WIP: Barriers - Mailing list pgsql-hackers

Hi hackers,

I would like to propose "barriers" for Postgres processes.  A barrier
is a very simple mechanism for coordinating parallel computation, as
found in many threading libraries.

First, you initialise a Barrier object somewhere in shared memory,
most likely in the DSM segment used by parallel query, by calling
BarrierInit(&barrier, nworkers).  Then workers can call
BarrierWait(&barrier) when they want to block until all workers arrive
at the barrier.  When the final worker arrives, BarrierWait returns in
all workers, releasing them to continue their work.  One arbitrary
worker receives a different return value as a way of "electing" it to
perform serial phases of computation.  For parallel phases of
computation, the return value can be ignored.  For example, there may
be preparation, merging, or post-processing phases which must be done
by just one worker, interspersed with phases where all workers do
something.

My use case for this is coordinating the phases of parallel hash
joins, but I strongly suspect there are other cases.  Parallel sort
springs to mind, which is why I wanted to post this separately and
earlier than my larger patch series, to get feedback from people
working on other parallel features.

A problem that I'm still grappling with is how to deal with workers
that fail to launch.  What I'm proposing so far is based on static
worker sets, where you can only give the number of workers at
initialisation time, just like pthread_barrier_init.  Some other
libraries allow for adjustable worker sets, and I'm wondering if a
parallel leader might need to be able to adjust the barrier when it
hears of a worker not starting.  More on that soon.

Please see the attached WIP patch.  I had an earlier version with its
own waitlists and signalling machinery etc, but I've now rebased it to
depend on Robert Haas's proposed condition variables, making this code
much shorter and sweeter.  So it depends on his
condition-variable-vX.patch[1], which in turn depends on my
lwlocks-in-dsm-vX.patch[2] (for proclist).

When Michaël Paquier's work on naming wait points[3] lands, I plan to
include event IDs as an extra argument to BarrierWait which will be
passed though so as to show up in pg_stat_activity.  Then you'll be
able to see where workers are waiting for each other!  For now I
didn't want to tangle this up with yet another patch.

I thought about using a different name to avoid colliding with
barrier.h and overloading the term: there are of course also compiler
barriers and memory barriers.  But then I realised that that header
was basically vacant real estate, and 'barrier' is the super-well
established standard term for this parallel computing primitive.

I'd be grateful for any thoughts, feedback, flames etc.

[1] https://www.postgresql.org/message-id/flat/CA%2BTgmoaj2aPti0yho7FeEf2qt-JgQPRWb0gci_o1Hfr%3DC56Xng%40mail.gmail.com
[2] https://www.postgresql.org/message-id/flat/CAEepm%3D0Vvr9zgwHt67RwuTfwMEby1GiGptBk3xFPDbbgEtZgMg%40mail.gmail.com
[3] https://www.postgresql.org/message-id/flat/CAB7nPqTGhFOUHag1eJrvsKn8-E5fpqvhM7aL0tAfsDzjQG_YKQ@mail.gmail.com

--
Thomas Munro
http://www.enterprisedb.com