Open Source Summit Japan

KaiGai Kohei presents a new feature built on PostgreSQL to accelerate analytic queries, with leverage of heterogeneous hardware. PostgreSQL has already allowed off-loading some CPU intensive SQL workloads onto GPU device, to utilize its multi-thousands cores to process tons of records. Not only massive parallel computing, GPU also has an interesting feature which runs peer-to-peer DMA from other PCIe device to GPU. It makes possible to load the data blocks of PostgreSQL on NVMe-SSD to GPU directly. Once data blocks get loaded, we already have a mechanism to run WHERE/JOIN/GROUP BY on GPU, then write back its execution results to host RAM. It is usually much smaller than raw data, thus less amount of records CPU must process. From the standpoint of applications, a couple of SSD and GPU look like an intelligent storage that understand and run SQL then returns minimum necessary data set.