Database test: dual Intel Xeon 5160

In September, AMD published some remarkable benchmarks, seemingly intended to draw the attention of the tech crowd to hardware based prefetchers. The SPECjbb2005 benchmark showed the server to be some 14% slower as soon as the prefetchers were turned on, in spite of Intel's attempts to avoid these negative side effects. One explanation is that software based and hardware based prefetchers may get in each other's ways, but that is little more than speculation. To examine how the prefetchers influenced our tests, we temporarily switched off the two options in the Apollo 5's BIOS labelled 'Hardware Prefetcher' and 'Adjacent Cache Line Prefetcher'. This catered for a different picture than we expected on the basis of SPECjbb2005: in our case, the prefetchers most certainly improved things, although not to the same extent in all cases. The gains are smallest for MySQL 5.0.20a while PostgreSQL 8.2 has the most to gain.

Influence of prefetchers	MySQL 4.1.20	MySQL 5.0.20a	PostgreSQL 8.2-dev
1x single-core	+13.1%	+5.1%	+9.9%
2x single-core	+8.8%	+8.8%	+11.7%
1x dual-core	+8.3%	+12.4%	+16.3%
2x dualcore	+6.2%	+2.6%	+10.6%
Average	+9.1%	+7.2%	+12.1%