On Fri, Sep 22, 2017 at 5:46 PM, Julien Rouhaud <rjuju123@gmail.com> wrote:
Hello, On Wed, Sep 13, 2017 at 3:01 AM, Haribabu Kommi <kommi.haribabu@gmail.com> wrote: > I ran the latest performance tests with and without IO times, there is an > overhead involved with IO time calculation and didn't observe any > performance > overhead with normal stats. May be we can enable the IO stats only in the > development environment to find out the IO stats? >
-1 for me to have these columns depending on a GUC *and* wether it's a debug or assert build (unless this behaviour already exists for other functions, but AFAIK that's not the case). > I added the other background stats to find out how much WAL write is > carried out by the other background processes. Now I am able to collect > the stats for the other background processes also after the pgbench test. > So I feel now the separate background stats may be useful. > > Attached latest patch, performance test results and stats details with > separate background stats and also combine them with backend including > the IO stats also. >
The results seem to vary too much to have a strong opinion, but it looks like the timing instrumentation can be too expensive, so I'd be -1 on adding this overhead to track_io_timing.
Thanks for the review.
I removed the time related columns from the view. As these columns are adding
an overhead and GUC controlled behavior is different to the other views.
Apart from above time columns, I removed walwriter_dirty_writes, as the
walwriter writers cannot be treated as dirty writes.
I have some minor comments on the last patch:
+ <row> + <entry><structfield>backend_writes</></entry> + <entry><type>bigint</type></entry> + <entry>Number of WAL writes that are carried out by the backend process</entry>