On Wed, Nov 17, 2021 at 08:04:41PM +0000, Jacob Champion wrote:
> On Wed, 2021-11-17 at 14:44 -0500, Jaime Casanova wrote:
> > I'm trying to add more parallelism by copying individual segments
> > of a relfilenode in different processes. Does anyone one see a big
> > problem in trying to do that? I'm asking because no one did it before,
> > that could not be a good sign.
>
> I looked into speeding this up a while back, too. For the use case I
> was looking at -- Greenplum, which has huge numbers of relfilenodes --
> spinning disk I/O was absolutely the bottleneck and that is typically
> not easily parallelizable. (In fact I felt at the time that Andres'
> work on async I/O might be a better way forward, at least for some
> filesystems.)
>
> But you mentioned that you were seeing disks that weren't saturated, so
> maybe some CPU optimization is still valuable? I am a little skeptical
> that more parallelism is the way to do that, but numbers trump my
> skepticism.
>
Sorry for being unresponsive too long. I did add a new --jobs-per-disk
option, this is a simple patch I made for the customer and ignored all
WIN32 parts because I don't know anything about that part. I was wanting
to complete that part but it has been in the same state two months now.
AFAIU, it seems there is a different struct for the parameters of the
function that will be called on the thread.
I also decided to create a new reap_*_child() function for using with
the new parameter.
Now, the customer went from copy 25Tb in 6 hours to 4h 45min, which is
an improvement of 20%!
> > - why we read()/write() at all? is not a faster way of copying the file?
> > i'm asking that because i don't actually know.
>
> I have idly wondered if something based on splice() would be faster,
> but I haven't actually tried it.
>
I tried and got no better result.
> But there is now support for copy-on-write with the clone mode, isn't
> there? Or are you not able to take advantage of it?
>
That's sadly not possible because those are different disks, and yes I
know that's something that pg_upgrade normally doesn't allow but is not
difficult to make it happen.
--
Jaime Casanova
Director de Servicios Profesionales
SystemGuards - Consultores de PostgreSQL