Abstract: Method and apparatus for managing data in a storage device, such as a solid-state drive (SSD). In some embodiments, write and read commands from a client device are placed into a command queue pending servicing to transfer data between the client device and a non-volatile memory (NVM). A write cache temporarily stores sets of writeback data pending transfer. A cache manager detects an overlap condition in which a subsequently received command at least partially overlaps a pending write command. In response, the cache manager enacts a change in caching policy that includes retention of the cached writeback data to aid in the servicing of the subsequently received command. The changes in caching policy can include an increase in the size of the write cache, delays in the writing of hot writeback data sets, the coalescing of different writeback data sets, cache hits using the cached writeback data, etc.

Abstract: Embodiments herein provide for a controller that is operable to soft read a data bit a plurality of times, to generate a bit set for the data bit from the soft reads, to logically operate on the bit set, and to generate a Hamming weight for the data bit based on the logical operation. The Hamming weight has fewer bits than the bit set and is operable to correct the data bit.

Abstract: Technologies are described herein for or reducing the size of the forward mapping table in an SSD or other storage device using hashing. A physical address of a storage location within a storage media is determined for the storage of data associated with a logical block address. The data is written to the storage location and a hash value is computed from a representation of the physical address using a hash function, where the size of the hash value is smaller than the representation of the physical address and the hash value points to a plurality of separate storage locations in the storage media. The hash value is stored in the forward mapping table associated with logical block address as opposed to the representation of the physical address.

Abstract: Technologies are described herein for or reducing the size of the forward mapping table in an SSD or other storage device using hashing. A physical address of a storage location within a storage media is determined for the storage of data associated with a logical block address. The data is written to the storage location and a hash value is computed from a representation of the physical address using a hash function, where the size of the hash value is smaller than the representation of the physical address and the hash value points to a plurality of separate storage locations in the storage media. The hash value is stored in the forward mapping table associated with logical block address as opposed to the representation of the physical address.

Abstract: Systems and methods presented herein provide for failure detection and data recovery in a storage system. In one embodiment, a method operable in a storage system comprises locating failures in data blocks in storage area of a storage device, categorizing the failures into block groups, each block group comprising one or more data blocks having failures, and halting input/output (I/O) operations to data blocks in a first of the block groups due to the failures of the first block group. The method also includes detecting additional failures in one or more data blocks of other block groups remaining in the storage area, and determining when to fail the storage area of the storage device based on the detected failures.

Abstract: Method and apparatus for decoding error correction code (ECC) code words. Reference voltages are used to extract a selected code word from a communication channel. The selected code word is processed by an ECC decoder, and an initial syndrome weight is determined indicative of unresolved parity errors. A coarse search operates to concurrently adjust, over a first succession of iterations, each of the reference voltages. A subsequent fine search operates, over a second succession of iterations, to individually adjust the reference voltages. Decoding and syndrome weight determination continues over each iteration until a minimum syndrome weight is obtained, after which a user data content of the code word is decoded. The coarse search may transition the decoder from a saturated operational region to a linear operational region. The decoder may be a low density parity check (LDPC) decoder.

Abstract: Method and apparatus for decoding error correction code (ECC) code words. Reference voltages are used to extract a selected code word from a communication channel. The selected code word is processed by an ECC decoder, and an initial syndrome weight is determined indicative of unresolved parity errors. A coarse search operates to concurrently adjust, over a first succession of iterations, each of the reference voltages. A subsequent fine search operates, over a second succession of iterations, to individually adjust the reference voltages. Decoding and syndrome weight determination continues over each iteration until a minimum syndrome weight is obtained, after which a user data content of the code word is decoded. The coarse search may transition the decoder from a saturated operational region to a linear operational region. The decoder may be a low density parity check (LDPC) decoder.

Abstract: A method operable with the storage device includes determining a workload to the storage device based on host Input/Output (I/O) requests to the storage device. When the workload is above a threshold, a first portion of the storage device is selected for garbage collection based on the I/O requests. Otherwise, when the workload is below the threshold, a second different portion of the storage device is selected for garbage collection based on a storage ability of the second portion of the storage device.

Abstract: A connection between a user device and a network server is established. Via the connection, a deep learning network is formed for a processing task. A first portion of the deep learning network operates on the user device and a second portion of the deep learning network operates on the network server. Based on cooperation between the user device and the network server, a boundary between the first portion and the second portion of the deep learning network is dynamically modified based on a change in a performance indicator that could affect the processing task.

Abstract: Embodiments herein provide for a controller that is operable to soft read a data bit a plurality of times, to generate a bit set for the data bit from the soft reads, to logically operate on the bit set, and to generate a Hamming weight for the data bit based on the logical operation. The Hamming weight has fewer bits than the bit set and is operable to correct the data bit.

Abstract: Systems and methods are disclosed for open block stability scanning. When a solid state memory block remains in an open state, where the block is only partially filled with written data, for a prolonged period of time, a circuit may perform a scan on the block to determine the stability of the stored data. When the scan indicates that the data is below a stability threshold, the data may be refreshed by reading the data and writing it to a new location. When the scan indicates that the data is above a stability threshold, the circuit may extend the time period in which the block may remain in the open state.

Abstract: A data storage system may have a number of data storage devices that each have a non-volatile memory connected to a memory buffer. The memory buffer can consist of a map unit having a predetermined size. In receipt of a data sector into the map unit of the memory buffer, the data sector may be identified as a runt with a runt module connected to the memory buffer and the non-volatile memory. The runt module can generate and subsequently execute a runt handling plan to fill the size of the map unit before storing the filled map unit in the non-volatile memory.

Abstract: Systems and methods presented herein provide for computing read voltages for a storage device. In one embodiment, a controller is controller is operable to soft read data from a portion of the storage device, and to iteratively test the soft read data a predetermined number of times. For example, the controller may test the soft read data a number of times by applying a different probability weight to the soft read data each time the soft read data is tested. The controller may then decode the soft read data based on the probability weight, and determine an error metric of the decoded soft read data. Then, the controller determines a read voltage for the portion of the storage device based on the probability weight and the error metric.

Abstract: An unrecoverable error is determined when attempting to access data stored in a target memory unit of a storage medium of a data storage device. Access to neighboring memory units that are proximate to the target memory unit is controlled in response. The controlling of the access affects a temperature of the target memory unit over a period of time. After the period of time, access to the data is attempted from the target memory unit.

Abstract: Method and apparatus for managing data in a memory. In accordance with some embodiments, a data object is stored in a first non-volatile tier of a multi-tier memory structure. A metadata unit is generated to describe the data object, the metadata unit having a selected granularity. The metadata unit is stored in a different, second non-volatile tier of the multi-tier memory structure responsive to the selected granularity.

Abstract: Systems and methods presented herein provide a controller is operable to increase a number of suspend operations during read Input/Output (I/O) operations of a storage device, and to detect an increase in response times for write commands due to the increased number of suspend operations. The controller is also operable to decrease the number of the suspend operations during the reads of the storage device to decrease the response times of the write commands.

Abstract: Embodiments herein provide for a controller that is operable to soft read a data bit a plurality of times, to generate a bit set for the data bit from the soft reads, to logically operate on the bit set, and to generate a Hamming weight for the data bit based on the logical operation. The Hamming weight has fewer bits than the bit set and is operable to correct the data bit.

Abstract: Systems and methods are disclosed for wear-based access optimization. An apparatus may comprise a circuit configured to perform a data access operation at a target location of a memory, and determine a wear value of the target location. The circuit may compare the wear value to global wear value of other locations of the drive, and adjust data access parameters for the target location based on the comparison.

Abstract: A method operable with the storage device includes determining a workload to the storage device based on host Input/Output (I/O) requests to the storage device. When the workload is above a threshold, a first portion of the storage device is selected for garbage collection based on the I/O requests. Otherwise, when the workload is below the threshold, a second different portion of the storage device is selected for garbage collection based on a storage ability of the second portion of the storage device.

Abstract: Systems and methods presented herein provide for computing read voltages for a storage device. In one embodiment, a controller is controller is operable to soft read data from a portion of the storage device, and to iteratively test the soft read data a predetermined number of times. For example, the controller may test the soft read data a number of times by applying a different probability weight to the soft read data each time the soft read data is tested. The controller may then decode the soft read data based on the probability weight, and determine an error metric of the decoded soft read data. Then, the controller determines a read voltage for the portion of the storage device based on the probability weight and the error metric.