Re: Spinlocks and semaphores in 9.2 and 9.3

On Mon, Apr 18, 2016 at 11:34 AM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> Andres Freund <andres(at)anarazel(dot)de> writes:
>> On 2016-04-18 11:24:07 -0400, Tom Lane wrote:
>>> (The thing that gave me pause about this was noticing that I could not
>>> start two such postmasters concurrently on my RHEL6 box, without changing
>>> the default system limits on number of SysV semaphores.)
>
>> Hm, is that different before/after the patch? Because afaics at around
>> NBuffers = 1000, the number of semaphores should be lower after the
>> patch?
>
> Yeah, that was probably the argument we used before. But it's true that a
> postmaster built with --disable-spinlocks is harder to test than one built
> without (because you're going from ~100 semas to ~1100 semas, at default
> configuration). If we believe that the main real use-case for this switch
> is testing, that's not a very nice property.
>
> The bottom line IMO is that --disable-spinlocks is actually not that awful
> for low-stress, low-concurrency scenarios; for example, it doesn't change
> the runtime of make installcheck-parallel very much for me. On the other
> hand, for high-concurrency scenarios you'd darn well better get a real
> spinlock implementation. Given that sort of scope of use, it seems like
> a hundred or so underlying semas should be plenty, and it'd be less likely
> to cause operational problems than 1024.

I suppose that's probably true. I thought surely any system worth its
salt wouldn't have a problem with 1024 semaphores, but a quick Google
search for "hp-ux semmns" turned up this page:

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