Re: sinval synchronization considered harmful

On Thu, Jul 21, 2011 at 9:19 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
>> On Thu, Jul 21, 2011 at 6:43 PM, Noah Misch <noah@2ndquadrant.com> wrote:
>>> On Wed, Jul 20, 2011 at 09:46:33PM -0400, Robert Haas wrote:
>>>> SIGetDataEntries() can pretty easily be made lock-free.  The only real
>>>> changes that seem to be are needed are (1) to use a 64-bit counter, so
>>>> you never need to decrement
>
>>> On second thought, won't this be inadequate on 32-bit systems, where updating
>>> the 64-bit counter produces two stores?  You must avoid reading it between those stores.
>
>> Now that is a potentially big problem.
>
> Could we do something similar to the xxid hacks?  That is, we have a lot
> of counters that should be fairly close to each other, so we store only
> the low-order 32 bits of each notional value, and separately maintain a
> common high-order word.  You probably would need some additional
> overhead each time the high-order word bumps, but that's reasonably
> infrequent.

Well, the trouble is figuring out what the shape of that additional
overhead needs to look like.  I think I have a simpler idea, though:
before acquiring any locks, just have SIGetDataEntries() do this:

+       if (stateP->nextMsgNum == segP->maxMsgNum && !stateP->resetState)
+               return 0;

Patch (with comment explaining why I think this is OK) attached.  If
the message numbers happen to be equal only because the counter has
wrapped, then stateP->resetState will be true, so we'll still realize
we need to do some work.

Test results, with the lazy vxid patch plus this patch, at 8 clients:

tps = 34028.144439 (including connections establishing)
tps = 34079.085935 (including connections establishing)
tps = 34125.295938 (including connections establishing)

And at 32 clients:

tps = 185521.605364 (including connections establishing)
tps = 188250.700451 (including connections establishing)
tps = 186077.847215 (including connections establishing)

And at 80 clients:

tps = 188568.886569 (including connections establishing)
tps = 191035.971512 (including connections establishing)
tps = 189363.019377 (including connections establishing)

Not quite as good as the unlocked version, but better than the
per-backend mutex, and a whole lot simpler.

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