Re: Tom Lane 2019-03-22 <4368.1553270...@sss.pgh.pa.us> > Christoph Berg <m...@debian.org> writes: > > I think, the next step in that direction would be to enable data > > checksums by default. They make sense in most setups, > > Well, that is exactly the point that needs some proof, not just > an unfounded assertion.
I run a benchmark with checksums disabled/enabled. shared_buffers is 512kB to make sure almost any read will fetch the page from the OS cache; scale factor is 50 (~750MB) to make sure the whole cluster fits into RAM. model name: Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz (8 threads) alter system set shared_buffers = '512kB'; pgbench -s 50 -i pgbench -P 5 -M prepared -c 8 -j 8 -T 60 --select-only without checksums: tps = 96893.627255 (including connections establishing) tps = 97570.587793 (including connections establishing) tps = 97455.484419 (including connections establishing) tps = 97533.668801 (including connections establishing) average: 97363 with checksums: tps = 91942.502487 (including connections establishing) tps = 92390.556925 (including connections establishing) tps = 92956.923271 (including connections establishing) tps = 92914.205047 (including connections establishing) average: 92551 select 92551.0/97363; 0.9506 So the cost is 5% in this very contrived case. In almost any other setting, the cost would be lower, I'd think. Christoph
