On Thu, Oct 9, 2014 at 3:52 AM, pinker <pinker@onet.eu> wrote:
Hello All, I have a brand new machine to tune: x3750 M4, 4xE5-4640v2 2.2GHz; 512GBRAM (32x16GB), 4x300GB SAS + SSD (Easy Tier) in RAID 10
What's particularly important now is to choose optimal configuration for write operations. We have discussion about checkpoint_segments and checkpoint_timeout parameters. Test, which was based on pg_replay, has shown that the biggest amount of data is written when checkpoint_segments are set to 10 000 and checkpoint_timeout to 30 min, but I'm worrying about amount of time needed for crash recovery.
Since you already have pg_replay running, kill -9 some backend (my favorite victim is the bgwriter) during the middle of pg_replay, and see how long it takes to recover.
You might want to try it with and without clobbering the FS cache, or simply rebooting the whole machine, depending on what kind of crash you think is more likely.
Recovering into a cold cache can be painfully slow. If your database mostly fits in memory, you can speed it up by using something (like "tar -cf - pgdata | wc -c" to) read the entire pg data directory in sequential fashion and hopefully cache it. If you find recovery too slow, you might want to try to this trick (and put it in your init scripts) rather than lowering checkpoint_segments.