# US 2019/0317901 A1 · System and method for optimizing performance of a solid-state drive using a deep neural network

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2019/0317901 A1
- Document ID: US20190317901A1
- Status: Published application
- Kind: A1
- Application number: 16/012,470
- Filed: 2018-06-19
- Published: 2019-10-17
- Earliest claimed: 2018-04-16
- Assignees: Samsung Electronics Co., Ltd.
- Applicants: Samsung Electronics Co., Ltd.
- Inventors: Ramdas P. KACHARE; Sompong Paul OLARIG; Vikas SINHA; Zvika GUZ
- Web page: https://patentgrep.com/patent/US20190317901A1

## CPC

- G06F 3/061: {Improving I/O performance}
- G06F 12/1009: using page tables, e.g. page table structures
- G06F 3/0604: {Improving or facilitating administration, e.g. storage management}
- G06F 3/064: {Management of blocks}
- G06F 3/0659: {Command handling arrangements, e.g. command buffers, queues, command scheduling}
- G06F 3/0679: {Non-volatile semiconductor memory device, e.g. flash memory, one time programmable memory [OTP]}
- G06F 3/0688: {Non-volatile semiconductor memory arrays}
- G06F 12/0246: {in block erasable memory, e.g. flash memory}
- G06F 12/0862: with prefetch
- G06F 12/0868: Data transfer between cache memory and other subsystems, e.g. storage devices or host systems
- G06F 12/109: for multiple virtual address spaces, e.g. segmentation (G06F12/1036 takes precedence)
- G06N 3/044: Recurrent networks, e.g. Hopfield networks
- G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
- G06N 3/045: Combinations of networks
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/088: Non-supervised learning, e.g. competitive learning
- G06N 3/0895: Weakly supervised learning, e.g. semi-supervised or self-supervised learning
- G06N 3/09: Supervised learning
- G06F 2212/1024: Latency reduction
- G06F 2212/214: Solid state disk
- G06F 2212/313: In storage device
- G06F 2212/502: using adaptive policy
- G06F 2212/6024: History based prefetching
- G06F 2212/7201: Logical to physical mapping or translation of blocks or pages
- G06F 2212/7203: Temporary buffering, e.g. using volatile buffer or dedicated buffer blocks
- G06N 20/00: Machine learning
- Y02D 10/00: Energy efficient computing, e.g. low power processors, power management or thermal management
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- Y02D: CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE

## Abstract

A controller of a data storage device includes: a host interface providing an interface to a host computer; a flash translation layer (FTL) translating a logical block address (LBA) to a physical block address (PBA) associated with an input/output (I/O) request; a flash interface providing an interface to flash media to access data stored on the flash media; and one or more deep neural network (DNN) modules for predicting an I/O access pattern of the host computer. The one or more DNN modules provide one or more prediction outputs to the FTL that are associated with one or more past I/O requests and a current I/O request received from the host computer, and the one or more prediction outputs include at least one predicted I/O request following the current I/O request. The FTL prefetches data stored in the flash media that is associated with the at least one predicted I/O request.

## Claims

### Claim 1

A controller of a data storage device comprising:
a host interface providing an interface to a host computer;
a flash translation layer (FTL) translating a logical block address (LBA) to a physical block address (PBA) associated with an input/output (I/O) request received from the host computer via the host interface;
a flash interface providing an interface to flash media of the data storage device to access data stored on the flash media of the data storage device; and
one or more deep neural network (DNN) module for predicting an I/O access pattern of the host computer,
wherein the host interface provides one or more input vectors to the one or more DNN modules, the one or more DNN modules provide one or more prediction outputs to the FTL that are associated with one or more past I/O requests and a current I/O request received from the host computer using the one or more input vectors, and wherein the one or more prediction outputs include at least one predicted I/O request following the current I/O request, and
wherein the FTL prefetches data stored in the flash media that is associated with the at least one predicted I/O request.

### Claim 2 (depends on 1)

The controller of claim 1, wherein the one or more prediction outputs identify an idle time between host operations, and the FTL performs background operations of the flash media during the idle time.

### Claim 3 (depends on 1)

The controller of claim 1, wherein the one or more DNN modules include a prediction table that stores a plurality of prediction outputs based on the current I/O request.

### Claim 4 (depends on 1)

The controller of claim 1, further comprising a prediction error estimator configured to provide a prediction error to the one or more DNN modules to perform training or learning the I/O access pattern associated with the one or more past I/O requests, the current I/O request, and the at least one predicted I/O request.

### Claim 5 (depends on 1)

The controller of claim 1, wherein the FTL determines a time and performs one or more of fetching or updating data structure of the FTL, performing an error correction, un-compressing the data, decrypt the data, performing background operations, providing statistics and logs to the host computer, and offloading a cache based on the I/O access pattern of the host computer.

### Claim 6 (depends on 1)

The controller of claim 1, wherein the controller is implemented in a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), or a combination of the FPGA and the ASIC.

### Claim 7 (depends on 1)

The controller of claim 1, wherein the one or more DNN modules employ a long short term memory (LSTM) network or a convolutional neural network (CNN).

### Claim 8 (depends on 1)

The controller of claim 1, wherein the data storage device is a non-volatile memory express (NVMe) solid-state drive (SSD) or an NVMe over fabrics (NVMe-oF) SSD.

### Claim 9 (depends on 1)

The controller of claim 1, wherein the one or more input vectors includes one or more of an LBA, a range of LBAes, a length of a host access, a number of a host access, a namespace ID, a host ID, a command opcode (OPC), an I/O type as indicated by the command opcode, a stream ID, an NVM set ID, a time delta and/or a time stamp.

### Claim 10 (depends on 1)

The controller of claim 1, wherein the one or more prediction outputs includes one or more of an LBA, an LBA cluster, a type of a host access, an imminent idle gap, an imminent access volume.

### Claim 11 (depends on 1)

The controller of claim 1, wherein the one or more DNN modules assign a probability of an imminent access to each block present in a cache, and the controller evicts a block with a lowest probability from the cache.

### Claim 12 (depends on 11)

The controller of claim 11, wherein the controller evicts a block that has an oldest time stamp from the cache if multiple blocks in the cache have a same lowest probability.

### Claim 13

A data storage device comprising:
one or more flash media;
a controller configured to control the one or more flash media; and
one or more deep neural network (DNN) modules for predicting an I/O access pattern of a host computer,
wherein the one or more DNN modules receive one or more input vectors from the controller, provide one or more prediction outputs to the controller that are associated with one or more past I/O requests and a current I/O request received from the host computer using the one or more input vectors, and the one or more prediction outputs include at least one predicted I/O request following the current I/O request, and
wherein the controller prefetches data stored in the flash media that is associated with the at least one predicted I/O request.

### Claim 14 (depends on 13)

The data storage device of claim 13, wherein the one or more prediction outputs identify an idle time between host operations, and the FTL performs background operations of the flash media during the idle time.

### Claim 15 (depends on 13)

The data storage device of claim 13, wherein the controller comprises a prediction error estimator configured to provide a prediction error to the one or more DNN modules to perform training or learning the host access pattern associated the one or more past I/O requests, the current I/O request, and the at least one predicted I/O request.

### Claim 16 (depends on 13)

The data storage device of claim 13, wherein the controller is implemented in a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), or a combination of the FPGA and the ASIC.

### Claim 17 (depends on 13)

The data storage device of claim 13, wherein the data storage device is a non-volatile memory express (NVMe) solid-state drive (SSD) or an NVMe over fabrics (NVMe-oF) SSD.

### Claim 18 (depends on 13)

The data storage device of claim 13, wherein the one or more DNN modules assign a number to each block present in a cache based on a probability of an imminent access, and the controller evicts a block with a lowest probability from the cache.

### Claim 19 (depends on 18)

The data storage device of claim 18, wherein the controller evicts a block that has an oldest time stamp from the cache if multiple blocks in the cache have a same lowest probability.

### Claim 20

A bridge device comprising:
one or more processors;
one or more deep neural network (DNN) modules;
a processor access module providing an interface between the one or more processors and the one or more DNN modules;
a host interface; and
a device interface to a solid-state drive (SSD),
wherein the one or more DNN modules receive one or more input vectors from the SSD via the host interface, provide one or more prediction outputs to the SSD that are associated with one or more past I/O requests from the remote initiator and a current I/O request received from the remote initiator using the one or more input vectors, and the one or more prediction outputs include at least one predicted I/O request following the current I/O request, and
wherein the SSD prefetches data stored in the flash media that is associated with the at least one predicted I/O request and provides the data to the bridge device via the device interface.

### Claim 21 (depends on 20)

The bridge device of claim 20, further comprising a plurality of buffers storing data to transfer between a remote initiator and the SSD.

### Claim 22 (depends on 20)

The bridge device of claim 20, wherein the interface to the SSD includes a PCIe Application (PAPP) layer and a PCIe root port (RP).

### Claim 23 (depends on 20)

The bridge device of claim 20, wherein the host interface is a fabric interface compatible with NVMe-oF, and the device interface is compatible with NVMe.

### Claim 24 (depends on 20)

The bridge device of claim 20, wherein the one or more DNN modules are implemented in a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), or a combination of the FPGA and the ASIC.
