Re: [PATCH] fs: Make sync_file_range(2) use WB_SYNC_NONE writeback

[Andres Freund <andres@anarazel.de>]

Hi,

On 2015-10-28 15:18:52 +0100, Jan Kara wrote:
> On Wed 28-10-15 10:30:40, Andres Freund wrote:
> > On 2015-10-24 21:28:17 +0200, Jan Kara wrote:
> > Thanks. Would scheduling a comparative benchmark of this be helpful
> > pushing htis forward ? Would probably only be early next week, I'm at
> > the european postgresql conference right now.
> 
> If you could run it, it would be nice. Thanks!

Sorry that it took longer. Had some $work deadline making a surprise
attack (sneaky bastard) and then difficulity getting time on a bigger
box. Hence I only have numbers from a single SSD drive (840 EVO 1TB)
laptop (16GB RAM, i7-4800MQ), with nothing else running.

I've compared 4.3.0-andres-07965-gd1e41ff, with/without the patch
applied. Best of three results:

Before:
transaction type: TPC-B (sort of)
scaling factor: 300
query mode: prepared
number of clients: 48
number of threads: 48
duration: 1000 s
number of transactions actually processed: 7873859
latency average: 6.094 ms
latency stddev: 35.376 ms
tps = 7871.887045 (including connections establishing)
tps = 7872.135871 (excluding connections establishing)


After:
transaction type: TPC-B (sort of)
scaling factor: 300
query mode: prepared
number of clients: 48
number of threads: 48
duration: 1000 s
number of transactions actually processed: 9370637
latency average: 5.119 ms
latency stddev: 24.423 ms
tps = 9369.212486 (including connections establishing)
tps = 9369.725595 (excluding connections establishing)

Pretty tidy improvement I'd say.

Feel free to add a
Tested-By: Andres Freund <andres@anarazel.de>

There's still quite some further room of improvement. E.g. some
quite massive peaks in latency, which don't coincide with background
work in postgres:
progress: 869.0 s, 9971.2 tps, lat 4.828 ms stddev 6.201
progress: 870.0 s, 8196.4 tps, lat 4.632 ms stddev 6.774
progress: 871.0 s, 497.0 tps, lat 89.598 ms stddev 308.398
progress: 872.0 s, 10360.6 tps, lat 5.917 ms stddev 40.041
progress: 873.0 s, 7005.5 tps, lat 6.743 ms stddev 21.985

Without controlling writeout via sync_file_range(), we frequently can
see stalls in the tens of seconds on busy machines though... So this is
still much better ;)

Greetings,

Andres Freund