Abstract: A system receives a request for a data block, in which the data block corresponds to a set of fragments, and the data block can be fully reconstructed using a threshold number of the set of fragments. A first subset of the set of fragments are stored in a first storage system, and a second subset of the set of fragments are stored in a second storage system characterized by a higher average latency than the first storage system. A system, responsive to receiving the request for the data block, requests the first subset of the set of fragments. A system receives at least the threshold number of the first subset of the fragments. A system reconstructs the data block using the received fragments of the first subset of the fragments. A system provides the reconstructed data block.

Abstract: The system obtains performance signals associated with respective hard disks of a volume of hard disks including a plurality of hard disks that are dedicated to activities of a service. The system determines a volume failure prediction for the volume of hard disks by, for each respective hard disk of the volume of hard disks, determining a hard disk failure prediction. The system determines a hard disk failure prediction by: inputting the respective performance signals into a supervised machine learning model; and receiving as output from the machine learning model the hard disk failure prediction for the respective hard disk. The system based on the received outputs, determines that the volume failure prediction is associated with a migration condition. The system, responsive to determining that the volume failure prediction is associated with the migration condition, migrates data from the volume of hard disks to a second volume of hard disks.

Abstract: Computer-implemented techniques for fair data scrubbing. The techniques can be used to balance a desire to verify recently stored data soon after it is stored on a target data storage media device, when the computing and networking cost of reconstructing the data in the event of a detected data storage media device error can be lower, against a desire to minimize the latency between rescrubbing data. By doing so, the techniques improve the operation of a data storage system that implements the techniques.

Abstract: Computer-implemented techniques for fair data scrubbing. The techniques can be used to balance a desire to verify recently stored data soon after it is stored on a target data storage media device, when the computing and networking cost of reconstructing the data in the event of a detected data storage media device error can be lower, against a desire to minimize the latency between rescrubbing data. By doing so, the techniques improve the operation of a data storage system that implements the techniques.

Abstract: Systems and techniques for performing a data transaction are disclosed that provide data redundancy using two or more cache devices. In some embodiments, a data transaction is received by a storage controller of a storage system from a host system. The storage controller caches data and/or metadata associated with the data transaction to at least two cache devices that are discrete from the storage controller. After caching, the storage controller provides a transaction completion response to the host system from which the transaction was received. In some examples, each of the at least two cache devices includes a storage class memory. In some examples, the storage controller caches metadata to the at least two cache devices and to a controller cache of the storage controller, while data is cached to the at least two cache devices without being cached in the controller cache.

Abstract: Systems and techniques for performing a data transaction are disclosed that provide data redundancy using two or more cache devices. In some embodiments, a data transaction is received by a storage controller of a storage system from a host system. The storage controller caches data and/or metadata associated with the data transaction to at least two cache devices that are discrete from the storage controller. After caching, the storage controller provides a transaction completion response to the host system from which the transaction was received. In some examples, each of the at least two cache devices includes a storage class memory. In some examples, the storage controller caches metadata to the at least two cache devices and to a controller cache of the storage controller, while data is cached to the at least two cache devices without being cached in the controller cache.

Abstract: Systems and techniques for performing a data transaction are disclosed that provide data redundancy using two or more cache devices. In some embodiments, a data transaction is received by a storage controller of a storage system from a host system. The storage controller caches data and/or metadata associated with the data transaction to at least two cache devices that are discrete from the storage controller. After caching, the storage controller provides a transaction completion response to the host system from which the transaction was received. In some examples, each of the at least two cache devices includes a storage class memory. In some examples, the storage controller caches metadata to the at least two cache devices and to a controller cache of the storage controller, while data is cached to the at least two cache devices without being cached in the controller cache.

Abstract: Systems and techniques are disclosed for the mirroring of cache data from a storage controller to a storage class memory (“SCM”) device. The storage controller receives a write request, caches the write data, and mirrors the write data to the SCM device instead of to a cache of another storage controller. The SCM device stores the mirrored data in the SCM device. The storage controller acknowledges the write to the host. If the storage controller later fails, an alternate controller assumes ownership of storage volumes associated with the failed controller. Upon receipt of a new read request to the failed controller, the alternate controller checks the SCM device for a cache hit. If there is, the data is read from the SCM device; otherwise, it is read from the storage volume(s). The read data is cached at the alternate controller and then sent on to the requesting host.

Abstract: A storage manager can reduce the overhead of parity based fault tolerance by leveraging the access performance of SSDs for the parities. Since reading a parity value can be considered a small read operation, the reading of parity from an SSD is an effectively “free” operation due to the substantially greater SSD read performance. With reading parity being an effectively free operation, placing parity on SSDs eliminates the parity read operations (in terms of time) from the parity based fault tolerance overhead. A storage manager can selectively place parity on SSDs from HDDs based on a criterion or criteria, which can relate to frequency of access to the data corresponding to the parity. The caching criterion can be defined to ensure the reduced overhead gained by reading parity values from a SSD outweighs any costs (e.g., SSD write endurance).

Abstract: Systems and techniques are disclosed for the mirroring of cache data from a storage controller to a storage class memory (“SCM”) device. The storage controller receives a write request, caches the write data, and mirrors the write data to the SCM device instead of to a cache of another storage controller. The SCM device stores the mirrored data in the SCM device. The storage controller acknowledges the write to the host. If the storage controller later fails, an alternate controller assumes ownership of storage volumes associated with the failed controller. Upon receipt of a new read request to the failed controller, the alternate controller checks the SCM device for a cache hit. If there is, the data is read from the SCM device; otherwise, it is read from the storage volume(s). The read data is cached at the alternate controller and then sent on to the requesting host.

Abstract: Systems and techniques for performing a data transaction are disclosed that provide data redundancy using two or more cache devices. In some embodiments, a data transaction is received by a storage controller of a storage system from a host system. The storage controller caches data and/or metadata associated with the data transaction to at least two cache devices that are discrete from the storage controller. After caching, the storage controller provides a transaction completion response to the host system from which the transaction was received. In some examples, each of the at least two cache devices includes a storage class memory. In some examples, the storage controller caches metadata to the at least two cache devices and to a controller cache of the storage controller, while data is cached to the at least two cache devices without being cached in the controller cache.

Abstract: A storage manager can reduce the overhead of parity based fault tolerance by leveraging the access performance of SSDs for the parities. Since reading a parity value can be considered a small read operation, the reading of parity from an SSD is an effectively “free” operation due to the substantially greater SSD read performance. With reading parity being an effectively free operation, placing parity on SSDs eliminates the parity read operations (in terms of time) from the parity based fault tolerance overhead. A storage manager can selectively place parity on SSDs from HDDs based on a criterion or criteria, which can relate to frequency of access to the data corresponding to the parity. The caching criterion can be defined to ensure the reduced overhead gained by reading parity values from a SSD outweighs any costs (e.g., SSD write endurance).