Ingres just published a TPC-H benchmark for VectorWise, an analytic database technology employing
1) SIMD processing (Intel SSE 4.2),
2) better memory optimizations to leverage on-chip cache,
3) compression,
4) Column-based storage.
Ingres originated as a research project at UC Berkeley (see Wikipedia) in the 1970s, and has since become a commercially supported, open source database system. Apparently, Ingres project people later founded Sybase. So Ingres in a sense, is the grandfather (or perhap grand-step-uncle?) to SQL Server. According to the Ingres whitepaper, VectorWise spun off from CWI in Amstersdam, with some roots in MonetDB? For the time being, Ingres VectorWise is not open-source. The interesting aspect of VectorWise among other column store database products is use of the SIMD, as in utilizing the special instructions and registers built into recent generation microprocessors. (In a previous blog, I proposed special SIMD instructions to improve the efficiency on accessing pages in traditional row-storage database engines.)
Below are TPC-H results at scale factor 100GB on HP ProLiant DL380 G7 with two Xeon 5680 six-core 3.33GHz processors for SQL Server 2008 R2 and Ingres VectorWise 1.5.
PowerThroughputQphHSQL Server99,626.355,038.273,974.6VectorWise257,142.9246,101.7251,561.7
Overall, the VectorWise columnar database has about 2.5X higher score on Power, and 5X on Throughput. The glaring weak point of SQL Server is that the throughput is (almost 2X) lower than the power. The SQL Server parallel execution engine is very powerful running a single stream with no other concurrent queries competing for CPU. It does not have a good strategy for sharing compute resources with other concurrent long running queries.
The general expectation set by all the new technology, columnar storage, memory optimization and vector processing was 10X, not 3X. However, we should consider that the TPC-H benchmark score is a composite geometric mean of 22 individual queries. The individual query performance differences between SQL Server and VectorWise are shown below, on a logarithmic scale (base 2, not 10, a cool feature in Excel 2007?).
