Abstract: Provided is a memory management system based on a non-uniform memory access, which includes: a first persistent memory disposed in a first node; a second persistent memory disposed in a second node physically distinguished from the first node; a first journaling process disposed in the first persistent memory; and a second journaling process disposed in the second persistent memory.

Abstract: Provided is a method of configuring a cluster, which is a method of assigning graphics processing unit (GPU) servers in a cloud in which a plurality of machine learning (ML) services are executed using an apparatus for configuring a cluster. The apparatus for configuring a cluster is configured to measure the power consumption characteristics of each of the GPU servers constituting the cloud for each of a plurality of different models processing the plurality of ML services and assign at least one GPU server to each of the plurality of models using power consumption characteristics of each of the GPU servers for each of the plurality of models to configure a GPU cluster for each of the plurality of models.

Abstract: Provides is a method for memory swap. The method for memory swap according to the present disclosure may comprise storing a plurality of pages in a volatile memory; intercepting at least one page among the plurality of pages before the at least one page is discarded from the volatile memory; and compressing the at least one intercepted page and storing the compressed page in a memory swap space, wherein the plurality of pages include a first page, which is a clean page among file-mapped pages, and a second page, which is an anonymous page.

Abstract: Provided is a memory management system based on a non-uniform memory access, which includes: a first persistent memory disposed in a first node; a second persistent memory disposed in a second node physically distinguished from the first node; a first journaling process disposed in the first persistent memory; and a second journaling process disposed in the second persistent memory.

Abstract: Provides is a method for memory swap. The method for memory swap according to the present disclosure may comprise storing a plurality of pages in a volatile memory; intercepting at least one page among the plurality of pages before the at least one page is discarded from the volatile memory; and compressing the at least one intercepted page and storing the compressed page in a memory swap space, wherein the plurality of pages include a first page, which is a clean page among file-mapped pages, and a second page, which is an anonymous page.

Abstract: Disclosed herein are a thermal management apparatus and method using a dynamic thermal margin, and a semiconductor processor device, a non-volatile data storage device and an access control method using the same. The thermal management apparatus and method using a dynamic thermal margin, and the semiconductor processor device, non-volatile data storage device and access control method using the same can guarantee required performance based on importance and priority by scaling a thermal margin based on the importance of a task or the priority of threads, can control the generation of heat in a software manner while being compatible with non-volatile memory interface standards, and can also provide performance varying depending on the importance of a request that is being processed.

Abstract: Disclosed herein are a thermal management apparatus and method using a dynamic thermal margin, and a semiconductor processor device, a non-volatile data storage device and an access control method using the same. The thermal management apparatus and method using a dynamic thermal margin, and the semiconductor processor device, non-volatile data storage device and access control method using the same can guarantee required performance based on importance and priority by scaling a thermal margin based on the importance of a task or the priority of threads, can control the generation of heat in a software manner while being compatible with non-volatile memory interface standards, and can also provide performance varying depending on the importance of a request that is being processed.

Abstract: A method of storing data in a distributed manner based on data compression ratio prediction, and a mass storage device and system using the method are disclosed. The device includes a compression ratio predicting unit, a compressing unit, and a control unit. When an address and first unit sized data are received, the compression ratio predicting unit estimates the predicted compression ratio of the first unit sized data. The compressing unit generates compressed data. The control unit calculates the benefit of compression based on at least the estimated predicted compression ratio, stores the compressed data in a first storage area if the calculated benefit of compression is higher than a predetermined benefit threshold value, and stores the first unit sized data in the second storage area if the calculated benefit of compression is equal to or lower than the predetermined benefit threshold value.

Abstract: A method of storing data in a distributed manner based on data compression ratio prediction, and a mass storage device and system using the method are disclosed. The device includes a compression ratio predicting unit, a compressing unit, and a control unit. When an address and first unit sized data are received, the compression ratio predicting unit estimates the predicted compression ratio of the first unit sized data. The compressing unit generates compressed data. The control unit calculates the benefit of compression based on at least the estimated predicted compression ratio, stores the compressed data in a first storage area if the calculated benefit of compression is higher than a predetermined benefit threshold value, and stores the first unit sized data in the second storage area if the calculated benefit of compression is equal to or lower than the predetermined benefit threshold value.