Distributed large model scheduling method based on deterministic network

Abstract

The invention discloses a distributed large model scheduling method based on a deterministic network, and the method comprises the following steps: S1, obtaining a network quality value of a single link through monitoring parameters in the deterministic network; establishing a whole network quality parameter map and a whole network quality weight map; obtaining a single-node network quality value and an overall network quality value of each node in the deterministic network; s2, calculating available values of all deterministic network large models in the deterministic network, and determining main large model nodes and standby large model nodes; and S3, allocating a large model task. The method has the advantages that on the basis of the real-time network quality map and node capability evaluation, the main / standby model nodes are dynamically elected, tasks are intelligently distributed, and the resource utilization rate is maximized; the periodic reelection mechanism can adapt to network fluctuation and node state change, and continuity and reliability of task execution are ensured.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence large model technology, and in particular to a distributed large model scheduling method based on a deterministic network. Background Art

[0002] Deterministic Networking (DetNet) is an emerging network technology designed to provide reliable end-to-end transmission services for various businesses. This type of network can ensure that data packets arrive at their destination within a specific time window and control key performance indicators such as latency, jitter, packet loss rate, and bandwidth, thereby meeting the stringent requirements for low latency, high reliability, and stability.

[0003] Large AI models are deep learning models with numerous parameters and complex structures, trained on large datasets. These models are typically based on the Transformer architecture and are capable of handling complex tasks such as natural language understanding, image recognition and generation, and speech recognition. Due to their large parameter size, these models typically rely on massively parallel computing, high-performance computing hardware, and high-bandwidth, low-latency, and reliable networks to ensure efficient communication between distributed computing nodes and reduce data transmission bottlenecks.

[0004] Distributed large AI models are powerful neural network models with massive parameters. These models are trained and inferred by distributing them across multiple computing nodes. These models typically contain billions or even trillions of parameters and demonstrate exceptional performance in fields such as natural language processing, computer vision, and speech recognition. Because a single computing device cannot meet the computational demands of these models, distributed computing technologies such as data parallelism, model parallelism, and pipeline parallelism are required to achieve efficient training and inference.

[0005] Distributed training of large models places extremely high demands on networks, as these models typically contain billions to trillions of parameters and require processing massive datasets. To effectively train, the network must possess high bandwidth, low latency, and support efficient overlap of computation and communication. However, when network fluctuations and node status changes occur, the continuity and reliability of task execution need to be further enhanced. Summary of the Invention

[0006] To this end, the present invention provides a distributed large model scheduling method based on a deterministic network to solve the problems in the prior art.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] A distributed large model scheduling method based on a deterministic network includes the following steps:

[0009] S1. Determine the network quality of individual links by monitoring parameters in the deterministic network. Build a network-wide quality parameter map based on the network quality of individual links and the network's routing topology.

[0010] Based on the full network quality parameter map, a full network quality weight map is established; through the full network quality parameter map of the deterministic network and the full network quality weight map, the single node network quality value of each node in the deterministic network and the output overall network quality value are obtained;

[0011] S2. Calculate the available values ​​of all deterministic network large models in the deterministic network and determine the main large model and backup large model nodes;

[0012] S3. After the active and standby large models are determined, they will establish connections with other large models. Upon receiving a client request, they will allocate tasks to each large model in proportion to the availability of the deterministic network large model.

[0013] Furthermore: the parameters in the deterministic network include: (1) parameters of the data communication layer, including delay, bandwidth, jitter, packet loss rate, bit error rate and frame error rate; (2) parameters of the optical transmission layer, including delay, bandwidth, optical power, dispersion, signal-to-noise ratio, loss and bit error rate.

[0014] Furthermore: the parameter map includes: (1) parameters of the data communication layer, including delay map, bandwidth map, jitter map, packet loss rate map, bit error rate map, and frame error rate map; (2) parameters of the optical transmission layer, including delay map, bandwidth map, optical power map, dispersion map, signal-to-noise ratio map, loss map, and bit error rate map.

[0015] Furthermore, the overall network quality value is calculated as follows:

[0016] (1) By establishing a full network quality parameter map and a full network quality weight map, the key network indicators and their weights of each link are clarified;

[0017] (2) Normalize the indicators, including negative and positive indicators;

[0018] (3) Calculate the normalized comprehensive score of each link;

[0019] The comprehensive score of each link = the sum of the normalized indicator scores and the corresponding weights;

[0020] (4) Calculate the overall network quality value based on the comprehensive score of each link;

[0021] The normalized composite scores of all links are averaged to obtain the overall network quality value.

[0022] Further: The normalization formula of the negative indicator is:

[0023] Q=(V_max-V) / (V_max-V_min);

[0024] The normalization formula for positive indicators is:

[0025] Q=(V-V_min) / (V_max-V_min);

[0026] Among them, V is the actual value, V_max and V_min are the maximum and minimum values ​​of the indicator respectively; the normalization range is determined by the minimum and maximum values ​​of each indicator in the entire network.

[0027] Further: the specific steps of S2 are:

[0028] (1) The CPU and GPU parameters of the large model are used to reflect the single-node large model capability value. The single-node large model capability value and the set single-node large model parameter weights are used to calculate the single-node large model measurement value.

[0029] (2) The large model in the deterministic network establishes a distributed large model alliance to calculate the available value of the deterministic network large model of all nodes;

[0030] Deterministic network large model availability value = single node network quality value * single node large model measurement value;

[0031] The node with the highest availability value of the deterministic network large model is used as the main large model, and the node with the second highest availability value of the deterministic network large model is used as the backup large model.

[0032] Furthermore: the election of the main big model and backup big model nodes changes according to the changes of the network and the big model, that is, the overall network quality value and the deterministic network big model measurement value are periodically calculated, and when the deterministic network big model availability value reaches the set threshold, re-election is performed.

[0033] The present invention has the following advantages: based on real-time network quality maps and node capability assessments (such as CPU / GPU performance), it dynamically selects active / standby model nodes and intelligently allocates tasks to maximize resource utilization; the periodic re-election mechanism can adapt to network fluctuations and node status changes to ensure the continuity and reliability of task execution.

[0034] Other features and advantages of the present invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] To more intuitively illustrate the prior art and the present application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be considered as limiting conditions for implementing the present application; for example, based on the technical concepts disclosed in this application and the exemplary drawings, those skilled in the art are capable of easily making routine adjustments or further optimizations to the addition / reduction / attribution division, specific shapes, positional relationships, connection methods, dimensional ratios, etc. of certain units (components).

[0036] Figure 1 A network diagram of a distributed large model scheduling method based on a deterministic network provided in an embodiment of the present invention.

[0037] Figure 2 A calculation flow chart is obtained for determining the available value of a network large model in a distributed large model scheduling method based on a deterministic network of the present invention. DETAILED DESCRIPTION

[0038] The following specific embodiments illustrate the implementation of the present invention. People familiar with this technology can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. It should be understood that these embodiments are only to further illustrate the present invention and cannot be understood as limiting the scope of protection of the present invention. Technical engineers in this field can make some non-essential improvements and adjustments to the present invention based on the content of the above invention; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0039] See also Figure 1 The large model nodes in the network include large model node A, large model node B, large model node C, and large model node D.

[0040] The large model node consists of a large model and a network device, that is, the large model node A includes the large model A and the network device A; therefore, the large model node includes the large model A, the network device A, the large model B, the network device B, the large model C, the network device C, the large model D and the network device D.

[0041] The network equipment consists of data communication equipment and optical transmission equipment, that is, network equipment A includes data communication equipment A and optical transmission equipment A. Therefore, the network equipment includes data communication equipment A, data communication equipment B, data communication equipment C, data communication equipment D, optical transmission equipment A, optical transmission equipment B, optical transmission equipment C and optical transmission equipment D.

[0042] A distributed large model scheduling method based on a deterministic network includes the following steps:

[0043] S1. Calculate the network quality value of the deterministic network;

[0044] (1) Awareness of network resource availability is achieved by monitoring parameters in deterministic networks. The parameters in deterministic networks include: 1) parameters of the data communication layer, including delay, bandwidth, jitter, packet loss rate, bit error rate, and frame error rate; 2) parameters of the optical transmission layer, including delay, bandwidth, optical power, dispersion, signal-to-noise ratio, loss, and bit error rate.

[0045] Deterministic networks are designed to provide a guaranteed level of network performance for data transmission. These parameters, such as latency and bandwidth, will not fluctuate significantly, and fluctuations will not exceed set acceptable thresholds. Deterministic networks are suitable for applications that require high-precision synchronization and high stability.

[0046] In this embodiment, in order to simplify the discussion, data communication layer delay, data communication layer bandwidth, data communication layer jitter, data communication layer packet loss rate, optical transmission layer delay, optical transmission layer bandwidth, and optical transmission layer optical attenuation are selected.

[0047] (2) Implementing a parameter map for the entire network based on network parameters. The parameter map includes: 1) parameters of the data communication layer, including delay map, bandwidth map, jitter map, packet loss rate map, bit error rate map, and frame error rate map; 2) parameters of the optical transmission layer, including delay map, bandwidth map, optical power map, dispersion map, signal-to-noise ratio map, loss map, and bit error rate map.

[0048] In this embodiment, to simplify the discussion, a data communication layer delay map, a data communication layer bandwidth map, a data communication layer jitter map, a data communication layer packet loss rate map, an optical transmission layer delay map, an optical transmission layer bandwidth map, and an optical transmission layer optical attenuation map are constructed. The network quality parameter map is shown in Table 1 below:

[0049] Table 1

[0050]

[0051]

[0052] (3) According to the parameter map, a weight map is established to form a network quality weight map as shown in Table 2 below:

[0053] Table 2

[0054]

[0055] For negative indicators (the lower the value, the better), the normalization method is opposite to that of positive indicators.

[0056] By establishing a full network quality parameter map and a full network quality weight map, the key network indicators and their weights of each link are clarified.

[0057] The processing method of indicator normalization includes negative indicators and positive indicators.

[0058] The normalization formula for negative indicators (delay, jitter, packet loss rate, and optical attenuation, the lower the value, the better) is:

[0059] Q=(V_max-V) / (V_max-V_min);

[0060] The normalization formula for positive indicators (bandwidth, the higher the value, the better) is:

[0061] Q=(V-V_min) / (V_max-V_min);

[0062] Among them, V is the actual value, V_max and V_min are the maximum and minimum values ​​of the indicator respectively.

[0063] The normalization range is determined by the minimum and maximum values ​​of each indicator in the entire network.

[0064] The calculation method of the link comprehensive score is as follows:

[0065] The comprehensive score of each link = the sum of the normalized indicator scores and the corresponding weights;

[0066] For ease of comparison, the comprehensive score is divided by the total weight (70) to obtain a link score in the range of 0 to 1.

[0067] The overall network quality value is calculated as follows:

[0068] The overall network quality value is obtained by taking the simple average of the normalized scores of all links. The closer the value is to 1, the higher the overall network quality.

[0069] Since the dimensions and evaluation directions of each indicator are different, for example, the dimension of delay is ms, and the evaluation direction is reverse, that is, the smaller the value, the better; the dimension of bandwidth is M, and the evaluation direction is forward, that is, the larger the value, the better, so normalization is required.

[0070] The minimum and maximum values ​​of each indicator are as follows:

[0071] Data communication layer delay (ms): minimum = 10, maximum = 20;

[0072] Data communication layer bandwidth (M): minimum = 100, maximum = 500;

[0073] Datacom layer jitter (UI): minimum = 0.05, maximum = 0.09;

[0074] Data communication layer packet loss rate (%): minimum = 0.001, maximum = 0.004;

[0075] Optical transmission layer delay (ms): minimum = 9, maximum = 12;

[0076] Optical transmission layer bandwidth (M): minimum = 1000, maximum = 3000;

[0077] Optical transmission layer optical attenuation (dB): minimum = 8, maximum = 20.

[0078] The link network quality values ​​include AB link, AC link, AD link, BC link, BD link, and CD link.

[0079] The link network quality values ​​of link AB are as follows:

[0080] Data communication layer delay: actual value = 10,

[0081] Q=(20-10) / (20-10)=10 / 10=1;

[0082] Data communication layer bandwidth: actual value = 100,

[0083] Q=(100-100) / (500-100)=0 / 400=0;

[0084] Data communication layer jitter: actual value = 0.05,

[0085] Q=(0.09-0.05) / (0.09-0.05)=0.04 / 0.04=1;

[0086] Data communication layer packet loss rate: actual value = 0.001, Q = (0.004-0.001) / (0.004-0.001) = 0.003 / 0.003 = 1; optical transmission layer delay: actual value = 9,

[0087] Q=(12-9) / (12-9)=3 / 3=1;

[0088] Optical transmission layer bandwidth: actual value = 1000,

[0089] Q=(1000-1000) / (3000-1000)=0 / 2000=0;

[0090] Light transmission layer light attenuation: actual value = 15,

[0091] Q=(20-15) / (20-8)=5 / 12≈0.4167.

[0092] The weighted calculation is as follows:

[0093] Data communication layer delay score: 1*10=10;

[0094] Data communication layer bandwidth score: 0*5=0;

[0095] Datacom layer jitter score: 1*5=5;

[0096] Data communication layer packet loss rate score: 1*10=10;

[0097] Optical transmission layer latency score: 1*20=20;

[0098] Optical transport layer bandwidth score: 0*10=0;

[0099] Light transmission layer light attenuation score: 0.4167*10≈4.167;

[0100] Total score = 10 + 0 + 5 + 10 + 20 + 0 + 4.167 ≈ 49.167; link network quality value = 49.167 / 70 ≈ 0.7024.

[0101] The link network quality values ​​of the AC link are as follows:

[0102] Data communication layer delay: actual value = 14,

[0103] Q=(20-14) / (20-10)=6 / 10=0.6;

[0104] Data communication layer bandwidth: actual value = 200,

[0105] Q=(200-100) / (500-100)=100 / 400=0.25;

[0106] Data communication layer jitter: actual value = 0.07,

[0107] Q=(0.09-0.07) / (0.09-0.05)=0.02 / 0.04=0.5;

[0108] Data communication layer packet loss rate: actual value = 0.002,

[0109] Q=(0.004-0.002) / (0.004-0.001)≈0.6667;

[0110] Optical transmission layer delay: actual value = 12,

[0111] Q=(12-12) / (12-9)=0 / 3=0;

[0112] Optical transmission layer bandwidth: actual value = 2000,

[0113] Q=(2000-1000) / (3000-1000)=1000 / 2000=0.5;

[0114] Light transmission layer light attenuation: actual value = 20,

[0115] Q=(20-20) / (20-8)=0 / 12=0.

[0116] The weighted calculation is as follows:

[0117] Data communication layer delay: 0.6*10=6;

[0118] Data communication layer bandwidth: 0.25*5=1.25;

[0119] Data communication layer jitter: 0.5*5=2.5;

[0120] Data communication layer packet loss rate: 0.6667*10≈6.667;

[0121] Optical transmission layer delay: 0*20=0;

[0122] Optical transmission layer bandwidth: 0.5*10=5;

[0123] Light attenuation of the optical transmission layer: 0*10=0;

[0124] Total score ≈ 6 + 1.25 + 2.5 + 6.667 + 0 + 5 + 0 ≈ 21.417; link network quality value = 21.417 / 70 ≈ 0.3060.

[0125] The link network quality values ​​of the AD link are as follows:

[0126] Data communication layer delay: actual value = 12,

[0127] Q=(20-12) / (20-10)=8 / 10=0.8;

[0128] Data communication layer bandwidth: actual value = 150,

[0129] Q=(150-100) / (500-100)=50 / 400=0.125;

[0130] Data communication layer jitter: actual value = 0.08,

[0131] Q=(0.09-0.08) / (0.09-0.05)=0.01 / 0.04=0.25;

[0132] Data communication layer packet loss rate: actual value = 0.004,

[0133] Q=(0.004-0.004) / (0.004-0.001)=0 / 0.003=0;

[0134] Optical transmission layer delay: actual value = 10,

[0135] Q=(12-10) / (12-9)≈2 / 3≈0.6667;

[0136] Optical transmission layer bandwidth: actual value = 3000,

[0137] Q=(3000-1000) / (3000-1000)=2000 / 2000=1;

[0138] Light transmission layer light attenuation: actual value = 10,

[0139] Q=(20-10) / (20-8)=10 / 12≈0.8333.

[0140] The weighted calculation is as follows:

[0141] Data communication layer delay: 0.8*10=8;

[0142] Data communication layer bandwidth: 0.125*5=0.625;

[0143] Data communication layer jitter: 0.25*5=1.25;

[0144] Data communication layer packet loss rate: 0*10=0;

[0145] Optical transmission layer delay: 0.6667*20≈13.334;

[0146] Optical transmission layer bandwidth: 1*10=10;

[0147] Light attenuation of optical transmission layer: 0.8333*10≈8.333;

[0148] Total score ≈ 8 + 0.625 + 1.25 + 0 + 13.334 + 10 + 8.333 ≈ 41.542; link network quality value = 41.542 / 70 ≈ 0.5935.

[0149] The link network quality values ​​of the BC link are as follows:

[0150] Data communication layer delay: actual value = 11,

[0151] Q=(20-11) / (20-10)=9 / 10=0.9;

[0152] Data communication layer bandwidth: actual value = 110,

[0153] Q=(110-100) / (500-100)=10 / 400=0.025;

[0154] Data communication layer jitter: actual value = 0.06,

[0155] Q=(0.09-0.06) / (0.09-0.05)=0.03 / 0.04=0.75;

[0156] Data communication layer packet loss rate: actual value = 0.001,

[0157] Q=(0.004-0.001) / (0.004-0.001)=0.003 / 0.003=1;

[0158] Optical transmission layer delay: actual value = 10,

[0159] Q=(12-10) / (12-9)≈2 / 3≈0.6667;

[0160] Optical transmission layer bandwidth: actual value = 2000,

[0161] Q=(2000-1000) / (3000-1000)=1000 / 2000=0.5;

[0162] Light transmission layer light attenuation: actual value = 8,

[0163] Q=(20-8) / (20-8)=12 / 12=1.

[0164] The weighted calculation is as follows:

[0165] Data communication layer delay: 0.9*10=9;

[0166] Data communication layer bandwidth: 0.025*5=0.125;

[0167] Data communication layer jitter: 0.75*5=3.75;

[0168] Data communication layer packet loss rate: 1*10=10;

[0169] Optical transmission layer delay: 0.6667*20≈13.334;

[0170] Optical transmission layer bandwidth: 0.5*10=5;

[0171] Light attenuation in the optical transmission layer: 1*10=10;

[0172] Total score ≈ 9 + 0.125 + 3.75 + 10 + 13.334 + 5 + 10 ≈ 51.209; link network quality value = 51.209 / 70 ≈ 0.7316.

[0173] The link network quality values ​​of BD links are as follows:

[0174] Data communication layer delay: actual value = 10,

[0175] Q = (20-10) / (20-10) = 1;

[0176] Data communication layer bandwidth: actual value = 125,

[0177] Q=(125-100) / (500-100)=25 / 400=0.0625;

[0178] Data communication layer jitter: actual value = 0.09,

[0179] Q=(0.09-0.09) / (0.09-0.05)=0 / 0.04=0;

[0180] Data communication layer packet loss rate: actual value = 0.002,

[0181] Q=(0.004-0.002) / (0.004-0.001)≈0.6667;

[0182] Optical transmission layer delay: actual value = 9,

[0183] Q=(12-9) / (12-9)=3 / 3=1;

[0184] Optical transmission layer bandwidth: actual value = 1000,

[0185] Q=(1000-1000) / (3000-1000)=0 / 2000=0

[0186] Optical transmission layer light attenuation: actual value = 16,

[0187] Q=(20-16) / (20-8)=4 / 12≈0.3333.

[0188] The weighted calculation is as follows:

[0189] Data communication layer delay: 1*10=10;

[0190] Data communication layer bandwidth: 0.0625*5=0.3125;

[0191] Data communication layer jitter: 0*5=0;

[0192] Data communication layer packet loss rate: 0.6667*10≈6.667;

[0193] Optical transmission layer delay: 1*20=20;

[0194] Optical transmission layer bandwidth: 0*10=0;

[0195] Light attenuation in the optical transmission layer: 0.3333*10≈3.333;

[0196] Total score ≈ 10 + 0.3125 + 0 + 6.667 + 20 + 0 + 3.333 ≈ 40.3125; link network quality value = 40.3125 / 70 ≈ 0.5759.

[0197] The link network quality values ​​of the CD link are as follows:

[0198] Data communication layer delay: actual value = 20,

[0199] Q=(20-20) / (20-10)=0 / 10=0;

[0200] Data communication layer bandwidth: actual value = 500,

[0201] Q=(500-100) / (500-100)=400 / 400=1;

[0202] Data communication layer jitter: actual value = 0.05,

[0203] Q=(0.09-0.05) / (0.09-0.05)=0.04 / 0.04=1;

[0204] Data communication layer packet loss rate: actual value = 0.001,

[0205] Q=(0.004-0.001) / (0.004-0.001)=0.003 / 0.003=1;

[0206] Optical transmission layer delay: actual value = 10,

[0207] Q=(12-10) / (12-9)≈2 / 3≈0.6667;

[0208] Optical transmission layer bandwidth: actual value = 1000,

[0209] Q=(1000-1000) / (3000-1000)=0 / 2000=0;

[0210] Light transmission layer light attenuation: actual value = 17,

[0211] Q=(20-17) / (20-8)=3 / 12=0.25.

[0212] The weighted calculation is as follows:

[0213] Data communication layer delay: 0*10=0;

[0214] Data communication layer bandwidth: 1*5=5;

[0215] Data communication layer jitter: 1*5=5;

[0216] Data communication layer packet loss rate: 1*10=10;

[0217] Optical transmission layer delay: 0.6667*20≈13.334;

[0218] Optical transmission layer bandwidth: 0*10=0;

[0219] Light attenuation of optical transmission layer: 0.25*10=2.5;

[0220] Total score ≈ 0+5+5+10+13.334+0+2.5 ≈ 35.834;

[0221] Link network quality value = 35.834 / 70≈0.5119.

[0222] Through the operation of the weight map and the parameter map, the network quality value of each node in the deterministic network is obtained.

[0223] To calculate the network quality value of node A in the large model, it is necessary to first calculate the network parameter matrix and the weight matrix of node A through the weight map and parameter map.

[0224] The network parameter matrix of node A is shown in Table 3 below.

[0225] Table 3

[0226]

[0227]

[0228] The network weight matrix of node A is shown in Table 4 below.

[0229] Table 4

[0230]

[0231] According to the calculation process of the overall network quality value, the scores from node A to each link are as follows:

[0232] Link AB: 0.7024;

[0233] Link AC: 0.3060;

[0234] Link AD: 0.5935;

[0235] Therefore, the average network quality of node A can be expressed as the sum of the quality of each link divided by the number of links, that is:

[0236] A=(0.7024+0.3060+0.5935) / 3≈0.5340.

[0237] To calculate the network quality value of node B in the large model, it is necessary to first calculate the network parameter matrix and the weight matrix of node B through the weight map and parameter map.

[0238] The Node B network parameter matrix is ​​shown in Table 5 below.

[0239] Table 5

[0240]

[0241] The network weight matrix of node B is shown in the following table.

[0242] Table 6

[0243]

[0244] According to the calculation process of the overall network quality value, the scores from node B to each link are as follows:

[0245] Link BA: 0.7024;

[0246] Link BC: 0.7316;

[0247] Link BD: 0.5759;

[0248] Therefore, the average network quality of Node B can be expressed as the sum of the quality of each link divided by the number of links, that is:

[0249] B=(0.7024+0.7316+0.5759) / 3≈0.6700.

[0250] To calculate the network quality value of node C in the large model, it is necessary to first calculate the network parameter matrix and the weight matrix of node C through the weight map and parameter map.

[0251] The network parameter matrix of node C is shown in Table 7 below.

[0252] Table 7

[0253]

[0254] The network weight matrix of node C is shown in Table 8 below.

[0255] Table 8

[0256]

[0257] According to the calculation process of the overall network quality value, the scores from node C to each link are as follows:

[0258] CA link: 0.3060;

[0259] CB link: 0.7316;

[0260] CD link: 0.5119;

[0261] Therefore, the average network quality of node C can be expressed as the sum of the quality of each link divided by the number of links, that is:

[0262] C=(0.3060+0.7316+0.5119) / 3≈0.5165.

[0263] To calculate the network quality value of node D in the large model, it is necessary to first calculate the network parameter matrix and the weight matrix of node D through the weight map and parameter map.

[0264] The network parameter matrix of node D is shown in Table 9 below.

[0265] Table 9

[0266]

[0267] The network weight matrix of node D is shown in Table 10 below.

[0268] Table 10

[0269]

[0270] According to the calculation process of the overall network quality value, the scores from node D to each link are as follows:

[0271] DA link: 0.5935;

[0272] DB link: 0.5759;

[0273] DC link: 0.5119;

[0274] Therefore, the average network quality of node D can be expressed as the sum of the quality of each link divided by the number of links, that is:

[0275] D=(0.5935+0.5759+0.5119) / 3≈0.5604.

[0276] The final scores of each node are shown in Table 11 below.

[0277] Table 11

[0278] Network Node Single node network quality value A 0.5340 B 0.6700 C 0.5165 D 0.5604

[0279] The simple averaging method is used to average the normalized scores of all links.

[0280] The scores of each link are: AB link 0.7024; AC link 0.3060; AD link 0.5935; BC link 0.7316; BD link 0.5759; CD link 0.5119.

[0281] The overall network quality value = (0.7024 + 0.3060 + 0.5935 + 0.7316 + 0.5759 + 0.5119) / 6 ≈ 0.5702.

[0282] That is, the average network quality value of the entire network link is 0.5702. If it is less than this value, the network quality is considered poor, and if it is greater than this value, the network quality is considered good.

[0283] S2. Establish a distributed large model alliance;

[0284] (1) Connect each large model to realize the large model alliance.

[0285] (2) Calculate the measurement value of the large model. Calculate the value based on the capability value table shown in Table 12 below.

[0286] Table 12

[0287]

[0288] The parameter weights of the large model are shown in Table 13 below.

[0289] Table 13

[0290]

[0291] In this embodiment,

[0292] CPU 16-bit floating-point calculations and GPU 16-bit floating-point calculations use TFLOPS (trillion floating-point operations per second) as units to reflect the computing power of the CPU / GPU.

[0293] Generally, GPU computing power far exceeds that of CPU, so GPU resources are more important in deep learning tasks.

[0294] CPU memory and GPU memory:

[0295] CPU memory (GB) affects data loading and preprocessing capabilities; GPU memory (GB): affects model size and batch processing capabilities, especially deep learning inference.

[0296] The large model score is weighted based on the CPU's computing power and memory, as well as the GPU's computing power and video memory.

[0297] The specific calculation formula is as follows:

[0298] S (score) = 0.2xCPU calculation + 0.2xCPU memory + 0.3xGPU calculation + 0.3xGPU display. In this implementation, the calculation process of large model A, large model B, large model C and large model D is as follows.

[0299] The measurement value of the large model node A is:

[0300] A=0.2×25+0.2×128+0.3×200+0.3×40=5+25.6+60+12=106.6.

[0301] The measurement value of the large model node B is:

[0302] B=0.2×30+0.2×256+0.3×250+0.3×48=6+51.2+75+14.4=136.8

[0303] The measurement value of the large model node C is:

[0304] C=0.2×22+0.2×96+0.3×180+0.3×32=4.4+19.2+54+9.6=90.8

[0305] The measurement value of the large model node D is:

[0306] D=0.2×35+0.2×512+0.3×300+0.3×80=7+102.4+90+24=203.4

[0307] (3) Deterministic network large model available value.

[0308] The large model in the deterministic network establishes a distributed large model alliance and selects the master and backup large models.

[0309] In this embodiment, to simplify the description, the weight value of the large model capability of each node is 1, that is, the weight is the same, and the large model node capability value is calculated.

[0310]

[0311]

[0312] In this embodiment, the final score of the deterministic network large model availability value is the single-node network quality value multiplied by the large model measurement value.

[0313] According to calculations, large model node D is the main large model node of the alliance, and large model node C is the backup large model node of the alliance.

[0314] The election of the active and standby large model nodes changes according to the changes in the network and the large model. That is, the overall network quality value and the measurement value of the large model are calculated periodically. When the availability value of the deterministic network large model reaches a certain threshold, a re-election is performed.

[0315] S3. Assigning tasks to large models.

[0316] After the main and backup large models are determined, they will establish connections with other large models. After receiving customer requests, they will allocate tasks to each large model in proportion based on the available value of the deterministic network large model.

[0317] When any node in the large model alliance receives a computing task, it passes the task to the large model node D, and the main large model node D distributes the computing task according to the available value ratio of the deterministic network large model.

[0318] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A distributed large model scheduling method based on deterministic network, characterized in that: The following steps are involved: S1. Determine the network quality of individual links by monitoring parameters in the deterministic network. Build a network-wide quality parameter map based on the network quality of individual links and the network's routing topology. Based on the full network quality parameter map, a full network quality weight map is established; through the full network quality parameter map and the full network quality weight map of the deterministic network, the single node network quality value and the overall network quality value of each node in the deterministic network are obtained; S2. Calculate the available values ​​of all deterministic network large models in the deterministic network and determine the main large model and backup large model nodes; S3. After the active and standby large models are determined, they will establish connections with other large models. Upon receiving a client request, they will allocate tasks to each large model in proportion to the availability of the deterministic network large model.

2. A distributed large model scheduling method based on a deterministic network according to claim 1, characterized in that: The parameters in the deterministic network include: (1) parameters of the data communication layer, including delay, bandwidth, jitter, packet loss rate, bit error rate and frame error rate; (2) parameters of the optical transmission layer, including delay, bandwidth, optical power, dispersion, signal-to-noise ratio, loss and bit error rate.

3. The distributed large model scheduling method based on a deterministic network according to claim 1, characterized in that: The parameter maps include: (1) parameters of the data communication layer, including delay map, bandwidth map, jitter map, packet loss rate map, bit error rate map, and frame error rate map; (2) parameters of the optical transmission layer, including delay map, bandwidth map, optical power map, dispersion map, signal-to-noise ratio map, loss map, and bit error rate map.

4. The distributed large model scheduling method based on a deterministic network according to claim 1, characterized in that: The overall network quality value is calculated as follows: (1) By establishing a full network quality parameter map and a full network quality weight map, the key network indicators and their weights of each link are clarified; (2) Normalize the indicators, including negative and positive indicators; (3) Calculate the normalized comprehensive score of each link; The comprehensive score of each link = the sum of the normalized indicator scores and the corresponding weights; (4) Calculate the overall network quality value based on the comprehensive score of each link; The normalized composite scores of all links are averaged to obtain the overall network quality value.

5. A distributed large model scheduling method based on a deterministic network according to claim 4, characterized in that: The normalization formula of the negative indicator is: Q=(V_max-V) / (V_max-V_min); The normalization formula for positive indicators is: Q=(V-V_min) / (V_max-V_min); Among them, V is the actual value, V_max and V_min are the maximum and minimum values ​​of the indicator respectively; the normalization range is determined by the minimum and maximum values ​​of each indicator in the entire network.

6. The distributed large model scheduling method based on a deterministic network according to claim 1, characterized in that: The specific steps of S2 are: (1) The CPU and GPU parameters of the large model are used to reflect the single-node large model capability value. The single-node large model capability value and the set single-node large model parameter weights are used to calculate the single-node large model measurement value. (2) The large model in the deterministic network establishes a distributed large model alliance to calculate the available value of the deterministic network large model of all nodes; Deterministic network large model availability value = single node network quality value * single node large model measurement value; The node with the highest availability value of the deterministic network large model is used as the main large model, and the node with the second highest availability value of the deterministic network large model is used as the backup large model.

7. A distributed large model scheduling method based on a deterministic network according to claim 6, characterized in that: The election of the main large model and backup large model nodes changes according to the changes of the network and the large model, that is, the overall network quality value and the deterministic network large model measurement value are periodically calculated. When the deterministic network large model availability value reaches the set threshold, re-election is performed.

Images