GPU GPU GPU memory scheduler and GPU memory preemption method using the same

摘要

The present invention relates to a method for preempting a GPU memory by a hooking module of a GPU memory scheduler. The method for preempting a GPU memory by a hooking module of a GPU memory scheduler comprises: hooking a calling function when a function calling a shared library in a user application is a GPU memory allocation requesting function; checking a requested GPU memory capacity from a variable of the hooked calling function and asking a node scheduler whether GPU memory allocation of the checked requested capacity to a user application is possible; and changing a first memory address, which is another variable of the hooked calling function, into a second memory address when the GPU memory allocation is approved from the node scheduler, and transferring the GPU memory allocation requesting function having the second memory address and the requested capacity as variables to the shared library so that a GPU memory is allocated from the second memory address as much as the requested capacity.