Abstract: Disclosed herein is a method for distributed training of an AI model in a channel-sharing network environment. The method includes determining whether data parallel processing is applied, calculating a computation time and a communication time when input data is evenly distributed across multiple computation devices, and unevenly distributing the input data across the multiple computation devices based on the computation time and the communication time.

Abstract: Disclosed herein are a method for machine-learning parallelization using host CPUs of a multi-socket structure and an apparatus therefor. The method, performed by the apparatus for machine-learning parallelization using host CPUs of a multi-socket structure, includes a compile phase in which a learning model is split at a layer level for respective pipeline stages and allocated to Non-Uniform Memory Access (NUMA) nodes for respective CPU sockets and a runtime phase in which parameters required for learning are initialized and multiple threads generated in consideration of a policy of each parallelism algorithm are executed by being allocated to respective cores included in the NUMA node.

Abstract: Disclosed herein are a hybrid memory management apparatus and method for an many-to-one virtualization environment. The hybrid memory management apparatus is implemented in an inverse-virtualization-based multi-node computing system including multiple physical nodes, each containing hybrid memory in which DRAM and NVRAM coexist, a virtual machine, and hypervisors, and includes memory for storing at least one program, and a processor for executing the program, wherein the program includes a remote request service module for processing a page-related request with reference to the hybrid memory and responding to the page-related request by transmitting a result of processing, an internal request service module for processing an internal page fault request with reference to a hybrid memory and responding to the internal page fault request, and a data arrangement module for responding to an inquiry request for a location at which a newly added page is to be arranged in the hybrid memory.

Abstract: Disclosed herein are an apparatus and method for page allocation in a many-to-one virtualization environment. The method may include determining whether a page fault interrupt is caused by page initialization for page allocation, sending an ownership change message to a node having ownership of the corresponding page when the page fault interrupt is determined to be caused by page initialization, and initializing the corresponding page upon receiving an ownership-change-processing-complete message.

Abstract: Disclosed herein are an apparatus for preloading data in a distributed computing environment and a method using the same. The method includes selecting a local preloading target that each of multiple computers connected over a network is to preload into the local memory thereof, registering a local preloading task corresponding to the local preloading target in local preloading metadata, and asynchronously starting the local preloading task at a preset time based on the local preloading metadata. The local preloading metadata is stored in a page other than the page in which remote preloading metadata for managing a remote preloading task is stored.

Abstract: Disclosed herein are an apparatus and method for processing fast persistent writing to nonvolatile memory. A method for processing fast persistent writing to a nonvolatile memory includes writing log data from a last storage position in a data block created in a certain size, reflecting the entire data block in a storage medium, increasing the last storage position by a size of the written log data, and reflecting the entire data block in a storage medium, wherein the data block is linearly linked to an additional data block in a creation order, and wherein the data block includes a log data storage region, information about a last storage position where the log data is stored in the log data storage region, and information about a linked next data block.

Abstract: Disclosed herein are a synchronization method for improving the concurrent read performance of a critical section in distributed shared memory and an apparatus for the same. The synchronization method, performed by a distributed-shared-memory management apparatus in a physical node of a multi-node system, includes checking whether a lock is held on each node based on a read-write lock having lock variables for respective nodes in a distributed shared memory environment, acquiring a lock for a read operation or a write operation in consideration of whether a lock is held on each node, and releasing the lock based on the lock variables for the respective nodes when the read operation or the write operation is terminated.

Abstract: Disclosed herein are a hybrid memory management apparatus and method for an many-to-one virtualization environment. The hybrid memory management apparatus is implemented in an inverse-virtualization-based multi-node computing system including multiple physical nodes, each containing hybrid memory in which DRAM and NVRAM coexist, a virtual machine, and hypervisors, and includes memory for storing at least one program, and a processor for executing the program, wherein the program includes a remote request service module for processing a page-related request with reference to the hybrid memory and responding to the page-related request by transmitting a result of processing, an internal request service module for processing an internal page fault request with reference to a hybrid memory and responding to the internal page fault request, and a data arrangement module for responding to an inquiry request for a location at which a newly added page is to be arranged in the hybrid memory.

Abstract: A distributed storage server includes a plurality of data server devices and a plurality of metadata server devices. The metadata server devices store metadata associated with data which is distributively stored in the data server devices. A selected metadata server device checks whether a dangling directory occurs by performing a rename operation, based on information associated with a full path of a parent directory of a source and a full path of a parent directory of a target included in a request of the rename operation. When it is determined that the dangling directory does not occur, the selected metadata server device processes transactions directed to the metadata such that the rename operation is performed. The checking whether the dangling directory occurs is performed before a transaction period for processing the transactions.

Abstract: A distributed file system includes metadata servers and data servers. The metadata server includes a selecting unit to select a data server from the data servers, a chunk allocation requesting unit to request that the selected data server perform chunk allocation, a chunk list managing unit to insert a list of chunks transmitted from the data server into a chunk list pool and determine an arrangement method of the chunk list pool, and a chunk fetching unit to fetch available chunk information from the chunk list pool. The data server includes a receiving unit to receive request for chunk allocation from the metadata server, a chunk allocating unit to allocate chunks in response to the request for chunk allocation and write a list of chunks based on information about the allocated chunk, and a transmitting unit to transmit the list of chunks to the metadata server.

Abstract: Disclosed herein is an apparatus for controlling synchronization of metadata, which includes a transaction creation unit for creating a transaction corresponding to a request for an operation for metadata, the request being received from a client; a journal management unit for storing the transaction in a journal; a journal synchronization unit for comparing the journal with an external journal in a node connected with the apparatus and transmitting and receiving only inconsistencies therebetween, thereby synchronizing the journal; and an operation-processing unit for processing the request for the operation or the transaction for the metadata.

Abstract: A distributed storage server includes a plurality of data server devices and a plurality of metadata server devices. The metadata server devices store metadata associated with data which is distributively stored in the data server devices. A selected metadata server device checks whether a dangling directory occurs by performing a rename operation, based on information associated with a full path of a parent directory of a source and a full path of a parent directory of a target included in a request of the rename operation. When it is determined that the dangling directory does not occur, the selected metadata server device processes transactions directed to the metadata such that the rename operation is performed. The checking whether the dangling directory occurs is performed before a transaction period for processing the transactions.

Abstract: A distributed file system includes metadata servers and data servers. The metadata server includes a selecting unit to select a data server from the data servers, a chunk allocation requesting unit to request that the selected data server perform chunk allocation, a chunk list managing unit to insert a list of chunks transmitted from the data server into a chunk list pool and determine an arrangement method of the chunk list pool, and a chunk fetching unit to fetch available chunk information from the chunk list pool. The data server includes a receiving unit to receive request for chunk allocation from the metadata server, a chunk allocating unit to allocate chunks in response to the request for chunk allocation and write a list of chunks based on information about the allocated chunk, and a transmitting unit to transmit the list of chunks to the metadata server.

Abstract: A method and system that can completely delete garbage data in a distributed network system are provided. Because it is impossible to initially access a data server, data to delete is not deleted, and thus when a garbage file is generated, a generated garbage file can be completely deleted. In this case, by performing a deletion operation of a garbage file in a distributed data server unit, operation efficiency can be maximized.

Abstract: Disclosed are an apparatus, a system, and a method for recovering meta data using fragmentary information. The system for recovering meta data includes: a data server storing recovery information for recovering at least one data file and meta data related to each of the data file; and a meta data server storing the meta data related to the data files and receiving the recovery information from the data server when the meta data are lost to recover the lost meta data. By this configuration, the exemplary embodiment of the present invention stably can the lost meta data of the meta data server by using the recovery the meta data and prevent the phenomenon that the meta data collides with each other on the same directory hierarchical structure.

Abstract: Provided are an apparatus and a method for acquiring a consistent result even if one client accesses any metadata server in a circumstance that the cluster configuring many metadata servers exist. The apparatus as mentioned above includes a deletion file list unit that stores information about a deleted metadata during a predetermined period; and a metadata server selector that selects a metadata server to receive a operation request about a metadata among a plurality of metadata servers based on information from the deletion file list unit.

Abstract: Disclosed are an apparatus, a system, and a method for recovering meta data using fragmentary information. The system for recovering meta data includes: a data server storing recovery information for recovering at least one data file and meta data related to each of the data file; and a meta data server storing the meta data related to the data files and receiving the recovery information from the data server when the meta data are lost to recover the lost meta data. By this configuration, the exemplary embodiment of the present invention stably can the lost meta data of the meta data server by using the recovery the meta data and prevent the phenomenon that the meta data collides with each other on the same directory hierarchical structure.

Abstract: Provided are a file storage system, and a method for managing duplicate files in the file storage system. The file storage system that performs file duplication for error recovery includes a client for requesting generation of a duplicate file of a new file being stored, a metadata server for storing information associated with prestored duplicate files and determining whether a file that is identical to the new file is present among the prestored duplicate files, to limit a number of identical duplicate files to a preset number, and a data server comprising a file duplication unit and a storage unit for respectively performing file duplication and file storage under the control of the metadata server. Accordingly, file recovery can be performed while duplicate files in the system are minimized.

Abstract: Provided is a system and method for updating a plurality of the same files consistently. The file update system comprises a first and a second data servers formed as pipe line type network for storing an original file and a duplicate file of the original file respectively, and for updating both the original file and the duplicate file in response to an update request, wherein the first and the second data servers mutually exchange file update information, correct the file update information so as to prevent overlapped update due to a plurality of update requests for the same file, and then update the original file or the duplicate file according to the corrected file update information.

Abstract: Provided are an apparatus and a method for acquiring a consistent result even if one client accesses any metadata server in a circumstance that the cluster configuring many metadata servers exist. The apparatus as mentioned above includes a deletion file list unit that stores information about a deleted metadata during a predetermined period; and a metadata server selector that selects a metadata server to receive a operation request about a metadata among a plurality of metadata servers based on information from the deletion file list unit.