Paper Accepted at the ACM/SIGAPP Symposium on Applied Computing (First author: Jung Min-gyo; corresponding author: Professor Kim Young-jae)

△ From left: Jung Min-gyo (master's, 4th semester; supervisor: Professor Park Sung-yong), Lee Chang-gyu (combined master's–doctoral, 7th semester; supervisor: Professor Kim Young-jae), Professor Park Sung-yong, Professor Kim Young-jae (corresponding author)

A paper titled 'GPUKV: An Integrated Framework with KVSSD and GPU through P2P Communication Support', written by graduate students Jung Min-gyo (master's, 4th semester; supervisor: Professor Park Sung-yong) and Lee Chang-gyu (combined master's–doctoral, 7th semester; supervisor: Professor Kim Young-jae) of the Department of Computer Science & Engineering under the supervision of Professors Park Sung-yong and Kim Young-jae (corresponding author), has been accepted for publication at the ACM/SIGAPP Symposium on Applied Computing, Operating Systems Track, listed at an adjusted IF of 1 among the outstanding international conferences in computer science under the BK21 Plus programme.

The research optimizes existing file system-based key-value stores for use without CPU intervention on GPGPU, which allows GPUs to be used for general-purpose computing. Data transfer and host system optimization were achieved through direct data transfer (PCIe P2P) between the GPU and KVSSD, an SSD acting as a key-value store. Compared with systems that perform GPU work using a file system-based key-value store, the approach achieved up to 18.7 times faster execution and up to 175 times greater savings in host CPU resources.

△ Figure 1. Data and control paths of existing systems for using a key-value store △ Figure 2. Data and control paths of the system proposed in this paper

Figure 1 shows the data transfer and control paths of an existing file system-based key-value store. To use data from the key-value store, communication must pass through the file system and the host kernel's NVMe driver, and data loaded into the CPU's main memory must then be moved again to the GPU. This process entails a great deal of unnecessary data copying and complex control paths. Figure 2 shows the data transfer and control paths of the new system proposed in this paper. Compared with the existing system, data copying is reduced and the host system's complex control paths need not be traversed.

The research was conducted jointly with researchers at SK hynix and carried out with SK hynix's support.

The ACM/SIGAPP SAC conference is a distinguished international conference covering a wide range of computing fields. It was to be held in Gwangju from 22 to 26 March next year, but owing to COVID-19 it will be held as an online conference.

● ACM/SIGAPP SAC conference site: https://sigapp.org/sac/sac2021/ ● 'GPUKV: An Integrated Framework with KVSSD and GPU through P2P Communication Support' paper: https://lass.sogang.ac.kr/pdf/2021/intl_conf/SAC_2021_MJeong.pdf