On Wed, Nov 30, 2016 at 10:22 PM, Michael Paquier
<michael.paquier@gmail.com> wrote:
> On Wed, Nov 30, 2016 at 8:51 PM, Heikki Linnakangas <hlinnaka@iki.fi> wrote:
>> On 11/30/2016 09:01 AM, Michael Paquier wrote:
>>> +bool
>>> +pg_backend_random(char *dst, int len)
>>> +{
>>> + int i;
>>> + char *end = dst + len;
>>> +
>>> + /* should not be called in postmaster */
>>> + Assert (IsUnderPostmaster || !IsPostmasterEnvironment);
>>> +
>>> + LWLockAcquire(BackendRandomLock, LW_EXCLUSIVE);
>>> Shouldn't an exclusive lock be taken only when the initialization
>>> phase is called? When reading the value a shared lock would be fine.
>
> Do we need to worry about performance in the case of application doing
> small transactions and creating new connections for each transaction?
> This would become a contention point when calculating cancel keys for
> newly-forked backends. It could be rather easy to measure a
> concurrency impact with for example pgbench -C with many concurrent
> transactions running something as light as SELECT 1.
I got curious about this point, so I have done a couple of tests with
my laptop using the following pgbench command:
pgbench -f test.sql -C -c 128 -j 4 -t 100
And test.sql is just that:
\set aid random(1,10)
In short, a backend is spawned and a cancel key is generated but
nothing is done with it to avoid any overhead. With HEAD and the patch
without/with --disable-strong-random, I am seeing pretty close
numbers. My laptop has only 4 cores, so we may see something on a
machine with a higher number of cores. But as far as things go I am in
the noise range.
--
Michael