Compact workload representation for memory systems

The compact representation of workloads in memory systems through the recurrent neural network compiler model solves the problem of difficult to analyze and separate workloads in the prior art, and improves the performance and reliability of flash memory.

CN114927157BActive Publication Date: 2024-11-26SK HYNIX INC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111207591.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2021-02-11
Filing Date
2021-10-18
Publication Date
2024-11-26
Estimated Expiration
2041-10-18

AI Technical Summary

Technical Problem

The prior art is difficult to effectively analyze and separate workloads in memory systems, especially for flash memory devices, where changes in workloads challenge performance and reliability improvements.

Method used

Using a compiler model based on recurrent neural network, a compact hidden state vector is generated through multiple recurrent encoding blocks to represent the workload. The activation function is used to convert the input command into a compact representation vector, and the original command is restored through the decoder to achieve work. Compact representation and visualization of load.

Benefits of technology

A compact representation of workloads in memory systems is achieved, helping with flash conversion layer (FTL) reconfiguration and firmware parameter adjustment, improving memory system performance and reliability, and able to detect multiple types of different workloads.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114927157B_ABST
    Figure CN114927157B_ABST
Patent Text Reader

Abstract

The present disclosure relates to a compact workload representation of a memory system. A compact representation of an input workload is generated in a memory system. The memory system includes a memory device; and a controller having a recurrent neural network compiler. The recurrent neural network compiler includes an encoder having a recurrent encoding block. Each recurrent encoding block: receives one of the input commands in the input workload associated with the memory device; and generates a hidden state vector corresponding to the received input command by applying a set of activation functions to the received input command. The final encoding block generates a final hidden state vector as a compact representation vector.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] Embodiments of the present disclosure relate to a scheme for analyzing workload in a memory system. Background Art

[0002] The computer environment paradigm has shifted to ubiquitous computing systems that can be used anywhere, anytime. As a result, the use of portable electronic devices such as mobile phones, digital cameras, and notebook computers has increased rapidly. These portable electronic devices typically use a memory system having (multiple) memory devices (i.e., (multiple) data storage devices). The data storage device is used as a primary memory device or an auxiliary memory device of the portable electronic device.

[0003] Since memory systems using memory devices have no moving parts, they have excellent stability, durability, high information access speed, and low power consumption. Examples of memory systems with these advantages include universal serial bus (USB) memory devices, memory cards with various interfaces (e.g., universal flash memory (UFS)), and solid-state drives (SSDs). The memory system can perform operations associated with one or more workloads from a host. Workload analysis becomes important for improving the performance and reliability of the memory system. In this case, an embodiment of the present invention is proposed. Summary of the invention

[0004] Aspects of the present invention include systems and methods for compactly representing input workloads in a memory system that are capable of separating the workloads by certain characteristics.

[0005] In one aspect, a system includes: a memory device; and a controller, including a recurrent neural network compiler (coder), the recurrent neural network compiler including an encoder having a plurality of recurrent encoding blocks, the encoder including a first encoding block to a last encoding block. Each recurrent encoding block is configured to: receive one of a plurality of input commands in an input workload associated with the memory device; and generate a hidden state vector corresponding to the received input command by applying a set of activation functions to the received input command. The last encoding block generates a final hidden state vector as a compact representation vector.

[0006] In another aspect, a method for operating a controller of a memory system includes providing a recurrent neural network compiler including an encoder having a plurality of recurrent encoding blocks; each recurrent encoding block receiving one of a plurality of input commands in an input workload associated with a memory device of the memory system; and each recurrent encoding block generating a hidden state vector corresponding to the received input command by applying a set of activation functions to the received input command. A final encoding block among the plurality of recurrent encoding blocks generates a final hidden state vector as a compact representation vector corresponding to the plurality of input commands.

[0007] Other aspects of the invention will become apparent from the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 is a block diagram showing a data processing system according to an embodiment of the present invention.

[0009] Figure 2 is a block diagram illustrating a memory system according to an embodiment of the present invention.

[0010] Figure 3 is a circuit diagram showing a memory block of a memory device according to an embodiment of the present invention.

[0011] Figure 4 is a diagram showing a data processing system according to an embodiment of the present invention.

[0012] Figure 5 is a diagram showing a recurrent neural network compiler according to an embodiment of the present invention.

[0013] Fig. 6A is a diagram illustrating a plurality of cycle blocks according to an embodiment of the present invention.

[0014] Figure 6B is a diagram illustrating a loop block according to an embodiment of the present invention.

[0015] Figure 7 It shows Figure 5 An illustration of the details of the recurrent neural network compiler.

[0016] Figure 8 is a diagram showing a recurrent neural network compiler according to an embodiment of the present invention.

[0017] Fig. 9 It shows Figure 8 An illustration of the details of the recurrent neural network compiler.

[0018] Fig.10 is a diagram showing a dataset of workloads compacted and visualized by a recurrent neural network compiler according to an embodiment of the present invention. DETAILED DESCRIPTION

[0019] Each embodiment is described in more detail below with reference to the accompanying drawings. However, the present invention can be implemented in different forms and should not be construed as being limited to the embodiments set forth herein. On the contrary, these embodiments are provided to make the present disclosure thorough and complete and to fully convey the scope of the present invention to those skilled in the art. In addition, references to "one embodiment", "another embodiment" and the like herein are not necessarily intended only for one embodiment, and different references to any such phrases are not necessarily intended for the same (multiple) embodiments. Throughout the disclosure, the same reference numerals in the drawings and embodiments of the present invention refer to the same parts.

[0020] The present invention can be implemented in a variety of ways, including being implemented as a process, a device, a system, a computer program product implemented on a computer-readable storage medium; and / or a processor, such as a processor suitable for running instructions stored on a memory connected to the processor and / or provided by the memory. In this specification, these embodiments or any other form that the present invention can take can be referred to as technology. In general, the order of the steps of the disclosed process can be changed within the scope of the present invention. Unless otherwise stated, components such as processors or memories described as being suitable for performing tasks can be implemented as general components temporarily configured to perform tasks at a given time or as specific components manufactured to perform tasks. As used herein, the term "processor" and the like refer to one or more devices, circuits, and / or processing cores suitable for processing data (e.g., computer program instructions).

[0021] A detailed description of embodiments of the invention and drawings illustrating aspects of the invention are provided below. The invention is described in conjunction with these embodiments, but the invention is not limited to any embodiment. The scope of the invention is limited only by the claims. The invention encompasses many alternatives, modifications, and equivalents within the scope of the claims. Many specific details are set forth in the following description in order to provide a thorough understanding of the invention. These details are provided for illustrative purposes only; the invention may be practiced according to the claims without some or all of these specific details. For the sake of clarity, technical materials known in the technical field related to the invention are not described in detail so as not to unnecessarily obscure the invention.

[0022] Figure 1 is a block diagram showing a data processing system 2 according to an embodiment of the present invention.

[0023] Reference Figure 1, the data processing system 2 may include a host device 5 and a memory system 10. The memory system 10 may receive a request from the host device 5 and operate in response to the received request. For example, the memory system 10 may store data to be accessed by the host device 5.

[0024] The host device 5 may be implemented using any of various types of electronic devices. In various embodiments, the host device 5 may include an electronic device such as a desktop computer, a workstation, a three-dimensional (3D) television, a smart television, a digital audio recorder, a digital audio player, a digital photo recorder, a digital photo player and / or a digital video recorder and a digital video player. In various embodiments, the host device 5 may include a portable electronic device such as a mobile phone, a smart phone, an e-book, an MP3 player, a portable multimedia player (PMP) and / or a portable game console.

[0025] The memory system 10 may be implemented using any of various types of storage devices such as a solid state drive (SSD) and a memory card. In various embodiments, the memory system 10 may be configured as one of various components in an electronic device such as a computer, an ultra mobile personal computer (UMPC), a workstation, a netbook, a personal digital assistant (PDA), a portable computer, a web tablet, a wireless phone, a mobile phone, a smart phone, an e-book reader, a portable multimedia player (PMP), a portable gaming device, a navigation device, a black box, a digital camera, a digital multimedia broadcasting (DMB) player, a three-dimensional television, a smart television, a digital audio recorder, a digital audio player, a digital photo recorder, a digital photo player, a digital video recorder, a digital video player, a storage device in a data center, a device capable of receiving and transmitting information in a wireless environment, a radio frequency identification (RFID) device, and one of various electronic devices in a home network, one of various electronic devices in a computer network, one of various electronic devices in a telematics network, or one of various components in a computing system.

[0026] The memory system 10 may include a memory controller 100 and a semiconductor memory device 200. The memory controller 100 may control the overall operations of the semiconductor memory device 200.

[0027] The semiconductor memory device 200 may perform one or more erase operations, program operations, and read operations under the control of the memory controller 100. The semiconductor memory device 200 may receive a command CMD, an address ADDR, and data DATA through an input / output line. The semiconductor memory device 200 may receive power PWR through a power line and receive a control signal CTRL through a control line. Depending on the design and configuration of the memory system 10, the control signal CTRL may include a command latch enable signal, an address latch enable signal, a chip enable signal, a write enable signal, a read enable signal, and other operation signals.

[0028] The memory controller 100 and the semiconductor memory device 200 may be integrated into a single semiconductor device such as a solid state drive (SSD). The SSD may include a storage device for storing data. When the memory system 10 is used for the SSD, the host device (e.g., Figure 1 The operating speed of the host device 5).

[0029] The memory controller 100 and the semiconductor memory device 200 may be integrated into a single semiconductor device such as a memory card. For example, the memory controller 100 and the semiconductor memory device 200 may be integrated to configure a personal computer (PC) card of the Personal Computer Memory Card International Association (PCMCIA), a compact flash (CF) card, a smart media (SM) card, a memory stick, a multimedia card (MMC), a reduced size multimedia card (RS-MMC), a micro version of MMC (MMCmicro), a secure digital (SD) card, a mini secure digital (miniSD) card, a micro secure digital (microSD) card, a secure digital high capacity (SDHC) and / or a universal flash memory (UFS).

[0030] Figure 2 is a block diagram showing a memory system according to an embodiment of the present invention. For example, Figure 2 The memory system can be described Figure 1 A memory system 10 is shown.

[0031] Reference Figure 2 , the memory system 10 may include a memory controller 100 and a semiconductor memory device 200. The memory system 10 may respond to a request from a host device (eg, Figure 1 The storage medium operates in response to a request from a host device 5) and in particular stores data to be accessed by the host device.

[0032] The memory device 200 may store data to be accessed by a host device.

[0033] The memory device 200 can be implemented using a volatile memory device (e.g., a dynamic random access memory (DRAM) and / or a static random access memory (SRAM)) or a non-volatile memory device (e.g., a read-only memory (ROM), a mask ROM (MROM), a programmable ROM (PROM), an erasable programmable ROM (EPROM), an electrically erasable programmable ROM (EEPROM), a ferroelectric random access memory (FRAM), a phase change RAM (PRAM), a magnetoresistive RAM (MRAM) and / or a resistive RAM (RRAM)).

[0034] The memory controller 100 may control storage of data in the memory device 200. For example, the memory controller 100 may control the memory device 200 in response to a request from a host device. The memory controller 100 may provide data read from the memory device 200 to the host device, and may store data provided from the host device in the memory device 200.

[0035] The memory controller 100 may include a storage device 110 connected via a bus 160, a control component 120 (which may be implemented as a processor such as a central processing unit (CPU)), an error correction code (ECC) component 130, a host interface (I / F) 140, and a memory interface (I / F) 150.

[0036] The memory device 110 may be used as a working memory of the memory system 10 and the memory controller 100, and store data for driving the memory system 10 and the memory controller 100. When the memory controller 100 controls the operation of the memory device 200, the memory device 110 may store data used by the memory controller 100 and the memory device 200 for operations such as a read operation, a write operation, a program operation, and an erase operation.

[0037] The storage device 110 may be implemented using a volatile memory such as a static random access memory (SRAM) or a dynamic random access memory (DRAM). As described above, the storage device 110 may store data used by the host device for read operations and write operations in the memory device 200. In order to store data, the storage device 110 may include a program memory, a data memory, a write buffer, a read buffer, a mapping buffer, etc.

[0038] The control component 120 can control the general operation of the memory system 10, and in particular, control the corresponding operation of the memory device 200 in response to a write request or a read request from a host device. The control component 120 can drive firmware called a flash translation layer (FTL) to control the general operation of the memory system 10. For example, the FTL can perform operations such as logical-physical (L2P) mapping, wear leveling, garbage collection, and / or bad block handling. L2P mapping is called logical block addressing (LBA).

[0039] The ECC component 130 may detect and correct errors in data read from the memory device 200 during a read operation. When the number of error bits is greater than or equal to a threshold number of correctable error bits, the ECC component 130 may not correct the error bits but may output an error correction failure signal indicating a failure to correct the error bits.

[0040] In various embodiments, the ECC component 130 can perform error correction operations based on coded modulation such as low density parity check (LDPC) codes, Bose-Chaudhuri-Hocquenghem (BCH) codes, turbo codes, turbo product codes (TPC), Reed-Solomon (RS) codes, convolutional codes, recursive systematic codes (RSC), trellis coded modulation (TCM), or block coded modulation (BCM). However, error correction is not limited to these techniques. Therefore, the ECC component 130 may include any and all circuits, systems, or devices for appropriate error correction operations.

[0041] The host interface 140 may communicate with the host device via one or more of a variety of interface protocols such as Universal Serial Bus (USB), MultiMediaCard (MMC), Peripheral Component Interconnect Express (PCI-e or PCIe), Small Computer System Interface (SCSI), Serial Attached SCSI (SAS), Serial Advanced Technology Attachment (SATA), Parallel Advanced Technology Attachment (PATA), Enhanced Small Disk Interface (ESDI), and / or Integrated Drive Electronics (IDE).

[0042] The memory interface 150 may provide an interface between the memory controller 100 and the memory device 200 to allow the memory controller 100 to control the memory device 200 in response to a request from a host device. The memory interface 150 may generate control signals for the memory device 200 and process data under the control of the control component 120. When the memory device 200 is a flash memory (e.g., a NAND flash memory), the memory interface 150 may generate control signals for the flash memory and process data under the control of the control component 120.

[0043] The memory device 200 may include a memory cell array 210, a control circuit 220, a voltage generating circuit 230, a row decoder 240, a page buffer 250 (which may be in the form of a page buffer array), a column decoder 260, and an input and output (input / output) circuit 270. The memory cell array 210 may include a plurality of memory blocks 211 that may store data. The voltage generating circuit 230, the row decoder 240, the page buffer array 250, the column decoder 260, and the input / output circuit 270 may form a peripheral circuit of the memory cell array 210. The peripheral circuit may perform a programming operation, a reading operation, or an erasing operation on the memory cell array 210. The control circuit 220 may control the peripheral circuit.

[0044] The voltage generating circuit 230 may generate various levels of operating voltages. For example, in an erase operation, the voltage generating circuit 230 may generate various levels of operating voltages, such as an erase voltage and a pass voltage.

[0045] The row decoder 240 may electrically communicate with the voltage generating circuit 230 and the plurality of memory blocks 211. The row decoder 240 may select at least one memory block among the plurality of memory blocks 211 in response to a row address generated by the control circuit 220, and transmit an operating voltage supplied from the voltage generating circuit 230 to the selected memory block.

[0046] The page buffer 250 can be connected through the bit line BL ( Figure 3 ) is coupled to the memory cell array 210. The page buffer 250 may precharge the bit line BL with a positive voltage in a program operation and a read operation, transmit data to a selected memory block, receive data from a selected memory block, or temporarily store the transmitted data in response to a page buffer control signal generated by the control circuit 220.

[0047] The column decoder 260 may transmit and receive data to and from the page buffer 250 , or transmit and receive data to and from the input / output circuit 270 .

[0048] The input / output circuit 270 can transmit data from an external device (e.g., Figure 1 The memory controller 100 transmits a command and an address received from the memory controller 100 to the control circuit 220, transmits data from an external device to the column decoder 260, or outputs data from the column decoder 260 to an external device through the input / output circuit 270.

[0049] The control circuit 220 may control the peripheral circuits in response to commands and addresses.

[0050] Figure 3is a circuit diagram showing a memory block of a semiconductor memory device according to an embodiment of the present invention. For example, Figure 3 The storage block can be Figure 2 Any one of the memory blocks 211 of the memory cell array 210 shown.

[0051] Reference Figure 3 , the memory block 211 may include a plurality of word lines WL0 to WLn-1, a drain select line DSL, and a source select line SSL coupled to the row decoder 240. These lines may be arranged in parallel with the plurality of word lines located between the DSL and the SSL.

[0052] The memory block 211 may further include a plurality of cell strings 221 connected to the bit lines BL0 to BLm-1, respectively. The cell string of each column may include one or more drain select transistors DST and one or more source select transistors SST. In the illustrated embodiment, each cell string has one DST and one SST. In the cell string, a plurality of memory cells or memory cell transistors MC0 to MCn-1 may be connected in series between the select transistors DST and SST. Each of the memory cells may be formed as a single-layer cell (SLC) storing one bit of data, a multi-layer cell (MLC) storing two bits of data, a triple-layer cell (TLC) storing three bits of data, or a quad-layer cell (QLC) storing four bits of data.

[0053] The source of the SST in each cell string can be connected to the common source line CSL, and the drain of each DST can be connected to the corresponding bit line. The gate of the SST in the cell string can be connected to the SSL, and the gate of the DST in the cell string can be connected to the DSL. The gates of the memory cells of different cell strings can be connected to their respective word lines. That is, the gate of the memory cell MC0 is connected to the corresponding word line WL0, the gate of the memory cell MC1 is connected to the corresponding word line WL1, and so on. A group of memory cells connected to a specific word line can be referred to as a physical page. Therefore, the number of physical pages in the storage block 211 can correspond to the number of word lines.

[0054] The page buffer array 250 may include a plurality of page buffers 251 coupled to the bit lines BL0 to BLm-1. The page buffers 251 may operate in response to a page buffer control signal. For example, the page buffers 251 may temporarily store data received through the bit lines BL0 to BLm-1 or sense a voltage or current of the bit lines during a read operation or a verification operation.

[0055] In some embodiments, the memory block 211 may include a NAND-type flash memory cell. However, the memory block 211 is not limited to this cell type, but may include a NOR-type flash memory cell. The memory cell array 210 may be implemented as a hybrid flash memory of a combination of two or more types of memory cells, or as a 1-NAND flash memory in which a controller is embedded inside a memory chip.

[0056] Figure 4 is a diagram showing a data processing system 2 according to an embodiment of the present invention.

[0057] Reference Figure 4 , the data processing system 2 may include a host device 5 and a memory system 10. The memory system 10 may include a memory controller 100 and a memory device 200. The memory controller 100 may include firmware (FW), which is a type of specific software for controlling various operations (e.g., read operations, write operations, and erase operations) of the memory device 200. In some embodiments, the firmware may reside in the storage device 110 and may be controlled by Figure 2 The control component 120 in runs.

[0058] The memory device 200 may include a plurality of memory cells (eg, NAND flash memory cells). The memory cells may be arranged in a manner such as Figure 3 The rows and columns shown are arranged in an array form. The cells in a particular row are connected to a word line (e.g., WL0), and the cells in a particular column are connected to a bit line (e.g., BL0). These word lines and bit lines are used for read operations and write operations. During a write operation, when the word line is specified (asserted), the data to be written ("1" or "0") is provided on the bit line. During a read operation, the word line is specified again, and the threshold voltage of each cell can then be obtained from the bit line. Multiple pages can share memory cells belonging to (i.e., connected to) the same word line.

[0059] In the memory system 10, the memory controller 100 can perform operations associated with one or more workloads from the host device 5. The workload can be defined as a data stream (or command) generated by an application of the host device 5, which is regarded as a collection of access patterns by the memory system 10. The memory system 10 can be implemented using a solid-state drive (SSD). According to the following report, by 2026, SSDs will occupy about 85% of enterprise storage capacity: David McIntyre "Annual Flash Controller Update", 2019 Flash Memory Summit Proceedings (Flash Memory Summit 2019, FMS'19). Compared with hard disk drives (HDDs), flash-based storage devices are also more advantageous in terms of cost per bit. However, due to the limited number of erase / write (EW) cycles and the complex logical to physical (L2P) mapping, the flash memory devices in SSDs are more sensitive to changes in input workloads. Therefore, workload analysis becomes important for improving the performance and reliability of flash memory devices (e.g., NAND flash memories).

[0060] The workload of an SSD is not limited to spatial characteristics (i.e., random or sequential access patterns), and therefore can be described using multiple features determined based on developer experience. For example, the workload analysis process can rely on multiple features, such as the number of commands and their duration, characteristics of access segments, input and output requests, intervals between requests, etc. However, even the most exhaustive feature sets may not take into account important workload characteristics. Moreover, certain features may be irrelevant to the flash translation layer (FTL) algorithm and should therefore be excluded. In order to address these issues, embodiments of the present invention are proposed. Therefore, embodiments provide a compact representation scheme for workloads in a memory system (e.g., an SSD (e.g., a NAND flash memory device)) that is capable of separating workloads by certain features.

[0061] According to an embodiment, Figure 4The memory controller 100 can provide a compiler (i.e., an encoder and a decoder) for a compact representation of an input workload and its encoding and decoding methods. Since the workload can be represented as a time series and further regarded as sequential data, a recurrent neural network should be able to efficiently process such data. Moreover, the internal representation of the workload should be compact so that it can be stored in a storage device (e.g., a volatile memory). Therefore, an embodiment provides a compact working representation based on a recurrent neural network model, and a recurrent neural network (RNN) autoencoder has been used as a model for representing the input workload, i.e., a compiler including an encoder and a decoder. The quality of the model depends on the diversity and size of the training set. The following describes the recurrent neural network model and recurrent neural network (RNN) autoencoder: T. Mikolov et al., "Linguistic Regularities in Continuous Space Word Representation", Proceedings of the 2013 Conference of the North American Chapter of Association for Computational Linguistics, pp. 746-751; and N. Srivastava et al., "Unsupervised Learning of Video Representations Using LSTMs", Proceedings of the 32nd international Conference on Machine Learning, PMLR 37: 843-852, 2015, the entire contents of each of which are incorporated herein by reference. The compact representation of the input workload facilitates FTL reconfiguration and more accurate FW parameter tuning. If the dataset used for training is extensive and diverse, embodiments will be able to detect many types of different workloads.

[0062] Figure 5 is a diagram illustrating a recurrent neural network compiler 500 according to an embodiment of the present invention. In some cases, the recurrent neural network compiler 500 can provide a compact workload representation and is referred to as a recurrent neural network (RNN) autoencoder.

[0063] Reference Figure 5, the recurrent neural network compiler 500 may include an encoder 510, a storage device 520, and a decoder 530. The encoder 510 may receive a signal including a signal transmitted by a memory device (e.g., Figure 4 The input workload is an input workload of multiple input commands associated with the memory device 200 of FIG. 1 . For example, the input workload is from Figure 4 In some embodiments, the workload may be represented as including N commands C 1 To C N Each command C i (where 1≤i≤N) can have at least two characteristics, namely command type (T i ), Logical Block Address (LBA i ). The features are not limited to the above two features. i is an integer corresponding to a possible command. For example, for a read command, T i The value can be 0. For a write command, T i The value can be 1. For the erase command, T i The value can be 2. LBA i Also an integer from a range of possible addresses. For example, LBA i Can range from 0 to 2 32 An integer of -1.

[0064] The encoder 510 may generate a compact representation vector corresponding to a plurality of input commands using a set of activation functions. The storage device 520 may be coupled to the encoder 510 and may store the compact representation vector. The decoder 530 may be coupled to the storage device 520 and may receive the compact representation vector from the storage device 520. The decoder 530 may generate a recovery command based on the compact representation vector and the plurality of input commands. FIG. 6A to FIG. 7 Details of the encoder 510, the storage device 520, and the decoder 530 are described.

[0065] like Fig. 6A As shown, each of the encoder 510 and the decoder 530 may utilize a first cycle block RB 1 To the Nth cycle block RB N The first cycle block RB 1 To the Nth cycle block RB N Cascade connection.

[0066] Figure 6B It is shown Fig. 6A Illustration of the general structure of the recurrent block in .

[0067] Reference Figure 6B , a cyclic block RB with index i (RB i) has two inputs and two outputs. The two inputs include the corresponding previous unit RB i-1 The hidden state vector h of the hidden state i-1 , and the input command vector X corresponding to the elements of the input sequence i The two outputs include corresponding to RB i The hidden state vector h of the hidden state i , and corresponding to RB i The output vector Y of the output values i In some embodiments, the vector X i The vector Y corresponds to the command of the input workload. i Corresponding to the recovery or prediction command of the input workload. Each RB shares three weight matrices W Y , W h and W X The output calculation of Rb is shown in Listing 1:

[0068] List 1:

[0069]

[0070] Referring to List 1, for the first input command X i and the first weight matrix W X Combination of X i W X Execute the activation function f(X i W X ) to generate the first vector. In some embodiments, the activation function includes a hyperbolic tangent function Sigmoid function and the rectified linear unit (ReLU) function f(x) = x + =one of max(0,x).

[0071] For the previous hidden state vector h i-1 and the second weight matrix W h The combination of the activation function f(h i-1 W h ) to generate a second vector. In some embodiments, the activation function here may be different from the previous activation function.

[0072] The sum of the first vector and the second vector f(X i W X )+f(h i-1 W h ) executes the activation function f(f(X i W X )+f(h i- 1 W h)) to generate the hidden state vector h i .

[0073] For the hidden state vector h i and the third weight matrix W Y The combination of activation function f(h i W Y ) can generate the output vector Y i =f(h i W Y ).

[0074] In some embodiments, the described RB can be implemented in different ways, including gated recurrent units (GRUs) and / or long short-term memories (LSTMs), as described in: J. Chung et al., "Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling," NIPS 2014 Workshop on Deep Learning, December 2014.

[0075] Figure 7 It shows Figure 5 An illustration of the details of the recurrent neural network compiler.

[0076] Reference Figure 7 , the recurrent neural network compiler may include an encoder 510, a storage device 520, and a decoder 530, and may be implemented using 2N recurrent blocks (RBs) to generate a compact representation vector R based on an input workload.

[0077] The encoder 510 can learn N commands C 1 -C N The internal representation vector R of the workload. In some embodiments, some workloads may be shorter or longer than N commands, so the workload should be padded (i.e., additional virtual commands are added to the workload) or pruned (i.e., the last command is removed from the workload). A weight matrix W may be used e X , W e h , W e Y The N recurrent blocks (RBs) of the proposed algorithm transform the workload into a compact representation R. Each RB can receive two inputs (i.e., the previous hidden state vector h from the previous recurrent block and the e i-1 , and command C i), and can output two values ​​(i.e., the command and the weighted hidden state vector h e i ). Matrix W e Y It may not be trainable because the encoder 510 accumulates information from the input workload and does not predict or recover arbitrary commands. Therefore, the output command is always equal to the zero vector.

[0078] exist Figure 7 In the example shown, the encoder 510 may include a plurality of cyclic coding blocks RB e 1 -RB e N Each cyclic encoding block is configured to receive one of multiple input commands. i , and generates a hidden state vector h corresponding to the received input command by applying a set of activation functions f to the received input command e i The last coded block RB among multiple cyclic coded blocks e N The final hidden state vector h can be generated e N as a compact representation vector R.

[0079] The first coded block RB e 1 is configured to: receive a first input command C among a plurality of input commands 1 ; For the first input command C 1 and the first weight matrix W e X The combination of performs the activation function f to generate the first vector f(C 1 W e X );For the initial hidden state vector h e 0 and the second weight matrix W e h The combination of performs an activation function to generate the second vector f(h e 0 W e h ); Perform an activation function on the sum of the first vector and the second vector to generate a first hidden state vector h e 1 ; and for the first hidden state vector h e 1 and the third weight matrix W e Y The combination of performs the activation function to generate the first output vector In some embodiments, the initial hidden state vector h e 0 The value can be 0, the first output vector The value of can be 0.

[0080] The second coding block RB e 2 is configured to: receive a second input command C among a plurality of input commands 2 ; For the second input command C 2 and the first weight matrix W e X The combination of performs the activation function f to generate the third vector f(C 2 W e X );For the first hidden state vector h e 1 and the second weight matrix W e h The combination of performs the activation function to generate the fourth vector f(h e 1 W e h ); Perform an activation function on the sum of the third vector and the fourth vector to generate a second hidden state vector h e 2 ; and for the second hidden state vector h e 2 and the third weight matrix W e Y The combination of performs the activation function to generate the second output vector In some embodiments, the second output vector The value of can be 0.

[0081] The remaining coding blocks among the plurality of cyclic coding blocks may perform the same operation as the first coding block RB e 1 and the second coding block RB e 2 Similar operation. The last coding block RB among multiple cyclic coding blocks e N The final hidden state vector h can be generated e N As the compact representation vector R. The compact representation vector R may be stored in the storage device 520. In some embodiments, the storage device 520 may be a volatile memory, such as a random access memory (RAM).

[0082] The decoder 530 may receive a compact representation vector R from the storage device 520 and decode the compact representation vector R based on the compact representation vector R and a plurality of input commands C 1 -C NGenerate recovery commands Decoder 530 can recover the workload command It may not be exactly the same as the original workload. The structure of the decoder 530 may be symmetrical with the encoder 510, but has a cascade connection structure in the opposite order in terms of input commands. The decoder 530 may utilize a weight matrix W d X , W d h , W d Y The N RB units are used to process the compact representation vector R, but in descending order from the last command to the first command. That is, multiple decoding blocks RB d 1 -RB d N The last decoded block RB d N It can be located at the first position of the decoder 530, the first decoding block RB d 1 may be located at the last position of the decoder 530. In some embodiments, the first input command C N can be the same as the zero vector, since the command should be recovered from the compact representation vector R. In some embodiments, the weight matrix W d X , W d h , W d Y Respectively with the weight matrix W e X , W e h , W e Y Same or different.

[0083] exist Figure 7 In the example shown, the last decoded block RB d N Configured to: receive multiple input commands C 1 -C N The last input command C N and compact representation vector R; for the last input command C N and the first weight matrix W d X The combination of performs the activation function f to generate the first vector f(C N W d X ); for the input hidden state vector h e N A compact representation vector R and a second weight matrix W dh The combination of performs an activation function to generate the second vector f(h e N W d h ); Execute the activation function on the sum of the first vector and the second vector to generate the final hidden state vector h d N ; and for the final hidden state vector h d N and the third weight matrix W d Y The combination of performs the activation function to generate the final output vector

[0084] The second decoding block RB d 2 Configured to: receive multiple input commands C 1 -C N The second input command C 2 ; For the second input command C 2 and the first weight matrix W d X The combination of performs the activation function f to generate the third vector f(C 2 W d X );From the third decoding block RB d 3 The received second hidden state vector h d 3 and the second weight matrix W d h The combination of performs the activation function to generate the fourth vector f(h d 3 W d h ); Perform an activation function on the sum of the third vector and the fourth vector to generate a second hidden state vector h d 2 ; and performing an activation function on the combination of the second hidden state vector and the third weight matrix to generate a second output vector

[0085] The first decoding block RB d 1 Configured to: receive multiple input commands C 1 -C N The first input command C 1 ; For the first input command C 1 and the first weight matrix W d X The combination of performs the activation function f to generate the third vector f(C 1 W dX );From the second decoding block RB d 2 The received second hidden state vector h d 2 and the second weight matrix W d h The combination of performs the activation function to generate the fourth vector f(h d 2 W d h ); Execute an activation function on the sum of the third vector and the fourth vector to generate a first hidden state vector h d 1 ; and performing an activation function on the combination of the first hidden state vector and the third weight matrix to generate a first output vector

[0086] Figure 7 The RNN compiler model in can be trained using a dataset containing M workloads, which can have different characteristics. The training process can adjust the weight matrix W e X , W e h , W e Y , W d X , W d h , W d Y , so that the source workload (C 1 , C 2 , ..., C N ) and recovered workloads The difference between them is minimized. Different optimization algorithms such as gradient descent, RMSProp, Adam, etc. can be used during training. The model can have two hyperparameters N and d. N represents the number of RBs in the encoder and decoder, and d represents the dimension of the target compact workload representation vector R.

[0087] Figure 8 is a diagram showing a recurrent neural network compiler 500 according to an embodiment of the present invention.

[0088] Reference Figure 8 ,Apart from Figure 5 In addition to the encoder 510, storage device 520 and decoder 530 in the cyclic neural network compiler 500, the cyclic neural network compiler 500 may also include a predictor 540. The predictor 540 may predict the next K commands after the input workload and generate predicted commands.

[0089] Fig. 9 It shows Figure 8 The encoder 510 and decoder 530 in the recurrent neural network compiler 500 may have the same Figure 7 The same structure as shown.

[0090] Reference Fig. 9 , the predictor 540 may include a plurality of cyclic prediction blocks RB p 1 -RB p K The predictor 540 may receive the compact representation vector R and, based on the compact representation vector R and the last input command C among the plurality of input commands, N Generate prediction command

[0091] The plurality of cyclic prediction blocks may include a first prediction block RB p 1 To the last prediction block RB p K , they have a cascade connection structure in ascending order (ie, from the first prediction block to the last prediction block).

[0092] The first prediction block RB p 1 is configured to: receive the last input command C among the plurality of input commands from the storage device 520 N and compact representation vector R; for the last input command and the first weight matrix W p X The combination of performs the activation function f to generate the first vector For the compact representation vector R(=h e N ) and the second weight matrix W p h The combination of performs an activation function to generate the second vector f(h e N W p h ); Perform an activation function on the sum of the first vector and the second vector to generate a first hidden state vector h p 1 ; and for the first hidden state vector h p 1 and the third weight matrix W p Y The combination of performs the activation function to generate the first output vector As the first predicted command after the last input command.

[0093] The second prediction block RB p 2is configured to: receive a first prediction command and the first hidden state vector h p 1 ; For the first prediction command and the first weight matrix W p X The combination of performs the activation function f to generate the third vector For the first hidden state vector h p 1 and the second weight matrix W p h The combination of performs the activation function to generate the fourth vector f(h p 1 W p h ); Perform an activation function on the sum of the third vector and the fourth vector to generate a second hidden state vector h p 2 ; and for the second hidden state vector h p 2 and the third weight matrix W p Y The combination of performs the activation function to generate the second output vector As the first prediction command After the second prediction command.

[0094] The remaining prediction blocks among the plurality of loop prediction blocks may perform the same operation as the first prediction block RB p 1 and the second prediction block RB p 2 Similar operation. The last prediction block RB among multiple loop prediction blocks p K The final hidden state vector h can be generated p K .

[0095] Therefore, predictor 540 requires an additional K RBs and a weight matrix W p X , W p h , W p Y In some embodiments, the weight matrix Wp X , W p h , W p Y Respectively with the weight matrix W e X 、We h , W e Y The output of predictor 540 is the input workload C1 -C N The next K commands after the command Each predicted command is fed into the next RB as an input command. Therefore, the next K commands after the input workload are predicted and generated.

[0096] An example of an implementation of a recurrent neural network (RNN) compiler according to an embodiment is described below.

[0097] Notably, a recurrent neural network (RNN) compiler for compact workload representation has been implemented using Long Short-Term Memory (LSTM) recurrent blocks and tested on datasets containing M=900 synthetic workloads, N=10,000 commands each.

[0098] The dataset has 9 types of workloads (each type has 100 samples), which are generated based on the following two parameters: queue depth (QD) and read / write ratio (RWR), where RWR represents the ratio of read to write commands in the workload. All workloads are random, and the 9 workload types are shown in Listing 2:

[0099] List 2:

[0100]

[0101] The workloads in the dataset have been represented as d = 25 dimensional vectors. These vectors have been processed by a dimensionality reduction algorithm (e.g., the t-SNE algorithm) in order to visualize them in three-dimensional (3-D) space. An implementation of the t-SNE algorithm is described below: Laurens van der Maaten and Geoffrey Hinton, "Visualizing Data using t-SNE," Journal of Machine Learning Research 9, pp. 2579-2605, 2008. For clarity, only 90 points are plotted (10 for each workload type) because the density of each workload type is high and the additional 810 points would hinder the graph separation between the points. The dataset of workloads compacted and visualized by the recurrent neural network compiler is shown in Fig.10 shown.

[0102] Listing 3 shows an example of a workload being converted into a 25-dimensional vector (ie, a compact representation vector):

[0103] List 3:

[0104]

[0105]

[0106] The average Euclidean distance between the centers of each workload type is shown in Table 1:

[0107] Table 1:

[0108] W1 <![CDATA[W 2 ]]> <![CDATA[W 3 ]]> <![CDATA[W 4 ]]> <![CDATA[W 5 ]]> <![CDATA[W 6 ]]> <![CDATA[W 7 ]]> <![CDATA[W 8 ]]> <![CDATA[W 9 ]]> W1 0.0000 0.7894 0.5607 0.0005 0.7894 0.5908 0.0003 0.7893 0.5800 W2 0.7894 0.0000 0.2307 0.7894 0.0002 0.2002 0.7894 0.0002 0.2109 W3 0.5607 0.2307 0.0000 0.5607 0.2307 0.0305 0.5607 0.2307 0.0199 W4 0.0005 0.7894 0.5607 0.0000 0.7894 0.5908 0.0002 0.7894 0.5800 W5 0.7894 0.0002 0.2307 0.7894 0.0000 0.2002 0.7894 0.0001 0.2109 W6 0.5908 0.2002 0.0305 0.5908 0.2002 0.0000 0.5908 0.2002 0.0108 W7 0.0003 0.7894 0.5607 0.0002 0.7894 0.5908 0.0000 0.7894 0.5800 W8 0.7893 0.0002 0.2307 0.7894 0.0001 0.2002 0.7894 0.0000 0.2102 W9 0.5800 0.2109 0.0199 0.5800 0.2109 0.0108 0.5800 0.2102 0.0000

[0109] like Fig.10 As shown in Table 1, most workloads are separated by a certain read / write ratio (RWR). 1 , W 4 , W 7 ) is separated by RWR = 0 / 100, (W 2 , W 5 , W 8 ) is separated by RWR=100 / 0, and (W 3 , W 6 , W 9 ) is separated by RWR=70 / 30. The last group (W 3 , W 6 , W 9 ) is more diverse, and the influence of the queue depth feature is greater. Therefore, it can be seen that the compact workload representation scheme according to the embodiment can separate the workloads by important features and estimate the differences between the input workloads in the NAND flash memory device.

[0110] As described above, embodiments provide a compact workload representation scheme to separate workloads by important features and estimate the differences between input workloads. Embodiments can facilitate FTL reconfiguration and more accurate FW parameter tuning. If the dataset used for training is extensive and diverse, embodiments will be able to detect multiple types of different workloads.

[0111] Although the foregoing embodiments have been shown and described in detail for the purpose of clarity and understanding, the present invention is not limited to the details provided. As will be appreciated by those skilled in the art based on the foregoing disclosure, there are many alternatives to implement the present invention. Therefore, the disclosed embodiments are exemplary, not restrictive. The present invention is intended to encompass all modifications and alternatives that fall within the scope of the appended claims.

Claims

1. A memory system, include: Memory device; as well as a controller, the controller comprising a recurrent neural network compiler, the recurrent neural network compiler comprising an encoder having a plurality of recurrent encoding blocks, the plurality of recurrent encoding blocks comprising a first encoding block to a last encoding block, Each loop encodes a block: receiving one of a plurality of input commands in an input workload associated with the memory device; as well as generating a hidden state vector corresponding to the received input command by applying a set of activation functions to the received input command, and The last encoding block generates the final hidden state vector as a compact representation vector, Wherein, the recurrent neural network compiler further comprises: a storage device coupled to the last encoded block and storing the compact representation vector; and a decoder comprising a plurality of cyclic decoding blocks and receiving the compact representation vector and generating a recovery command based on the compact representation vector and the plurality of input commands, Wherein, the recurrent neural network compiler further comprises: a predictor comprising a plurality of cyclic prediction blocks and receiving the compact representation vector and generating a prediction command based on the compact representation vector and a last input command among the plurality of input commands, The plurality of cyclic prediction blocks include a first prediction block to a last prediction block, and the first prediction block to the last prediction block have an ascending cascade connection structure.

2. The memory system according to claim 1, in, The activation function among the set of activation functions includes one of a hyperbolic tangent function, a sigmoid function, and a rectified linear unit function.

3. The memory system according to claim 1, in, The first coding block to the last coding block are cascade-connected.

4. The memory system according to claim 3, in, The first coding block: receiving a first input command among the plurality of input commands; executing the activation function on a combination of the first input command and a first weight matrix to generate a first vector; executing the activation function on a combination of the initial hidden state vector and the second weight matrix to generate a second vector; performing the activation function on the sum of the first vector and the second vector to generate a first hidden state vector; and The activation function is performed on a combination of the first hidden state vector and a third weight matrix to generate a first output vector.

5. The memory system according to claim 4, in, A second coding block among the plurality of cyclic coding blocks: receiving a second input command among the plurality of input commands; executing the activation function on a combination of the second input command and the first weight matrix to generate a third vector; performing the activation function on a combination of the first hidden state vector and the second weight matrix to generate a fourth vector; executing the activation function on the sum of the third vector and the fourth vector to generate a second hidden state vector; and The activation function is performed on a combination of the second hidden state vector and the third weight matrix to generate a second output vector.

6. The memory system according to claim 5, in, The plurality of cyclic decoding blocks include a first decoding block to a last decoding block, and the first decoding block to the last decoding block are symmetrical to the first encoding block to the last encoding block and have a cascade connection structure in descending order.

7. The memory system according to claim 6, in, The last decoded block, receiving a last input command among the plurality of input commands and the compact representation vector; executing the activation function on a combination of the last input command and a first weight matrix to generate a first vector; performing the activation function on a combination of the compact representation vector and a second weight matrix to generate a second vector; performing the activation function on the sum of the first vector and the second vector to generate a final hidden state vector; and The activation function is performed on a combination of the final hidden state vector and a third weight matrix to generate a final output vector.

8. The memory system according to claim 7, in, The first decoding block: receiving a first input command among the plurality of input commands; executing the activation function on a combination of the first input command and the first weight matrix to generate a third vector; performing the activation function on a combination of a second hidden state vector received from a second decoding block and the second weight matrix to generate a fourth vector; executing the activation function on the sum of the third vector and the fourth vector to generate a first hidden state vector; and The activation function is performed on a combination of the first hidden state vector and the third weight matrix to generate a first output vector.

9. The memory system according to claim 7, in, The first weight matrix, the second weight matrix, and the third weight matrix are trained so that differences between the plurality of input commands and restoration commands are minimized.

10. The memory system according to claim 1, in, The first prediction block: receiving a last input command among the plurality of input commands and the compact representation vector; executing the activation function on a combination of the last input command and a first weight matrix to generate a first vector; performing the activation function on a combination of the compact representation vector and a second weight matrix to generate a second vector; performing the activation function on the sum of the first vector and the second vector to generate a first hidden state vector; and The activation function is performed on a combination of the first hidden state vector and a third weight matrix to generate a first output vector as a first predicted command after the last input command.

11. The memory system according to claim 10, in, A second prediction block among the plurality of cyclic prediction blocks: Receiving the first prediction command and the first hidden state vector; executing the activation function on a combination of the first prediction command and the first weight matrix to generate a third vector; performing the activation function on a combination of the first hidden state vector and the second weight matrix to generate a fourth vector; executing the activation function on the sum of the third vector and the fourth vector to generate a second hidden state vector; and The activation function is performed on a combination of the second hidden state vector and the third weight matrix to generate a second output vector as a second prediction command subsequent to the first prediction command.

12. The memory system according to claim 1, in, The loop encoding block, the loop decoding block and the loop prediction block are implemented in different ways.

13. A method for operating a controller of a memory system, the method include: providing a recurrent neural network compiler, the recurrent neural network compiler comprising an encoder having a plurality of recurrent encoding blocks; Each loop encoding block receives one of a plurality of input commands in an input workload associated with a memory device of the memory system; and Each recurrent encoding block generates a hidden state vector corresponding to the received input command by applying a set of activation functions to the received input command. wherein a last encoding block among the plurality of cyclic encoding blocks generates a final hidden state vector as a compact representation vector corresponding to the plurality of input commands, Further including: providing a storage device and a decoder, the storage device being coupled to the last encoded block and storing the compact representation vector, the decoder comprising a plurality of loop decoding blocks, and receiving the compact representation vector and generating a recovery command based on the compact representation vector and the plurality of input commands, Further including: providing a predictor comprising a plurality of loop prediction blocks and receiving the compact representation vector and generating a prediction command based on the compact representation vector and a last input command among the plurality of input commands, The plurality of cyclic prediction blocks include a first prediction block to a last prediction block, and the first prediction block to the last prediction block have an ascending cascade connection structure.

14. The method according to claim 13, in, The activation function among the set of activation functions includes one of a hyperbolic tangent function, a sigmoid function, and a rectified linear unit function, and The first coding block to the last coding block of the plurality of cyclic coding blocks are cascade-connected.

15. The method according to claim 13, in, The plurality of cyclic decoding blocks include a first decoding block to a last decoding block, which are symmetrical to the first encoding block to the last encoding block of the plurality of cyclic encoding blocks and have a cascade connection structure in descending order.

16. The method according to claim 13, in, Generating a hidden state vector corresponding to the received input command includes performing the activation function on a combination of the received input command and one or more set matrices, wherein the one or more set matrices are trained to minimize the difference between the multiple input commands and the restored command.

17. The method according to claim 13, The loop encoding block, the loop decoding block and the loop prediction block are implemented in different ways.