Re: [PATCH] Refactoring of LWLock tranches

On Tue, Jan  5, 2016 at 04:31:15PM +0100, Andres Freund wrote:
> On 2016-01-05 10:28:25 -0500, Bruce Momjian wrote:
> > On Sun, Dec 13, 2015 at 12:35:34PM +0100, Andres Freund wrote:
> > > > > One thing to call out is that an "oversized" s_lock can now make
> > > > > BufferDesc exceed 64 bytes, right now that's just the case when it's
> > > > > larger than 4 bytes.  I'm not sure if that's cause for real concern,
> > > > > given that it's not very concurrent or ancient platforms where that's
> > > > > the case.
> > > > > http://archives.postgresql.org/message-id/20150915020625.GI9666%40alap3.anarazel.de
> > > > > would alleviate that concern again, as it collapses flags, usage_count,
> > > > > buf_hdr_lock and refcount into one 32 bit int...
> > > > 
> > > > I don't think that would be worth worrying about even if we didn't
> > > > have a plan in mind that would make it go away again, and even less so
> > > > given that we do have such a plan.
> > > 
> > > Ok cool. I'm not particularly concerned either, just didn't want to slip
> > > that in without having it called out.
> > 
> > Uh, didn't you and I work in 9.5 to make sure the BufferDesc was 64-byte
> > aligned to avoid double-CPU cache invalidation that was causing
> > performance problems on a server you were testing?
> 
> Yes? But it's ok sizewise on the common platforms?

What is the uncommon part?  I guess I missed that.

--  Bruce Momjian          http://momjian.us EnterpriseDB                             http://enterprisedb.com

+ As you are, so once was I. As I am, so you will be. +
+ Roman grave inscription                             +