DOE: database offloading engine for accelerating SQL processing

DOE: database offloading engine for accelerating SQL processing

연구 분야: Databases

학회: Distributed and Parallel Databases

The CPU-Accelerator heterogeneous systems have demonstrated performance and efficiency benefits on DBMSs. However, the CPU-Cache-DRAM architecture can not fully utilize the bandwidth of DRAMs such that in-memory approach get limited improvement. To overcome this drawback, it is non-trivial to customize efficient domain-specific accelerators and efficiently shuttle data between the host memory space and accelerator. But even if high-performance accelerators are available for DBMS, it is challenging to integrate the software with accelerator non-intrusively. To address these problems, we propose a hardware-software co-designed system, database offloading engine (DOE), which contains hardware accelerator architecture—Conflux for effective SQL operation offloading, and a software DOE programming platform—DP2 for application integration and seamless harness of the computing power. We subtly partition each well-known relational operator, such as filter, join, group by, aggregate, and sort, and dynamically map these operators on multiple kernels in parallel. The DOE kernels work in streaming processing mode, over which the microarchitecture aggressively exploits data parallelism and memory bandwidth. The experiment results show that DOE achieves more than 100x and 10x performance improvement compared with PostgreSQL and MonetDB respectively.