On Mon, May 8, 2017 at 1:39 PM, David Rowley
<david.rowley@2ndquadrant.com> wrote:
> On 6 May 2017 at 13:44, Thomas Munro <thomas.munro@enterprisedb.com> wrote:
>> Experimentation required...
>
> Indeed. I do remember long discussions on this before Parallel seq
> scan went in, but I don't recall if anyone checked any OS kernels to
> see what they did.
>
> We really need a machine with good IO concurrency, and not too much
> RAM to test these things out. It could well be that for a suitability
> large enough table we'd want to scan a whole 1GB extent per worker.
I did a bunch of simple experiments this morning to try to observe RA
effects, using a couple of different EDB machines running Linux. I
wrote a simple program to read large files sequentially using lseek +
read, but rotate the reads over N file descriptors to simulate
parallel workers. I was surprised to find that I couldn't change
cache-cold read performance that way, up to very large numbers of N.
I did manage to break it by introducing some artificial disorder,
reversing/scrambling the read order of small groups of blocks, but
even that required groups over about 16 blocks before performance
started to drop (possibly related to the window size which I can't see
due to permissions right now). I've also learned that RAID cards
sometimes do read-ahead of their own, making matters more complicated.
I hope to report more when I figure out all the moving parts...
> I did post a patch to have heap_parallelscan_nextpage() use atomics
> instead of locking over in [1], but I think doing atomics there does
> not rule out also adding batching later. In fact, I think it
> structures things so batching would be easier than it is today.
+1
--
Thomas Munro
http://www.enterprisedb.com