Re: measuring spinning

On Fri, Jun 15, 2012 at 9:43 AM, Merlin Moncure <mmoncure@gmail.com> wrote:
> On Thu, Jun 14, 2012 at 4:39 PM, Robert Haas <robertmhaas@gmail.com> wrote:
>> On Wed, Jan 11, 2012 at 8:48 PM, Robert Haas <robertmhaas@gmail.com> wrote:
>>> I've had cause, a few times this development cycle, to want to measure
>>> the amount of spinning on each lwlock in the system.  To that end,
>>> I've found the attached patch useful.  Note that if you don't define
>>> LWLOCK_STATS, this changes nothing except that the return value from
>>> s_lock becomes int rather than void.  If you do define LWLOCK_STATS,
>>> then LWLockAcquire() counts the number of pg_usleep() calls that are
>>> required to acquire each LWLock, in addition to the other statistics.
>>> Since this has come up for me a few times now, I'd like to proposing
>>> including it in core.
>>
>> Well, this fell through the cracks, because I forgot to add it to the
>> January CommitFest.  Here it is again, rebased.
>
> +1.  It might be too awkward to add, but it would be nice to be able
> to fetch the number of spins as well as number of delays (aside, it's
> a bit confusing that in s_lock.c 'delay' is used both for the hardware
> sleep as well as the yielding sleep).

Yeah, I'm inclined to keep it simple for now.  We can always change it
again later if someone comes up with something better.  I suspect that
delays probably tracks "contention bad enough that you should be
worried" pretty well, but of course I might be wrong.  The only thing
I know for sure is that I've found this useful in my own testing, and
therefore others testing with LWLOCK_STATS might also find it useful.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company