On Wed, Sep 28, 2016 at 7:32 PM, Tsunakawa, Takayuki
<tsunakawa.takay@jp.fujitsu.com> wrote:
> From: Thomas Munro [mailto:thomas.munro@enterprisedb.com]
>> > huge_pages=off: 70412 tps
>> > huge_pages=on : 72100 tps
>>
>> Hmm. I guess it could be noise or random code rearrangement effects.
>
> I'm not the difference was a random noise, because running multiple set of three runs of pgbench (huge_pages = on,
off,on, off, on...) produced similar results. But I expected a bit greater improvement, say, +10%. There may be
betterbenchmark model where the large page stands out, but I think pgbench is not so bad because its random data access
wouldcause TLB cache misses.
Your ~2.4% number is similar to what was reported for Linux with 4GB
shared_buffers:
https://www.postgresql.org/message-id/20130913234125.GC13697%40roobarb.crazydogs.org
Later in that thread there was a report of a dramatic ~15% increase in
"best result" TPS, but that was with 60GB of shared_buffers on a
machine with 256GB of RAM:
https://www.postgresql.org/message-id/20131024060313.GA21888%40toroid.org
--
Thomas Munro
http://www.enterprisedb.com
