> On Sat, Feb 12, 2011 at 9:48 PM, Marcus Grimm > <mgr...@medcom-online.de>wrote: > >> > I should've realized it wasn't running this fast but the small 5000 >> record >> > size got me. >> > Test it yourself. >> > I do have a 7200RPM drive. My 261.4 numer is still 2+X your >> theoretical. >> >> I don't want to be a smart-arse, but I still think your 261.4 is to >> fast. >> On a 7200 RPM drive one will have 125 chances/sec to see a sector to be >> written. Since sqlite, under normal journal mode, will need 3 syncs >> per commit as far as I can recall, the maximum number drops further >> down to 41 commit/sec. This is theoretical, in reality one will see >> maybe 20 commits/sec. Not sure if a disc write-cache will interfere >> with that caluclation, though. >> Am I wrong ? :-) >> > > > Interesting, I did a test on a 7200 file and the best I could do was 50 > commits per second (a simple base/table with only id, journalling off and > no > extra code since the tool I use has "a repeated query" option with > accurate > timing). You mentioned 3 syncs per commit, but I tried to look at the log > of > Process Monitor for the tool process and I saw only one entry with > 'FlushBuffersFile' that is as I suppose was a mirror name for > FlushFileBuffers in winSync of sqlite so probably 3 syncs was a too big an > estimate, wasn't it? In this case 50 commits per second looks reasonable > limit
I found the number of 3 syncs some time ago, I can't proofe it right now. As far as I remeber I simply activated the debug logout put in sqlite and have seen these syncs. sqlite needs to sync the journal file once it has copied the pages from the DB, then it needs to sync the DB file itselve after written pages to the main DB, then it probably needs to clear and sync the journal file to indicate that the operation is finished. This sums up to 3 -- but I'm guessing here, maybe DRH is willing to confirm... :-) Note that this handling is changed via PRAGMA synchronous; I'm talking about windows, not sure if that applies on unix or mac. Marcus > > Max Vlasov > _______________________________________________ sqlite-users mailing list sqlite-users@sqlite.org http://sqlite.org:8080/cgi-bin/mailman/listinfo/sqlite-users
