Abstract: In one embodiment, a device includes interface circuitry and processing circuitry. The interface circuitry communicates with a plurality of storage devices associated with a storage system. The processing circuitry receives a request to write a data object to the storage system. The data object includes a set of data elements, and the storage system is organized into blocks and shards, which are distributed across the storage devices. The processing circuitry determines a storage layout for the data object, which arranges the set of data elements across a set of blocks and shards with padding to align each data element within block and shard boundaries. The processing circuitry writes the data object to the storage system based on the storage layout.

Abstract: In one embodiment, a system comprises a host processor and a storage system. The storage system comprises one or more storage devices, and each storage device comprises a non-volatile memory and a compute offload controller. The non-volatile memory stores data, and the compute offload controller performs compute tasks on the data based on compute offload commands from the host processor.

Abstract: Techniques for storage and processing for distributed file systems are disclosed. In the illustrative embodiment, padding is placed between data elements in a file to be stored on a distributed file system. The file is to be split into several objects in order to be stored in the distributed file system, and the padding is used to prevent a data element from being split across two different objects. The objects are stored on data nodes, which analyze the objects to determine which data elements are present in the object as well at the location of those objects. The location of the objects is saved on the data storage device, and those locations can be used to perform queries on the data elements in the object on the data storage device itself. Such an approach can reduce transfer of data elements from data storage to local memory of the data node.

Abstract: Embodiments of the present disclosure may relate to a compute offload controller that may include a parser to parse a received compute offload command, and identify a block-based compute descriptor based at least in part on the compute offload command. In some embodiments, the compute offload controller may further include an offload executor to perform an operation on data in a block-based storage device based at least in part on the block-based compute descriptor. In some embodiments, the block-based compute descriptor may include a virtual input object, a virtual output object, and a compute type identifier. Other embodiments may be described and/or claimed.

Abstract: Embodiments of the present disclosure may relate to a compute offload controller that may include a parser to parse a received compute offload command, and identify a block-based compute descriptor based at least in part on the compute offload command. In some embodiments, the compute offload controller may further include an offload executor to perform an operation on data in a block-based storage device based at least in part on the block-based compute descriptor. In some embodiments, the block-based compute descriptor may include a virtual input object, a virtual output object, and a compute type identifier. Other embodiments may be described and/or claimed.

Abstract: A multiple data protection scheme logic enables applications to use objects encoded with different data protection schemes in a single namespace. Instead of configuring data protection policy supporting only one data protection scheme in a single namespace, flexible data protection policies that allow different data protection schemes in the single namespace promote more efficient use of storage and processor resources. Smaller objects can use data protection schemes that favor more efficient processor performance over increased storage costs such as replication, whereas larger objects can use data protection schemes that favor decreased storage costs over less efficient processor performance such as erasure coding.

Abstract: Methods and apparatus related to dynamic management of relationships in distributed object stores are described. In one embodiment, one or more links are generated between two or more objects of the object stores. A single request directed at a first object may return data corresponding to the first object and one or more other objects. Other embodiments are also disclosed and claimed.