RE: [Patch] Optimize dropping of relation buffers using dlist

On Friday, September 25, 2020 6:02 PM, Tsunakawa-san wrote:

> From: Jamison, Kirk/ジャミソン カーク <k.jamison@fujitsu.com>
> > [Results]
> > Recovery/Failover performance (in seconds). 3 trial runs.
> >
> > | shared_buffers | master | patch  | %reg    |
> > |----------------|--------|--------|---------|
> > | 128MB          | 32.406 | 33.785 | 4.08%   |
> > | 1GB            | 36.188 | 32.747 | -10.51% |
> > | 2GB            | 41.996 | 32.88  | -27.73% |
> 
> Thanks for sharing good results.  We want to know if we can get as
> significant results as you gained before with hundreds of GBs of shared
> buffers, don't we?

Yes. But I don't have a high-spec machine I could use at the moment.
I'll try if I can get one by next week. Or if someone would like to reproduce the
test with their available higher spec machines, it'd would be much appreciated.
The test case is upthread [1]

> > I also did similar benchmark performance as what Tomas did [1], simple
> > "pgbench -S" tests (warmup and then 15 x 1-minute runs with 1, 8 and
> > 16 clients, but I'm not sure if my machine is reliable enough to
> > produce reliable results for 8 clients and more.
> 
> Let me confirm just in case.  Your patch should not affect pgbench
> performance, but you measured it.  Is there anything you're concerned
> about?
> 

Not really. Because In the previous emails, the argument was the BufferAlloc 
overhead. But we don't have it in the latest patch. But just in case somebody
asks about benchmark performance, I also posted the results.

[1] https://www.postgresql.org/message-id/OSBPR01MB2341683DEDE0E7A8D045036FEF360%40OSBPR01MB2341.jpnprd01.prod.outlook.com

Regards,
Kirk Jamison