Described apparatuses and methods track access metadata pertaining to activity within respective address ranges. The access metadata can be used to inform prefetch operations within the respective address ranges. The prefetch operations may involve deriving access patterns from access metadata covering the respective ranges. Suitable address range sizes for accurate pattern detection, however, can vary significantly from region to region of the address space based on, inter alia, workloads produced by programs utilizing the regions. Advantageously, the described apparatuses and methods can adapt the address ranges covered by the access metadata for improved prefetch performance. A data structure may be used to manage the address ranges in which access metadata are tracked. The address ranges can be adapted to improve prefetch performance through low-overhead operations implemented within the data structure. The data structure can encode hierarchical relationships that ensure the resulting address ranges are distinct.

Described is a computing system for vector prefetching which includes a hierarchical memory including multiple caches, a missing address storage unit (MASU) associated with each cache which stores prefetch requests suffering a cache miss, a prefetcher which sends prefetch requests towards the hierarchical memory, and a vector prefetch unit. The vector prefetch unit determines existence of at least one of a relationship between a cache block associated with the prefetch request and cache blocks associated with one or more entries in a MASU, or a relationship between cache blocks associated with different entries in a MASU, and sends a vector prefetch request based on related prefetch requests including indicators indicating a starting cache block and a number of related cache blocks to a higher memory level to obtain data associated with each cache block. The hierarchical memory stores the data received in at least one response message from the higher memory level if available.

A multi-memory apparatus that uses machine learning is described. The apparatus may include an interface controller, a non-volatile memory, and a volatile memory. The interface controller may cause the apparatus to receive a first command from a host device. The interface controller may cause the apparatus to communicate the first command to a machine learning engine and to circuitry configured to store and manage commands for the non-volatile memory and the volatile memory. The interface controller may further cause the apparatus to communicate a second command generated by the machine learning engine to the circuitry. The second command may be based on information determined by the machine learning engine during a training mode.

A computing system having memory components, including first memory and second memory. The computing system further includes a processing device, operatively coupled with the memory components, to: receive, in a prediction engine, usage history of pages in the second memory; train a prediction model based on the usage history; predict, by the prediction engine using the prediction model, likelihood of the pages being used in a subsequent period of time; and responsive to the likelihood predicted by the prediction engine, copy by a controller data in a page in the second memory to the first memory.

Methods, systems, and apparatus, including computer programs encoded on computer storage media, relating to multi-task recurrent neural networks. One of the methods includes maintaining data specifying, for a recurrent neural network, a separate internal state for each of a plurality of memory regions; receiving a current input; identifying a particular memory region of the memory access address defined by the current input; selecting, from the internal states specified in the maintained data, the internal state for the particular memory region; processing, in accordance with the selected internal state for the particular memory region, the current input in the sequence of inputs using the recurrent neural network to: generate an output, the output defining a probability distribution of a predicted memory access address, and update the selected internal state of the particular memory region; and associating the updated selected internal state with the particular memory region in the maintained data.

Disclosed in some examples are improved address prediction and memory preloading that leverages next-delta prediction and/or far-delta prediction for scheduling using a DNN. Previous memory access sequence data that identify one or more memory addresses previously accessed by one or more processors of a system may be processed and then converted into a sequence of delta values. The sequence of delta values are then mapped to one or more classes that are then input to a DNN. The DNN then outputs a predicted future class identifier sequence that represents addresses that the DNN predicts will be accessed by the processor in the future. The predicted future class identifier sequence is then converted back to a predicted delta value sequence and back into a set of one or more predicted addresses.

An indication that a virtual machine is starting is received. Requested data blocks associated with the virtual machine are identified. Based on identifiers of the requested data blocks, a trained learning model is used to predict one or more subsequent data blocks likely to be requested while the virtual machine is starting. The one or more subsequent data blocks are caused to be preloaded in a cache storage.

Described apparatuses and methods track access metadata pertaining to activity within respective address ranges. The access metadata can be used to inform prefetch operations within the respective address ranges. The prefetch operations may involve deriving access patterns from access metadata covering the respective ranges. Suitable address range sizes for accurate pattern detection, however, can vary significantly from region to region of the address space based on, inter alia, workloads produced by programs utilizing the regions. Advantageously, the described apparatuses and methods can adapt the address ranges covered by the access metadata for improved prefetch performance. A data structure may be used to manage the address ranges in which access metadata are tracked. The address ranges can be adapted to improve prefetch performance through low-overhead operations implemented within the data structure. The data structure can encode hierarchical relationships that ensure the resulting address ranges are distinct.

Methods, systems, and devices supporting techniques for pre-fetching information using pattern detection are described. Some memory systems may support pre-fetching information, such as logical-to-physical (L2P) mapping tables, data, or both, if a sequential pattern of read commands is detected. In some examples, the memory system may store a list of logical addresses indicated by received read commands and may determine whether the list corresponds to a sequential pattern independent of intervening write-alike commands. The list may store previous logical addresses for read commands, allowing the memory system to determine whether subsequent read commands form a sequential pattern. Additionally or alternatively, the memory system may track a ratio of hibernate commands to other commands (e.g., sequential read commands) and may refrain from pre-fetching L2P mapping tables for a detected sequence if the tracked ratio satisfies (e.g., exceeds) a threshold ratio.

A system and method improve caching efficiency in a data storage system by performing machine learning processes on metadata relating to extents of data blocks, rather than individual blocks themselves. Thus, once the storage devices are divided into extents, various metadata regarding access to the blocks within each extent are aggregated, and per-extent features are extracted. These features are used to train a data regression model that is subsequently used to infer a most likely “hotness” value for each extent at a future time. These predicted values, which may be further classified as e.g. “hot”, “warm”, and “cold” using thresholds, are used to implement the cache replacement policy. Embodiments scale to large and multi-layered caches, and may avoid common caching problems like thrashing, by adjusting the extent size. Policy goal functions may be optimized by dynamically adjusting the classification thresholds.