PM-DB: Partition-based Multi-instance Database System for Multicore Platforms
Xi Fang, Takeshi Mishima, Haruo Yokota · 2015
The continued evolution of modern hardware has brought several new challenges to database management systems (DBMSs).Multicore CPUs are now mainstream, while the future lies in massively parallel computing performed on many-core processors.However, because they were developed originally for single-core processors, DBMSs cannot take full advantage of the parallel computing that uses so many cores.Several components in the traditional database engines become new bottlenecks on multicore platforms.In this paper, we analyze the bottlenecks in existing database engines on a modern multicore platform using the mixed workload of the TPC-W benchmark and describe strategies for higher scalability and throughput for existing DBMSs on multicore platforms.First, we show how to overcome the limitations of the database engine by introducing a partition-based multi-instance database system on a single multicore platform without any modification of existing DBMSs.Second, we analyze the possibility of further improving the performance by optimizing the cache performance of concurrent queries.Implemented by middleware, our proposed PM-DB can avoid the challenging work of modifying existing database engines.Performance evaluation using the TPC-W benchmark revealed that our proposal can achieve at most 2.5 times higher throughput than the existing engine of PostgreSQL.