On Wed, Jul 28, 2010 at 9:18 AM, Yeb Havinga <yebhavi...@gmail.com> wrote:
> Yeb Havinga wrote: > >> Due to the LBA remapping of the SSD, I'm not sure of putting files that >> are sequentially written in a different partition (together with e.g. >> tables) would make a difference: in the end the SSD will have a set new >> blocks in it's buffer and somehow arrange them into sets of 128KB of 256KB >> writes for the flash chips. See also http://www.anandtech.com/show/2899/2 >> >> But I ran out of ideas to test, so I'm going to test it anyway. >> > Same machine config as mentioned before, with data and xlog on separate > partitions, ext3 with barrier off (save on this SSD). > > pgbench -c 10 -M prepared -T 3600 -l test > starting vacuum...end. > transaction type: TPC-B (sort of) > scaling factor: 300 > query mode: prepared > number of clients: 10 > duration: 3600 s > number of transactions actually processed: 10856359 > tps = 3015.560252 (including connections establishing) > tps = 3015.575739 (excluding connections establishing) > > This is about 25% faster than data and xlog combined on the same > filesystem. > > The trick may be in kjournald for which there is 1 for each ext3 journalled file system. I learned back in Red Hat 4 pre U4 kernels there was a problem with kjournald that would either cause 30 second hangs or lock up my server completely when pg_xlog and data were on the same file system plus a few other "right" things going on. Given the multicore world we have today, I think it makes sense that multiple ext3 file systems, and the kjournald's that service them, is faster than a single combined file system. Greg