The present disclosure provides systems, methods, and computer program products for providing efficient embedding table storage and lookup in machine-learning models. A computer-implemented method may include obtaining an embedding table comprising a plurality of embeddings respectively associated with a corresponding index of the embedding table, compressing each particular embedding of the embedding table individually allowing each respective embedding of the embedding table to be decompressed independent of any other embedding in the embedding table, packing the embedding table comprising individually compressed embeddings with a machine-learning model, receiving an input to use for locating an embedding in the embedding table, determining a lookup value based on the input to search indexes of the embedding table, locating the embedding based on searching the indexes of the embedding table for the determined lookup value, and decompressing the located embedding independent of any other embedding in the embedding table.

The present disclosure provides systems, methods, and computer program products for providing efficient embedding table storage and lookup in machine-learning models. A computer-implemented method may include obtaining an embedding table comprising a plurality of embeddings respectively associated with a corresponding index of the embedding table, compressing each particular embedding of the embedding table individually allowing each respective embedding of the embedding table to be decompressed independent of any other embedding in the embedding table, packing the embedding table comprising individually compressed embeddings with a machine-learning model, receiving an input to use for locating an embedding in the embedding table, determining a lookup value based on the input to search indexes of the embedding table, locating the embedding based on searching the indexes of the embedding table for the determined lookup value, and decompressing the located embedding independent of any other embedding in the embedding table.

A method for operating a RAID storage system includes configuring the RAID storage devices to receive in a read or write command a byte count, receiving a first data block to write to the storage system, compressing the received first data block to generate a first compressed data block, and then storing the first compressed data block memory. The method additionally includes executing a set of RAID operations to perform a partial stripe update, including: retrieving a second compressed data block from memory; determining a physical size of the second compressed data block; generating, based on the second compressed data block and the physical size, redundant data corresponding with the second compressed data block; and writing the second compressed data block and the redundant data by transmitting a write command including the second compressed data block, the redundant data, and the physical size to the set of RAID storage devices.

A storage system comprises a plurality of enclosures and a storage controller. Each enclosure comprises at least one processing device and a plurality of drives configured in accordance with a redundant array of independent disks (RAID) arrangement. The storage controller obtains data pages associated with an input-output request, provides the data pages to a processing device of a given enclosure, and issues a command to the processing device to perform at least one operation based at least in part on the data pages. The processing device of the given enclosure receives the data pages from the storage controller, generates compressed data pages based at least in part on the received data pages, stores one or more of the compressed data pages on the plurality of drives according to the RAID arrangement and returns information associated with the storage of the compressed data pages to the storage controller.

A method for more efficiently utilizing storage space in a redundant array of independent disks (RAID) is disclosed. In one embodiment, such a method implements a RAID from multiple storage drives. The RAID utilizes data striping with distributed parity values to provide desired data protection/redundancy. The distributed parity values are placed on selected storage drives of the RAID in accordance with a designated parity rotation. The method further adaptively alters the parity rotation of the RAID to provide an increased concentration of parity values in certain storage drives of the RAID compared to other storage drives of the RAID. This parity rotation may be adapted based on residual storage capacity in each storage drive, consumed space in each storage drive, or the like. A corresponding system and computer program product are also disclosed.

The present disclosure provides systems, methods, and computer program products for providing efficient embedding table storage and lookup in machine-learning models. A computer-implemented method may include obtaining an embedding table comprising a plurality of embeddings respectively associated with a corresponding index of the embedding table, compressing each particular embedding of the embedding table individually allowing each respective embedding of the embedding table to be decompressed independent of any other embedding in the embedding table, packing the embedding table comprising individually compressed embeddings with a machine-learning model, receiving an input to use for locating an embedding in the embedding table, determining a lookup value based on the input to search indexes of the embedding table, locating the embedding based on searching the indexes of the embedding table for the determined lookup value, and decompressing the located embedding independent of any other embedding in the embedding table.

A method for more efficiently utilizing storage space in a set of storage drives is disclosed. In one embodiment, such a method implements, in a set of storage drives, a first RAID utilizing data striping with distributed parity values. The method further implements, in a subset of the set of storage drives, a second RAID using residual storage space in storage drives belonging to the subset. Storage drives belonging to the subset may have a storage capacity that is larger than storage drives not belonging to the subset. In certain embodiments, the method adaptively alters a parity rotation of the first RAID to provide an increased concentration of parity values in certain storage drives of the first RAID compared to other storage drives of the first RAID. A corresponding system and computer program product are also disclosed.

A disclosed method is performed at a fault-tolerant object-based storage system including M data storage entities, each is configured to store data on an object-basis. The method includes obtaining a request to store N copies of a data object and in response, storing the N copies of the data object across the M data storage entities, where the N copies are distributed across the M data storage entities. The method additionally includes generating a first parity object for a first subset of M copies of the N copies of the data object, where the first parity object is stored on a first parity storage entity separate from the M data storage entities. The method also includes generating a manifest linking the first parity object with one or more other subsets of M copies of the N copies of the data object.

In one aspect, a method includes splitting empty RAID stripes into sub-stripes and storing pages into the sub-stripes based on a compressibility score. In another aspect, a method includes reading pages from 1-stripes, storing compressed data in a temporary location, reading multiple stripes, determining compressibility score for each stripe and filling stripes based on the compressibility score. In a further aspect, a method includes scanning a dirty queue in a system cache, compressing pages ready for destaging, combining compressed pages in to one aggregated page, writing one aggregated page to one stripe and storing pages with same compressibility score in a stripe.

A storage system comprises a plurality of enclosures and a storage controller. Each enclosure comprises at least one processing device and a plurality of drives configured in accordance with a redundant array of independent disks (RAID) arrangement. The storage controller obtains data pages associated with an input-output request, provides the data pages to a processing device of a given enclosure, and issues a command to the processing device to perform at least one operation based at least in part on the data pages. The processing device of the given enclosure receives the data pages from the storage controller, generates compressed data pages based at least in part on the received data pages, stores one or more of the compressed data pages on the plurality of drives according to the RAID arrangement and returns information associated with the storage of the compressed data pages to the storage controller.