The Design Of CUDA Based Table Scan Accelerator For PostgreSQL

Due to technology advancements and business growth,modern enterprise and academes are receiving more data than any time in history in their area.To gain an advantage in business or academic research,data owners need to find valuable business insight or research progress by analyzing large scale of data.Thus,the ability to store and manage big data is required by modern database.In traditional database manage systems(DBMS),the workloads are all handled by CPU resources.With the lack of massive thread parallel processing ability in CPUs,the data-intensive workloads from DBMS are challenging the computing performance of nowadays CPU(Central Processing Unit).On the other hand,GPUs(Graphics Processing Units)are designed naturally for massive thread parallelism along with a high memory bandwidth.When scanning tables in a DBMS,traditional CPU method need to process all the data in tables serially.However,if the scan task is handled properly in the GPU case,massive rows from the table can be scanned in parallel.This can be a significate speed improvement in database queries.In this paper,the main effort is to achieve an GPU-based accelerator for DBMS PostgreSQL.The interface of data transferring between PostgreSQL and GPU device is the extension of PostgreSQL.And the table scan task in PostgreSQL can be divided into massive parallel scan threads handled in GPU.With the help of native interface in PostgreSQL for developers,the GPU accelerator works with on influence on original PostgreSQL functions.At the same time,the GPU accelerator provides a platform for further development.In this paper,how data transmitting time and calculation time is affecting the efficiency of GPU accelerator is also discussed.The final result shows an overall time decrease from 45% to 50% in PostgreSQL table scan queries with large enough data table.