Re: measuring lwlock-related latency spikes

On Mon, Apr 2, 2012 at 8:16 PM, Simon Riggs <simon(at)2ndquadrant(dot)com> wrote:
> Agreed, though I think it means the fsync is happening on a filesystem
> that causes a full system fsync. That time is not normal.

I don't know what you mean. It looks like there are two cases where
this code path executes. Either more than 16 clog files are being
flushed by the SimpleLRUFlush() during a checkpoint or a dirty page is
being evicted by SlruSelectLRUPage().

I don't know that 16 is so crazy a number of clog files to be touching
between checkpoints any more on a big machine like this. The number of
clog files active concurrently in pgbench should be related to how
quickly xids are being used up and how large the database is -- both
of which are pretty big in these tests. Perhaps the 16 should have
been raised to 32 when CLOGShmemBuffers was raised to 32.

--
greg