channel with SDN based on the Secure Shell protocol and the transport layer security protocol, and communicates through the NETCONF network configuration protocol. SDN verifies the identity of computing power nodes and determines the device type of computing power nodes based on certificates or pre-shared keys, and records the device address of computing power nodes; SDN uses the NETCONF network configuration protocol to send the master clock address to the computing power node and receive the time synchronization confirmation information from the computing power node. SDN distributes DNS server addresses and resource registration templates to computing nodes, which are used to configure DNS server resolution rules for the computing nodes and complete the management and initialization of the computing nodes.

3. The resource coordination and time synchronization method for large-scale computing power clusters according to claim 2, characterized in that, Step S1, which involves registering resources with an authoritative DNS using computing power nodes, is as follows: The computing nodes generate DNS resource records based on the DNS server address and resource registration template issued by SDN, and use transaction signature TSIG authentication to ensure the legitimacy of the request. The DNS resource records include SID-INFO records and SRV records. The SID-INFO records are used to record resource status, including: the IP address of the computing node, the device type of the computing node, the real-time load rate, and the geographical location of the computing node. The SRV records are used to locate the standard record type of the service and point to the service port. The computing node sends a DNS dynamic update protocol request to the authoritative DNS server. After the request is approved, the resource status is registered with the DNS server in real time using the DNS dynamic update protocol. The authoritative DNS server verifies whether the IP address of the computing node sending the DNS dynamic update protocol request is in the SDN pre-authorized list and verifies the validity of the transaction signature TSIG. If both verifications pass, the registration is successful. If either verification fails, the registration fails, and the computing node re-registers according to the exponential backoff strategy until SDN intervenes to repair the issue.

4. The resource coordination and time synchronization method for large-scale computing power clusters according to claim 3, characterized in that, The state synchronization and maintenance of computing power nodes in step S1 includes constructing a state synchronization and maintenance mechanism for computing power nodes to synchronize the resource status of computing power nodes; The state synchronization and maintenance mechanism of the computing nodes includes an SDN periodic monitoring mechanism and a short-cycle monitoring mechanism; The process of the SDN periodic monitoring mechanism is as follows: The computing nodes periodically report their load and health status to the SDN. The SDN determines whether the load of the computing nodes exceeds a pre-set threshold. If it does, the SDN triggers a DNS record weight adjustment, reducing the priority of the SRV record. If the limit is not exceeded, SDN continues to monitor the load and health status of the computing power nodes; The short-cycle monitoring mechanism includes: setting the TTL of the SID-INFO record to a short cycle; in each short cycle, the computing node sends a DNS UPDATE request to SDN at least once to refresh its SID-INFO record; otherwise, SDN considers the computing node offline and uses NETCONF to delete the node configuration and calls the DNS API to delete the SID-INFO record.

5. The resource coordination and time synchronization method for large-scale computing power clusters according to claim 4, characterized in that, The process of step S2 is as follows: S21.SDN divides the task area based on task requirements, which include the type and number of computing power nodes, and generates a task area identifier (Task-ID) and a list of computing power node members. S22. Determine whether the task requires specific computing resources. If so, the computing node sends a resource request to the SDN, the SDN sends a resource query request to the local DNS, the local DNS returns a list of matching candidate resource nodes, and S23 is executed. If the task does not require specific computing resources, S23 is executed directly. S23. Select the optimal computing node to execute the task, use a multi-factor weighting algorithm to calculate the node selection weight, select the node with the highest weight as the target computing node, and generate an explicit path containing the IP address of the target computing node and the SID sequence of intermediate forwarding nodes. S24.SDN explicitly sends the path containing the target compute node's IP address and the SID sequence of intermediate forwarding nodes to the switch of the compute node, thus fulfilling the task requirements.

6. The resource coordination and time synchronization method for large-scale computing power clusters according to claim 5, characterized in that, Step S3 describes the process of SDN dividing the computing cluster into several clock synchronization domains based on the task area range: SDN uses the computing power node located at the center of the physical topology within each task area as the core of the clock synchronization domain. It jumps outward from the core of the time synchronization domain until the preset maximum number of node jumps is reached. The computing power node where the preset maximum number of node jumps is reached is defined as the boundary of the time synchronization domain, thus completing the division of the time synchronization domain.

7. The resource coordination and time synchronization method for large-scale computing power clusters according to claim 6, characterized in that, The master clock selection process described in step S3 is as follows: The operating parameters of all computing nodes within the same time synchronization domain are collected as the scoring criteria. The operating parameters include in-degree, throughput, clock precision level and link latency. The weights of the scoring criteria are initialized based on the network type. Based on the scoring criteria, the score of each computing node is calculated, and based on the network status, the score of the computing node is updated according to a preset time period. The clock state of the highest-scoring computing power node is set as the master clock. The highest-scoring computing power node is the master clock node, and the other computing power nodes are slave clock nodes. If two or more computing power nodes have the same score, they are compared level by level based on the preset scoring criteria priority until the unique and optimal computing power node is selected. The priority of the scoring criteria is as follows: in-degree, clock accuracy level, throughput, and link latency.

8. The resource coordination and time synchronization method for large-scale computing power clusters according to claim 7, characterized in that, The process of generating the optimal clock synchronization path identifier is as follows: A minimum spanning tree is constructed using Dijkstra's algorithm, with the root node of the minimum spanning tree being the master clock node. This tree is used to plan the shortest path from the master clock computing node to all other nodes in the same time synchronization domain. Generate the optimal clock synchronization path identifier based on the shortest path.

9. A resource coordination and time synchronization system for large-scale computing power clusters, characterized in that, The system is used to implement the method according to any one of claims 1 to 8, comprising: The computing node initialization unit is used to manage and initialize computing nodes in the computing cluster, register resources with the authoritative DNS using computing nodes, and synchronize and maintain the status of computing nodes. The data transmission path planning unit is used to receive task requirements, divide the task area based on the task requirements, determine the optimal target computing power node that responds to the task requirements within the task area, and plan the optimal task data transmission path. The master clock timestamp generation unit is used to divide the computing power cluster into several corresponding clock synchronization domains based on the task area range, select a computing power node in each clock synchronization domain as the master clock node, use its clock state as the master clock, and generate the master clock timestamp. The clock synchronization path generation unit is used to plan the optimal clock synchronization path starting from the master clock node and generate the optimal clock synchronization path identifier. The resource coordination and time calibration unit is used to encapsulate the optimal task data transmission path, the master clock timestamp, and the optimal clock synchronization path identifier, and forward them to all other computing power nodes in the clock synchronization domain using the master clock node, thereby achieving resource coordination and time synchronization.

10. A computer device, characterized in that, The computer device includes: a memory, a processor, and a computer program stored in the memory that can be run on the processor, wherein the processor executes the computer program to implement the method according to any one of claims 1 to 8.

Images