Techniques for cache management involve accessing, when a first data block to be accessed is missing in a first cache, the first data block from a storage device storing the first data block; selecting, when the first cache is full and based on a plurality of parameters associated with a plurality of eviction policies, an eviction policy for evicting a data block in the first cache from the plurality of eviction policies, the plurality of parameters indicating corresponding possibilities that the plurality of eviction policies are selected; evicting a second data block in the first cache to a second cache based on the selected eviction policy, the second cache being configured to record the data block evicted from the first cache; and caching the accessed first data block in the first cache. Such techniques can improve the cache hit rate, thereby improving the access performance of a system.

The present disclosure provides methods, apparatuses, and servers for cache-coherence. In some embodiments, an apparatus includes a plurality of compute express link (CXL) devices, and a switch. Each CXL device of the plurality of CXL devices includes a memory in which a portion of the memory is allocated as a cache buffer, to which different cache eviction policies are allocated. The different cache eviction policies are modified according to a cache hit ratio of the cache buffer. The switch is configured to connect the plurality of CXL devices to each other.

A data storage system with interconnected compute nodes includes a shared memory with volatile and non-volatile portions. Data tracks evicted from the volatile portion are moved to the non-volatile portion based on a cache-miss interarrival rate threshold that is calculated based on capacity and fall-through time of the non-volatile portion of the shared memory. Data extents on non-volatile drives are characterized based on dominant modes of extent-level cache-miss interarrival histograms generated using countdown timers of most recent backend accesses of sub-extents. The dominant mode of the extent in which a backend track evicted from the volatile portion of the shared memory resides is compared with the threshold in order to determine whether to move the backend track to the non-volatile portion of the shared memory.

Various embodiments include methods and devices for implementing a criterion aware cache replacement policy by a computing device. Embodiments may include updating a staling counter, writing a value of a local counter to a system cache in association with a location in the system cache for with data, in which the value of the local counter includes a value of the staling counter when (i.e., at the time) the associated data is written to the system cache, and using the value of the local counter of the associated data to determine whether the associated data is stale.

Apparatuses, systems, and techniques to control operation of a memory cache. In at least one embodiment, cache guidance is specified within application source code by associating guidance with declaration of a memory block, and then applying specified guidance to source code statements that access said memory block.

Method and apparatus for managing a front-end cache formed of ferroelectric memory element (FME) cells. Prior to storage of writeback data associated with a pending write command from a client device, an intelligent cache manager circuit forwards a first status value indicative that sufficient capacity is available in the front-end cache for the writeback data. Non-requested speculative readback data previously transferred to the front-end cache from the main NVM memory store may be jettisoned to accommodate the writeback data. A second status value may be supplied to the client device if insufficient capacity is available to store the writeback data in the front-end cache, and a different, non-FME based cache may be used in such case. Mode select inputs can be supplied by the client device specify a particular quality of service level for the front-end cache, enabling selection of suitable writeback and speculative readback data processing strategies.

A method of processing a workload in a graphics processing unit (GPU) may include detecting a work item of the workload in the GPU, determining a cache policy for the work item, and operating at least a portion of a cache memory hierarchy in the GPU for at least a portion of the work item based on the cache policy. The work item may be detected based on information received from an application and/or monitoring one or more performance counters by a driver and/or hardware detection logic. The method may further include monitoring one or more performance counters, wherein the cache policy for the work item may be determined and/or changed based on the one or more performance counters. The cache policy for the work item may be selected based on a runtime learning model.

Broadly speaking, embodiments of the present technique provide apparatus and methods for improved wear-levelling in (volatile and non-volatile) memories. In particular, the present wear-levelling techniques comprise moving static memory states within a memory, in order to substantially balance writes across all locations within the memory.

Techniques for cache management involve: determining respective elimination scores of a plurality of entries in a cache based at least in part on compression rates of data blocks corresponding to the plurality of entries, the elimination score being proportional to the compression rate; and removing, from the cache, at least one entry with a relatively low elimination score among the plurality of entries. Such techniques are able to optimize the retention and elimination strategies for entries in a cache, thus increasing payoffs for using the cache.

Provided is a resource manager. The resource manager may be a memory manager which is configured to determine a plurality of eviction rankings, wherein each of the plurality of eviction rankings assigns a position within the eviction ranking to each of a plurality of eviction candidates based on at least one eviction criterion. The memory manager may then select one of said eviction candidates by applying a voting algorithm to the plurality of eviction rankings and cause the selected eviction candidate to be evicted.