Operation method of storage device configured to support multi-streams
By employing a machine learning-based stream manager to classify data streams, the storage device enhances performance and lifespan by optimizing data management in semiconductor memory devices.
Patent Information
- Application Number
- KR1020210032001
- Authority / Receiving Office
- KR · KR
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2021-03-11
- Publication Date
- 2026-11-06
- Estimated Expiration
- 2041-03-11
AI Technical Summary
Existing semiconductor memory devices, such as SSDs, face challenges in improving operational performance and lifespan due to inadequate management of data streams without host control.
A storage device employs machine learning-based methods to classify and manage data into multiple streams by using a stream manager that utilizes multiple machine learning models to determine stream identifiers, reducing maintenance operations and enhancing performance and lifespan.
The method improves storage device performance by effectively managing data streams, reducing maintenance operations, and extending the lifespan of non-volatile memory devices.
Smart Images

Figure 112021028834866-PAT00006_ABST
Abstract
Description
Technology Field
[0001] The present invention relates to a semiconductor memory, and more specifically, to a method of operation of a storage device configured to support multiple streams. Background Technology
[0002] Semiconductor memory is classified into volatile memory devices, such as SRAM and DRAM, in which stored data is lost when the power supply is cut off, and non-volatile memory devices, such as flash memory devices, PRAM, MRAM, RRAM, and FRAM, which retain stored data even when the power supply is cut off.
[0003] Recently, flash memory-based solid-state drives (SSDs) are widely used as mass storage media for computing systems. Hosts using SSDs can generate various types of data depending on the application. To improve the operational performance of the SSD, the host can provide information about the data to the storage device along with the data. However, if the host does not control the information, problems may arise in improving the performance and lifespan of the SSD. The problem to be solved
[0004] The objective of the present invention is to provide a method of operation for a storage device configured to support multiple streams with enhanced performance and enhanced lifespan by separating data from a host into stream units. means of solving the problem
[0005] According to an embodiment of the present invention, a method of operation of a storage device supporting multiple streams comprises: receiving an input / output request from an external host; generating a plurality of stream identifier candidates by performing machine learning on the input / output request based on a plurality of machine learning models; generating a model ratio based on the characteristics of the input / output request; assigning a final stream identifier for the input / output request by reflecting the model ratio on the plurality of stream identifier candidates; and storing write data corresponding to the input / output request in a non-volatile memory device of the storage device based on the final stream identifier, wherein each of the plurality of machine learning models is based on a different machine learning algorithm.
[0006] According to an embodiment of the present invention, a method of operation of a storage device supporting multiple streams comprises: receiving a sequential input / output request from an external host; performing machine learning on the sequential input / output request based on first and second machine learning models to generate first and second result values; generating a first model ratio based on the characteristics of the sequential input / output request; generating a first stream identifier by reflecting the first model ratio on the first and second result values; assigning the first stream identifier to the sequential input / output request; receiving a random input / output request from the external host; performing machine learning on the random input / output request based on first and second machine learning models to generate third and fourth result values; generating a second model ratio different from the first model ratio based on the characteristics of the random input / output request; generating a second stream identifier by reflecting the second model ratio on the third and fourth result values; and assigning the second stream identifier to the random input / output request.
[0007] According to an embodiment of the present invention, a method of operation of a storage device supporting multiple streams comprises: receiving an input / output request from an external host; performing machine learning on the input / output request based on a plurality of machine learning models to generate a plurality of stream identifier candidates; performing inference on the plurality of stream identifier candidates based on an ensemble model to generate a final stream identifier; and storing write data corresponding to the input / output request in a non-volatile memory device of the storage device based on the final stream identifier, wherein each of the plurality of machine learning models is based on a different machine learning algorithm. Effects of the invention
[0008] According to the present invention, a storage device can allocate a stream for a write request received from a host based on a plurality of machine learning models. At this time, the storage device can determine a reflection ratio for each of the plurality of machine learning models based on the characteristics of the received write request. Accordingly, since write requests having similar characteristics are managed as the same stream, a method of operation of a storage device configured to support multiple streams with enhanced performance is provided. Brief explanation of the drawing
[0009] FIG. 1 is a block diagram illustrating an exemplary storage system according to an embodiment of the present invention. Figure 2 is a diagram illustrating the software layer of the storage system of Figure 1 in an exemplary manner. FIG. 3 is a block diagram exemplarily showing the non-volatile memory device of FIG. 1. Figure 4 is a diagram illustrating a stream managed in a storage device. Figure 5 is a diagram showing the stream manager of Figure 1. Figure 6 is a block diagram showing an example of the model ratio calculator of Figure 5. FIGS. 7 to 9 are drawings for explaining the operation of the stream manager of FIG. 5. Figure 10 is a flowchart showing the operation of the storage device of Figure 1. Figure 11 is a diagram showing the stream manager of Figure 1. FIG. 12 is a diagram illustrating an exemplary storage system to which a storage device according to the present invention is applied. FIG. 13 is a block diagram illustrating an exemplary storage system according to an embodiment of the present invention. FIG. 14 is a block diagram exemplarily showing a data center to which a storage device according to an embodiment of the present invention is applied. Specific details for implementing the invention
[0010] In the following, embodiments of the present invention will be described clearly and in detail so that a person skilled in the art can easily practice the present invention.
[0011] Components described by reference to terms such as part or unit, module, engine, etc. used in the detailed description and functional blocks illustrated in the drawings may be implemented in the form of software, hardware, or a combination thereof. For example, software may be machine code, firmware, embedded code, and application software. For example, hardware may include electrical circuits, electronic circuits, processors, computers, integrated circuits, integrated circuit cores, pressure sensors, inertial sensors, MEMS (microelectromechanical systems), passive components, or a combination thereof.
[0012] Furthermore, unless otherwise defined, all terms used herein, including those with a technical or scientific meaning, shall have a meaning that can be understood by those skilled in the art to which this invention pertains. Generally, terms defined in dictionaries shall be interpreted to have a meaning equivalent to their contextual meaning in the relevant technical field, and shall not be interpreted to have an ideal or overly formal meaning unless explicitly defined in the text.
[0013] FIG. 1 is a block diagram illustrating an exemplary storage system according to an embodiment of the present invention. Referring to FIG. 1, the storage system (10) may include a host (11) and a storage device (100). In one embodiment, the storage system (10) may be one of information processing devices configured to process various information and store the processed information, such as a personal computer, laptop, server, workstation, smartphone, tablet PC, digital camera, black box, etc.
[0014] The host (11) can control the general operations of the storage system (100). For example, the host (11) can send a request (RQ) to the storage device (100) to store data (DATA) in the storage device (100) or to read data (DATA) stored in the storage device (100). In one embodiment, the host (11) may be a processor core, such as a central processing unit (CPU) or an application processor (AP), configured to control the storage system (10), or a computing node connected via a network.
[0015] In one embodiment, the host (11) may include a host controller (12) and a host memory (13). The host controller (12) may be a device configured to control the general operations of the host (11) or to control the storage device (100) on the host (11) side. The host memory (13) may be a buffer memory, cache memory, or operation memory used by the host (11).
[0016] The storage device (100) may operate under the control of the host (11). The storage device (100) may include a storage controller (110) and a non-volatile memory device (120). The storage controller (110) may store data in the non-volatile memory device (120) or read data stored in the non-volatile memory device (120) under the control of the host (11). In one embodiment, the storage controller (110) may perform various management operations to efficiently use the non-volatile memory device (120).
[0017] The storage controller (110) may include a central processing unit (CPU) (111), a flash translation layer (FTL) (112), an error correction code (ECC) engine (113), an advanced encryption standard (AES) engine (114), a buffer memory (115), a host interface circuit (116), a memory interface circuit (117), and a stream manager (118).
[0018] The CPU (111) can control the general operations of the storage controller (110). The FTL (112) can perform various operations to efficiently use the non-volatile memory device (120). For example, the host (11) can manage the storage space of the storage device (100) as a logical address. The FTL (112) can be configured to manage address mapping between the logical address from the host (11) and the physical address of the storage device (100). The FTL (112) can perform wear leveling operations to prevent excessive degradation of specific memory blocks among the memory blocks of the non-volatile memory device (120). The lifespan of the non-volatile memory device (120) can be improved by the wear leveling operations of the FTL (112). The FTL (112) can perform garbage collection on the non-volatile memory device (120) to secure free memory blocks.
[0019] In one embodiment, the FTL (112) may be implemented in software or hardware form. If the FTL (112) is implemented in software form, program code or information related to the FTL (112) may be stored in a buffer memory (115) and executed by the CPU (111). If the FTL (112) is implemented in hardware form, a hardware accelerator configured to perform the operation of the FTL (112) may be provided separately from the CPU (112).
[0020] The ECC engine (113) can perform error detection and error correction on data read from the non-volatile memory device (120). For example, the ECC engine (113) can generate an error correction code (or parity bit) for data to be written to the non-volatile memory device (120). The generated error correction code (or parity bit) can be stored in the non-volatile memory device (120) along with the data to be written. Subsequently, when the data written from the non-volatile memory device (120) is read out, the ECC engine (113) can detect and correct errors in the read-out data based on the read-out data and the corresponding error correction code (or corresponding parity bit).
[0021] The AES engine (114) can perform an encryption operation or a decryption operation on data received from a host (11) or a non-volatile memory device (120). In one embodiment, the encryption operation or the decryption operation may be performed based on a symmetric-key algorithm.
[0022] The buffer memory (115) may be a write buffer or a read buffer configured to temporarily store data input to the storage controller (110). Alternatively, the buffer memory (115) may be configured to store various information necessary for the operation of the storage controller (110). For example, the buffer memory (115) may store a mapping table managed by the FTL (112). Alternatively, the buffer memory (115) may store software, firmware, or information related to the FTL (112).
[0023] In one embodiment, the buffer memory (115) may be an SRAM, but the scope of the present invention is not limited thereto, and the buffer memory (115) may be implemented as various types of memory devices such as DRAM, MRAM, PRAM, etc. For the sake of brevity of the drawings and convenience of explanation, the buffer memory (115) is shown in FIG. 1 as being included in the storage controller (110), but the scope of the present invention is not limited thereto. The buffer memory (115) may be located outside the storage controller (110), and the storage controller (110) may communicate with the buffer memory through a separate communication channel or interface.
[0024] The host interface circuit (116) may be configured to communicate with the host (11) according to a predetermined interface protocol. In one embodiment, the predetermined interface protocol may include at least one of various interface protocols such as an ATA (Advanced Technology Attachment) interface, a SATA (Serial ATA) interface, an e-SATA (external SATA) interface, a SCSI (Small Computer Small Interface) interface, a SAS (Serial Attached SCSI) interface, a PCI (Peripheral Component Interconnection) interface, a PCIe (PCI express) interface, an NVMe (NVM express) interface, IEEE 1394, a USB (universal serial bus) interface, an SD (secure digital) card, a MMC (multi-media card) interface, an eMMC (embedded multi-media card) interface, a UFS (Universal Flash Storage) interface, an eUFS (embedded Universal Flash Storage) interface, a CF (compact flash) card interface, or a network interface. The host interface circuit (116) can receive a signal based on a predetermined interface protocol from the host (11) and operate based on the received signal. Alternatively, the host interface circuit (116) can transmit a signal based on a predetermined interface protocol to the host (11).
[0025] The memory interface circuit (117) may be configured to communicate with the non-volatile memory device (120) according to a predetermined interface protocol. In one embodiment, the predetermined interface protocol may include at least one of various interface protocols, such as a toggle interface, an ONFI interface, etc. In one embodiment, the memory interface circuit (117) may communicate with the non-volatile memory device (120) based on a toggle interface. In this case, the memory interface circuit (117) may communicate with the non-volatile memory device (120) through a plurality of channels (CHs). In one embodiment, each of the plurality of channels (CHs) may include a plurality of signal lines configured to transmit various control signals (e.g., / CE, CLE, ALE, / WE, / RE, R / B, etc.), data signals (DQ), and data strobe signals (DQS).
[0026] The stream manager (118) can separate data or input / output requests received from the host (11) into multiple streams or manage them on a stream basis. For example, data received from the host (11) may have various attributes or various characteristics. In this case, if the data is stored in the non-volatile memory device (120) without separate classification of the data, maintenance operations (e.g., garbage collection) on the non-volatile memory device (120) may occur frequently. As an example, frequent maintenance operations on the non-volatile memory device (120) may increase the write amplification factor (WAF) characteristics of the storage device (100) (i.e., decrease performance).
[0027] A stream manager (118) of a storage controller (110) according to an embodiment of the present invention can classify data into stream units based on various attributes of data received from a host (11) (e.g., logical block address, data size, data frequency, etc.) and can store data in a non-volatile memory device (120) based on the classified streams. In this case, since data having similar characteristics is classified into the same stream, the number of maintenance operations for the non-volatile memory device (120) can be reduced. Therefore, overall performance degradation of the storage device (100) can be prevented. In one embodiment, the stream manager (118) according to the present invention can perform stream allocation for data based on a plurality of machine learning models. Each of the plurality of machine learning models can perform stream classification for data based on different machine learning algorithms, and the stream manager (118) can determine the final stream for data by combining results from the plurality of machine learning models based on the characteristics of the data. The structure and operation of the stream manager (118) according to the present invention will be described in more detail with reference to the drawings below.
[0028] As described above, according to an embodiment of the present invention, the storage device (100) can classify and manage data into multiple streams based on various attribute information of the data, even without information about the characteristics of the data (e.g., explicit stream designation) from the host (11). In this case, since the number of maintenance operations for the non-volatile memory device (120), such as garbage collection, is reduced, the storage device (100) having improved performance and improved lifespan is provided.
[0029] FIG. 2 is a diagram illustrating the software layer of the storage system of FIG. 1 as an example. Referring to FIG. 1 and FIG. 2, the software layer of the storage system (10) may include an application layer (APP), a file system layer (FS), a device driver layer (DD), a stream manager (118), and a flash conversion layer (112).
[0030] The application layer (APP) may include various applications running on the host (11). The file system (FS) may be configured to organize files or data used by the application layer (APP). For example, the file system (FS) may manage the storage space of the storage device (100) as logical block addresses (LBA). The file system (FS) may assign and manage logical block addresses to data stored in the storage device (100). In one embodiment, the file system (FS) may have a different form depending on the operating system of the host (11). The file system (FS) may include at least one of various file system types such as FAT (File Allocation Table), FAT32, NTFS (NT File System), HFS (Hierarchical File System), JSF2 (Journaled File System2), XFS, ODS-5 (On-Disk Structure-5), UDF, ZFS, UFS (Unix File System), ext2, ext3, ext4, ReiserFS, Reiser4, ISO 9660, Gnome VFS, BFS, WinFS, etc. The device driver (DD) may perform an operation of converting information from the file system layer (FS) or the application layer (APP) into information recognizable by the storage device (100). In one embodiment, the application layer (APP), the file system layer (FS), and the device driver (DD) may be implemented in software form and may run on the host (11).
[0031] The stream manager (118) may be configured to allocate a stream for a request received from the host (11). In one embodiment, the stream manager (118) may be configured to allocate a stream for data based on machine learning.
[0032] The FTL (112) may be configured to convert the logical block address of a request received from the host (11) into a physical block address (or physical address) used in the non-volatile memory device (120). In one embodiment, the FTL (112) may receive information about a stream identifier assigned from the stream manager (118) and perform address mapping so that the request received from the host (11) is stored in the non-volatile memory device (120) according to the assigned stream.
[0033] FIG. 3 is a block diagram illustrating an exemplary non-volatile memory device of FIG. 1. FIG. 4 is a diagram illustrating a stream managed by a storage device. Referring to FIG. 1, FIG. 3, and FIG. 4, the non-volatile memory device (120) may include a plurality of non-volatile memories (NVM11 to NVM44). Each of the plurality of non-volatile memories (NVM11 to NVM44) may be implemented as a single semiconductor chip, a single semiconductor die, or a single semiconductor package.
[0034] The non-volatile memory (NVM11) may include a plurality of planes (PL1, PL2) and peripheral circuits (PERI). Each of the plurality of planes (PL1, PL2) may include a plurality of memory blocks (BLK11~BLK14, BLK21~BLK24). Each of the plurality of memory blocks (BLK11~BLK14, BLK21~BLK24) may include a plurality of pages. In one embodiment, a plurality of memory blocks (e.g., BLK11~BLK14) included in the same plane (e.g., PL1) may be configured to share the same bitline, but the scope of the present invention is not limited thereto.
[0035] Peripheral circuitry (PERI) of the non-volatile memory (NVM11) may be connected to a corresponding channel (e.g., CH1) among multiple channels (CH1 to CH4). In response to various signals received through the corresponding channel, the peripheral circuitry (PERI) may store data received through the corresponding channel in multiple planes (PL1, PL2) or output data stored in the multiple planes (PL1, PL2) through the corresponding channel. For the above-described operation, the peripheral circuitry (PERI) may include various components such as an address decoder, a voltage generator, a page buffer circuit, an input / output circuit, and a control logic circuit.
[0036] For the sake of brevity in the drawings, one nonvolatile memory (NVM11) is depicted as comprising two planes (PL2) and one plane as comprising four memory blocks; however, the scope of the invention is not limited thereto, and the number of planes, the number of memory blocks, or the number of pages can be varied. In one embodiment, the remaining nonvolatile memories (NVM12 to NVM44) may have a structure similar to the nonvolatile memory (NVM11) described above, and a detailed description thereof is omitted.
[0037] Among the plurality of non-volatile memories (NVM11~NVM44), a first portion (NVM11, NVM12, NVM13, NVM14) can communicate with the storage controller (1100) through a first channel (CH1), a second portion (NVM21, NVM22, NVM23, NVM24) can communicate with the storage controller (1100) through a second channel (CH2), a third portion (NVM31, NVM32, NVM33, NVM34) can communicate with the storage controller (1100) through a third channel (CH3), and a fourth portion (NVM41, NVM42, NVM43, NVM44) can communicate with the storage controller (1100) through a fourth channel (CH4). Among the plurality of non-volatile memories (NVM11~NVM44), a first portion (NVM11, NVM21, NVM31, NVM41) can form a first way (WAY1), a second portion (NVM12, NVM22, NVM32, NVM42) can form a second way (WAY2), a third portion (NVM13, NVM23, NVM33, NVM43) can form a third way (WAY3), and a fourth portion (NVM14, NVM24, NVM34, NVM44) can form a fourth way (WAY4). That is, the non-volatile memory device (1200) can have a multi-way / multi-channel structure, and it will be understood that the scope of the present invention is not limited to the structure shown in FIG. 3.
[0038] In one embodiment, the storage device (1000) can manage a plurality of memory blocks included in a non-volatile memory device (1200) based on a plurality of streams. For example, as shown in FIG. 4, the storage device (1100) can manage a first memory block (BLK1) among a plurality of memory blocks as a stream corresponding to a first stream identifier (SID1), a second memory block (BLK2) as a stream corresponding to a second stream identifier (SID2), a third memory block (BLK3) as a stream corresponding to a third stream identifier (SID3), and a first memory block (BLK4) as a stream corresponding to a fourth stream identifier (SID4).
[0039] In one embodiment, memory blocks (e.g., first memory blocks (BLK1)) corresponding to the same stream identifier (e.g., SID1) may be contained in the same plane, contained in the same non-volatile memory, contained in non-volatile memories connected to the same channel, or contained in non-volatile memories contained in the same way. Alternatively, memory blocks (e.g., first memory blocks (BLK1)) corresponding to the stream identifier (e.g., SID1) may be distributed among a plurality of non-volatile memories. However, the foregoing is merely an example, and the scope of the present invention is not limited thereto.
[0040] As described above, the stream manager (118) can manage data having similar characteristics as the same stream, and accordingly, data having similar characteristics will be stored in memory blocks corresponding to the same stream. In this case, since the data stored in the same stream has similar characteristics, performance degradation due to maintenance operations (e.g., garbage collection) of the storage device (100) may be reduced or the write amplification factor (WAF) may be reduced.
[0041] FIG. 5 is a diagram showing the stream manager of FIG. 1. In the following, to facilitate the explanation of embodiments of the present invention, it is assumed that a storage device (100) receives a write request (RQ) from a host (11) and assigns a stream identifier (SID) based on the received write request. At this time, the write request (RQ) may not include separate information regarding the attributes or characteristics of the data (e.g., stream). In the following embodiments, stream assignment for a single write request (RQ) is described, but the scope of the present invention is not limited thereto. For example, a stream manager (118) may accumulate multiple write requests and perform stream assignment for the accumulated multiple write requests. In one embodiment, multiple write requests may be accumulated based on logical block addresses, and the sum of the data of the multiple write requests may have a predetermined unit (e.g., a chunk unit).
[0042] In one embodiment, with reference to the drawings below, a method for allocating a stream for a write request (RQ) is described in detail, but the scope of the present invention is not limited thereto. For example, a storage device (100) may perform subsequent operations (i.e., address mapping and program operations) so that write data corresponding to a write request is stored in a non-volatile memory device (120) for each allocated stream based on the method described with reference to FIGS. 1 to FIGS. 4.
[0043] In the following, terms such as "stream" and "stream identifier" are used to facilitate the description of embodiments of the present invention. A stream may refer to a set of data having identical or similar characteristics. Alternatively, a stream may refer to a set of data consisting of a series of logical block addresses managed by a host (11). A stream identifier may be unique information for distinguishing each of a plurality of streams by being assigned to each of the plurality of streams.
[0044] Referring to FIGS. 1 and FIGS. 5, the stream manager (118) may include a plurality of machine learning models (ML1 to MLn), a model ratio calculator (MRC), and a stream identifier discriminator (DTM). In one embodiment, each of the stream manager (118) or the components included in the stream manager (118) may be implemented based on a software module, a hardware block, a hardware IP (intellectual property), a hardware accelerator, or a combination thereof.
[0045] Each of the multiple machine learning models (ML1 to MLn) can perform machine learning on a write request (RQ) and output result values (a_1 to a_n). The result values of the multiple machine learning models (ML1 to MLn) can be used as stream identifier candidates for the write request (RQ). For example, various characteristics of the write request (RQ) can be provided as input values to the multiple machine learning models (ML1 to MLn). Each of the multiple machine learning models (ML1 to MLn) can perform machine learning on the various characteristics of the write request (RQ) to derive result values (a_1 to a_n). In one embodiment, the characteristics or attributes of the write request (RQ) may include information such as a logical block address range (LBA range), a request size, etc. Alternatively, the characteristics or attributes of the write request (RQ) may include information such as the continuity of logical block addresses of adjacent write requests (i.e., workloads), an overwrite rate, etc.
[0046] In one embodiment, each of the plurality of machine learning models (ML1 to MLn) may use a different machine learning algorithm. That is, for the same write request (RQ), the result values (a_1 to a_n) of each of the plurality of machine learning models (ML1 to MLn) may have different values. In one embodiment, the types of machine learning algorithms may include various machine learning algorithms such as linear regression algorithm, support vector machine algorithm, deep neural network algorithm, deep stream algorithm, K-means algorithm, clustering algorithm, autoencoder algorithm, convolutional neural network algorithm, Siamese network algorithm, etc.
[0047] The stream manager (118) can determine a stream identifier (SID) for a write request (RQ) based on one or a combination of result values (a_1 to a_n) of a plurality of machine learning models (ML1 to MLn).
[0048] The model ratio calculator (MRC) of the stream identifier (118) can calculate a model ratio (R_x) based on a write request (RQ). The model ratio (R_x) may represent a ratio or weight reflected in the result values (a_1 to a_n) of multiple machine learning models (ML1 to MLn). For example, among the multiple machine learning models (ML1 to MLn), the first machine learning model (ML1) can classify sequential patterns relatively accurately, and the second machine learning model (ML2) can classify random patterns relatively accurately. If the write request (RQ) has a sequential pattern, the model ratio calculator (MRC) can calculate the model ratio (R_x) such that the first result value (a_1) of the first machine learning model (ML1) is reflected in the stream identifier (SID) at a relatively high ratio. Alternatively, if the write request (RQ) has a random pattern, the model ratio operator (MRC) can compute a model ratio (R_x) such that the second result value (a_2) of the second machine learning model (ML2) is reflected in the stream identifier (SID) at a relatively high rate.
[0049] In one embodiment, the model ratio calculator (MRC) can calculate the model ratio (R_x) based on the characteristics or attributes of the write request (RQ). The characteristics or attributes of the write request (RQ) may include information such as the logical block address range (LBA range) and the request size. Alternatively, the characteristics or attributes of the write request (RQ) may include information such as the continuity of logical block addresses of adjacent write requests (i.e., workloads) and the overwrite ratio. The operation and configuration of the model ratio calculator (MRC) are described in more detail with reference to FIG. 6.
[0050] The stream identifier discriminator (DTM) may receive multiple result values (a_1 to a_n) from multiple machine learning models (ML1 to MLn) and receive a model ratio (R_x) from a model ratio calculator (MRC). The stream identifier discriminator (DTM) may determine or assign a stream identifier (SID) for a write request (RQ) by applying the model ratio (R_x) to multiple result values (a_1 to a_n) from multiple machine learning models (ML1 to MLn). For example, the stream identifier (SID) may be computed as in Equation 1.
[0051]
[0052] Referring to mathematical formula 1, n indicates the number of machine learning models, and a i represents the result value of the i-th machine learning model (MLi), and r i represents the coefficient value corresponding to the i-th machine learning model (MLi) among the coefficients included in the model ratio (R_x). As described above, the stream identifier (SID) can be determined by reflecting the model ratio (R_x) in the result values (a_1 to a_n) of a plurality of machine learning models (ML1 to MLn).
[0053] In one embodiment, when a single machine learning model is used for stream classification or assignment, the accuracy or performance of stream classification in a specific logical block address range may be reduced depending on the characteristics of the single machine learning model.
[0054] On the other hand, as described above, the stream manager (118) according to an embodiment of the present invention can compute stream identifier candidates (i.e., result values (a_1 to a_n)) through a plurality of machine learning models (ML1 to MLn) for a write request (RQ), and compute a stream identifier (SID) by reflecting a model ratio (R_x) based on the write request (RQ) to the computed identifier candidates. In this case, for various characteristics of the write request (RQ), result values computed from different machine learning models can be predominantly reflected in the stream identifier (SID). Accordingly, the stream classification accuracy of the storage device (100) is improved, and accordingly, the performance of the storage device (100) or WAF characteristics can be improved.
[0055] FIG. 6 is a block diagram showing an example of the model ratio calculator of FIG. 5. Referring to FIG. 1, FIG. 5, and FIG. 6, the model ratio calculator (MRC) of the stream manager (118) may include a feature extractor (IOCX) and a look-up table (LUT).
[0056] The feature extractor (IOCX) may be configured to extract features of a write request (RQ). For example, the features of a write request (RQ) may include information such as the logical block address, logical block address range, and request size of the write request (RQ). Alternatively, the features of a write request (RQ) may include information such as the logical block address continuity of adjacent write requests (i.e., workloads), the overwrite rate, etc. The features used to select the features of the write request (RQ) or the model rate (R_x) described above are merely examples and the scope of the invention is not limited thereto. The features of a write request (RQ) may include one or more combinations of the various features described above or various features not mentioned.
[0057] A lookup table (LUT) may include multiple entries. Each of the multiple entries may include coefficients (r_11 to r_mn) for multiple machine learning models (ML1 to MLn) for each of the multiple features (CR1 to CRm). A model ratio operator (MRC) may select a model ratio (R_x) corresponding to a feature (CRx) of a write request (RQ) from the lookup table (LUT). For example, if the feature extracted from the feature extractor (IOCX) is the first feature (CR1), the model ratio operator (MRC) may select [r_11, r_12, … r1n] as the model ratio (R_x). Or, if the feature extracted from the feature extractor (IOCX) is the second feature (CR2), the model ratio operator (MRC) may select [r_21, r_22, … r2n] as the model ratio (R_x). The selected model ratio (R_x) can be provided as a stream identifier discriminator (DTM) described with reference to FIG. 5.
[0058] In one embodiment, the lookup table (LUT) may be a result learned through testing on various workloads and may be stored as firmware of the storage device (100) during the manufacturing phase of the storage device (100). In one embodiment, the lookup table (LUT) may be stored in a non-volatile memory device (120) and may be loaded from the non-volatile memory device (120) to the storage controller (110) upon booting up the storage device (100). In one embodiment, the lookup table (LUT) may be replaced with another hardware accelerator or hardware computer.
[0059] FIGS. 7 through 9 are drawings for explaining the operation of the stream manager of FIG. 5. For convenience of explanation, it is assumed that the stream manager (118) performs stream classification using two machine learning models (ML1, ML2). However, the scope of the invention is not limited thereto, and the stream manager (118) may use more machine learning models. For convenience of explanation, the stream manager (118) assumes that the characteristic of the write request (RQ) is the logical block address of the write request (RQ).
[0060] In the embodiments of FIGS. 7 to 9, the first and second machine learning models (ML1, ML2) are assumed to be trained based on the workload illustrated in FIGS. 7 to 9. The workload may include a sequential pattern and a random pattern. In the workload, the first and third logical block address ranges (LBA_r1, LBA_r3) may be a set of I / O requests having a random pattern (RAND.), and the second and fourth logical block address ranges (LBA_r2, LBA_r4) may be a set of I / O requests having a sequential pattern (SEQ.). In the embodiments of FIGS. 7 to 9, the characteristics of the workload or the characteristics of the write request are classified into sequential patterns and random patterns based on the logical block address ranges, but the scope of the present invention is not limited thereto.
[0061] As illustrated in FIG. 7, the stream classification performance or accuracy of the first machine learning model (ML1) may be relatively good for random patterns (RAND.) but relatively poor for sequential patterns (SEQ). Conversely, the stream classification performance or accuracy of the second machine learning model (ML2) may be relatively poor for random patterns (RAND.) but good for sequential patterns (SEQ). That is, for the same write request, the stream classification performance may differ depending on the machine learning model or machine learning algorithm. The stream manager (118) according to an embodiment of the present invention can improve stream classification accuracy for all ranges according to the characteristics of the write request (RQ) by applying different model ratios to each of the plurality of machine learning models (ML1 to MLn) based on the range according to the characteristics of the write request (RQ).
[0062] For example, referring to FIGS. 1, FIGS. 5, and FIGS. 8, a first write request (RQ1) received from a host (11) may be included in a second logical block address range (LBA_r2). Each of the first and second machine learning models (ML1, ML2) of the stream manager (118) may perform machine learning on the first write request (RQ1) to derive result values. The model ratio calculator (MRC) of the stream manager (118) may select or calculate a first model ratio (R_1) as a model ratio corresponding to the second logical block address range (LBA_r2). The second logical block address range (LBA_r2) corresponds to a sequential pattern (SEQ.), and with respect to the sequential pattern (SEQ.), the second machine learning model (ML2) may have better stream classification performance and accuracy than the first machine learning model (ML1). Accordingly, in the first model ratio (R_1), the coefficient (r11) corresponding to the first machine learning model (ML1) can be set to 0.05, and the coefficient (r12) corresponding to the second machine learning model (ML2) can be set to 0.95. When a stream identifier (SID) is computed based on the first model ratio (R_1), the result value of the second machine learning model (ML2), which has better performance and accuracy for the sequential pattern (SEQ.), will be predominantly reflected.
[0063] On the other hand, referring to FIGS. 1, FIGS. 5, and FIGS. 9, a second write request (RQ2) received from a host (11) may be included in a third logical block address range (LBA_r3). Each of the first and second machine learning models (ML1, ML2) of the stream manager (118) may perform machine learning on the second write request (RQ2) to derive result values. The model ratio calculator (MRC) of the stream manager may select or compute a second model ratio (R_2) as a model ratio corresponding to the third logical block address range (LBA_r3). The third logical block address range (LBA_r3) corresponds to a random pattern (RAND.), and for the random pattern (RAND.), the first machine learning model (ML1) may have better stream classification performance and accuracy than the second machine learning model (ML2). Accordingly, in the second model ratio (R_2), the coefficient (r21) corresponding to the first machine learning model (ML1) can be set to 0.90, and the coefficient (r22) corresponding to the second machine learning model (ML2) can be set to 0.10. When a stream identifier (SID) is calculated based on the second model ratio (R_2), the result value of the first machine learning model (ML2), which has better performance and accuracy for the random pattern (RAND.), will be predominantly reflected.
[0064] As described above, a storage device (100) according to an embodiment of the present invention can perform stream classification for write requests or input / output requests using a plurality of machine learning models. At this time, the storage device (100) can select or calculate the ratio of models reflected in the plurality of machine learning models based on the characteristics of the write request or input / output request. Accordingly, the performance and accuracy of stream classification can be improved for all ranges according to the characteristics of the write request or input / output request, and accordingly, the overall performance of the storage device can be improved.
[0065] FIG. 10 is a flowchart illustrating the operation of the storage device of FIG. 1. Referring to FIG. 1 and FIG. 10, in step S110, the storage device (100) may receive an input / output request from the host (11). For example, the storage device (100) may receive an input / output request from the host (11) through the host interface circuit (116). In one embodiment, the input / output request may be the write request described above. In one embodiment, the input / output request may not include information such as a stream identifier. Or, the input / output request may include a logical stream identifier. In one embodiment, the logical stream identifier may be information referring to a stream managed by the host (11), separate from the stream identifier managed by the storage device (100) (i.e., physical stream identifier). In one embodiment, the number of logical streams managed by the host (11) may be greater than the number of physical streams managed by the storage device (100). In this case, the storage device (100) requires separate management or stream allocation for mapping logical streams to physical streams, and the storage device (100) can perform mapping between logical streams and physical streams based on the operation method described above.
[0066] In one embodiment, an input / output request may include a plurality of accumulated write requests. For example, a storage device (100) may receive a plurality of write requests from a host (11) and accumulate the received plurality of write requests. The storage device (100) may perform stream allocation in units of input / output requests, which are units in which a plurality of write requests are accumulated. In one embodiment, the number of accumulated write requests included in a single input / output request (i.e., the unit in which stream allocation is performed) may be a predetermined number. Alternatively, the size of the accumulated write data corresponding to the accumulated write requests included in a single input / output request (i.e., the unit in which stream allocation is performed) may be a predetermined size.
[0067] The storage device (100) may perform the operations of steps S120 through S150 after the operation of step S110. In one embodiment, the storage device (100) may perform the operations of steps S120 through S150 for each of the input / output requests. Alternatively, the storage device (100) may accumulate the input / output requests in predetermined size units through the operation of step S110 and perform the operations of steps S120 through S150 in predetermined size units. In one embodiment, the storage device (100) may distinguish and accumulate input / output requests according to a logical block address range. For convenience of explanation below, it is assumed that the operations of steps S120 through S150 are performed for a single input / output request.
[0068] In step S120, the storage device (100) may generate candidates for stream identifiers using at least two machine learning models (or multiple machine learning models). For example, the storage device (100) may perform machine learning on an input / output request based on multiple machine learning models to derive result values. The result values may be candidates for stream identifiers. In one embodiment, each of the multiple machine learning models may be based on a different machine learning algorithm, and each of the multiple machine learning models may have different stream classification performance and accuracy for each of the intervals according to the characteristics of the input / output request.
[0069] In step S130, the storage device (100) can extract characteristics of an input / output request. In one embodiment, the characteristics of the input / output request may include information such as a logical block address, a logical block address range, a request size, etc. Alternatively, the characteristics of the input / output request may include information such as the continuity of logical block addresses of adjacent input / output requests (i.e., workloads), an overwrite rate, etc. In one embodiment, the continuity of logical block addresses of the workloads and the overwrite rate may be obtained by sampling the input / output request or the write request.
[0070] In step S140, the storage device (100) can determine the model ratio based on the extracted characteristics. For example, the stream manager (118) of the storage device (100) can determine the model ratio to be reflected in each of the plurality of machine learning models based on the configuration described with reference to FIG. 6.
[0071] In step S150, the storage device (100) can determine a stream identifier (SID) for an input / output request based on candidates for stream identifiers and model ratios. For example, the stream manager (118) of the storage device (100) can determine a stream identifier (SID) for an input / output request by reflecting the model ratios to each of the candidates for stream identifiers (i.e., result values of multiple machine learning models). In this case, since the machine learning model having better stream classification performance with respect to the characteristics of the input / output request among the multiple machine learning models is predominantly reflected in the finally determined stream identifier (SID), the performance and accuracy of the assignment and classification of stream identifiers are improved.
[0072] FIG. 11 is a diagram showing the stream manager of FIG. 1. For convenience of explanation, a detailed description of the previously described components is omitted. Referring to FIG. 1 and FIG. 11, the stream manager (118-A) may include a plurality of machine learning models (ML1 to MLn) and a stream identifier discriminator. The plurality of machine learning models (ML1 to MLn) may perform machine learning on a write request (RQ) to derive result values (a_1 to a_n). Since the plurality of machine learning models (ML1 to MLn) have been previously described, a detailed description thereof is omitted.
[0073] In one embodiment, the stream identifier discriminator may be implemented as a zero machine learning model (ML0). The zero machine learning model (ML0) may be an ensemble model configured to combine result values (a_1 to a_n) from a plurality of machine learning models (ML1 to MLn) based on a write request (RQ). That is, unlike the embodiment of FIG. 6, the stream identifier (118-A) can determine the final stream identifier (SID) by combining result values (a_1 to a_n) of a plurality of machine learning models (ML1 to MLn) using the zero machine learning model (ML0) (or ensemble model) without a separate look-up table (LUT) containing the relationship between model ratios (R_x) and features (CRx).
[0074] In one embodiment, the 0th machine learning model (ML0) can infer or derive a final stream identifier (SID) by using the result values (a_1 to a_n) of a plurality of machine learning models (ML1 to MLn) as input values, without information about the characteristics of a separate write request (RQ0). In one embodiment, the 0th machine learning model (ML0) can be trained during the inference process described above.
[0075] As described above, according to the present invention, a storage device (100) can manage input / output requests from a host (11) by separating or classifying them into stream units. For example, the storage device (100) can generate candidates for stream identifiers for input / output requests using a plurality of machine learning models and determine model ratios based on the characteristics of the input / output requests. The storage device (100) can determine a final stream identifier (SID) by reflecting the model ratios on the candidates for stream identifiers. Compared to the case where a stream is classified using a single machine learning model, since the stream is classified using a plurality of machine learning models and different model ratios are reflected in the plurality of machine learning models according to the characteristics of the input / output requests, the stream classification performance and accuracy are improved across the entire range according to the characteristics of the input / output requests.
[0076] FIG. 12 is a diagram illustrating an exemplary storage system to which a storage device according to the present invention is applied. FIG. 12 may have a configuration similar to the storage device (10) of FIG. 1. However, to explain the effects according to the present invention more clearly, the storage system is illustrated in a conceptual form.
[0077] Referring to FIGS. 1 and FIGS. 12, a plurality of applications (APP1 to APP4) can be run in a storage system (10). In one embodiment, the plurality of applications (APP1 to APP4) can be run on a host (11). In one embodiment, the host (11) can be implemented as a multi-tenant, and each of the plurality of applications (APP1 to APP4) can be run on a different tenant.
[0078] Multiple applications (APP1 to APP4) may be configured to access the storage device (100) through a file system (FS). In one embodiment, files or data generated by each of the multiple applications (APP1 to APP4) may have different characteristics. However, when multiple applications (APP1 to APP4) access the storage device (100) through the same file system (FS), the storage device (100) will generally manage the files or data generated by each of the multiple applications (APP1 to APP4) unless separate information (e.g., stream information managed by the host (11)) is provided. In this case, since the files or data generated by each of the multiple applications (APP1 to APP4) are not distinguished within the non-volatile memory device (120), maintenance operations will occur frequently.
[0079] On the other hand, the stream manager (118) of the storage device (100) according to an embodiment of the present invention can separate files or data generated by a plurality of applications (APP1 to APP4) into a plurality of streams (SID_1 to SID_4) based on the stream allocation method described with reference to FIGS. 1 to 12. That is, files or data generated by the first application (APP_1) can be classified into the first stream (SID_1), files or data generated by the second application (APP_2) can be classified into the second stream (SID_2), files or data generated by the third application (APP_3) can be classified into the third stream (SID_3), and files or data generated by the fourth application (APP_4) can be classified into the fourth stream (SID_4). In this case, since files or data generated by one application are managed as the same stream, maintenance operations for the non-volatile memory device (120) may be reduced or WAF characteristics may be improved.
[0080] FIG. 13 is a block diagram illustrating an exemplary storage system according to an embodiment of the present invention. For convenience of explanation, detailed descriptions of the components described above are omitted. Referring to FIG. 13, the storage system (1000) may include a host (1100) and a storage device (1200). The host (1100) may include a host controller (1110), a host memory (1120), a device driver (1130), and a stream manager (1140). The storage device (1200) may include a storage controller (1210) and a non-volatile memory device (1220).
[0081] In one embodiment, in the embodiments described with reference to FIGS. 1 through 12, the storage device managed the stream for write requests, but in the embodiment of FIG. 13, the host (1100) may be configured to manage the stream for data to be stored in the storage device (1200). In one embodiment, the stream manager (1140) of the host (1100) may be the stream manager described with reference to FIGS. 1 through 12 or may operate based on the stream allocation method described with reference to FIGS. 1 through 12.
[0082] In one embodiment, information about a stream allocated by a stream manager (1140) of a host (1100) may be provided to a storage device (1200) along with a write request.
[0083] In one embodiment, the storage device (1200) may be configured to manage data based on stream information provided by the host (1100). In one embodiment, the storage device (1200) may be configured to manage internal streams for write requests internally, separately from the stream information provided by the host (1100). In this case, the storage device (1200) may manage internal streams based on the stream allocation method described with reference to FIGS. 1 through 12.
[0084] FIG. 14 is a block diagram exemplarily showing a data center to which a storage device according to an embodiment of the present invention is applied. The data center (2000) is a facility that maintains and manages various data and provides various services for various data, and may be called a data storage center. The data center (200) may be a system for operating a search engine or database, and may be a computing system used by various organizations. The data center (2000) may include a plurality of application servers (2100_1 to 2100_n) and a plurality of storage servers (2200_1 to 2200_m). The number of the plurality of application servers (2100_1 to 2100_n) and the number of the plurality of storage servers (2200_1 to 2200_m) may be varied in various ways.
[0085] For convenience of explanation, an example of a first storage server (2200_1) is described below. Each of the remaining storage servers (2200_2 to 2200_m) and the plurality of application servers (2100_1 to 2100_n) may have a structure similar to that of the first storage server (2200_1).
[0086] The first storage server (2200_1) may include a processor (2210_1), memory (2220_1), a switch (2230_1), a network interface connector (NIC) (2240_1), and a storage device (2250_1). The processor (2210_1) may control the overall operation of the first storage server (2200_1). The memory (2220_1) may store various instructions or data under the control of the processor (2210_1). The processor (2210_1) may be configured to access the memory (2220_1) to execute various instructions or process data. In one embodiment, the memory (2220_1) may include at least one of various types of memory devices such as DDR SDRAM (Double Data Rate Synchronous DRAM), HBM (High Bandwidth Memory), HMC (Hybrid Memory Cube), DIMM (Dual In-line Memory Module), Optane DIMM, or NVDIMM (Non-Volatile DIMM).
[0087] In one embodiment, the number of processors (2210_1) and the number of memories (2220_1) included in the first storage server (2200_1) may vary. In one embodiment, the processors (2210_1) and memories (2220_1) included in the first storage server (2200_1) may form processor-memory pairs, and the number of processor-memory pairs included in the first storage server (2200_1) may vary. In one embodiment, the number of processors (2210_1) and the number of memories (2220_1) included in the first storage server (2200_1) may differ from each other. The processor (2210_1) may include a single-core processor or a multi-core processor.
[0088] The switch (2230_1) can selectively connect the processor (2210_1) and the storage device (2250_1) or selectively connect the NIC (2240_1) and the storage device (2250_1) according to the control of the processor (2210_1).
[0089] The NIC (2240_1) may be configured to connect the first storage server (2200_1) to a network (NT). The NIC (2240_1) may include a network interface card, a network adapter, etc. The NIC (2240_1) may be connected to the network (NT) by a wired interface, a wireless interface, a Bluetooth interface, an optical interface, etc. The NIC (2240_1) may include internal memory, a DSP, a host bus interface, etc., and may be connected to a processor (2210_1) or a switch (2230_1), etc., through the host bus interface. The host bus interface may include at least one of various interfaces such as ATA (Advanced Technology Attachment), SATA (Serial ATA), e-SATA (external SATA), SCSI (Small Computer Small Interface), SAS (Serial Attached SCSI), PCI (Peripheral Component Interconnection), PCIe (PCI express), NVMe (NVM express), IEEE 1394, USB (universal serial bus), SD (secure digital) card, MMC (multi-media card), eMMC (embedded multi-media card), UFS (Universal Flash Storage), eUFS (embedded Universal Flash Storage), CF (compact flash) card interface, etc. In one embodiment, the NIC (2240_1) may be integrated with at least one of a processor (2210_1), a switch (2230_1), and a storage device (2250_1).
[0090] The storage device (2250_1) can store data or output stored data under the control of the processor (2210_1). The storage device (2250_1) may include a controller (2251_1), non-volatile memory (2252_1), DRAM (2253_1), and an interface (2254_1). In one embodiment, the storage device (2250_1) may further include a Secure Element (SE) for security or privacy.
[0091] The controller (2251_1) can control the general operations of the storage device (2250_1). In one embodiment, the controller (2250_1) may include an SRAM. The controller (2251_1) can store data in the non-volatile memory (2252_1) or output data stored in the non-volatile memory (2252_1) in response to signals received through the interface (2254_1). In one embodiment, the controller (2251_1) may be configured to control the non-volatile memory (2252_1) based on a toggle interface or an ONFI interface.
[0092] The DRAM (2253_1) may be configured to temporarily store data to be stored in the non-volatile memory (2252_1) or data read from the non-volatile memory (2252_1). The DRAM (2253_1) may be configured to store various data (e.g., metadata, mapping data, etc.) required for the controller (2251_1) to operate. The interface (2254_1) may provide a physical connection between the processor (2210_1), the switch (2230_1), or the NIC (2240_1) and the controller (2251_1). In one embodiment, the interface (2254_1) may be implemented in a Direct Attached Storage (DAS) manner, which directly connects the storage device (2250_1) via a dedicated cable. In one embodiment, the interface (2254_1) may be configured based on at least one of the various interfaces described above via the host interface bus.
[0093] The configurations of the first storage server (2200_1) described above are exemplary and the scope of the invention is not limited thereto. The configurations of the first storage server (2200_1) described above may be applied to other storage servers or each of a plurality of application servers. In one embodiment, in each of the plurality of application servers (2100_1 to 2100_n), the storage device (2150_1) may be optionally omitted.
[0094] Multiple application servers (2100_1 to 2100_n) and multiple storage servers (2200_1 to 2200_m) can communicate with each other via a network (NT). The network (NT) can be implemented using Fibre Channel (FC) or Ethernet, etc. In this case, FC is a medium used for relatively high-speed data transmission, and an optical switch providing high performance / high availability can be used. Depending on the access method of the network (NT), the storage servers (2200_1 to 2200_m) can be provided as file storage, block storage, or object storage.
[0095] In one embodiment, the network (NT) may be a storage-dedicated network such as a Storage Area Network (SAN). For example, the SAN may be an FC-SAN that utilizes an FC network and is implemented according to the FC Protocol (FCP). Alternatively, the SAN may be an IP-SAN that utilizes a TCP / IP network and is implemented according to the iSCSI (SCSI over TCP / IP or Internet SCSI) protocol. In one embodiment, the network (NT) may be a general network such as a TCP / IP network. For example, the network (NT) may be implemented according to protocols such as FC over Ethernet (FCoE), Network Attached Storage (NAS), and NVMe over Fabrics (NVMe-oF).
[0096] In one embodiment, at least one of the plurality of application servers (2100_1 to 2100_n) may be configured to access at least one other of the plurality of application servers (2100_1 to 2100_n) or at least one of the plurality of storage servers (2200_1 to 2200_m) through a network (NT).
[0097] For example, the first application server (2100_1) can store data requested by a user or client in at least one of a plurality of storage servers (2200_1 to 2200_m) via a network (NT). Alternatively, the first application server (2100_1) can obtain data requested by a user or client from at least one of a plurality of storage servers (2200_1 to 2200_m) via a network (NT). In this case, the first application server (2100_1) can be implemented as a web server or a DBMS (Database Management System), etc.
[0098] That is, the processor (2110_1) of the first application server (2100_1) can access the memory (2120_n) or storage device (2150_n) of another application server (e.g., 2100_n) through the network (NT). Alternatively, the processor (2110_1) of the first application server (2100_1) can access the memory (2220_1) or storage device (2250_1) of the first storage server (2200_1) through the network (NT). Through this, the first application server (2100_1) can perform various operations on data stored in other application servers (2100_2 to 2100_n) or multiple storage servers (2200_1 to 2200_m). For example, the first application server (2100_1) may execute or issue a command to move or copy data between other application servers (2100_2 to 2100_n) or multiple storage servers (2200_1 to 2200_m). In this case, the data to be moved or copied may be moved from the storage devices (2250_1 to 2250_m) of the storage servers (2200_1 to 2200_m) through the memories (2220_1 to 2220_m) of the storage servers (2200_1 to 2200_m) or directly to the memories (2120_1 to 2120_n) of the application servers (2100_1 to 2100_n). Data transmitted over the network (NT) may be encrypted data for security or privacy.
[0099] In one embodiment, the storage servers (2200_1 to 2200_m) or storage devices (2150_1 to 2150_n, 2250_1 to 2250_m) described above may include a stream manager according to an embodiment of the present invention. That is, at least one of the storage servers (2200_1 to 2200_m) or storage devices (2150_1 to 2150_n, 2250_1 to 2250_m) may be configured to allocate and manage a stream for input data based on the method described with reference to FIGS. 1 to 12.
[0100] The above description describes specific embodiments for implementing the present invention. The present invention will include not only the embodiments described above, but also embodiments that can be simply modified or easily modified. Furthermore, the present invention will include technologies that can be easily modified and implemented using the embodiments. Accordingly, the scope of the present invention should not be limited to the embodiments described above, but should be defined by the claims set forth below as well as equivalents to the claims of this invention.
Claims
Claim 1 A method of operation for a storage device that supports multiple streams, comprising: receiving an input / output request from an external host; performing machine learning on the input / output request based on a plurality of machine learning models to generate a plurality of stream identifier candidates from each of the plurality of machine learning models; generating a model ratio based on the characteristics of the input / output request; reflecting the model ratio on the plurality of stream identifier candidates to assign a final stream identifier for the input / output request; and storing write data corresponding to the input / output request in a non-volatile memory device of the storage device based on the final stream identifier, wherein each of the plurality of machine learning models is based on a different machine learning algorithm. Claim 2 A method of operation according to claim 1, wherein the characteristics of the input / output request include at least one of the logical block address, logical block address range, and request size of the input / output request. Claim 3 A method of operation according to claim 1, wherein the characteristics of the input / output request include at least one of the continuity of logical block addresses of the input / output request and the overwrite ratio of the input / output request adjacent to the input / output request. Claim 4 A method of operation according to claim 1, wherein the model ratio includes a plurality of coefficients corresponding to the plurality of machine learning models, and the plurality of coefficients have different values. Claim 5 In claim 4, the plurality of coefficients in the case where the characteristic of the input / output request corresponds to the first logical block address range are different from the plurality of coefficients in the case where the characteristic of the input / output request corresponds to the second logical block address range. Claim 6 In claim 1, each of the plurality of machine learning models is a method of operation based on at least one algorithm among a linear regression algorithm, a support vector machine algorithm, a deep neural network algorithm, a deep stream algorithm, a K-means algorithm, a clustering algorithm, an autoencoder algorithm, a convolutional neural network algorithm, and a sham network algorithm. Claim 7 A method of operation according to claim 1, wherein the step of generating a model ratio based on the characteristics of the input / output request comprises: the step of extracting the characteristics of the input / output request; and the step of selecting and outputting the model ratio corresponding to the characteristics from a lookup table, wherein the lookup table comprises a plurality of entries, and each of the plurality of entries comprises information on a model ratio corresponding to each of the plurality of characteristics. Claim 8 In claim 7, the look-up table is a method of operation included in the firmware of the storage device. Claim 9 A method of operation according to claim 1, wherein if the final stream identifier is a first stream identifier, the write data is stored in a first memory block of the non-volatile memory device, and if the final stream identifier is a second stream identifier different from the first stream identifier, the write data is stored in a second memory block different from the first memory block of the non-volatile memory device. Claim 10 A method of operation according to claim 1, wherein the input / output request includes accumulated write requests, and the size of the accumulated write data corresponding to the accumulated write requests is a predetermined size.
Citation Information
Patent Citations
Enhanced multi-streaming through statistical analysis
KR1020170063340A
Access parameter based multi-stream storage device access
KR1020170133256A
Storage device that performs machine learning and method thereof
KR1020190115303A
Stream selection for multi-stream storage devices
KR1020190119080A
Method of operating storage device, storage device performing the same and storage system including the same
KR1020200054600A
