With the rise of today's GPU supercomputers, the database manager becomes more important for many real-time systems. The problem is that the memory size is limited by the each single GPU device on system. This leads to the requirement of such database that can access for each device and the uniformity of all device's database on system (Toan and Yasushi, Audio fingerprint hierarchy searching on massively parallel with multi-GPGPUs using k-modes and lsh, 49-54, 2016) [1]. In this research, we focus the management of unstructured dataset on single GPU device and the communication between sub-database that are stored separate GPU device. Specifically, we considered the database manager for audio fingerprint storage for a real-time retrieval system, where each audio fingerprint is a binary sequence that holds the audio information for a source audio track. The system uses the memory on GPGPU devices for storing the sub-databases, that can support fast selecting, inserting, removing. With the significant result, our method can be applied to different datasets with high performance in the field of parallel processing on GPGPU.
