On May 22, 2008, at 9:40 AM, Shawn Wilsher wrote: > It was mentioned in the bug that opening the file with the O_SYNC flag > would no longer require fsyncs. Has this been looked into before by > sqlite? >
I don't think this would work well. The O_SYNC flag causes each write() system call to block until the data is on oxide. But sometimes SQLite will do a sequence of small consecutive writes of a few bytes each with the anticipation that the filesystem will buffer these writes and only do a single write to oxide at the end. So if we just blindly turned on O_SYNC, the number of disk writes would increase dramatically. We could update the unix backend so that some amount of buffering was done in the backend itself to coalesce a sequence of small writes into a single write() system call. This would reduce the amount of I/O back to what it was without the O_SYNC. But it still does not reduce the amount of I/O and it also deprives the filesystem the opportunity to reorder write requests, which will likely result in an efficiency loss, not a gain. Nevertheless, these kinds of things often defy logic, so I will give it a try and see what happens.... D. Richard Hipp [EMAIL PROTECTED] _______________________________________________ sqlite-users mailing list sqlite-users@sqlite.org http://sqlite.org:8080/cgi-bin/mailman/listinfo/sqlite-users
