A pipeline and industry service deployment method and system for a distributed intelligent computing center network

Abstract

This invention relates to a pipelined parallel service deployment method and system for distributed intelligent computing center networks in the field of digital information transmission technology. The method includes obtaining a set of services to be deployed. For each service in the set, the following processing is performed: traversing the micro-batch set of the current service in forward order, determining whether the current service has completed forward propagation deployment; if not, propagating the micro-batch in forward order; if completed, updating the current service to a reverse propagation state, traversing the micro-batch set of the current service in reverse order, determining whether the current service has completed reverse propagation deployment; if not, propagating the micro-batch in reverse order; if completed, determining that the current service deployment is complete. This invention achieves efficient mapping between pipeline stages and computing nodes, and completes service deployment based on this mapping, reducing idle waiting time during pipeline execution and improving overall system throughput and resource utilization efficiency.

Description

Technical Field

[0001] This invention relates to a pipelined parallel service deployment method and system for distributed intelligent computing center networks, belonging to the field of digital information transmission technology. Background Technology

[0002] With the rapid development of large-scale artificial intelligence models, single data centers can no longer meet their demands for computing resources and data exchange capabilities. Currently, multiple intelligent computing centers are interconnected through high-speed optical networks to form a distributed intelligent computing center network, supporting cross-regional collaborative computing. Among these, the pipelined parallel strategy, by dividing the model into multiple stages and distributing them across different computing nodes, achieves overlap between computing and communication, and is an important method for cross-intelligent computing center deployment.

[0003] In existing technologies, the scheduling problem for distributed computing tasks is mainly addressed from two aspects: computing power allocation and network resource allocation. On the one hand, regarding computing power scheduling, various scheduling strategies based on task completion time or load balancing have been proposed to improve the efficiency of computing resource utilization. On the other hand, regarding optical network resource management, various heuristic or optimization algorithms have been proposed to improve network resource utilization, focusing on routing and spectrum allocation issues.

[0004] In distributed intelligent computing center network environments, existing technologies still have several shortcomings in supporting pipelined parallel service deployments. First, existing methods generally lack deep collaborative optimization of computing and network resources. Typically, computing node selection is performed first, followed by network path and spectrum allocation. This phased approach easily leads to local optima, resulting in both idle computing resources and network congestion, reducing overall resource utilization efficiency. Second, during pipelined parallel deployment, traditional stage division and mapping strategies are mainly designed based on computing load balancing, without fully considering cross-node communication overhead and network latency. When pipeline stages are distributed across different intelligent computing centers, a large amount of intermediate data needs to be transmitted between nodes. If network resources are limited or paths are long, communication latency will increase significantly, thereby reducing pipeline execution efficiency and generating noticeable pipeline bubbles. Summary of the Invention

[0005] The purpose of this invention is to propose a pipelined parallel service deployment method and system for distributed intelligent computing center networks. It aims to achieve efficient mapping between pipeline stages and computing nodes through joint optimization of computing resources and network spectrum resources, and on this basis, complete service deployment, reduce idle waiting time in pipeline execution, and improve the overall system throughput and resource utilization efficiency.

[0006] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0007] In a first aspect, this invention proposes a pipelined parallel service deployment method for distributed intelligent computing center networks, comprising:

[0008] Obtain a set of services to be deployed, wherein each service to be deployed includes a set of stages and a set of micro-batches corresponding to that service; each stage in the set of stages corresponds to a deployment node;

[0009] For each service to be deployed in the set of services to be deployed, the following processing is performed:

[0010] Iterate through the micro-batch set of the current business in forward order, and determine whether the current business has completed forward propagation deployment. If not, propagate the micro-batch in forward order.

[0011] If completed, update the current business to the reverse propagation state, traverse the micro-batch set of the current business in reverse, and determine whether the current business has completed the reverse propagation deployment. If not completed, propagate the micro-batch in reverse order; if completed, determine that the current business deployment is complete.

[0012] Specifically, when micro-batches are propagated in either forward or reverse order...

[0013] If the current micro-batch is in a pending computation state, allocate computing power to the current micro-batch; if the allocated computing power meets the computation requirements, perform computation on the current micro-batch and update the current micro-batch's state to a pending transmission state; if the allocated computing power does not meet the computation requirements, determine that the current micro-batch is blocked and process the next micro-batch.

[0014] If the current micro-batch is in a pending transmission state, the deployment node and target transmission route for the next stage are determined by the routing algorithm, and spectrum resources are allocated for the target transmission route; it is determined whether the spectrum resources have been successfully occupied; if yes, an optical channel is established to perform cross-stage transmission of the current micro-batch, and the status of the current micro-batch is updated to a pending calculation state; if no, the next micro-batch is processed until all micro-batches have completed cross-stage transmission.

[0015] Furthermore, the multiple deployment nodes constitute a node set; each deployment node in the node set is an optical switching node in the elastic optical network, and each optical switching node carries an intelligent computing center node, which is used to provide the computing power resources and spectrum resources required for load balancing.

[0016] Furthermore, when propagating microbatches in a forward order, two constraints ensure that each microbat performs the computation task only once at different stages of the forward propagation and is then passed on to the next stage, as follows:

[0017] (1);

[0018] (2);

[0019] In the formula, Represents a set of nodes. Represents a set of business functions. Represents the micro-batch set, Represents a set of stages. Represents a set of time slots. Represents the set of time slots during forward sequential propagation. internal nodes Business micro-batch In the stage The number of calculations performed. Indicates nodes during forward sequence propagation Business micro-batch In the time slot via routing From the stage Transmit to the next stage Number of times, This is the final stage;

[0020] When propagating microbatches in reverse order, two constraints ensure that each microbat performs its computational task only once in different stages of reverse propagation and is then passed to the next stage, as follows:

[0021] (3);

[0022] (4);

[0023] In the formula, Represents the set of time slots during reverse propagation. internal nodes Business micro-batch In the stage The number of calculations performed. Indicates nodes during reverse propagation Business micro-batch In the time slot via routing From the stage Transmit to the next stage The number of times.

[0024] Furthermore, the following constraints are also satisfied when propagating micro-batches in the forward order:

[0025] (5);

[0026] (6);

[0027] (7);

[0028] In the formula, Nodes during forward sequence propagation Business micro-batch In the stage Transmit to the next stage The corresponding route, For nodes Business micro-batch In the stage Corresponding nodes, and These are the nodes during the forward sequential propagation. Business micro-batch In the stage The execution time for computational tasks and forward transfer tasks. Nodes during forward sequence propagation Business micro-batch In the stage The duration of the calculation; For nodes Business micro-batch In the stage Corresponding node; Nodes during forward sequence propagation Business micro-batch In the stage The execution time of the forward transmission task. Nodes during forward sequence propagation Business micro-batch In the stage Transmit to the next stage The duration; For nodes Business micro-batch In the stage Corresponding node; Nodes during forward sequence propagation Business micro-batch In the stage The execution time of the forward transmission task; Equation (5) indicates that in any stage of forward sequential propagation, a micro-batch must complete the computation task of this stage before it can be transmitted to the next stage; Equation (6) indicates that a micro-batch only begins to execute the computation task of this stage after it has arrived at this stage through forward sequential propagation; Equation (7) indicates that for multiple micro-batches of the same service, they are executed sequentially during forward sequential propagation, and the current micro-batch is not allowed to exceed the execution progress of the previous micro-batch of the service:

[0029] Furthermore, the following constraints are also satisfied when propagating micro-batches in reverse order:

[0030] (8);

[0031] (9);

[0032] (10);

[0033] (11);

[0034] In the formula, For nodes Business micro-batch In the stage Corresponding nodes, Nodes during reverse propagation Business micro-batch In the stage Transmit to the next stage The corresponding route; For nodes Business micro-batch In the stage Corresponding nodes, Nodes during reverse propagation Business micro-batch In the final stage The execution time for performing computational tasks; For nodes Business micro-batch In the stage Corresponding nodes, Nodes during forward sequence propagation Business The last micro-batch In the final stage The execution time of the transmission task; Nodes during forward sequence propagation Business The last micro-batch In the final stage The duration of the calculation; and These are the nodes during reverse propagation. Business micro-batch In the stage The execution time for computation and transmission tasks; Nodes during reverse propagation Business micro-batch In the stage The duration of the calculation; For nodes Business micro-batch In the stage Corresponding nodes, Nodes during reverse propagation Business micro-batch In the stage The execution time for performing computational tasks; Nodes during reverse propagation Business micro-batch In the stage Transmit to the next stage The duration; For nodes Business micro-batch In the stage Corresponding nodes, Nodes during reverse propagation Business micro-batch In the stage The execution time for the computation task; Equation (8) indicates that only the last micro-batch of the business is processed. After the forward computation task is completed in the last stage P', reverse sequential propagation is allowed; Equation (9) indicates that in any stage of reverse sequential propagation, a micro-batch must complete the computation task of this stage before it can be transferred to the next stage; Equation (10) indicates that a micro-batch only starts to execute the computation task of this stage after it has arrived at this stage through reverse sequential propagation; Equation (11) indicates that for multiple micro-batches of the same service, they are executed sequentially during reverse sequential propagation, and the current micro-batch is not allowed to exceed the execution progress of the previous micro-batch of the service.

[0035] Furthermore, in the forward sequential propagation, the allocation of spectrum resources for the target transmission route satisfies the following constraints:

[0036] (12);

[0037] (13);

[0038] (14);

[0039] In the formula, and These are the nodes during the forward sequential propagation. Business micro-batch In the stage Transmission to the next stage occupies the route The start and end indices of the frequency slot. Indicates nodes during forward sequence propagation Business micro-batch In the time slot via routing From the stage Transmit to the next stage state, =1 indicates that the transmission has been completed. =0 indicates no transmission. Let f be the maximum range of frequency slots in the link, and f be the frequency slot. It is the maximum value. Indicates nodes during forward sequence propagation Business micro-batch In the stage In the time slot Route occupied frequency gap state, =1 indicates that the status is occupied. =0 indicates an unoccupied state; Equations (12) and (13) indicate that the frequency slot range occupied during the forward sequential propagation of the micro-batch does not exceed the maximum frequency slot range of the link; Equation (14) indicates that the frequency slot range occupied during the forward sequential propagation of the service is the same as the actual occupied frequency slot range.

[0040] Furthermore, in reverse sequential propagation, the allocation of spectrum resources for the target transmission route satisfies the following constraints:

[0041] (15);

[0042] (16);

[0043] (17);

[0044] In the formula, and These are the nodes during reverse propagation. Business micro-batch In the stage Transmission to the next stage occupies the route The start and end indices of the frequency slot; Nodes during reverse propagation Business micro-batch In the time slot via routing From the stage Transmit to the next stage state, =1 indicates that the transmission has been completed. =0 indicates no transmission; Indicates nodes during reverse propagation Business micro-batch In the stage In the time slot Route occupied frequency gap state, =1 indicates that the status is occupied. =0 indicates an unoccupied state; Equations (15) and (16) indicate that the frequency slot range occupied during the reverse sequential propagation of the micro-batch does not exceed the maximum frequency slot range of the link; Equation (17) indicates that the frequency slot range occupied during the reverse sequential propagation of the service is the same as the actual occupied frequency slot range.

[0045] Furthermore, the step of determining the next-stage node and target transmission route through a routing algorithm includes:

[0046] The load value of each node is calculated according to the pre-configured load calculation formula; the node with the smallest load value is selected as the next stage node, and the target transmission route to the next stage node is determined. The load calculation formula is as follows:

[0047] (18);

[0048] In the formula, For load values, The current starting node The set of K routes to the backup intelligent computing center node. Indicates the links traversed by the route. Indicates the current link The number of frequency slots occupied. Indicates the maximum number of bandwidths in the link. This indicates the number of physical links traversed by the route. This indicates the amount of computing resources that have been used by the backup node. This represents the maximum computing power resources of the standby node. This represents the weighting coefficient.

[0049] Secondly, this invention proposes a pipelined parallel service deployment system for distributed intelligent computing center networks, including a data acquisition module and a service deployment module;

[0050] The data acquisition module is used to acquire a set of services to be deployed. Each service to be deployed in the set of services to be deployed includes a set of stages and a set of micro-batches corresponding to the service to be deployed. Each stage in the set of stages corresponds to a deployment node.

[0051] The service deployment module is used to perform the following processing for each service to be deployed in the set of services to be deployed:

[0052] Iterate through the micro-batch set of the current business in forward order, and determine whether the current business has completed forward propagation deployment. If not, propagate the micro-batch in forward order.

[0053] If completed, update the current business to the reverse propagation state, traverse the micro-batch set of the current business in reverse, and determine whether the current business has completed the reverse propagation deployment. If not completed, propagate the micro-batch in reverse order; if completed, determine that the current business deployment is complete.

[0054] Specifically, when micro-batches are propagated in either forward or reverse order...

[0055] If the current micro-batch is in a pending computation state, allocate computing power to the current micro-batch; if the allocated computing power meets the computation requirements, perform computation on the current micro-batch and update the current micro-batch's state to a pending transmission state; if the allocated computing power does not meet the computation requirements, determine that the current micro-batch is blocked, and process the next stage and the next micro-batch.

[0056] If the current micro-batch is in a pending transmission state, the deployment node and target transmission route for the next stage are determined by the routing algorithm, and spectrum resources are allocated for the target transmission route; it is determined whether the spectrum resources have been successfully occupied; if yes, an optical channel is established to perform cross-stage transmission of the current micro-batch, and the status of the current micro-batch is updated to a pending calculation state; if no, the next stage and the next micro-batch are processed until all micro-batches have completed cross-stage transmission.

[0057] Compared with the prior art, the present invention has the following advantages and technical effects:

[0058] This invention provides a pipelined parallel service deployment method and system based on network-computer collaborative load balancing. According to the distribution of computing resources and the amount of network spectrum resources, a suitable optical channel is established for the service during the transmission stage, and multi-dimensional resources are flexibly allocated to effectively improve the average computing resource utilization rate, reduce the total service completion time, and solve the multi-dimensional resource scheduling problem of deploying services in a distributed intelligent computing center network. Attached Figure Description

[0059] Figure 1 This is a flowchart of the pipeline parallel service deployment method based on computing network collaborative load balancing proposed in an embodiment of the present invention;

[0060] Figure 2 This is a flowchart of the deployment method for propagating micro-batches in a forward or reverse order, as proposed in an embodiment of the present invention. Detailed Implementation

[0061] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use.

[0062] Example 1:

[0063] This embodiment proposes a pipelined parallel service deployment method for distributed intelligent computing center networks, such as... Figure 1 As shown, it includes:

[0064] 1. Construct a distributed intelligent computing center network.

[0065] This embodiment constructs a distributed intelligent computing center network interconnected by a flexible optical network, with each optical switching node carrying one intelligent computing center node. All intelligent computing centers and their computing and spectrum resources in the network are managed uniformly, with global visibility and centralized scheduling capabilities. In this mode, computing and network resources can be uniformly orchestrated and optimized, facilitating efficient collaborative scheduling between computing and the network.

[0066] 2. Obtain the pipeline parallel business sequence.

[0067] The time required to deploy pipelined parallel services is defined as the basic time unit, represented by a time slot (TS). Given a set of pipelined parallel services, each service is represented by a time slot (TS). It means that, among them, As the source node for the business, This is the micro-batch set for this business. This is a set of stages for the service, with each stage corresponding to a different ICC node in the network. This refers to the time required for the business to complete deployment. Define micro-batch as... ,in This refers to the set of computing power required for micro-batch processing at different stages. This is the set of frequency slots required for micro-batch transmission across stages.

[0068] 3. Forward deployment of pipelined parallel services.

[0069] For each service to be deployed in the set of services to be deployed, traverse the micro-batch set of the current service in forward order, such as... Figure 2 As shown;

[0070] Determine whether the current business has completed forward propagation deployment. If not, propagate the micro-batch in the forward order.

[0071] If the current micro-batch is in a pending computation state, computing power is allocated to the current micro-batch through the current intelligent computing center node; if the allocated computing power meets the computation requirements, computation is performed on the current micro-batch, and the status of the current micro-batch is updated to a pending transmission state; if the allocated computing power does not meet the computation requirements, the current micro-batch is determined to be blocked, and the next stage and the next micro-batch are processed.

[0072] If the current micro-batch is in a pending transmission state, the next stage node and target transmission route are determined through the routing algorithm, and spectrum resources are allocated for the target transmission route; it is then determined whether the spectrum resources have been successfully occupied; if so, an optical channel is established to perform cross-stage transmission of the micro-batch, and the current micro-batch status is updated to pending calculation state; if not, the next stage and the next micro-batch are processed until all micro-batches have completed cross-stage transmission, and the service is updated to forward deployment.

[0073] 4. Forward deployment of pipelined parallel services.

[0074] When a business is in the reverse propagation state, the micro-batch set of the current business is traversed in reverse to determine whether the current business has completed the reverse propagation deployment. If it has not been completed, the micro-batch is propagated in reverse order; if it has been completed, the business deployment is considered to be completed and the business deployment is finished.

[0075] If the current micro-batch is in a pending computation state, the current intelligent computing center node needs to allocate sufficient computing power to it. If the current intelligent computing center node's computing power is insufficient, the current micro-batch will be blocked and queued for processing according to the queue order. Otherwise, if the current micro-batch is in a pending transmission state, the source-destination node pair for cross-stage transmission will be determined based on the current stage and the next stage. A suitable route will be selected for the micro-batch in the network, and the next stage node and route will be determined according to the load formula to avoid excessive computing power load on the initially reserved nodes, which could lead to micro-batch computation blocking and affect the overall business deployment progress.

[0076] In this embodiment, the next-stage node and target transmission route are determined through a routing algorithm, including:

[0077] The load value of each node is calculated according to the pre-configured load calculation formula; the node with the smallest load value is selected as the next stage node, and the target transmission route to the next stage node is determined. The load calculation formula is as follows:

[0078] (1);

[0079] In the formula, For load values, The current starting node The set of K routes to the backup intelligent computing center node. Indicates the links traversed by the route. Indicates the current link The number of frequency slots occupied. Indicates the maximum number of bandwidths in the link. This indicates the number of physical links traversed by the route. This indicates the amount of computing resources that have been used by the backup node. This represents the maximum computing power resources of the standby node. This represents the weighting coefficient.

[0080] To avoid node overload, the amount of service carried by each node needs to be limited. In this embodiment, the calculation formula for computing power constraint is as follows:

[0081] (2);

[0082] In the formula, Represents a set of nodes. Represents a set of business functions. Represents the micro-batch set, Represents a set of stages. Represents a set of time slots. This indicates the node during the forward propagation phase. Business micro-batch In the stage Perform calculations. This indicates the node during the forward propagation phase. Business micro-batch In the stage Required computing resources. Indicates the node in the reverse propagation phase. Business micro-batch In the stage Perform calculations. Indicates the node in the reverse propagation phase. Business micro-batch In the stage The required computing resources. This formula guarantees the nodes. The business demands it undertakes do not exceed its available computing resources. .

[0083] To avoid redundant computation or transmission of microbatches during forward propagation, this embodiment sets microbat deployment constraints: when propagating microbatches in forward order, two constraints ensure that each microbatchment performs its computation task only once at different stages of forward propagation and is then transmitted to the next stage, as follows:

[0084] (3);

[0085] (4);

[0086] In the formula, Represents a set of nodes. Represents a set of business functions. Represents the micro-batch set, Represents a set of stages. Represents a set of time slots. Represents the set of time slots during forward sequential propagation. internal nodes Business micro-batch In the stage The number of calculations performed. Indicates nodes during forward sequence propagation Business micro-batch In the time slot via routing From the stage Transmit to the next stage Number of times, This is the final stage;

[0087] When propagating microbatches in reverse order, two constraints ensure that each microbat performs its computational task only once in different stages of reverse propagation and is then passed to the next stage, as follows:

[0088] (5);

[0089] (6);

[0090] In the formula, Represents the set of time slots during reverse propagation. internal nodes Business micro-batch In the stage The number of calculations performed. Indicates nodes during reverse propagation Business micro-batch In the time slot via routing From the stage Transmit to the next stage The number of times.

[0091] In this embodiment, the following constraints must also be satisfied when propagating micro-batches in a forward order:

[0092] (7);

[0093] (8);

[0094] (9);

[0095] In the formula, Nodes during forward sequence propagation Business micro-batch In the stage Transmit to the next stage The corresponding route, For nodes Business micro-batch In the stage Corresponding nodes, and These are the nodes during the forward sequential propagation. Business micro-batch In the stage The execution time for computational tasks and forward transfer tasks. Nodes during forward sequence propagation Business micro-batch In the stage The duration of the calculation; For nodes Business micro-batch In the stage Corresponding node; Nodes during forward sequence propagation Business micro-batch In the stage The execution time of the forward transmission task. Nodes during forward sequence propagation Business micro-batch In the stage Transmit to the next stage The duration; For nodes Business micro-batch In the stage Corresponding node; Nodes during forward sequence propagation Business micro-batch In the stage The execution time of the forward transmission task; Equation (7) indicates that in any stage of forward sequential propagation, a micro-batch must complete the computation task of this stage before it can be transmitted to the next stage; Equation (8) indicates that a micro-batch only begins to execute the computation task of this stage after it has arrived at this stage through forward sequential propagation; Equation (9) indicates that for multiple micro-batches of the same service, they are executed sequentially during forward sequential propagation, and the current micro-batch is not allowed to exceed the execution progress of the previous micro-batch of the service:

[0096] When propagating micro-batches in reverse order, the following constraints must also be met:

[0097] (10);

[0098] (11);

[0099] (12);

[0100] (13);

[0101] In the formula, For nodes Business micro-batch In the stage Corresponding nodes, Nodes during reverse propagation Business micro-batch In the stage Transmit to the next stage The corresponding route; For nodes Business micro-batch In the stage Corresponding nodes, Nodes during reverse propagation Business micro-batch In the final stage The execution time for performing computational tasks; For nodes Business micro-batch In the stage Corresponding nodes, Nodes during forward sequence propagation Business The last micro-batch In the final stage The execution time of the transmission task; Nodes during forward sequence propagation Business The last micro-batch In the final stage The duration of the calculation; and These are the nodes during reverse propagation. Business micro-batch In the stage The execution time for computation and transmission tasks; Nodes during reverse propagation Business micro-batch In the stage The duration of the calculation; For nodes Business micro-batch In the stage Corresponding nodes, Nodes during reverse propagation Business micro-batch In the stage The execution time for performing computational tasks; Nodes during reverse propagation Business micro-batch In the stage Transmit to the next stage The duration; For nodes Business micro-batch In the stage Corresponding nodes, Nodes during reverse propagation Business micro-batch In the stage The execution time for the computation task. Equation (10) indicates that only the last micro-batch of the business is processed. After the forward computation task is completed in the last stage P', reverse sequential propagation is allowed; Equation (11) indicates that in any stage of reverse sequential propagation, a micro-batch must complete the computation task of this stage before it can be transferred to the next stage; Equation (12) indicates that a micro-batch only starts to execute the computation task of this stage after it arrives at this stage through reverse sequential propagation; Equation (13) indicates that for multiple micro-batches of the same service, they are executed sequentially during reverse sequential propagation, and the current micro-batch is not allowed to exceed the execution progress of the previous micro-batch of the service.

[0102] To ensure that the spectrum resource allocation for micro-batch forward propagation meets link capacity constraints and conforms to spectrum continuity rules, this embodiment introduces forward propagation spectrum range constraints:

[0103] (14);

[0104] (15);

[0105] (16);

[0106] (17);

[0107] (18);

[0108] In the formula, and These are the nodes during the forward sequential propagation. Business micro-batch In the stage Transmission to the next stage occupies the route The start and end indices of the frequency slot. Indicates nodes during forward sequence propagation Business micro-batch In the time slot via routing From the stage Transmit to the next stage , Let f be the maximum range of frequency slots in the link, and f be the frequency slot. It is the maximum value. Indicates nodes during forward sequence propagation Business micro-batch In the stage Towards the stage The frequency slot resources required for transmission; Equation (14) represents the starting index of the occupied frequency slot. Not greater than the terminating index Equations (15) and (16) indicate that the frequency slot range occupied by the micro-batch during forward sequential propagation does not exceed the maximum frequency slot range of the link; Equations (17) and (18) indicate that the number of frequency slots occupied by the forward sequential propagation of the service micro-batch is equal to the number of frequency slots required for micro-batch transmission.

[0109] Similarly, to ensure that the spectrum resource allocation for micro-batch reverse propagation meets link capacity constraints and conforms to spectrum continuity rules, a reverse propagation spectrum range constraint is introduced:

[0110] (19);

[0111] (20);

[0112] (twenty one);

[0113] (twenty two);

[0114] (twenty three);

[0115] In the formula, and These are the nodes during reverse propagation. Business micro-batch In the stage Transmission to the next stage occupies the route The start and end indices of the frequency slot; This represents the maximum range of frequency slots in the link. Nodes during reverse propagation Business micro-batch In the time slot via routing From the stage Transmission to stage ; Indicates nodes during reverse propagation Business micro-batch In the stage Towards the stage The frequency slot resources required for transmission; Equation (19) represents the starting index of the occupied frequency slot. Not greater than the terminating index Equations (20) and (21) indicate that the frequency slot range occupied by the micro-batch during reverse sequential propagation does not exceed the maximum frequency slot range of the link; Equation (22) indicates that the number of frequency slots occupied by the reverse sequential propagation of the service micro-batch is equal to the number of frequency slots required for micro-batch transmission.

[0116] To avoid two micro-batches occupying the same time slot of the same physical link at the same time and to achieve non-overlapping spectrum, this embodiment introduces a non-overlapping spectrum constraint:

[0117] (twenty four);

[0118] (25);

[0119] (26);

[0120] (27);

[0121] (28);

[0122] In the formula, Represents a node Business Medium and micro batches In the stage The corresponding set of forward and backward node pairs, Represents node pairs The set of links traversed by the route between them. As a binary indicator variable, when at node When the micro-batch b of service d in time slot t occupies the frequency slot f on the route corresponding to node pair r in stage p, its value is 1. and These represent the nodes during the forward propagation. Business micro-batch In the stage In the time slot And time slot t2 occupies the route frequency gap It is 1 if it is true, otherwise it is 0; and These represent the nodes during reverse propagation. Business micro-batch In the stage In the time slot And time slot t2 occupies the route frequency gap =1 otherwise =0; Equation (24) indicates that the frequency slot of each physical link can only be occupied once in any time slot; Equations (25) and (26) indicate that the number of frequency slots occupied by the micro-batch forward or reverse sequential propagation is the same as the number of frequency slots actually occupied; Equations (27) and (28) indicate that the frequency slots not occupied by the micro-batch during the micro-batch forward or reverse sequential propagation will not be marked as occupied.

[0123] In this embodiment, the spectrum occupied by forward sequential propagation and reverse sequential propagation does not overlap, satisfying the constraint:

[0124] (29);

[0125] (30);

[0126] (31);

[0127] (32);

[0128] Equations (29) and (30) indicate that the frequency slot range occupied by the forward sequential propagation of the service is the same as the actual frequency slot range occupied, and Equations (31) and (32) indicate that the frequency slot range occupied by the reverse sequential propagation of the service is the same as the actual frequency slot range occupied.

[0129] In this embodiment, by comparing the average computing power utilization, average spectrum utilization, and total service completion time under different service request volumes, the pipelined parallel service deployment algorithm based on computing network collaborative load balancing is compared with the service deployment method without task scheduling strategy. The differences between the two in the above three indicators are analyzed to evaluate the performance of the algorithm. The formula for calculating the average computing power utilization is:

[0130] (33);

[0131] The formula for calculating average spectrum utilization is:

[0132] (34);

[0133] The formula for calculating the total business completion time is:

[0134] (35);

[0135] In the formula, This represents the average computing power utilization rate. Indicates the current Time source node The amount of computing resources already consumed. Represents a set of network nodes. Represents a set of network links. This represents the amount of computing resources available to a node. Indicates the number of network nodes. Indicates the maximum processing time; Indicates average spectrum utilization. Indicates the current time link frequency gap If it is occupied, the value is 1. Represents the set of network slots. Represents a set of network links. Represents the set of network slots. Indicates the number of network slots. Indicates the number of network links. Y represents the maximum deployment time for a given service; Y represents the total completion time for the service, used to ensure that the total completion time is not less than the completion time of any single service. For the source node Business Micro-batch 1 in reverse phase 1 at the source node The time required for calculation. For the source node Business The duration of the computation of microbatch 1 in reverse phase 1.

[0136] Example 2:

[0137] This embodiment, based on the pipelined parallel service deployment method for distributed intelligent computing center networks in Embodiment 1, proposes a specific implementation method, including the following steps:

[0138] 1. In this embodiment, a 6-node, 9-link (n6s9) network is used for simulation. Each node has 100 PFLOPS of available computing power resources, and each physical link has 30 available frequency slots. The number of backup nodes and candidate paths for each node are set to 2.

[0139] 2. The number of parallel service requests in the pipeline of each node is 1 to 5. Each pipeline service has 2 micro-batches and needs to go through 3 stages of processing. The amount of computing power resources required by each micro-batch in different stages is [5, 15] PFLOPS, and the bandwidth requirement in the transmission stage is [50, 100] Gb / s, generating a computing power service sequence.

[0140] 3. By deploying services in the n6s9 network using a pipelined parallel service deployment method based on a distributed intelligent computing center network, the performance of average computing power utilization, average spectrum utilization, and total service completion time was evaluated. As the number of pipelined parallel service requests increases, this deployment method performs best, with a 25% increase in computing power utilization, a 60% increase in spectrum utilization, and a 30% reduction in total service completion time.

[0141] Example 3:

[0142] Based on Embodiment 1, this embodiment proposes a pipelined parallel service deployment system for distributed intelligent computing center networks, including a data acquisition module and a service deployment module.

[0143] The data acquisition module is used to acquire the set of services to be deployed. Each service to be deployed in the set includes its corresponding stage set and micro-batch set.

[0144] The business deployment module performs the following processing for each business to be deployed in the set of businesses to be deployed:

[0145] Iterate through the micro-batch set of the current business in forward order, and determine whether the current business has completed forward propagation deployment. If not, propagate the micro-batch in forward order.

[0146] If completed, update the current business to the reverse propagation state, traverse the micro-batch set of the current business in reverse, and determine whether the current business has completed the reverse propagation deployment. If not completed, propagate the micro-batch in reverse order; if completed, determine that the current business deployment is complete.

[0147] Specifically, when micro-batches are propagated in either forward or reverse order...

[0148] If the current microbatch is in a pending computation state, allocate computing power to the current microbatch; if the allocated computing power meets the computation requirements, perform computation on the current microbatch and update the current microbatch's state to a pending transmission state; if the allocated computing power does not meet the computation requirements, determine that the current microbatch is blocked, and process the next stage and the next microbatch.

[0149] If the current micro-batch is in a pending transmission state, the next stage node and target transmission route are determined through the routing algorithm, and spectrum resources are allocated for the target transmission route; it is then determined whether the spectrum resources have been successfully occupied; if so, an optical channel is established to perform cross-stage transmission of the micro-batch, and the current micro-batch status is updated to a pending calculation state; if not, the next stage and the next micro-batch are processed until all micro-batches have completed cross-stage transmission.

[0150] It should be noted that the above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Those skilled in the art can make various modifications and improvements without departing from the essence of the technical solution of the present invention, and these modifications and improvements also fall within the scope of protection of the present invention.

Claims

1. A pipelined parallel service deployment method for distributed intelligent computing center networks, characterized in that, include: Obtain a set of services to be deployed, wherein each service to be deployed in the set includes a stage set and a micro-batch set corresponding to the service to be deployed; Each stage in the set of stages corresponds to a deployment node; For each service to be deployed in the set of services to be deployed, the following processing is performed: Iterate through the micro-batch set of the current business in forward order, and determine whether the current business has completed forward propagation deployment. If not, propagate the micro-batch in forward order. If completed, update the current business to the reverse propagation state, traverse the micro-batch set of the current business in reverse, and determine whether the current business has completed the reverse propagation deployment. If not completed, propagate the micro-batch in reverse order; if completed, determine that the current business deployment is complete. Specifically, when micro-batches are propagated in either forward or reverse order... If the current micro-batch is in a pending computation state, allocate computing power to the current micro-batch; if the allocated computing power meets the computation requirements, perform computation on the current micro-batch and update the current micro-batch's state to a pending transmission state; if the allocated computing power does not meet the computation requirements, determine that the current micro-batch is blocked and process the next micro-batch. If the current micro-batch is in a pending transmission state, the deployment node and target transmission route for the next stage are determined by the routing algorithm, and spectrum resources are allocated for the target transmission route; it is determined whether the spectrum resources have been successfully occupied; if yes, an optical channel is established to perform cross-stage transmission of the current micro-batch, and the status of the current micro-batch is updated to a pending calculation state; if no, the next micro-batch is processed until all micro-batches have completed cross-stage transmission.

2. The pipelined parallel service deployment method for distributed intelligent computing center networks according to claim 1, characterized in that, Multiple deployment nodes constitute a node set; each deployment node in the node set is an optical switching node in the elastic optical network, and each optical switching node carries an intelligent computing center node, which is used to provide the computing power resources and spectrum resources required for load balancing.

3. The pipelined parallel service deployment method for distributed intelligent computing center networks according to claim 1, characterized in that, When propagating microbatches in forward order, two constraints ensure that each microbat performs its computational task only once at different stages of forward propagation and is then carried over to the next stage, as follows: （1）； （2）； In the formula, Represents a set of nodes. Represents a set of business functions. Represents the micro-batch set, Represents a set of stages. Represents a set of time slots. Represents the set of time slots during forward sequential propagation. internal nodes Business micro-batch In the stage The number of calculations performed. Indicates nodes during forward sequence propagation Business micro-batch In the time slot via routing From the stage Transmit to the next stage Number of times, This is the final stage; When propagating microbatches in reverse order, two constraints ensure that each microbat performs its computational task only once in different stages of reverse propagation and is then passed to the next stage, as follows: (3)； (4)； In the formula, Represents the set of time slots during reverse propagation. internal nodes Business micro-batch In the stage The number of calculations performed. Indicates nodes during reverse propagation Business micro-batch In the time slot via routing From the stage Transmit to the next stage The number of times.

4. The pipelined parallel service deployment method for distributed intelligent computing center networks according to claim 3, characterized in that, The following constraints must also be satisfied when propagating micro-batches in the forward order: (5)； (6)； (7)； In the formula, Nodes during forward sequence propagation Business micro-batch In the stage Transmit to the next stage The corresponding route, For nodes Business micro-batch In the stage Corresponding nodes, and These are the nodes during the forward sequential propagation. Business micro-batch In the stage The execution time for computational tasks and forward transfer tasks. Nodes during forward sequence propagation Business micro-batch In the stage The duration of the calculation; For nodes Business micro-batch In the stage Corresponding node; Nodes during forward sequence propagation Business micro-batch In the stage The execution time of the forward transmission task. Nodes during forward sequence propagation Business micro-batch In the stage Transmit to the next stage The duration; For nodes Business micro-batch In the stage Corresponding node; Nodes during forward sequence propagation Business micro-batch In the stage The execution time of the forward transmission task; Equation (5) indicates that in any stage of forward propagation, a micro-batch must complete the computational tasks of this stage before it can be transferred to the next stage; Equation (6) indicates that a micro-batch only begins to execute the computational tasks of this stage after it has arrived at this stage through forward propagation; Equation (7) indicates that for multiple micro-batches of the same service, they are executed sequentially during forward propagation, and the current micro-batch is not allowed to exceed the execution progress of the previous micro-batch of the service.

5. The pipelined parallel service deployment method for distributed intelligent computing center networks according to claim 4, characterized in that, The following constraints must also be satisfied when propagating micro-batches in reverse order: (8)； (9)； (10)； (11)； In the formula, Nodes during reverse propagation Business micro-batch In the stage Transmit to the next stage The corresponding route; For nodes Business micro-batch In the stage Corresponding nodes, Nodes during reverse propagation Business micro-batch In the final stage The execution time for performing computational tasks; For nodes Business micro-batch In the stage Corresponding nodes, Nodes during forward sequence propagation Business The last micro-batch In the final stage The execution time of the transmission task; Nodes during forward sequence propagation Business The last micro-batch In the final stage The duration of the calculation; and These are the nodes during reverse propagation. Business micro-batch In the stage The execution time for computation and transmission tasks; Nodes during reverse propagation Business micro-batch In the stage The duration of the calculation; For nodes Business micro-batch In the stage Corresponding nodes, Nodes during reverse propagation Business micro-batch In the stage The execution time for performing computational tasks; Nodes during reverse propagation Business micro-batch In the stage Transmit to the next stage The duration; For nodes Business micro-batch In the stage Corresponding nodes, Nodes during reverse propagation Business micro-batch In the stage The execution time for the computation task; Equation (8) indicates that only the last micro-batch of the business is processed. After the forward computation task is completed in the last stage P', reverse sequential propagation is allowed; Equation (9) indicates that in any stage of reverse sequential propagation, a micro-batch must complete the computation task of this stage before it can be transferred to the next stage; Equation (10) indicates that a micro-batch only starts to execute the computation task of this stage after it has arrived at this stage through reverse sequential propagation; Equation (11) indicates that for multiple micro-batches of the same service, they are executed sequentially during reverse sequential propagation, and the current micro-batch is not allowed to exceed the execution progress of the previous micro-batch of the service.

6. The pipelined parallel service deployment method for distributed intelligent computing center networks according to claim 5, characterized in that, In the forward sequential propagation, the allocation of spectrum resources for the target transmission route satisfies the following constraints: (12)； (13)； (14)； In the formula, and These are the nodes during the forward sequential propagation. Business micro-batch In the stage Transmission to the next stage occupies the route The start and end indices of the frequency slot. Indicates nodes during forward sequence propagation Business micro-batch In the time slot via routing From the stage Transmit to the next stage state, =1 indicates that the transmission has been completed. =0 indicates no transmission. Let f be the maximum range of frequency slots in the link, and f be the frequency slot. It is the maximum value. Indicates nodes during forward sequence propagation Business micro-batch In the stage In the time slot Route occupied frequency gap state, =1 indicates that the status is occupied. =0 indicates an unoccupied state; Equations (12) and (13) indicate that the frequency slot range occupied during the forward sequential propagation of the micro-batch does not exceed the maximum frequency slot range of the link; Equation (14) indicates that the frequency slot range occupied during the forward sequential propagation of the service is the same as the actual occupied frequency slot range.

7. The pipelined parallel service deployment method for distributed intelligent computing center networks according to claim 6, characterized in that, In reverse-order propagation, the allocation of spectrum resources for the target transmission route satisfies the following constraints: (15)； (16)； (17)； In the formula, and These are the nodes during reverse propagation. Business micro-batch In the stage Transmission to the next stage occupies the route The start and end indices of the frequency slot; Nodes during reverse propagation Business micro-batch In the time slot via routing From the stage Transmit to the next stage state, =1 indicates that the transmission has been completed. =0 indicates no transmission; Indicates nodes during reverse propagation Business micro-batch In the stage In the time slot Route occupied frequency gap state, =1 indicates that the status is occupied. =0 indicates an unoccupied state; Equations (15) and (16) indicate that the frequency slot range occupied during the reverse sequential propagation of the micro-batch does not exceed the maximum frequency slot range of the link; Equation (17) indicates that the frequency slot range occupied during the reverse sequential propagation of the service is the same as the actual occupied frequency slot range.

8. The pipelined parallel service deployment method for distributed intelligent computing center networks according to claim 1, characterized in that, The step of determining the deployment nodes and target transmission routes for the next stage using routing algorithms includes: The load value of each node is calculated according to the pre-configured load calculation formula; the node with the smallest load value is selected as the deployment node for the next stage, and the target transmission route to the deployment node for the next stage is determined. The load calculation formula is as follows: （18）； In the formula, For load values, The current starting node The set of K routes to the backup intelligent computing center node. Indicates the links traversed by the route. Indicates the current link The number of frequency slots occupied. Indicates the maximum number of bandwidths in the link. This indicates the number of physical links traversed by the route. This indicates the amount of computing resources that have been used by the backup node. This represents the maximum computing power resources of the standby node. This represents the weighting coefficient.

9. A pipelined parallel service deployment system for distributed intelligent computing center networks, characterized in that, This includes a data acquisition module and a business deployment module; The data acquisition module is used to acquire a set of services to be deployed. Each service to be deployed in the set of services to be deployed includes a set of stages and a set of micro-batches corresponding to the service to be deployed. Each stage in the set of stages corresponds to a deployment node. The service deployment module is used to perform the following processing for each service to be deployed in the set of services to be deployed: Iterate through the micro-batch set of the current business in forward order, and determine whether the current business has completed forward propagation deployment. If not, propagate the micro-batch in forward order. If completed, update the current business to the reverse propagation state, traverse the micro-batch set of the current business in reverse, and determine whether the current business has completed the reverse propagation deployment. If not completed, propagate the micro-batch in reverse order; if completed, determine that the current business deployment is complete. Specifically, when micro-batches are propagated in either forward or reverse order... If the current micro-batch is in a pending computation state, allocate computing power to the current micro-batch; if the allocated computing power meets the computation requirements, perform computation on the current micro-batch and update the current micro-batch's state to a pending transmission state; if the allocated computing power does not meet the computation requirements, determine that the current micro-batch is blocked, and process the next stage and the next micro-batch. If the current micro-batch is in a pending transmission state, the deployment node and target transmission route for the next stage are determined by the routing algorithm, and spectrum resources are allocated for the target transmission route; it is determined whether the spectrum resources have been successfully occupied; if yes, an optical channel is established to perform cross-stage transmission of the current micro-batch, and the status of the current micro-batch is updated to a pending calculation state; if no, the next stage and the next micro-batch are processed until all micro-batches have completed the cross-stage transmission.

Images