Ok so I ran 6 parallel pgbench initializations at a relatively high scale 2500 on 6 different pgbench databases on the same server and was able to push disk io to around 380MB/sec write with pretty low cpu wio (around 5-15%). Progress! As I hoped, I just haven't been pushing hard enough.
On Wed, May 11, 2016 at 10:16 AM, Mike Broers <mbroers@gmail.com> wrote:
I've set this parameter to 256 to test, still getting the same results. The only way I reliably see a spike in IO at all is on the pgbench initialization (still peaking around 110MB/sec writes), so I'm thinking about playing around with that or duplicating its approach with larger rows and running it from multiple sessions in different initialization databases on the same volume. If I cant push it further using that technique then there must be some dumb bottleneck in config I am overlooking.
On Tue, May 10, 2016 at 3:24 PM, Kevin Grittner <kgrittn@gmail.com> wrote:
On Tue, May 10, 2016 at 2:02 PM, Mike Broers <mbroers@gmail.com> wrote:
> If someone has an idea for how to set up pgbench to > really push the io to prove out ssd potential, or a different > script/approach that would be awesome.
You might want to play with some large effective_io_concurrency settings, starting around 256.
