WO2020259543A1 - 一种基于联合优化回声状态网络的热数据预测方法 - Google Patents

一种基于联合优化回声状态网络的热数据预测方法 Download PDF

Info

Publication number
WO2020259543A1
WO2020259543A1 PCT/CN2020/097950 CN2020097950W WO2020259543A1 WO 2020259543 A1 WO2020259543 A1 WO 2020259543A1 CN 2020097950 W CN2020097950 W CN 2020097950W WO 2020259543 A1 WO2020259543 A1 WO 2020259543A1
Authority
WO
WIPO (PCT)
Prior art keywords
particle
storage layer
echo state
state network
fitness value
Prior art date
Application number
PCT/CN2020/097950
Other languages
English (en)
French (fr)
Inventor
罗旗舞
王玥童
阳春华
桂卫华
周灿
Original Assignee
中南大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 中南大学 filed Critical 中南大学
Publication of WO2020259543A1 publication Critical patent/WO2020259543A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F3/00—Input arrangements for transferring data to be processed into a form capable of being handled by the computer; Output arrangements for transferring data from processing unit to output unit, e.g. interface arrangements
    • G06F3/06—Digital input from, or digital output to, record carriers, e.g. RAID, emulated record carriers or networked record carriers
    • G06F3/0601—Interfaces specially adapted for storage systems
    • G06F3/0602—Interfaces specially adapted for storage systems specifically adapted to achieve a particular effect
    • G06F3/0614—Improving the reliability of storage systems
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F3/00—Input arrangements for transferring data to be processed into a form capable of being handled by the computer; Output arrangements for transferring data from processing unit to output unit, e.g. interface arrangements
    • G06F3/06—Digital input from, or digital output to, record carriers, e.g. RAID, emulated record carriers or networked record carriers
    • G06F3/0601—Interfaces specially adapted for storage systems
    • G06F3/0628—Interfaces specially adapted for storage systems making use of a particular technique
    • G06F3/0638—Organizing or formatting or addressing of data
    • G06F3/064—Management of blocks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F3/00—Input arrangements for transferring data to be processed into a form capable of being handled by the computer; Output arrangements for transferring data from processing unit to output unit, e.g. interface arrangements
    • G06F3/06—Digital input from, or digital output to, record carriers, e.g. RAID, emulated record carriers or networked record carriers
    • G06F3/0601—Interfaces specially adapted for storage systems
    • G06F3/0668—Interfaces specially adapted for storage systems adopting a particular infrastructure
    • G06F3/0671—In-line storage system
    • G06F3/0673—Single storage device
    • G06F3/0679—Non-volatile semiconductor memory device, e.g. flash memory, one time programmable memory [OTP]

Definitions

  • the invention belongs to the technical field of chaotic time series prediction, and specifically relates to a thermal data prediction method based on a joint optimization echo state network.
  • NAND flash memory As a non-volatile storage technology, NAND flash memory is widely used in communication systems and consumer electronic products. Compared with hard disk drives, NAND flash memory has higher access speed and power efficiency. In consumer electronic devices based on NAND flash memory, a large number of applications rely on NAND flash memory for data exchange, file storage and video storage. NAND flash memory is mainly used to store large-capacity data. The NAND structure can provide extremely high cell density, which can achieve high storage density, high writing and high erasing speed; therefore, NAND flash memory is mostly used for large-capacity data storage, such as solid state drives. The demand for NAND flash memory will continue to grow in the future, mainly in cloud computing, Internet of Things and data centers.
  • NAND flash memory faces at least two challenges, namely, remote update and limited durability, which limits its large-scale application.
  • NAND flash memory has the defect that the operation cannot be overwritten, that is, a new write operation cannot be performed on a page before the flash memory is erased. Therefore, improper updates will generate many invalid pages and dead pages, which will reduce efficiency and performance.
  • NAND flash memory has a limited lifespan because the flash memory block can only withstand a limited number of erasing. If the block's erasing number is greater than the block's maximum erasable number, it will be unusable.
  • Garbage Collection (GC) and Wear Leveling (Wear Leveling, WL) allocate frequently written data (i.e.
  • HDI hot data identification
  • the essence of HDI is to try to understand the access behavior of hot data well so as to intelligently allocate different data to the appropriate blocks.
  • traditional HDI has the following two problems. One is that the memory overhead is large. At present, most of the hot data identification mechanisms are based on the idea of identifying hot data pages in NAND flash memory.
  • the core principle of these mechanisms is the page counter, which records the number of read and write operations of the logical page corresponding to the NAND flash memory page within a certain period of time. If the number of reads and writes is greater than the set threshold, the requested page is judged as a hot page, otherwise it is a cold page. Another serious problem is that the recognition accuracy is not high.
  • the hot data recognition mechanism based on the Bloom filter is widely used in the recognition of the hot and cold data of the SSD, but the inherent defect of the Bloom filter is the false positive, which means that the data does not belong to the set. Wrongly judged to be in the set.
  • the hot data identification method based on load request size and load access mode has a single consideration factor, and the local characteristics of the load are not fully considered, and the accuracy of hot data identification is not high.
  • the purpose of the present invention is to provide a thermal data prediction method based on a joint optimization echo state network, which creatively proposes to replace the traditional thermal data recognition with thermal data prediction, and constructs a joint optimization echo state network, so that the predicted thermal data More real-time and reliable.
  • the present invention provides a hot data prediction method based on a joint optimization echo state network, which includes the following steps:
  • the position information of the particle includes the initial position and the position range of the particle, and the position of each particle is represented by a storage layer parameter in the echo state network;
  • the quantum particle swarm algorithm is used to update the particle position based on the position range of each particle.
  • the L2+ adaptive L1/2 regularization constrained echo state network is used to calculate the output weight and calculate the global optimal fitness value.
  • the particle position corresponding to the global best fitness value at the end of the iteration is used as the best storage layer parameter;
  • y represents the obtained predicted logical block address
  • the data on the predicted logical block address is the thermal data
  • x is the logical block address where the input historical thermal data is located
  • W out represents the output weight.
  • the historical thermal data The address of the logical block is used in the echo state network training process in step S2 and step S3.
  • the present invention innovatively replaces the thermal data recognition module in the flash memory conversion layer with thermal data prediction while continuing to use the solid state hard disk structure framework, and uses the joint optimization echo state network to predict thermal data.
  • the joint optimization includes two parts. The first part uses quantum The particle swarm algorithm iterative optimization determines the optimal storage layer parameters of the echo state network. The second part uses the L2+adaptive L1/2 regularization constrained ESN to obtain high sparsity output weights.
  • the present invention iteratively searches the quantum particle swarm algorithm Optimal and L2+adaptive L1/2 regularization constraints are combined to obtain the best storage layer parameters, and the joint optimized echo state network used for prediction is more real-time and reliable.
  • the present invention uses the logical block address where the historical hot data is located to train the echo state network to obtain the final output weight, and then uses it to predict the logical block address where the hot data is located.
  • step S2 the execution process of iterative optimization and determination of optimal storage layer parameters in step S2 is as follows:
  • the current position of each particle is sequentially used as the storage layer parameter in the echo state network and the output weight is calculated;
  • S23 Based on the principle of minimum fitness value, select the individual best fitness value, individual best parameter, and global best fitness value and global best parameter of each particle according to the fitness value of each particle;
  • the particle position selected as the global best fitness value is the global best parameter
  • step S25 Determine whether the number of iterations reaches the maximum number of iterations, if not, return to step S24 for the next iteration calculation; otherwise, use the current global optimal parameter as the optimal storage layer parameter.
  • any particle j is updated according to the following formula:
  • P j (t+1) and P j (t) respectively represent the position of particle j after and before update
  • u j are random numbers
  • sbest j and sbest i represent the best individual parameters of the jth and ith particles
  • mbest is the average value of the current individual best parameters of all particles
  • iter and iter max are the current iteration times respectively
  • the maximum number of iterations ⁇ max and ⁇ min are the inertia factors respectively
  • N is the total number of particles.
  • the calculation formula for the fitness value of any particle j is:
  • Fitness represents the fitness value of the current particle j, ⁇ 1 and ⁇ 2 are both regularization coefficients, and W out is the output weight corresponding to the current particle j;
  • Y represents the logical block address of the historical thermal data used for network training
  • X represents the state information of the storage layer updated based on the previous segment of the logical block address of the historical hot data based on network training, and
  • X*W out represents the prediction result corresponding to the latter segment of the logical block address of the historical thermal data.
  • the process of calculating the output weights using the L2+adaptive L1/2 regularization constraint echo state network is as follows:
  • the input layer-storage layer weight matrix and the storage layer internal connection weight matrix are related to the storage layer parameters in the echo state network;
  • the state information X is composed of state node information X(t);
  • U(t) represents the t-th data in the input variable U
  • X(t) and X(t-1) respectively represent the t-th and t-1th state node information
  • T of t is determined by the input
  • the data length of the variable U is determined
  • W in and W x respectively represent the input layer-storage layer weight matrix in the echo state network, the storage layer internal connection weight matrix, logsig( ⁇ ) represents the activation function
  • E represents the loss function
  • ⁇ 1 and ⁇ 2 are both regularization coefficients.
  • step U403 is: simplify the loss function, and then calculate the output weight by using the coordinate descent algorithm;
  • I is the identity matrix
  • the method of solving matrix W′ out is to calculate each element separately, the value of the k-th element in the m-th row of W′ out is as follows:
  • Y′ k (t) represents the t-th element in the k-th row of Y′
  • X′ j (t) represents the t-th element in the j-th row of X′
  • L is the number of output layer nodes
  • n is the number of storage layer nodes.
  • step U403 it also includes adaptive optimization of the output weight obtained in step U403, and the optimization process is as follows:
  • the converted loss function is:
  • K is the number of nodes in the input layer.
  • the storage layer parameters in the echo state network include four key parameters: internal connection spectrum radius, storage layer scale, input layer scale factor and storage layer sparsity.
  • the parameters required to initialize the quantum particle swarm algorithm in step S1 include the particle swarm size N, the maximum number of iterations iter max , and the inertia factors ⁇ max and ⁇ min .
  • the particle position parameter is set to a boundary value corresponding to the exceeding position range.
  • the present invention innovatively proposes to replace traditional thermal data recognition with thermal data prediction.
  • the public thermal data prediction technology can predict the nature of the next data one or even several beats in advance based on historical access behavior, and actively allocate storage to
  • the solid state drive (SSD) corresponding block (hot/cold data block) is more active than the traditional hot data recognition, and the implementation process of the present invention uses joint optimization to improve the accuracy of network prediction, thereby obtaining better Accurate thermal data recognition effect and efficient thermal data prediction will better serve garbage collection and wear leveling technologies, and ultimately improve the life of solid state drives.
  • the neural network method retains more characteristic information for the input, and more comprehensively classifies the thermal data.
  • the present invention performs joint optimization on the echo state network.
  • the L2 regularization constraint obtains a good generalization ability through the trade-off between the model deviation and the prediction variance so as to obtain the weight of continuous shrinkage, but it cannot produce sparse solution; L1/2 regularization can generate very sparse solutions, but when there is a high degree of correlation between the predictor variables, L1/2 cannot play a good regulatory role.
  • the present invention uses L2+adaptive L1/2 regularization to train the least square Multiplication can obtain the advantages of two kinds of regularization, thereby improving the prediction accuracy of thermal data.
  • optimizing the parameters of the echo state network storage layer based on the QPSO algorithm can solve the problem that the storage layer parameters cannot be determined when building the model.
  • this algorithm removes the velocity information of the particles based on the wave-particle duality and only retains the position information, which can effectively reduce the complexity of the calculation, and at the same time obtain the storage layer parameters of the adaptive model, thereby further improving Prediction accuracy; further, the present invention combines L2+adaptive L1/2 regularization and QPSO algorithm to obtain the best storage layer parameters and improve prediction accuracy.
  • Figure 1 is a typical architecture of a NAND flash memory system
  • FIG. 2 is a flowchart of a method for predicting hot data based on a joint optimized echo state network according to an embodiment of the present invention
  • Fig. 3 is a specific algorithm flow chart of the iterative optimization of the quantum particle swarm algorithm of the present invention; wherein, after the execution of step U304 in Fig. 3A is completed, it turns to step U305 in Fig. 3B.
  • Fig. 4 is a specific algorithm flow chart of the present invention using L2+adaptive L1/2 constrained echo state network to calculate output weights.
  • Fig. 5 is a performance comparison diagram of four actual workloads according to an embodiment of the present invention.
  • the present invention provides a hot data prediction method based on a joint optimization echo state network, which is mainly applied to a NAND flash memory system.
  • the typical architecture of a NAND flash memory system includes module B101 (user operation) and module B102. (File system) and module B103 (solid state drive). The actual operation of the user will affect the solid state drive through the file system.
  • the solid state drive also includes a flash memory conversion layer, a flash controller, and a NAND flash array.
  • the flash memory conversion layer includes an address allocation unit, a garbage collection unit, a wear leveling unit, and a thermal data prediction unit.
  • the present invention innovatively proposes to use the thermal data prediction unit
  • the traditional hot data identification method usually passively analyzes user access behavior, and allocates and stores the corresponding data to the corresponding block (hot/cold data area) of the solid state drive (SSD) through the Flash Transport Layer Protocol (FTL) Block), this method has high hot data missed detection or false alarms when responding to requests with complex access behaviors.
  • the hot data prediction technology disclosed in the present invention can predict the nature of the next data one or even a few beats in advance based on historical access behavior, and actively allocate and store it to the corresponding block (hot/cold data block) of the solid state drive (SSD).
  • the thermal data prediction method proposed by the present invention is essentially "predictive thermal data identification".
  • the predicted logical block address information finally obtained by the present invention is used for garbage collection and wear leveling processing.
  • the traditional thermal data recognition is to accurately and efficiently distinguish which data is valid data.
  • the present invention provides a thermal data prediction method based on a joint optimization echo state network, which replaces thermal data identification with thermal data prediction, and has high-precision prediction, which specifically includes the following steps:
  • the position information of the particle includes the initial position and the position range of the particle.
  • the position of each particle is represented by the storage layer parameters in the echo state network (ESN).
  • the storage layer parameters in the echo state network include the internal connection spectrum radius, the storage layer scale, and the input.
  • the layer scale factor and the storage layer sparsity, the dimension of each particle is initialized to 4 in this example, that is, each particle is a 1*4 matrix, which represents the 4 parameters of the ESN storage layer.
  • Determine the range of the 4 parameters set the parameter range as the position range of all particles, and randomly assign a value to each particle within the position range during initialization. In the subsequent update process, it can be regarded as the particle continuously moving towards the maximum within the specified range. If the particle moves beyond the specified range, the particle position information is updated to the boundary value.
  • Each particle position represents a specific value of ESN storage layer parameters.
  • the parameters required by the quantum particle swarm algorithm include the particle swarm size N, the maximum number of iterations Itermax, the inertia factors ⁇ max and ⁇ min (used to update the particle position information later).
  • the quantum particle swarm algorithm is used to update the particle position based on the position range of each particle, and the echo state network with L2+adaptive L1/2 regularization constraint is used to calculate the output weight during each update process to obtain the global optimal fitness value.
  • the particle position corresponding to the global optimal fitness value is used as the optimal storage layer parameter.
  • the current position of each particle is sequentially used as the storage layer parameter in the echo state network and the output weight is calculated;
  • S23 Based on the principle of minimum fitness value, select the individual best fitness value, individual best parameter, and global best fitness value and global best parameter of each particle according to the fitness value of each particle;
  • the particle position selected as the global best fitness value is the global best parameter
  • step S25 Determine whether the number of iterations reaches the maximum number of iterations, if not, return to step S24 for the next iteration calculation; otherwise, use the current global optimal parameter as the optimal storage layer parameter.
  • FIG. 3 provides an example flowchart as shown in FIG. 3, which includes the following steps:
  • U302 Set the location of the j-th particle as the ESN storage layer parameter, and use L2+adaptive L1/2 regularization to constrain the least square calculation that appears in the training to obtain a higher sparsity output weight Wout.
  • the detailed steps of calculating the output weight Wout with the ESN of the L2+adaptive L1/2 regularization constraint are shown in Figure 4, which will be described in detail below.
  • U303 Calculate the fitness value corresponding to the j-th particle based on the output weight Wout corresponding to the j-th particle.
  • the calculation formula is as follows:
  • ⁇ 1 and ⁇ 2 are regularization coefficients, and W out is the output weight corresponding to the current particle j; Y represents the last part of the logical block address of the historical hot data used for network training, and X represents network-based training The state information of the storage layer updated before the logical block address where the historical thermal data is located, X*W out represents the prediction result corresponding to the latter section of the logical block address where the historical thermal data is located.
  • D301 Determine whether all particles have completed the fitness value calculation, if not, add 1 to j, and return to step U302 to calculate the fitness value of the next particle. If all particles have completed the fitness value calculation, proceed to step U304.
  • the individual best fitness value, individual best parameter, and global best fitness value and global best parameter of each particle are selected according to the fitness value of each particle. After all particles have calculated their fitness values, compare and judge, record the fitness value of each particle as the individual best fitness value fsbest, and the position of each particle as the individual best parameter sbest; record the smallest particle fitness value among all particles The fitness value is the global best fitness value fgbest, and its corresponding position is the global best parameter gbest. These obtained parameters will be used in subsequent iterations to optimize.
  • sbest i represents the individual best parameter of the i-th particle
  • mbest is the average value of the current individual best parameters of all particles, that is, the average value of each dimension parameter of all particles is used to update the particle position information.
  • P j (t+1) and P j (t) respectively represent the position of the particle j after and before the update
  • u j are random numbers between (0,1), where ⁇ is calculated as:
  • the parameter ⁇ representing the step length of the particle movement is larger, and the particles can move to the optimal position faster; while the smaller ⁇ in the later stage of the iteration means that the parameter is in the optimal position.
  • the nearby particles decrease the step size, and move closer to the best position more accurately each time.
  • D302, D303, U308, U309 are: update the individual according to the newly calculated fitness value Best and global best. If the newly calculated fitness value is less than the individual best fitness value of the particle, the individual best fitness value of the particle is updated to the newly calculated fitness value, and the individual best fitness value of the particle is updated at the same time.
  • the best parameter is the parameter of the current particle; if the newly calculated fitness value is less than the global best fitness value, the fitness value is updated to the global best fitness value, and the global best parameter is the parameter of the particle at the same time.
  • D304 Judge whether all particles have been updated, if not, j+1 and return to U306, use the updated particle parameters to recalculate mbest, and update the position information of the next particle. If all particles have been updated, proceed to D305 .
  • D305 Judge whether the number of iterations has reached the maximum number of iterations, if not, add 1 to iter and return to U305 for the next iteration. If the maximum number of iterations has been reached, the final global optimal parameters are derived for subsequent training to jointly optimize the echo state network. Predict the logical block address.
  • y represents the obtained predicted logical block address
  • the data on the predicted logical block address is the thermal data
  • x is the logical block address where the input historical thermal data is located
  • W out represents the output weight.
  • y is the predicted access address. It is worth noting that both x and Wout can be multi-dimensional variables, and the obtained y is a one-dimensional variable.
  • the data on the obtained logical block address will be classified as hot data for Garbage collection and wear leveling treatment.
  • a set of storage layer parameters are determined, namely, the internal connection spectrum radius, storage layer scale, input layer scale factor and storage layer sparsity.
  • the process of calculating output weights using the echo state network with L2+adaptive L1/2 regularization constraints in the present invention is as follows:
  • U401 Obtain the input layer-storage layer weight matrix in the echo state network, the internal connection weight matrix of the storage layer, and use the previous segment of the logical block address where the historical thermal data is located as the input variable U, and the latter segment as the actual result Y.
  • the Echo State Network is a low-complexity and fast-converging calculation scheme, which is suitable for temporal data classification and prediction tasks.
  • the ESN network architecture includes three layers: input layer, storage layer and output layer.
  • the weight of the input layer-storage layer is Win
  • the internal connection weight of the storage layer is Wx
  • the storage layer-output layer is Wout.
  • the input layer-storage layer weight W in ⁇ R n ⁇ K is determined based on the storage layer parameters, and the storage layer internal connection weight W x ⁇ R n ⁇ n .
  • the logical block address of historical thermal data selected in the embodiment of the present invention is the logical block address of historical thermal data recorded by the user. Other feasible embodiments
  • the selected length can be other, and the present invention does not specifically limit it.
  • the general idea is to use the previous address to predict the latter address, and then compare the predicted latter address with the actual address to adjust the network. Part of it is the original feature of the echo state network, which is not described in detail in the present invention.
  • the state information X is composed of state node information X(t);
  • U(t) represents the t-th data in the input variable U
  • X(t) and X(t-1) represent the t-th and t-1 state node information respectively
  • the number of nodes is determined by the input variable U Determined by the data length
  • W in and W x respectively represent the input layer-storage layer weight matrix in the echo state network, and the internal connection weight matrix of the storage layer.
  • Logsig( ⁇ ) represents the activation function, which can approximate any non- Linear function, and then the neural network can be applied to the nonlinear model.
  • the activation function we directly multiply the input amount by the input layer scale coefficient and transform it into the corresponding range of the activation function. Since the input is sequentially calculated, t can be understood as the time.
  • E represents the loss function
  • ⁇ 1 and ⁇ 2 are both regularization coefficients.
  • the present invention simplifies the loss function E, and then uses the coordinate descent algorithm to calculate the output weight;
  • the method of solving matrix W′ out is to calculate each element separately, the value of the k-th element in the m-th row of W′ out is as follows:
  • Y′ k (t) represents the t-th element in the k-th row of Y′
  • X′ j (t) represents the t-th element in the j-th row of X′
  • the output weight W out is calculated using the relationship between the matrix W'out and the output weight W out .
  • This embodiment also includes adaptive optimization of the output weight obtained in step U403, and the optimization is U404:
  • the converted loss function is:
  • n is the number of storage layer nodes
  • K is the number of output layer nodes.
  • the present invention innovatively uses thermal data prediction instead of thermal data identification, which improves the accuracy of thermal data discrimination.
  • Financial1 is a write-intensive tracking file.
  • MSR is a common workload for large-scale enterprise servers. Distilled represents a typical use mode of personal computers.
  • MillSSD is collected from industrial automatic optical inspection instruments and has Runcore RCS hardware configuration -V- T25SSD (512GB, SATA2), Intel X2 7400 and 2G DDR3. MillSSD is also a write-intensive tracking file because it has the effect of substantial image backup. The performance comparison result of this embodiment is shown in Figure 5.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Human Computer Interaction (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

一种基于联合优化回声状态网络的热数据预测方法，将量子粒子群优化用于计算回声状态网络的存储层参数得到最佳存储层参数，在粒子位置更新过程中结合L2+自适应L1/2正则化约束的回声状态网络计算输出权值并计算出全局最佳适应值，迭代结束时将全局最佳适应值对应的粒子位置作为最佳存储层参数，最后基于最佳存储层参数采用L2+自适应L1/2正则化约束的回声状态网络计算最终输出权值，并利用最终输出权值和输入的历史热数据所在逻辑区块地址预测出热数据，预测逻辑区块地址上的数据为热数据，该方法将热数据预测替代了热数据识别，能更好地服务于固态硬盘的垃圾回收和磨损均衡任务。

Description

一种基于联合优化回声状态网络的热数据预测方法
本申请要求于2019年6月27日提交中国专利局、申请号为201910566123.4、发明名称为“一种基于联合优化回声状态网络的热数据预测方法”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本发明属于混沌时间序列预测技术领域，具体涉及一种基于联合优化回声状态网络的热数据预测方法。
背景技术
作为一种非易失性存储技术，NAND闪存被广泛用于通信系统和消费电子产品，与硬盘驱动器相比，具有更高的访问速度和功率效率。在基于NAND闪存的消费电子设备中，大量应用依赖于NAND闪存来进行数据交换、文件存储和视频存储。NAND闪存主要用来存储大容量数据，NAND结构能提供极高的单元密度，可以达到高存储密度、高写入和高擦除速度；因此NAND闪存多应用于大容量数据存储，例如固态硬盘。未来NAND闪存的需求量还将持续增长，主要应用在云计算、物联网及数据中心等领域。
然而，NAND闪存至少面临两个挑战，即异地更新和有限的耐久性，这限制了其大规模应用。NAND闪存存在无法覆盖操作的缺陷，即在一个页面闪存被擦除之前无法在该页面上执行新的写操作。因此不当的更新将生成许多无效页面和死页面，这会降低效率和性能。此外，NAND闪存具有有限的寿命，因为闪存块只能承受有限次数的擦除，如果块的擦除次数大于块的最大可擦除次数，则它将无法使用。垃圾回收(Garbage Collection,GC)和磨损均衡(Wear Leveling,WL)是将频繁写入的数据(即热数据)分配到具有较少擦除次数的块中，并将最近最少使用的数据(即冷数据)分配到具有较多擦除次数的块中的设计思想，对解决这两个挑战有着重要的影响，而GC和WL的效率和性能较大程度上依赖于热数据识别(Hot Data Identification,HDI)。HDI的本质是试图很好地理解热数据的访问行为，以 便智能地将不同的数据分配给适当的块，但传统的HDI具有以下两个问题，一个是内存开销很大。目前，大部分的热数据识别机制均是采用识别NAND闪存中热数据页的思想，这些机制的核心原理就是页计数器，在一定时间内记录与NAND闪存页相对应的逻辑页的读写操作次数，如果读写次数大于设定的阈值，则该请求页被判断为热页，否则为冷页。另一个严重的问题时识别的精确度不高，基于Bloom过滤器的热数据识别机制广泛适用于SSD的冷热数据识别，但Bloom过滤器固有的缺陷就是假阳性，即将不属于集合内的数据错误地判定为在集合内。另外，基于负载请求大小和负载访问模式等热数据识别方式考虑因素比较单一，没能完全综合考虑负载的局部性特征，热数据识别的准确度不高。
发明内容
本发明的目的是提供一种基于联合优化回声状态网络的热数据预测方法，其创造性地提出用热数据预测代替传统的热数据识别，并构建了联合优化回声状态网络，从而使预测的热数据更具有实时性和可靠性。
本发明提供一种基于联合优化回声状态网络的热数据预测方法，包括如下步骤：
S1：初始化量子粒子群算法所需参数以及每个粒子的位置信息；
其中，粒子的位置信息包括粒子初始位置以及位置范围，每个粒子的位置以回声状态网络中存储层参数表示；
S2：利用量子粒子群算法迭代寻优确定最佳存储层参数；
其中，基于每个粒子的位置范围采用量子粒子群算法更新粒子位置，每次更新过程中利用L2+自适应L1/2正则化约束的回声状态网络计算输出权值并计算出全局最佳适应值，迭代结束时全局最佳适应值对应的粒子位置作为最佳存储层参数；
S3：基于回声状态网络中最佳存储层参数采用L2+自适应L1/2正则化约束的回声状态网络计算最终输出权值；
S4：利用最终输出权值和输入的历史热数据所在逻辑区块地址预测出热数据，预测公式如下：
y＝x*W out
式中，y表示获取的预测逻辑区块地址，预测逻辑区块地址上的数据为热数据，x为输入的历史热数据所在逻辑区块地址，W out表示输出权值，所述历史热数据所在的逻辑区块地址用于步骤S2和步骤S3中回声状态网络训练过程。
本发明在沿用固态硬盘结构框架的同时将闪存转换层中的热数据识别模块创新性的用热数据预测代替，使用联合优化回声状态网络预测热数据，联合优化包括两个部分，第一部分使用量子粒子群算法迭代寻优确定回声状态网络的最佳存储层参数，第二部分使用L2+自适应L1/2正则化约束的ESN获取高稀疏度的输出权值，本发明将量子粒子群算法迭代寻优以及L2+自适应L1/2正则化约束相结合得到最佳存储层参数，用于预测的联合优化回声状态网络更具有实时性和可靠性。本发明利用历史热数据所在的逻辑区块地址训练回声状态网络得到最终输出权值，再利用其预测为热数据所在的逻辑区块地址。
进一步优选，步骤S2中迭代寻优确定最佳存储层参数的执行过程如下：
S21：将每个粒子的位置依次作为回声状态网络中存储层参数，并分别采用L2+自适应L1/2正则化约束的回声状态网络计算出每个粒子对应输出权值；
其中，将每个粒子的当前位置依次作为回声状态网络中的存储层参数并计算输出权值；
S22：利用每个粒子对应的输出权值计算出每个粒子的适应值；
S23：基于最小适应值原则根据每个粒子的适应值选择出每个粒子的个体最佳适应值、个体最佳参数，以及全局最佳适应值以及全局最佳参数；
其中，选为全局最佳适应值的粒子位置为全局最佳参数；
S24：在粒子的位置范围内更新每个粒子的位置，并基于每个粒子的更新位置重新计算每个粒子的适应值，并基于最小适应值原则更新每个粒子的个体最佳适应值、个体最佳参数，以及全局最佳适应值以及全局最佳参数；
S25：判断迭代次数是否达到最大迭代次数，若未达到，返回步骤S24进行下一次迭代计算；否则，将当前的全局最佳参数作为最佳存储层参数。
进一步优选，按照如下公式更新任一粒子j的位置：
Figure PCTCN2020097950-appb-000001
其中，
Figure PCTCN2020097950-appb-000002
式中，P j(t+1)、P j(t)分别表示更新后、更新前的粒子j的位置，
Figure PCTCN2020097950-appb-000003
和u j均为随机数，sbest j、sbest i表示第j个、第i个粒子的个体最佳参数，mbest为所有粒子当前个体最佳参数的平均值，iter、iter max分别为当前迭代次数以及最大迭代次数，ω max、ω min分别为惯性因子，N为粒子总数。
进一步优选，任一粒子j的适应值的计算公式为：
Figure PCTCN2020097950-appb-000004
式中，Fitness表示当前粒子j的适应值，λ 1、λ 2均为正则化系数，W out为当前粒子j对应的输出权值；Y表示用于网络训练的历史热数据所在逻辑区块地址的后段，X表示基于网络训练的历史热数据所在逻辑区块地址的前段更新的存储层的状态信息，X*W out表示历史热数据所在逻辑区块地址的后段对应的预测结果。
进一步优选，采用L2+自适应L1/2正则化约束的回声状态网络计算输出权值的过程如下：
U401：获取回声状态网络中输入层-存储层权值矩阵，存储层内部连接权值矩阵，以及利用历史热数据所在逻辑区块地址的前段作为输入变量U，后段作为实际结果Y；
其中，输入层-存储层权值矩阵以及存储层内部连接权值矩阵与回声状态网络中存储层参数相关；
U402：基于输入变量U更新存储层的状态信息X，状态信息X由状态节点信息X(t)构成；
X(t)＝log sig(U(t)W in+X(t-1)W x)
式中，U(t)表示输入变量U中第t个数据，X(t)、X(t-1)分别表示第t 个、第t-1个状态节点信息，t的最大值T由输入变量U的数据长度决定，W in、W x分别表示回声状态网络中输入层-存储层权值矩阵，存储层内部连接权值矩阵，logsig(·)表示激活函数；
U403：基于L2+自适应L1/2正则化约束下的损失函数获取损失函数最小值下的输出权值；
Figure PCTCN2020097950-appb-000005
式中，E表示损失函数，λ 1、λ 2均为正则化系数。
进一步优选，步骤U403的过程为：将损失函数进行简化，再采用坐标下降算法计算出输出权值；
其中，简化后的损失函数表示为：
Figure PCTCN2020097950-appb-000006
存在：
Figure PCTCN2020097950-appb-000007
其中，I为单位矩阵；
求解矩阵W′ out的方法为对其中每个元素分别计算，W′ out第m行第k个元素的值如下：
Figure PCTCN2020097950-appb-000008
其中，
Figure PCTCN2020097950-appb-000009
Figure PCTCN2020097950-appb-000010
式中，Y′ k(t)表示Y′第k行第t个元素，X′ j(t)表示X′第j行第t个元素；
Figure PCTCN2020097950-appb-000011
表示矩阵W′ out的第j行第k个元素，j>m时，
Figure PCTCN2020097950-appb-000012
为零；L为输出层节点个数，n为存储层节点个数。
进一步优选，还包括对步骤U403得到的输出权值进行自适应优化，优化过程如下：
对损失函数进行转换，并采用坐标下降算法计算出权值W″ out，再计算出优化后的输出权值；
转换后的损失函数为：
Figure PCTCN2020097950-appb-000013
权值W″ out与输出权值W out的关系为：
Figure PCTCN2020097950-appb-000014
其中，
Figure PCTCN2020097950-appb-000015
K为输入层节点个数。
进一步优选，回声状态网络中存储层参数包括内部连接谱半径，存储层规模，输入层比例系数和存储层稀疏度四个关键参数。
进一步优选，步骤S1中初始化量子粒子群算法所需参数包括粒子群规模N，最大迭代次数iter max、以及惯性因子ω max和ω min。
进一步优选，更新粒子位置时，若粒子移动距离超过了粒子对应的位置范围，则将粒子位置参数设置为超过位置范围对应的边界值。
有益效果：
1、本发明创新性地提出了用热数据预测替代传统的热数据识别，公开的热数据预测技术可根据历史访问行为，提前一个甚至几个节拍预测下一个数据的性质，主动地分配存储至固态硬盘(SSD)对应区块(热/冷数据区块)，相较于传统的热数据识别更具有主动性，且本发明实现过程使用联合优化来提升网络预测的准确度，从而获得了更精确的热数据识别效果，高效的热数据预测将更好的服务于垃圾回收和磨损均衡技术，最终提升固态硬盘的寿命。同时神经网络的方法对于输入量保留了更多的特征信息，更全面的对热数据进行分类。
2、本发明对回声状态网络进行联合优化，L2正则化约束通过模型偏差和预测方差之间的权衡获得了良好的泛化能力从而能够获得连续收缩的权重，然而它不能产生稀疏解；自适应L1/2正则化可以生成非常稀疏的解，但是当预测变量之间存在高度相关性时，L1/2不能很好地发挥调节作用， 本发明综合采用L2+自适应L1/2正则化训练最小二乘法，可以获得两种正则化的优点，进而提高了对热数据的预测精度。另外，基于QPSO算法对回声状态网络存储层参数进行优化，能够解决搭建模型时无法确定存储层参数的问题。相较于传统的PSO算法，该算法基于波粒二象性去除了粒子的速度信息，仅保留位置信息，可以有效的降低计算的复杂性，同时将获取适应模型的存储层参数，从而进一步提升预测精度；进而，本发明将L2+自适应L1/2正则化以及QPSO算法相结合，得到了最佳存储层参数，提升预测精度。
附图说明
图1是NAND闪存系统的典型体系结构；
图2是本发明一个实施例提供的基于联合优化回声状态网络的热数据预测方法的流程图；
图3是本发明量子粒子群算法迭代寻优的具体算法流程图；其中，图3A中步骤U304执行完毕后转向图3B中步骤U305。
图4是本发明使用L2+自适应L1/2约束回声状态网络计算输出权值的具体算法流程图。
图5是本发明一实施例四种实际工作量的性能对比图。
具体实施方式
下面将结合实施例对本发明做进一步的说明。
本发明提供一种基于联合优化回声状态网络的热数据预测方法主要是应用于NAND闪存系统，如图1所示为NAND闪存系统的典型体系结构，其包括模块B101(用户的操作)、模块B102(文件系统)和模块B103(固态硬盘)。用户的实际操作将通过文件系统对固态硬盘产生影响。固态硬盘又包括闪存转换层、flash控制器和NAND flash阵列，其中闪存转换层包含地址分配单元、垃圾回收单元、磨损均衡单元、热数据预测单元，其中本发明创新性地提出将热数据预测单元替代传统的热数据识别单元，传统的热数据识别方法通常被动地分析用户访问行为，通过Flash传输层协议 (FTL)将对应数据分配存储至固态硬盘(SSD)对应区块(热/冷数据区块)，该方法在应对访问行为复杂的请求时存在较高的热数据漏检或虚警。本发明公开的热数据预测技术可根据历史访问行为，提前一个甚至几个节拍预测下一个数据的性质，主动地分配存储至固态硬盘(SSD)对应区块(热/冷数据区块)，同时，兼容传统热数据识别方案的二次校验。据此，本发明提出的热数据预测方法实质为"预测性热数据识别"。本发明最终获取的预测逻辑区块地址信息用于垃圾回收、磨损均衡处理。
上述过程中可以看出磨损均衡和垃圾收集在固态硬盘中具有较大的影响力。传统的热数据识别就是为了准确，高效的分辨出哪些数据属于有效数据。本发明提供一种基于联合优化回声状态网络的热数据预测方法是将热数据预测来代替热数据识别，具有高精准度的预测，其具体包括如下步骤：
S1：初始化量子粒子群算法所需参数以及每个粒子的位置信息。
其中，粒子的位置信息包括粒子初始位置以及位置范围，每个粒子的位置用回声状态网络(ESN)中存储层参数表示，回声状态网络中存储层参数包括内部连接谱半径，存储层规模，输入层比例系数和存储层稀疏度，每个粒子的维数在本例中初始化为4，即每个粒子是1*4的矩阵，分别代表ESN存储层的4个参数。确定4个参数的范围，将参数的范围定为所有粒子的位置范围，初始化时在位置范围内给每个粒子随机赋值，在随后的更新过程中可视为粒子在规定范围内不断朝着最优方向移动，若粒子在移动中超出了规定范围则将粒子位置信息更新为边界值。其中每个粒子位置都代表ESN存储层参数具体数值。
量子粒子群算法所需参数包括粒子群规模N、最大迭代次数Itermax、惯性因子ωmax和ωmin(用于后续更新粒子位置信息)。
S2：利用量子粒子群算法迭代寻优确定最佳存储层参数；
其中，基于每个粒子的位置范围采用量子粒子群算法更新粒子位置，并每次更新过程中利用L2+自适应L1/2正则化约束的回声状态网络计算输出权值再得到全局最佳适应值，迭代结束时全局最佳适应值对应的粒子位置作为最佳存储层参数。其具体过程包括如下步骤：
S21：将每个粒子的位置作为回声状态网络中存储层参数，并分别采用L2+自适应L1/2正则化约束的回声状态网络计算出每个粒子对应输出权值；
其中，将每个粒子的当前位置依次作为回声状态网络中的存储层参数并计算输出权值；
S22：利用每个粒子对应的输出权值计算出每个粒子的适应值；
S23：基于最小适应值原则根据每个粒子的适应值选择出每个粒子的个体最佳适应值、个体最佳参数，以及全局最佳适应值以及全局最佳参数；
其中，选为全局最佳适应值的粒子位置为全局最佳参数；
S24：在粒子的位置范围内更新每个粒子的位置，并基于每个粒子的更新位置重新计算每个粒子的适应值，并基于最小适应值原则更新每个粒子的个体最佳适应值、个体最佳参数，以及全局最佳适应值以及全局最佳参数；
S25：判断迭代次数是否达到最大迭代次数，若未达到，返回步骤S24进行下一次迭代计算；否则，将当前的全局最佳参数作为最佳存储层参数。
基于上述逻辑，本发明实施例提供如图3所示的实例流程图，其包括如下步骤：
U301：迭代初始化，设置当前迭代次数iter为1，设置粒子标号j为1。
U302：将第j个粒子所在位置设置为ESN存储层参数，利用L2+自适应L1/2正则化约束训练中出现的最小二乘计算，以获取更高稀疏度的输出权值Wout。用L2+自适应L1/2正则化约束的ESN计算输出权值Wout的详细步骤如图4所示，下文再具体描述。
U303：基于第j个粒子对应的输出权值Wout计算第j个粒子对应的适应值，计算公式如下：
Figure PCTCN2020097950-appb-000016
式中，λ 1、λ 2均为正则化系数，W out为当前粒子j对应的输出权值；Y表示用于网络训练的历史热数据所在逻辑区块地址的后段，X表示基于网络训练的历史热数据所在逻辑区块地址的前段更新的存储层的状态信息，X*W out表示历史热数据所在逻辑区块地址的后段对应的预测结果。
D301：判断是否所有粒子均已完成适应值计算，若没有则将j加1，返回步骤U302计算下一个粒子的适应值，若所有粒子均已计算完适应值，则进行步骤U304。
U304：基于最小适应值原则根据每个粒子的适应值选择出每个粒子的个体最佳适应值、个体最佳参数，以及全局最佳适应值以及全局最佳参数。在所有粒子均已计算完适应值后进行比较判断，记录每个粒子的适应值为个体最佳适应值fsbest，每个粒子的位置为个体最佳参数sbest；记录所有粒子适应值中最小的粒子适应值为全局最佳适应值fgbest，其对应的位置为全局最佳参数gbest。获得的这些参数将用于后续迭代寻优。
U305：迭代开始，将粒子标号j重新置为1。
U306：计算第j个粒子对应的mbest，计算公式如下：
Figure PCTCN2020097950-appb-000017
式中，sbest i表示第i个粒子的个体最佳参数，mbest为所有粒子当前个体最佳参数的平均值，即所有粒子的每一维参数分别取平均值，用于更新粒子位置信息。
U307：更新粒第j个粒子的位置信息，更新公式如下：
Figure PCTCN2020097950-appb-000018
其中，P j(t+1)、P j(t)分别表示更新后、更新前的粒子j的位置，
Figure PCTCN2020097950-appb-000019
和u j均为(0,1)之间的随机数，其中，β的计算公式为：
Figure PCTCN2020097950-appb-000020
从β的计算公式中可以看出，在迭代前期，代表粒子移动的步长的参数β较大，粒子可以更快的向最佳位置移动；而迭代后期β较小，意味着在最佳位置附近粒子减小步长，每次移动更精确的靠近最佳位置。
更新完位置信息后，对新得到的存储层参数再次利用L2+自适应L1/2正则化约束的ESN重新计算适应值，因此D302、D303、U308、U309为：根据新计算出的适应值更新个体最佳和全局最佳，若新计算出的适应值小于该粒子的个体最佳适应值，则将该粒子的个体最佳适应值更新为新计算 出的适应值，同时更新该粒子的个体最佳参数为当前粒子的参数；若新计算出的适应值小于全局最佳适应值，则更新该适应值为全局最佳适应值，同时更新全局最佳参数为该粒子的参数。
D304：判断是否所有粒子都更新完，若没有，j+1并返回U306，利用已更新的粒子参数重新计算mbest，并更新下一个粒子的位置信息，若所有粒子都已更新完，则进行D305。
D305：判断迭代次数是否达到最大迭代次数，若没有则将iter加1返回U305，进行下一次迭代，若已达到最大迭代次数，导出最终的全局最佳参数用于后续训练联合优化回声状态网络来预测逻辑区块地址。
S3：基于回声状态网络中最佳存储层参数采用L2+自适应L1/2正则化约束的回声状态网络计算最终输出权值。其中，计算最终输出权值的过程如图4所示，下文将对此进行详细描述。
S4：利用最终输出权值和输入的历史热数据所在逻辑区块地址预测出热数据，预测公式如下：
y＝x*W out
式中，y表示获取的预测逻辑区块地址，预测逻辑区块地址上的数据为热数据，x为输入的历史热数据所在逻辑区块地址，W out表示输出权值。其中，y就是预测访问地址，值得注意的是，x和Wout均可以是多维变量，而得到的y则是一维变量，得到的逻辑区块地址上的数据将被归为热数据，用于垃圾回收和磨损均衡处理。
计算输出权值时，均为确定了一组存储层参数，即内部连接谱半径，存储层规模，输入层比例系数和存储层稀疏度。如图4所示，本发明中采用L2+自适应L1/2正则化约束的回声状态网络计算输出权值的过程如下：
U401：获取回声状态网络中输入层-存储层权值矩阵，存储层内部连接权值矩阵，以及利用历史热数据所在逻辑区块地址的前段作为输入变量U，后段作为实际结果Y。
具体的，回声状态网络(Echo State Network,ESN)是一种计算复杂度低且快速收敛的计算方案，适用于时间数据分类和预测任务。ESN网络架构包括三层：输入层，存储层和输出层，其中，输入层-存储层权值为Win， 存储层内部连接权值为Wx，存储层-输出层权值为Wout。初始化输入层、存储层、输出层中节点个数为K、n、L，存储层节点个数n即为存储层参数中存储层规模决定。以及初始化输入层-存储层权值W in∈R n×K，即随机赋值；初始化存储层内部连接权值W x∈R n×n，即n×n×存储层稀疏度得到非零个数，再对连接权值Wx中的非零元素的位置和大小随机赋值，其他元素均为零。且当存储层稀疏度越大时，非线性逼近能力越强；之后利用内部连接谱半径确定内部连接权值Wx的最大特征值，只有当内部连接谱半径小于1时才能确保网络的稳定。因此，基于存储层的参数来确定输入层-存储层权值W in∈R n×K，存储层内部连接权值W x∈R n×n。本实施例中，还初始化L1/2和L2系数λ1＝5*10-7,λ2＝1*10-5用于正则化计算，利用利用输入的历史热数据所在逻辑区块地址的前2/3构造成输入变量U，后1/3构造成实际结果Y，本发明实施例中选用的历史热数据所在逻辑区块地址为用户记载的历史热数据所在逻辑区块地址，其他可行的实施例中，选取的长度可以是其他，本发明对其不进行具体的限定，总体思路为，利用前段地址来预测后段地址，再将预测的后段地址与实际地址进行比对来调节网络，此部分为回声状态网络原有特性，本发明对其不进行具体的描述。
U402：基于输入变量U更新存储层的状态信息X，状态信息X由状态节点信息X(t)构成；
X(t)＝log sig(U(t)W in+X(t-1)W x)
式中，U(t)表示输入变量U中第t个数据，X(t)、X(t-1)分别表示第t个、第t-1个状态节点信息，节点个数由输入变量U的数据长度决定，W in、W x分别表示回声状态网络中输入层-存储层权值矩阵，存储层内部连接权值矩阵，logsig(·)表示激活函数，它将神经网络可以任意逼近任何非线性函数，之后神经网络就可以应用到非线性模型中，在使用激活函数时我们将输入量直接乘以输入层比例系数变换到激活函数相应的范围内。由于依次输入进行计算，因此t可以理解为时刻。
U403：基于L2+自适应L1/2正则化约束下的损失函数获取损失函数最小值下的输出权值；
Figure PCTCN2020097950-appb-000021
式中，E表示损失函数，λ 1、λ 2均为正则化系数。
为了实现计算，本发明将损失函数E进行简化，再采用坐标下降算法计算出输出权值；
其中，简化后的损失函数表示为：
Figure PCTCN2020097950-appb-000022
存在：
Figure PCTCN2020097950-appb-000023
其中，I为单位矩阵？
求解矩阵W′ out的方法为对其中每个元素分别计算，W′ out第m行第k个元素的值如下：
Figure PCTCN2020097950-appb-000024
其中，
Figure PCTCN2020097950-appb-000025
Figure PCTCN2020097950-appb-000026
式中，Y′ k(t)表示Y′第k行第t个元素，X′ j(t)表示X′第j行第t个元素；
Figure PCTCN2020097950-appb-000027
表示矩阵W′ out的第j行第k个元素，j>m时，
Figure PCTCN2020097950-appb-000028
为零。
最后利用矩阵W′ out与输出权值W out的关系计算出输出权值W out。本实施例中还包括对步骤U403得到的输出权值进行自适应优化，优化为U404：
U404：对损失函数进行转换，并采用坐标下降算法计算出权值W″ out，再利用权值W″ out计算出优化后的输出权值；
转换后的损失函数为：
Figure PCTCN2020097950-appb-000029
权值W″ out与输出权值W out的关系为：
Figure PCTCN2020097950-appb-000030
其中，
Figure PCTCN2020097950-appb-000031
n为存储层节点个数，K为输出层节点个数。
为了验证本发明所述方法的可靠性，本发明创新性地用热数据预测代替热数据识别，提高了热数据判别准确率。我们采用了四个实际工作量进行客观评估。Financial1是一个写密集型跟踪文件，MSR是大型企业服务器的常见工作负载，Distilled代表个人计算机的典型使用模式，最后，MillSSD是从工业自动光学检测仪器收集的，具有Runcore RCS的硬件配置-V-T25SSD(512GB，SATA2)，Intel X2 7400和2G DDR3。MillSSD也是一个写密集型跟踪文件，因为它具有实质性映像备份的作用。本实施例的性能对比结果见图5。从试验结果可知：以WDAC为基准，HOESN热比率曲线在大多数情况下几乎与WDAC重叠。在所有四种工作负载下都可以清楚地发现这一主要趋势，特别是对于更多写密集型MSR和MillSSD。很明显，在四个工作负载下，我们的HOESN的FIR最低，其次是DL-MBF_s。尽管MBF经历了相对较高的FIR，但它仍然是用于SSD的良好HDI方案，其中提出了WDAC，其成为以下研究的经典基准。值得注意的是，在四种工作负载中，HOESN的改进程度对于MillSSD来说是最令人印象深刻的(从4.08％到2.23％)。这些初步测试也证明了我们最初的想法，即理解NAND闪存的热数据的访问行为可以被认为是时间序列的预测，HOESN正是针对这个想法提出的。结果表明，我们的预测方法可以很好地了解磁盘工作负载的访问行为，这是为GC和WL提供可靠服务的基本前提。
需要强调的是，本发明所述的实例是说明性的，而不是限定性的，因此本发明不限于具体实施方式中所述的实例，凡是由本领域技术人员根据本发明的技术方案得出的其他实施方式，不脱离本发明宗旨和范围的，不论是修改还是替换，同样属于本发明的保护范围。

Claims (10)

  1. 一种基于联合优化回声状态网络的热数据预测方法，其特征在于：包括如下步骤：
    S1：初始化量子粒子群算法所需参数以及每个粒子的位置信息；
    其中，粒子的位置信息包括粒子初始位置以及位置范围，每个粒子的位置以回声状态网络中存储层参数表示；
    S2：利用量子粒子群算法迭代寻优确定最佳存储层参数；
    其中，基于每个粒子的位置范围采用量子粒子群算法更新粒子位置，每次更新过程中利用L2+自适应L1/2正则化约束的回声状态网络计算输出权值并计算出全局最佳适应值，迭代结束时全局最佳适应值对应的粒子位置作为最佳存储层参数；
    S3：基于回声状态网络中最佳存储层参数采用L2+自适应L1/2正则化约束的回声状态网络计算最终输出权值；
    S4：利用最终输出权值和输入的历史热数据所在逻辑区块地址预测出热数据，预测公式如下：
    y＝x*W out
    式中，y表示获取的预测逻辑区块地址，预测逻辑区块地址上的数据为热数据，x为输入的历史热数据所在逻辑区块地址，W out表示输出权值，所述历史热数据所在的逻辑区块地址用于步骤S2和步骤S3中回声状态网络训练过程。
  2. 根据权利要求1所述的方法，其特征在于：步骤S2中迭代寻优确定最佳存储层参数的执行过程如下：
    S21：将每个粒子的位置依次作为回声状态网络中存储层参数，并分别采用L2+自适应L1/2正则化约束的回声状态网络计算出每个粒子对应输出权值；
    其中，将每个粒子的当前位置依次作为回声状态网络中的存储层参数并计算输出权值；
    S22：利用每个粒子对应的输出权值计算出每个粒子的适应值；
    S23：基于最小适应值原则根据每个粒子的适应值选择出每个粒子的个 体最佳适应值、个体最佳参数，以及全局最佳适应值以及全局最佳参数；
    其中，选为全局最佳适应值的粒子位置为全局最佳参数；
    S24：在粒子的位置范围内更新每个粒子的位置，并基于每个粒子的更新位置重新计算每个粒子的适应值，并基于最小适应值原则更新每个粒子的个体最佳适应值、个体最佳参数，以及全局最佳适应值以及全局最佳参数；
    S25：判断迭代次数是否达到最大迭代次数，若未达到，返回步骤S24进行下一次迭代计算；否则，将当前的全局最佳参数作为最佳存储层参数。
  3. 根据权利要求2所述的方法，其特征在于：按照如下公式更新任一粒子j的位置：
    Figure PCTCN2020097950-appb-100001
    其中，
    Figure PCTCN2020097950-appb-100002
    式中，P j(t+1)、P j(t)分别表示更新后、更新前的粒子j的位置，
    Figure PCTCN2020097950-appb-100003
    和u j均为随机数，sbest j、sbest i表示第j个、第i个粒子的个体最佳参数，mbest为所有粒子当前个体最佳参数的平均值，iter、iter max分别为当前迭代次数以及最大迭代次数，ω max、ω min分别为惯性因子，N为粒子总数。
  4. 根据权利要求2所述的方法，其特征在于：任一粒子j的适应值的计算公式为：
    Figure PCTCN2020097950-appb-100004
    式中，Fitness表示当前粒子j的适应值，λ 1、λ 2均为正则化系数，W out为当前粒子j对应的输出权值；Y表示用于网络训练的历史热数据所在逻辑区块地址的后段，X表示基于网络训练的历史热数据所在逻辑区块地址的前段更新的存储层的状态信息，X*W out表示历史热数据所在逻辑区块地址的后段对应的预测结果。
  5. 根据权利要求1所述的方法，其特征在于：采用L2+自适应L1/2正则化约束的回声状态网络计算输出权值的过程如下：
    U401：获取回声状态网络中输入层-存储层权值矩阵，存储层内部连接 权值矩阵，以及利用历史热数据所在逻辑区块地址的前段作为输入变量U，后段作为实际结果Y；
    其中，输入层-存储层权值矩阵以及存储层内部连接权值矩阵与回声状态网络中存储层参数相关；
    U402：基于输入变量U更新存储层的状态信息X，状态信息X由状态节点信息X(t)构成；
    X(t)＝log sig(U(t)W in+X(t-1)W x)
    式中，U(t)表示输入变量U中第t个数据，X(t)、X(t-1)分别表示第t个、第t-1个状态节点信息，t的最大值T由输入变量U的数据长度决定，Win、Wx分别表示回声状态网络中输入层-存储层权值矩阵，存储层内部连接权值矩阵，logsig(·)表示激活函数；
    U403：基于L2+自适应L1/2正则化约束下的损失函数获取损失函数最小值下的输出权值；
    Figure PCTCN2020097950-appb-100005
    式中，E表示损失函数，λ 1、λ 2均为正则化系数。
  6. 根据权利要求5所述的方法，其特征在于：步骤U403的过程为：将损失函数进行简化，再采用坐标下降算法计算出输出权值；
    其中，简化后的损失函数表示为：
    Figure PCTCN2020097950-appb-100006
    存在：
    Figure PCTCN2020097950-appb-100007
    其中，I为单位矩阵；
    求解矩阵W′ out的方法为对其中每个元素分别计算，W′ out第m行第k个元素的值如下：
    Figure PCTCN2020097950-appb-100008
    其中，
    Figure PCTCN2020097950-appb-100009
    Figure PCTCN2020097950-appb-100010
    式中，Y′ k(t)表示Y′第k行第t个元素，X′ j(t)表示X′第j行第t个元素；
    Figure PCTCN2020097950-appb-100011
    表示矩阵W′ out的第j行第k个元素，j>m时，
    Figure PCTCN2020097950-appb-100012
    为零；L为输出层节点个数，n为存储层节点个数。
  7. 根据权利要求6所述的方法，其特征在于：还包括对步骤U403得到的输出权值进行自适应优化，优化过程如下：
    对损失函数进行转换，并采用坐标下降算法计算出权值W″ out，再计算出优化后的输出权值；
    转换后的损失函数为：
    Figure PCTCN2020097950-appb-100013
    权值W″ out与输出权值W out的关系为：
    Figure PCTCN2020097950-appb-100014
    其中，
    Figure PCTCN2020097950-appb-100015
    K为输入层节点个数。
  8. 根据权利要求1所述的方法，其特征在于：回声状态网络中存储层参数包括内部连接谱半径，存储层规模，输入层比例系数和存储层稀疏度四个关键参数。
  9. 根据权利要求1所述的方法，其特征在于：步骤S1中初始化量子粒子群算法所需参数包括粒子群规模N，最大迭代次数iter max、以及惯性因子ω max和ω min。
  10. 根据权利要求1所述的方法，其特征在于：更新粒子位置时，若粒子移动距离超过了粒子对应的位置范围，则将粒子位置参数设置为超过位置范围对应的边界值。
PCT/CN2020/097950 2019-06-27 2020-06-24 一种基于联合优化回声状态网络的热数据预测方法 WO2020259543A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910566123.4 2019-06-27
CN201910566123.4A CN110554838B (zh) 2019-06-27 2019-06-27 一种基于联合优化回声状态网络的热数据预测方法

Publications (1)

Publication Number Publication Date
WO2020259543A1 true WO2020259543A1 (zh) 2020-12-30

Family

ID=68735438

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/097950 WO2020259543A1 (zh) 2019-06-27 2020-06-24 一种基于联合优化回声状态网络的热数据预测方法

Country Status (2)

Country Link
CN (1) CN110554838B (zh)
WO (1) WO2020259543A1 (zh)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110554838B (zh) * 2019-06-27 2020-08-14 中南大学 一种基于联合优化回声状态网络的热数据预测方法
CN112448697A (zh) * 2020-10-30 2021-03-05 合肥工业大学 一种基于量子粒子群算法的有源滤波器优化方法及系统
CN112731019B (zh) * 2020-12-21 2022-10-14 合肥工业大学 一种anpc三电平逆变器故障诊断方法
CN115841067A (zh) * 2022-10-12 2023-03-24 大连理工大学 针对航空发动机故障预警的量子回声状态网络模型构建方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103020434A (zh) * 2012-11-30 2013-04-03 南京航空航天大学 一种基于粒子群优化的最小二乘支持向量机组合预测方法
US20170046079A1 (en) * 2015-08-11 2017-02-16 International Business Machines Corporation Read distribution in a three-dimensional stacked memory based on thermal profiles
CN109901800A (zh) * 2019-03-14 2019-06-18 重庆大学 一种混合内存系统及其操作方法
CN110554838A (zh) * 2019-06-27 2019-12-10 中南大学 一种基于联合优化回声状态网络的热数据预测方法

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10307109B2 (en) * 2011-04-20 2019-06-04 Novo Nordisk A/S Glucose predictor based on regularization networks with adaptively chosen kernels and regularization parameters
US10014026B1 (en) * 2017-06-20 2018-07-03 Seagate Technology Llc Head delay calibration and tracking in MSMR systems
CN109726858A (zh) * 2018-12-21 2019-05-07 新奥数能科技有限公司 基于动态时间规整的热负荷预测方法及装置
CN109656485A (zh) * 2018-12-24 2019-04-19 合肥兆芯电子有限公司 区分热数据和冷数据的方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103020434A (zh) * 2012-11-30 2013-04-03 南京航空航天大学 一种基于粒子群优化的最小二乘支持向量机组合预测方法
US20170046079A1 (en) * 2015-08-11 2017-02-16 International Business Machines Corporation Read distribution in a three-dimensional stacked memory based on thermal profiles
CN109901800A (zh) * 2019-03-14 2019-06-18 重庆大学 一种混合内存系统及其操作方法
CN110554838A (zh) * 2019-06-27 2019-12-10 中南大学 一种基于联合优化回声状态网络的热数据预测方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
ZHU, HONGQIU ET AL.: "Hybrid-QPSO-based Parameters Optimization of LS-SVM and its Application", JOURNAL OF CENTRAL SOUTH UNIVERSITY (SCIENCE AND TECHNOLOGY), vol. 42, no. 4, 30 April 2011 (2011-04-30), DOI: 20200901141252A *

Also Published As

Publication number Publication date
CN110554838A (zh) 2019-12-10
CN110554838B (zh) 2020-08-14

Similar Documents

Publication Publication Date Title
WO2020259543A1 (zh) 一种基于联合优化回声状态网络的热数据预测方法
US10896126B2 (en) Storage device, method and non-volatile memory device performing garbage collection using estimated number of valid pages
Yang et al. Reducing garbage collection overhead in {SSD} based on workload prediction
WO2021143883A1 (zh) 神经网络的自适应搜索方法及装置
CN110968272B (zh) 基于时间序列预测的海量小文件存储性能优化方法及系统
CN107025071A (zh) 非易失性存储器装置及其垃圾收集方法
US11734103B2 (en) Behavior-driven die management on solid-state drives
CN103150245A (zh) 确定数据实体的访问特性的方法和存储控制器
CN106569962A (zh) 一种基于增强时间局部性的热点数据识别方法
Wang et al. A novel buffer management scheme based on particle swarm optimization for SSD
EP3651024B1 (en) Method of operating storage device, storage device performing the same and storage system including the same
Gupta et al. Relevance feedback based online learning model for resource bottleneck prediction in cloud servers
CN108829343B (zh) 一种基于人工智能的缓存优化方法
Lima et al. Evaluation of recurrent neural networks for hard disk drives failure prediction
Poczeta et al. Analysis of fuzzy cognitive maps with multi-step learning algorithms in valuation of owner-occupied homes
WO2023202484A1 (zh) 神经网络模型的修复方法和相关设备
CN115827651B (zh) 一种低能耗的机载嵌入式数据库内存事务管理方法及系统
Cao et al. A loan default discrimination model using cost-sensitive support vector machine improved by PSO
Liu et al. Semi‐supervised breast histopathological image classification with self‐training based on non‐linear distance metric
Ha et al. Dynamic hot data identification using a stack distance approximation
Ushasukhanya et al. Development and optimization of deep convolutional neural network using taguchi method for real-time electricity conservation system
Reddy et al. Analysis of Stock Market Value Prediction using Simple Novel Long Short Term Memory Algorithm in Comparison with Back Propagation Algorithm for Increased Accuracy Rate
Lee et al. File Access Characteristics of Deep Learning Workloads and Cache-Friendly Data Management
Chen et al. Pruning deep feature networks using channel importance propagation
Kargar et al. E2-NVM: A Memory-Aware Write Scheme to Improve Energy Efficiency and Write Endurance of NVMs using Variational Autoencoders.

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20833324

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20833324

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205 DATED 25/05/2022)

122 Ep: pct application non-entry in european phase

Ref document number: 20833324

Country of ref document: EP

Kind code of ref document: A1