Re: wait events for disk I/O

On Tue, Jan 31, 2017 at 8:54 AM, Michael Paquier <michael(dot)paquier(at)gmail(dot)com>
wrote:

> On Mon, Jan 30, 2017 at 10:01 PM, Rushabh Lathia
> <rushabh(dot)lathia(at)gmail(dot)com> wrote:
> > Attached is the patch, which extend the existing wait event
> infrastructure
> > to implement the wait events for the disk I/O. Basically
> pg_stat_activity's
> > wait event information to show data about disk I/O as well as IPC
> primitives.
> >
> > Implementation details:
> >
> > - Added PG_WAIT_IO to pgstat.h and a new enum WaitEventIO
> > - Added a wait_event_info argument to FileRead, FileWrite, FilePrefetch,
> > FileWriteback, FileSync, and FileTruncate. Set this wait event just
> before
> > performing the file system operation and clear it just after.
> > - Pass down an appropriate wait event from caller of any of those
> > functions.
> > - Also set and clear a wait event around standalone calls to read(),
> > write(), fsync() in other parts of the system.
> > - Added documentation for all newly added wait event.
>
> Looks neat, those are unlikely to overlap with other wait events.
>

Thanks.

>
> > Open issue:
> > - Might missed few standalone calls to read(), write(), etc which need
> > to pass the wait_event_info.
>
> It may be an idea to use LD_PRELOAD with custom versions of read(),
> write() and fsync(), and look at the paths where no flags are set in
> MyProc->wait_event_info, and log information when that happens.
>
>
Yes, may be I will try this.

> > Thanks to my colleague Robert Haas for his help in design.
> > Please let me know your thought, and thanks for reading.
>
> Did you consider a wrapper of the type pg_read_event() or
> pg_write_event(), etc.?
>

I thought on that, but eventually stick to this approach as it looks
more neat and uniform with other wait event implementation.

> --
> Michael
>

Thanks,
Rushabh Lathia
www.EnterpriseDB.com