One tiny change I'd suggest here: if you look at the code for checkpoint
buffer writing there are traces for two points in the process:
CheckPointBuffers(int flags) {
+ TRACE_POSTGRESQL_BUFFER_CHECKPOINT_START(flags); CheckpointStats.ckpt_write_t = GetCurrentTimestamp();
BufferSync(flags); CheckpointStats.ckpt_sync_t = GetCurrentTimestamp(); smgrsync();
CheckpointStats.ckpt_sync_end_t= GetCurrentTimestamp();
+ TRACE_POSTGRESQL_BUFFER_CHECKPOINT_DONE(); }
Note how the existing code also tracks how long the sync phase took
compared to the write one, and reports both numbers in the checkpoint
logs. It would be nice to add another probe at that same point (just
after ckpt_sync_t is set) so that dtrace users could instrument all these
possibilities as well: just buffer write time/resources, just sync ones,
or both.
--
* Greg Smith gsmith@gregsmith.com http://www.gregsmith.com Baltimore, MD