Electronics and Telecommunications Research Institute

Eun-Ji Lim

Shin-Young Ahn

Wan Gyun Choi

Yong-Seok Choi

Young-Choon Woo

Chei-Yol Kim

Sung-In Jung

Young-Ho Kim

Sung Hoon Son

Seung-Jo Bae

Gyu-Il Cha

Jin-Mee Kim

Sung-Hoon Son

Disclosed herein are a parameter server and a method for sharing distributed deep-learning parameters using the parameter server. The method for sharing distributed deep-learning parameters using the parameter server includes initializing a global weight parameter in response to an initialization request by a master process; performing an update by receiving a learned local gradient parameter from the worker process, which performs deep-learning training after updating a local weight parameter using the global weight parameter; accumulating the gradient parameters in response to a request by the master process; and performing an update by receiving the global weight parameter from the master process that calculates the global weight parameter using the accumulated gradient parameters of the one or more worker processes.

Parameter server and method for sharing distributed deep learning parameter using the same

Disclosed herein are an apparatus and method for interfacing with common memory. The apparatus for interfacing with common memory includes a computer-input/output (I/O)-interface-protocol-processing unit for receiving a packet for accessing common memory from a computer; a direct memory access unit for transforming the packet into an instruction for performing any one of reading from and writing to the common memory; and a common memory interface unit for transmitting the instruction to the common memory and receiving information about whether execution of the instruction is completed from the common memory.

Apparatus and method for interfacing with common memory

A parameter-sharing apparatus includes a memory allocation unit for managing allocation of a memory area, in which a parameter is to be stored, to a memory box, and updating a mapping table stored in the memory box based on allocation management of the memory area, and an operation processing unit for providing the memory allocation unit with parameter information required for the allocation management of the memory area in which the parameter is to be stored and sharing the parameter stored in the memory box.

Parameter-sharing apparatus and method

Disclosed herein are a resource manager node and a resource management method. The resource manager node includes a resource management unit, a resource policy management unit, a shared resource capability management unit, a shared resource status monitoring unit, and a shared resource allocation unit. The resource management unit performs an operation necessary for resource allocation when a resource allocation request is received. The resource policy management unit determines a resource allocation policy based on the characteristic of the task, and generates resource allocation information. The shared resource capability management unit manages the topology of nodes, information about the capabilities of resources, and resource association information. The shared resource status monitoring unit monitors and manages information about the status of each node and the use of allocated resources. The shared resource allocation unit sends a resource allocation request to at least one of the plurality of nodes.

Apparatus and method for managing resources in cluster computing environment

There is provided a socket compatibility layer for a TOE in a high quality server system. The socket compatibility layer for a TOE provides a standard socket API to an application program and provides binary-level compatibility so that a networking function of a TOE can be used without modifying the conventional socket application program thereby supporting simultaneous use of a TOE and a general Ethernet-based NIC. In a kernel structure of the Linux operation system, the socket compatibility layer for a TOE is located below a BSD socket layer receiving a socket system call from an application program through a standard socket API, is located above an INET socket layer and a TOE driver to control connection between an upper layer and a lower layer, and transmits the socket system call received from the BSD socket layer to the TOE driver or the INET socket layer according to call contents of the socket system call. The socket compatibility layer for a TOE includes a TOE socket generation module, a virtual TOE socket, a TOE STREAM module unit and a TOE DGRAM module unit.

Socket compatibility layer for TOE

The present invention propose a new mechanism, IP check layer which enables to use TOE cards and legacy NICs simultaneously within same protocol family in a Linux or a UNIX computing environment. The IP check layer is inserted between a BSD socket layer and an INET socket layer. IP check layer maintains a routing information and, information of IP addresses of each network interfaces and each socket's bind state. Using this information, IP check layer determines how to process a socket call and which layer (TOE layer, host INET layer) will take over the socket call. By means of the above mechanism, the present invention makes existing socket applications to use TOE cards and legacy NICs at the same time without modification or recompiling.

Growing community of inventors

Daejeon, South Korea

Eun-Ji Lim