Re: Parallel Seq Scan vs kernel read ahead

On Mon, 22 Jun 2020 at 16:54, David Rowley <dgrowleyml@gmail.com> wrote:
> I also tested this an AMD machine running Ubuntu 20.04 on kernel
> version 5.4.0-37.  I used the same 100GB table I mentioned in [1], but
> with the query "select * from t where a < 0;", which saves having to
> do any aggregate work.

I just wanted to add a note here that Thomas and I just discussed this
a bit offline. He recommended I try setting the kernel readhead a bit
higher.

It was set to 128kB, so I cranked it up to 2MB with:

sudo blockdev --setra 4096 /dev/nvme0n1p2

I didn't want to run the full test again as it took quite a long time,
so I just tried with 32 workers.

The first two results here are taken from the test results I just
posted 1 hour ago.

Master readhead=128kB = 89921.283 ms
v2 patch readhead=128kB = 36085.642 ms
master readhead=2MB = 60984.905 ms
v2 patch readhead=2MB = 22611.264 ms

There must be a fairly large element of reading from the page cache
there since 22.6 seconds means 4528MB/sec over the 100GB table. The
maximum for a PCIe 3.0 x4 slot is 3940MB/s

David