Multi-port SSD low-power collaborative control methods, systems, media, solid-state drives, and devices

Abstract

This invention relates to the field of solid-state drive (SSD) technology, specifically to a low-power collaborative control method, system, medium, SSD, and device for multi-port SSDs. The method includes the following steps: S1, obtaining the load intensity trend and service type probability distribution of each port within a future set time window; S2, calculating energy-saving priority and response priority for each port, and generating a collaborative power consumption control strategy that allows ports to be in different power consumption states; S3, triggering a state synchronization protocol through hardware control logic; S4, controlling the target port to switch power consumption states through hardware control logic; S5, restoring the port state and performing consistency verification through hardware control logic. Compared with existing technologies, this method integrates AI prediction, dynamic priority arbitration, and hardware collaborative control, enabling intelligent load prediction, collaborative scheduling of port resources, and ensuring smooth and reliable state switching, achieving collaborative optimization of power consumption, performance, and reliability.

Description

Technical Field

[0001] This invention relates to the field of solid-state drive technology, specifically to a method, system, medium, solid-state drive, and device for low-power collaborative control of multi-port SSDs. Background Technology

[0002] With the increasing demands for storage performance and energy efficiency in data centers, AI servers, and autonomous driving, high-availability SSDs supporting multiple ports are being widely adopted. To reduce system power consumption, PCIe links typically support low-power states such as Active State Power Management (ASPM), including L0s and L1 states. However, in multi-port SSD scenarios, simply applying single-port low-power strategies can lead to serious compatibility and reliability issues.

[0003] Existing technologies primarily focus on power consumption control based on fixed rules. For example, patent CN119356873B discloses a method for adjusting dual-port power consumption based on ASPM link idle state and cache margin value to detect service pressure. While this method can achieve some energy-saving effects, it has significant limitations: 1) Passive and coarse-grained decision-making: Reactive control based solely on the current idle state and cache level cannot predict future load trends, easily leading to frequent and unnecessary state switching. 2) Lack of port coordination mechanism: Its control logic is essentially a simple superposition of two independent single-port controls, or requires ports to "advance and retreat together," failing to consider the need for differentiated and coordinated power consumption management when the load between ports is unbalanced. 3) Neglecting the side effects of state switching: No dedicated mechanism is designed to ensure data consistency, transaction integrity, and response latency stability during port state switching. At the moment of switching, data loss, command out-of-order delivery, or service jitter of tens of milliseconds perceptible to the host may occur, failing to meet the stringent requirements of AI training, autonomous driving, and other scenarios that are extremely sensitive to latency and data consistency. The core contradiction of ensuring high availability and high performance of multi-port SSDs under complex and dynamic loads while achieving deep energy savings remains unresolved. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a low-power collaborative control method, system, medium, solid-state drive, and device for multi-port SSDs. It integrates artificial intelligence (AI) prediction, dynamic priority arbitration, and hardware collaborative control, which can intelligently predict load, collaboratively schedule port resources, and ensure smooth and reliable state switching, thereby achieving collaborative optimization of power consumption, performance, and reliability.

[0005] To achieve the above objectives, a low-power collaborative control method for multi-port SSDs is designed, including the following steps:

[0006] S1 collects real-time time-series characteristic data of each port of the solid-state drive and inputs it into the AI ​​prediction engine to obtain the load intensity trend and business type probability distribution of each port within a future set time window.

[0007] S2: Based on the load prediction results of step S1, calculate the energy saving priority and response priority for each port, and generate a collaborative power consumption control strategy that allows the port to be in different power consumption states.

[0008] S3: Before executing the cooperative power consumption control strategy and switching the state of the target port, the state synchronization protocol is triggered through hardware control logic to exchange state synchronization frames containing transaction context information between ports and confirm them.

[0009] S4: After the state synchronization is confirmed, the target port is controlled by the hardware control logic to switch the power consumption state.

[0010] S5: When the port wakes up from a low-power state, the port state is restored and consistency verification is performed based on the saved context information through hardware control logic.

[0011] In step S1, the timing characteristic data of each port includes:

[0012] I / O behavior characteristics: historical I / O request sequence, command type distribution, average command size;

[0013] Queue status characteristics: depth of each submission queue and completion queue, slope of fill rate change, and frequency of contention between ports;

[0014] Physical environment characteristics: SSD temperature change rate, power supply voltage fluctuation.

[0015] In step S1, the specific method for pre-training the AI ​​prediction engine is as follows:

[0016] S11, offline general pre-training: collects the historical running days of several multi-port SSDs in multiple scenarios; utilizes a high-performance GPU cluster in the cloud, labeled with future actual load, to supervise the learning of LSTM or Transformer neural network models and minimize the prediction error MSE; adopts knowledge distillation technology to compress large models into lightweight student models with less than 50KB of parameters, and generates a general base weight file to be pre-loaded into the SSD firmware.

[0017] S12, Online Fine-tuning: When the SSD is running, it uses idle computing power to incrementally update the base model based on locally collected feature data.

[0018] The specific method for step S2 is as follows:

[0019] S21, based on the prediction result of S1, by constructing a multi-factor weighted scoring function, quantitatively calculate the response priority score X and the energy-saving priority score Y of each port, and finally obtain the comprehensive decision vector Z;

[0020] S22, generate an asymmetric power consumption control strategy according to the scores;

[0021] In S21, for the response priority score X of the i-th port, X(i)=α×Pload(i)+β×Qdepth(i)+γ×Wresp(i), for the energy-saving priority score Y of the i-th port, Y(i)=(1-Pload(i)) ×δ×Ttrend(i) ×Wsave(i), Z(i)=X(i) -k×Y(i), wherein:

[0022] Pload(i) represents the AI-predicted load intensity score of the i-th port, with a value ranging from 0.0 to 1.0;

[0023] Qdepth(i) represents the normalized current queue depth of the i-th port, Qdepth(i)=current depth / maximum depth;

[0024] Ttrend(i) represents the temperature change trend coefficient of the i-th port, and its value is determined based on the temperature change rate dT / dt;

[0025] Wresp(i) represents the service response weight set by the host for the i-th port;

[0026] Wsave(i) represents the current energy-saving strategy weight set by the system for the i-th port;

[0027] α, β, γ, δ are dynamic adjustment coefficients, wherein α, β, γ ∈ [0,1] and α+β+γ=1, δ ∈ [0.5, 2.0];

[0028] k is a balance coefficient, which is set by the system power consumption mode;

[0029] In S22, according to the comparison result between the comprehensive decision vector Z(i) and the preset dynamic threshold [Th_high, Th_low], the asymmetric power consumption control strategy is generated:

[0030] If Z(i)>Th_high: forcibly maintain the Active state of L0;

[0031] If Th_low<Z(i)≤Th_high: enter the shallow sleep state of L0s;

[0032] If Z(i)≤Th_low: enter the deep sleep state of L1.1 / L1.2;

[0033] Where Th_high∈[0.6,0.8], Th_low∈[0.2,0.4].

[0034] In step S3, the port whose state is to be switched constructs a state synchronization frame and sends it to the peer port through the controller's internal high-speed bus. After the peer port confirms receipt, it temporarily locks the relevant shared resources to ensure that no potentially conflicting access is initiated during the switch. The synchronization frame includes: a list of currently valid transaction queue IDs for the port to be switched, a flag of incomplete atomic operations, and a range of critical metadata addresses in the cache that have not yet been refreshed.

[0035] Step S4 further includes: simultaneously activating the jitter suppression controller, pre-reading key data that may be affected by the switching delay into the cache within a microsecond-level time window before and after the switching command is issued, and intelligently rearranging the queued transactions to mask the physical delay.

[0036] To achieve the above objectives, a system for a low-power collaborative control method for multi-port SSDs is designed, comprising:

[0037] Feature acquisition module: used to collect time-series feature data from each port in real time;

[0038] AI Prediction Engine Module: Built-in lightweight time series model pre-trained offline and fine-tuned online, outputs load prediction results for future time windows;

[0039] Dynamic arbitration and strategy generation module: Based on the multi-factor weighted scoring formula, calculate the response and energy-saving priority of each port, and generate an asymmetric power consumption control strategy;

[0040] State synchronization protocol module: used to manage the exchange and acknowledgment of state synchronization frames between ports and resource locking;

[0041] Jitter Suppression Control Module: Used to manage transaction scheduling during state transitions to suppress jitter;

[0042] Power consumption control execution module: Used to execute instructions for switching power consumption states.

[0043] To achieve the above objectives, a computer-readable storage medium is designed, on which a computer program is stored, which, when executed by a processor, implements the above method.

[0044] To achieve the above objectives, a solid-state drive (SSD) is designed, including the aforementioned system.

[0045] To achieve the above objectives, a computing device is designed, including the aforementioned solid-state drive.

[0046] Compared with the prior art, the present invention has the following advantages:

[0047] 1. Superior energy efficiency and intelligent operation: AI-based prediction and advance planning avoid blind switching based on instantaneous states, reducing unnecessary switching overhead and achieving lower overall power consumption under the same business load.

[0048] 2. Significantly improved compatibility and reliability: The state synchronization protocol fundamentally solves the data consistency problem that may be caused by asynchronous power switching of multiple ports, making it particularly suitable for scenarios with extremely high data integrity requirements, such as financial transactions and autonomous driving.

[0049] 3. Seamless Business Response: Jitter suppression technology minimizes the impact of state transitions on upper-layer applications, ensuring smooth business responses and improving user experience and system stability.

[0050] 4. Flexible and universal architecture: It can be widely used in dual-port or multi-port SSDs with PCIe / NVMe interfaces, and can be adapted to various high-requirement application environments such as AI servers, edge computing, and automotive storage. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the process of the present invention.

[0052] Figure 2 This is a schematic diagram of the architecture of the present invention. Detailed Implementation

[0053] The present invention will now be further described with reference to the accompanying drawings.

[0054] like Figure 1 As shown, the AI-driven multi-port SSD low-power collaborative control method in the first aspect embodiment of the present invention includes the following steps:

[0055] S1 collects the time-series characteristic data of each port of the solid-state drive in real time through hardware control logic and inputs it into the AI ​​prediction engine to obtain the load intensity trend and business type probability distribution of each port within a future set time window.

[0056] S2: Based on the load prediction results of step S1, the energy-saving priority and response priority of each port are calculated and determined by AI, and a collaborative power consumption control strategy that allows the port to be in different power consumption states is generated.

[0057] S3: Before executing the cooperative power consumption control strategy and switching the state of the target port, the state synchronization protocol is triggered through hardware control logic to exchange state synchronization frames containing transaction context information between ports and confirm them.

[0058] S4: After the state synchronization is confirmed, the target port is controlled by the hardware control logic to switch the power consumption state.

[0059] S5: When the port wakes up from a low-power state, the port state is restored and consistency verification is performed based on the context information saved in the state synchronization frame through hardware control logic.

[0060] In step S1, the timing characteristic data of each port includes:

[0061] I / O behavior characteristics: historical I / O request sequence, i.e., read / write / Trim ratio, command type distribution, and average command size;

[0062] Queue status characteristics: depth of each submission queue and completion queue, slope of fill rate change, and frequency of contention between ports;

[0063] Physical environment characteristics: SSD temperature change rate, power supply voltage fluctuation, temperature change rate is ΔT / Δt, ΔT is the temperature difference within Δt, and Δt is the set time period.

[0064] In step S1, the future time window is set to 50-200 milliseconds.

[0065] In step S1, the specific method for pre-training the AI ​​prediction engine is as follows:

[0066] S11, offline general pre-training, builds the base model: collects historical operation logs of several multi-port SSDs in multiple scenarios and performs anonymization processing; utilizes a high-performance GPU cluster in the cloud, labeled with future actual load, to perform supervised learning on LSTM or Transformer neural network models and minimize prediction error MSE; adopts knowledge distillation technology to compress large models into lightweight student models with less than 50KB of parameters, and generates a general base weight file to be pre-loaded into the SSD firmware;

[0067] S12, online fine-tuning, adaptive to scenarios: When the SSD is running, it uses idle computing power, such as during GC intervals, to incrementally update the base model based on locally collected feature data, so that it can adapt to the specific business mode of the current host, such as night backup mode or burst video stream mode.

[0068] Step S12 also includes: In a cluster deployment scenario, each SSD only uploads the model gradient update amount to the central server for aggregation to generate a new generation of global model, achieving collective intelligent evolution without leaking user data privacy.

[0069] The AI ​​prediction engine in this embodiment is obtained through a two-stage strategy of offline general pre-training followed by online fine-tuning, ensuring that the model possesses both generalization ability and adaptability to specific scenarios. The AI ​​prediction engine employs an LSTM or Transformer neural network model, which is updated and optimized through federated learning.

[0070] The specific method for step S2 is as follows:

[0071] S21. Based on the prediction results of S1, the dynamic priority arbitrator constructs a multi-factor weighted scoring function to quantify the response priority score X and energy-saving priority score Y of each port, and finally obtains the comprehensive decision vector Z.

[0072] S22, the dynamic priority arbitrator generates an asymmetric power consumption control strategy based on the score and outputs it to the power consumption control execution module;

[0073] In S21, the response priority score X of the i-th port is X(i) = α × Pload(i) + β × Qdepth(i) + γ × Wresp(i), and the energy saving priority score Y of the i-th port is Y(i) = (1 - Pload(i)) × δ × Ttrend(i) × Wsave(i), Z(i) = X(i) - k × Y(i), where:

[0074] Pload(i) represents the load intensity score predicted by AI for the i-th port, with a value of 0.0-1.0. In step S1, the load intensity trend value predicted by AI is itself a scalar of [0.0,1.0] output by the AI ​​model, which is directly used as the value of Pload(i).

[0075] Qdepth(i) represents the normalized current queue depth of the i-th port, where Qdepth(i) = current depth / maximum depth. The current depth is the sum of the number of pending requests in all submission queues (SQ) of the port, read in real-time from the SSD controller's internal register. The maximum depth is determined by the NVMe protocol and SSD design and is a fixed constant. That is, Qdepth(i) = (sum of the current number of pending requests in all submission queues of port i) / (the sum of the maximum queue depths supported by this port).

[0076] Ttrend(i) represents the temperature change trend coefficient of the i-th port, based on the temperature change rate dT / dt; the unit is °C / s. When dT / dt<5, Ttrend(i)=1.0; when 5≤dT / dt<10, Ttrend(i)=1.2; when dT / dt≥10, Ttrend(i)=1.5; when dT / dt<-2, ​​the temperature drops, Ttrend(i)=0.8.

[0077] Wresp(i) represents the service response weight set by the host for the i-th port; the specific weight is set according to the QoS policy, with a default of 1.0, 2.0 for critical system ports, 1.5 for real-time service ports, and 0.5 for background maintenance ports.

[0078] Wsave(i) represents the weight of the current energy-saving policy set by the system for the i-th port; in specific application, it is set according to the battery mode, with a default value of 1.0, 2.5 when the battery level is <20%, 1.5 when the battery level is sufficient, and 0.5 when the battery is in high-performance mode.

[0079] α, β, γ, δ are dynamic adjustment coefficients, wherein α, β, γ ∈ [0,1] and α+β+γ=1, δ ∈ [0.5,2.0]; in this embodiment, α=0.5, β=0.3, γ=0.2, δ=1.0.

[0080] k is a balance coefficient, set by the system power consumption mode; wherein, k is 0.0 in the extreme performance mode, 0.5~0.8 in the balance mode, 1.0~1.5 in the energy-saving mode powered by AC, and 1.5~2.0 in the extreme energy-saving mode powered by battery.

[0081] In S22, an asymmetric power consumption control strategy is generated according to the comparison result between the comprehensive decision vector Z(i) and the preset dynamic threshold [Th_high, Th_low]:

[0082] If Z(i)>Th_high: the Active state of L0 is forcibly maintained;

[0083] If Th_low<Z(i)≤Th_high: enter the shallow sleep state of L0s;

[0084] If Z(i)≤Th_low: enter the deep sleep state of L1.1 / L1.2;

[0085] Wherein Th_high∈[0.6,0.8], Th_low∈[0.2,0.4]. In this embodiment, Th_high and Th_low is 0.3.

[0086] In step S22, according to the load prediction result, ports with light load are assigned higher energy-saving priority and scheduled to enter a deeper low-power state, while ports with heavy load or load that is about to increase are assigned higher response priority and scheduled to remain active or enter a shallow sleep state. Further, a high predicted load, queue accumulation or high service importance leads to a high response score; the higher the score, the more necessary it is to remain active. A port with a high response priority weight tends to remain in an active state or a shallow sleep state, such as L0 state or L0s state. A higher energy-saving priority score means the port can enter sleep more easily, and a port with a high energy-saving priority weight tends to be scheduled to enter a deeper low-power state, such as L1 state.

[0087] In step S3, the port whose state is about to be switched constructs a state synchronization frame and sends it to the peer port or shared management unit through the controller's internal high-speed bus. After the peer port or shared management unit confirms receipt, it temporarily locks the relevant shared resources, such as the DRAM cache, to ensure that no potentially conflicting access is initiated during the switch.

[0088] The synchronization frame includes: a list of currently valid transaction queue IDs for the port to be switched, a flag for incomplete atomic operations, and a range of critical metadata addresses in the cache that have not yet been refreshed.

[0089] Step S4 also includes: simultaneously activating the jitter suppression controller, and within a microsecond-level time window before and after the switching command is issued, such as 15-20 microseconds, pre-reading critical data that may be affected by the switching delay into a high-speed cache, and intelligently rearranging queued transactions to mask physical latency, so that the additional latency perceived by the host is less than 1 millisecond. That is, adjusting transaction scheduling to suppress performance jitter perceived by the host side.

[0090] In step S5, when it is necessary to wake up the port from a low-power state, the physical link is restored first. Therefore, based on the context information saved in the synchronization frame in S3, the transaction state and cache view of the port are quickly reconstructed, and consistency verification is performed to ensure that subsequent services can be seamlessly taken over after waking up without data errors or omissions.

[0091] like Figure 2 As shown, the system for implementing the above-mentioned AI-driven multi-port SSD low-power collaborative control method, as proposed in the second aspect embodiment of the present invention, includes:

[0092] Feature acquisition module: used to collect time-series feature data from each port in real time;

[0093] AI Prediction Engine Module: Built-in lightweight time series model pre-trained offline and fine-tuned online, outputs load prediction results for future time windows;

[0094] Dynamic arbitration and strategy generation module: Based on the multi-factor weighted scoring formula, calculate the response and energy-saving priority of each port, and generate an asymmetric power consumption control strategy;

[0095] State synchronization protocol module: used to manage the exchange and acknowledgment of state synchronization frames between ports and resource locking;

[0096] Jitter Suppression Control Module: Used to manage transaction scheduling during state transitions to suppress jitter;

[0097] Power consumption control execution module: Used to execute instructions for switching power consumption states.

[0098] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0099] A fourth aspect of the present invention provides a solid-state drive, including the system described above.

[0100] A fifth aspect of the present invention provides a computing device including the solid-state drive described above.

[0101] In one embodiment of the invention, the system is integrated within the controller of a multi-port SSD. The feature acquisition module continuously monitors ports 0 and 1, and the acquired data is packaged every 100 milliseconds and sent to the AI ​​prediction engine module.

[0102] The I / O sequence for port 0 is [read, read, write, read, idle], with the cache queue depth fluctuating between 70% and 85%, and the port contention count is 3 times; the I / O sequence for port 1 is mainly [idle, idle, read], with the cache queue depth remaining stable at 20%.

[0103] The AI ​​prediction engine uses an LSTM model to output the estimated IOPS and read / write ratios for two ports within the next 150 milliseconds. Analysis revealed that port 0 experiences short-cycle, high-volume bursts of read and write requests, with a consistently high cache level. Based on historical patterns, it is estimated that port 0 has an 80% probability of maintaining a high load or continuing to increase within the next 150ms. Meanwhile, the model identifies long idle intervals in port 1's traffic flow, predicting a greater than 90% probability that port 1 will maintain a low load within the next 150ms.

[0104] After receiving the prediction result from S1, the dynamic priority arbitrator performs the calculation.

[0105] The system is in equilibrium mode, k=1, α=0.5, β=0.3, γ=0.2, δ=1.0.

[0106] For port 0, Pload=0.85, Qdepth=0.7.

[0107] Then X(0) = 0.5 × 0.85 + 0.3 × 0.7 + 0.2 × 1.0 = 0.835; Y(0) = (1 - 0.85) × 1.0 = 0.15;

[0108] Z(0) = 0.835 - 0.15 = 0.685, which is greater than the high threshold, so it is determined to be L0.

[0109] For port 1, Pload=0.15, Qdepth=0.1.

[0110] Then X(1) = 0.5 × 0.15 + 0.3 × 0.1 + 0.2 × 0.5 = 0.205; Y(1) = (1 - 0.15) × 1.0 = 0.85;

[0111] Z(1) = 0.205 - 0.85 = -0.645, which is less than the low threshold, so it is determined to be L1.

[0112] Before putting port 1 into the L1 state, the state synchronization protocol module is invoked. It constructs a synchronization frame containing information about all currently active I / O queues for port 1 and sends it to the logical unit representing port 0 via the controller's internal bus. After port 0 confirms, it temporarily suspends write operations to certain areas of the shared buffer.

[0113] Subsequently, the power control execution module sends an L1 inbound command to the PCIe physical layer. Simultaneously, the jitter suppression control module intervenes, analyzing the transaction queue and discovering an important metadata query command ID:102 from the host, positioned late in the queue. If processed in its original order, this command might be significantly affected by the physical latency of entering L1 from port 1. Therefore, before issuing the L1 inbound command, a 15-microsecond buffer window is inserted. Within this window, command 102 is rearranged and pre-fetched with a preceding mergeable read command ID:105, preloading the data into a faster cache to mask the physical latency.

[0114] When the host needs to access port 1, port 1 receives a wake-up command and the context saved in the state synchronization frame in S3, such as the transaction queue ID list and atomic operation flags. The initial state of the port after wake-up is quickly compared with the expected state saved in the synchronization frame. The context saved by the state synchronization protocol is used to verify and restore cache consistency, and port 1 subsequently resumes service without being aware of it.

[0115] This invention achieves intelligent collaborative management of multi-port power consumption by constructing a closed-loop control chain of "perception-prediction-arbitration-execution-verification". Specifically, this invention uses an AI model to accurately predict the future load of multiple ports, realizing a paradigm shift in power management from passive response to proactive planning. A dynamic priority arbitration mechanism is designed to implement differentiated and asymmetric low-power strategies for each port based on the prediction results, achieving optimal overall power consumption. A lightweight state synchronization protocol is constructed to ensure that the cross-port transaction context remains consistent before and after any low-power state switch on any port, preventing data loss and out-of-order delivery. Jitter suppression control is introduced to minimize the performance impact of state switching on host applications and ensure smooth service response.

Claims

1. A low-power collaborative control method for multi-port SSDs, characterized in that: Includes the following steps: S1 collects real-time time-series characteristic data of each port of the solid-state drive and inputs it into the AI ​​prediction engine to obtain the load intensity trend and business type probability distribution of each port within a future set time window. S2: Based on the load prediction results of step S1, calculate the energy saving priority and response priority for each port, and generate a collaborative power consumption control strategy that allows the port to be in different power consumption states. S3: Before executing the cooperative power consumption control strategy and switching the state of the target port, the state synchronization protocol is triggered through hardware control logic to exchange state synchronization frames containing transaction context information between ports and confirm them. S4: After the state synchronization is confirmed, the target port is controlled by the hardware control logic to switch the power consumption state. S5: When the port wakes up from a low-power state, the port state is restored and consistency verification is performed based on the saved context information through hardware control logic; The specific method for step S2 is as follows: S21, based on the prediction results of S1, by constructing a multi-factor weighted scoring function, the response priority score X and energy saving priority score Y of each port are quantitatively calculated, and finally the comprehensive decision vector Z is obtained; S22, generate an asymmetric power consumption control strategy based on the score; In S21, the response priority score X of the i-th port is X(i) = α × Pload(i) + β × Qdepth(i) + γ × Wresp(i), and the energy saving priority score Y of the i-th port is Y(i) = (1 - Pload(i)) × δ × Ttrend(i) × Wsave(i), Z(i) = X(i) - k × Y(i), where: Pload(i) represents the load intensity score predicted by AI for the i-th port, with a value ranging from 0.0 to 1.0; Qdepth(i) represents the i-th port, normalized to the current queue depth, Qdepth(i) = current depth / maximum depth; Ttrend(i) represents the temperature change trend coefficient of the i-th port, which is based on the temperature change rate dT / dt. Wresp(i) represents the service response weight set by the host for the i-th port; Wsave(i) represents the current energy-saving strategy weight set by the system for the i-th port; α, β, γ, δ are dynamic adjustment coefficients, where α, β, γ ∈ [0, 1] and α + β + γ = 1, δ ∈ [0.5, 2.0]. k is the balance coefficient, which is set by the system power consumption mode.

2. The multi-port SSD low-power collaborative control method according to claim 1, characterized in that: In step S1, the timing characteristic data of each port includes: I / O behavior characteristics: historical I / O request sequence, command type distribution, average command size; Queue status characteristics: depth of each submission queue and completion queue, slope of fill rate change, and frequency of contention between ports; Physical environment characteristics: SSD temperature change rate, power supply voltage fluctuation.

3. The multi-port SSD low-power collaborative control method according to claim 1, characterized in that: In step S1, the specific method for pre-training the AI ​​prediction engine is as follows: S11, offline general pre-training: collect historical operation data of a plurality of multi-port SSDs in multiple scenarios; use a high-performance GPU cluster in the cloud, take the actual future load as a label, perform supervised learning on the LSTM or Transformer neural network model, and minimize the prediction error MSE; adopt knowledge distillation technology to compress the large model into a lightweight student model with a parameter volume <50KB, generate a general base weight file and pre-place it in the SSD firmware; S12, online fine-tuning: when the SSD is running, use idle computing power to perform incremental update on the base model based on feature data collected locally in real time.

4. The low-power collaborative control method for multi-port SSD according to claim 1, characterized in that: In said step S22, an asymmetric power control strategy is generated according to the comparison result between the comprehensive decision vector Z(i) and preset dynamic thresholds [Th_high,Th_low]: If Z(i)>Th_high: force the port to keep the Active state of L0; If Th_low<Z(i)≤Th_high: enter the shallow sleep state of L0s; If Z(i)≤Th_low: enter the deep sleep state of L1.1 / L1.2; wherein Th_high∈[0.6,0.8], Th_low∈[0.2,0.4].

5. The multi-port SSD low-power collaborative control method according to claim 1, characterized in that: In said step S3, the port to be switched constructs a state synchronization frame and sends it to the opposite-end port via the high-speed internal bus of the controller; after the opposite-end port confirms reception, it temporarily locks related shared resources to ensure that no access that may cause conflicts is initiated during the switching process; Said synchronization frame includes: the currently valid transaction queue ID list of the port to be switched, unfinished atomic operation flags, and the address range of key metadata that has not been refreshed in the cache.

6. The multi-port SSD low-power collaborative control method according to claim 1, characterized in that: Said step S4 further comprises: simultaneously starting a jitter suppression controller, pre-reading key data that may be affected by switching delay into the high-speed cache within a microsecond time window before and after the switching instruction is issued, and intelligently reordering transactions in the queue to cover physical delay.

7. A system for a multi-port SSD low-power collaborative control method according to any one of claims 1-6, characterized in that: comprising: Feature acquisition module: configured to collect timing feature data of each port in real time; AI prediction engine module: with a built-in lightweight timing model that has undergone offline pre-training and online fine-tuning, and outputs a load prediction result for a future time window; Dynamic arbitration and strategy generation module: configured to calculate the response and energy saving priority of each port based on a multi-factor weighted scoring formula, and generate an asymmetric power control strategy; State synchronization protocol module: configured to manage the exchange, confirmation and resource locking of state synchronization frames between ports; Jitter suppression control module: configured to manage transaction scheduling during state switching to suppress jitter; Power control execution module: configured to execute power state switching instructions.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: said program, when executed by a processor, implements the method according to any one of claims 1-6.

9. A solid-state drive, characterized in that: comprising the system according to claim 7.

10. A computing device, characterized in that: comprising the solid state drive according to claim 9.

Images