Re: [HACKERS] Re: [QUESTIONS] Does Storage Manager support >2GB tables? - Mailing list pgsql-hackers

Bruce Momjian writes:
> I am adding this to the TODO list:
>
>     * Do async I/O to do better read-ahead of data

Good.

> Because we are not threaded, we really can't do anything else while we
> are waiting for I/O, but we can pre-request data we know we will need.

Threading is a bit like raw devices. It sounds like a really good idea,
particularly with M$ banging the "NT, now with threads" drum, but in real
life there are some very good reasons not to thread. Particularly with an
extensible product like Postgres where J-Random routine gets loaded at
runtime. In a threaded system, J-Random routine needs to be pretty well
perfect or the whole system comes down. In a process based system, unless
it trashes something in the shared memory, only the one connection instance
needs to come down. My experience with Illustra says that this is fairly
important.

The other big problem with threading is that now the stacks and all dynamically
allocated data are in the shared memory and are not easily extendable. So,
if some recursive procedure (like in the rewriter) uses a bit of extra stack
some other thread gets its stack trashed. This is never pretty.

Or if some user function loads a giant large object (like an mpeg say), that
memory has to come out of the shared memory, now if they leak that memory,
it is gone for good. In a per process system, it just ends up using a little
more swap space.

The other thing threading does is introduce new synchronization requirements
into things that never needed it before. For example, Postgres is pretty
free with memory allocation and deallocation (think Nodes!). With threading
each palloc() and pfree() is going to have to take a round trip through a
mutex. This makes an expensive operation even more costly. By and large,
the dbms's that are threaded have had pretty static (ie pre-allocate every
thing in arrays at boot time) memory models. Postgres does not fit this
picture very well.

Ultimately, threading may buy some performance, but not very much compared
to how hard it is to get right and how brittle it tends to make the system.

Unless I have misunderstood the state of Postgres, there is a vast amount
of performance improvement to be had without even thinking about threading.
If it were me, I would pick up the easy stuff, then the middle-hard stuff
with the really big wins like a proper transaction log, and leave the very
hard stuff like threading until last.

-dg


David Gould            dg@illustra.com           510.628.3783 or 510.305.9468
Informix Software  (No, really)         300 Lakeside Drive  Oakland, CA 94612
"Of course, someone who knows more about this will correct me if I'm wrong,
 and someone who knows less will correct me if I'm right."
               --David Palmer (palmer@tybalt.caltech.edu)