Re: POSIX shared memory redux

On Thu, Apr 14, 2011 at 7:26 AM, A.M. <agentm(at)themactionfaction(dot)com> wrote:
> From what I understood, the primary purpose of the SysV check was to protect the shared memory from multiple stompers. The interlock was a neat side-effect.

Not really - the purpose of the interlock is to protect the underlying
data files. The nattch interlock allows us to be very confident that
there isn't another postmaster running on the same data directory on
the same machine, and that is extremely important.

You've just about convinced me that it might not be a bad idea to add
the fcntl() interlock in addition because, as you say, that has a
chance of working even over NFS. But the interlock we have now is
*extremely* reliable, and I think we'd need to get some other
amazingly compelling benefit to consider changing it (even if we were
convinced that the alternate method was also reliable). I don't see
that there is one. Anyone running an existing version of PostgreSQL
in an environment where they care *at all* about performance has
already adjusted their SysV shm settings way up. The benefit of using
POSIX shm is that in, say, PostgreSQL 9.2, it might be possible for
shared buffers to have a somewhat higher default setting out of the
box, and be further increased from there without kernel parameter
changes. And there might be more benefits besides that, but certainly
those by themselves seem pretty worthwhile. SysV shm is extremely
portable, so I don't think that we're losing anything by continuing to
allocate a small amount of it (a few kilobytes, perhaps) and just push
everything else out into POSIX shm.

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