Apparatus and method for cache-coherence
AssigneeSamsung Electronics Co., Ltd.
InventorSoo-Young Ji
Life of this patent
Estimating the term…
Abstract
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.
Claims19
An apparatus, comprising:
- a plurality of compute express link (CXL) devices comprising first CXL devices and second CXL devices different from the first CXL devices, wherein each CXL device of the plurality of CXL devices comprises a memory in which a portion of the memory is allocated as a cache buffer, wherein each cache buffer is initially allocated a different cache eviction policy based on a corresponding measured latency of that cache buffer, wherein each cache buffer of the first CXL devices is initially allocated a first cache eviction policy and each cache buffer of the second CXL devices is initially allocated a second cache eviction policy different from the first cache eviction policy, and wherein allocation of the different cache eviction policies to the first CXL devices is modified based on a calculated ratio of a first average hit ratio of the cache buffer of the first CXL devices to a second average hit ratio of the cache buffer of the second CXL devices being different from an initial allocation ratio;
- and
- a switch configured to connect the plurality of CXL devices to each other.
A server, comprising:
- a plurality of first devices comprising a cache buffer that comprises a plurality of areas;
- and
- a second device configured to:
- initially allocate, based on a corresponding measured latency, a corresponding cache eviction policy from among a plurality of cache eviction policies to each of one or more first areas allocated among the plurality of areas of the plurality of first devices;
- and
- modify allocation of the corresponding cache eviction policy allocated to each of the one or more first areas based on a calculated ratio of a first average cache hit ratio of the cache buffer of the plurality of first devices to a second average cache hit ratio of one or more second areas of the plurality of areas different from the one or more first areas being different from an initial allocation ratio.
A method performed by a computing system, comprising:
- determining latencies of a plurality of compute express link (CXL) devices;
- allocating a cache eviction policy of a plurality of cache eviction policies to each of the plurality of CXL devices, according to the latencies of the plurality of CXL devices, the plurality of cache eviction policies comprising a first cache eviction policy and a second cache eviction policy, the plurality of CXL devices comprising a first quantity of first CXL devices to which the first cache eviction policy has been allocated and a second quantity of second CXL devices to which the second cache eviction policy has been allocated;
- checking cache hit ratios of the plurality of CXL devices;
- and
- modifying allocation of the cache eviction policy of at least one of the plurality of CXL devices to at least one of the first cache eviction policy and the second cache eviction policy, based on a first ratio of an average cache hit ratio of the first CXL devices and an average cache hit ratio of the second CXL devices, and a second ratio of the first quantity of the first CXL devices and the second quantity of the second CXL devices.
Dependencies are inferred from the claim text.