From: | Gregory Stark <stark(at)enterprisedb(dot)com> |
---|---|
To: | Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> |
Cc: | Greg Smith <gsmith(at)gregsmith(dot)com>, Robert Haas <robertmhaas(at)gmail(dot)com>, ITAGAKI Takahiro <itagaki(dot)takahiro(at)oss(dot)ntt(dot)co(dot)jp>, Postgres <pgsql-hackers(at)postgresql(dot)org> |
Subject: | Re: posix_fadvise v22 |
Date: | 2009-01-03 04:15:10 |
Message-ID: | 87mye9t48h.fsf@oxford.xeocode.com |
Views: | Raw Message | Whole Thread | Download mbox | Resend email |
Thread: | |
Lists: | pgsql-hackers |
Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> writes:
> The point of the suggestion is to prove that the patch works as
> advertised. How wide the sweet spot is for this test isn't nearly as
> interesting as proving that there *is* a sweet spot. If you can't
> find one it suggests that either the patch or the local posix_fadvise
> doesn't work.
I posted tons of reproducible test cases with graphs of results for various
raid stripe widths a while back. There was a very slight benefit on a single
spindle at some prefetch depths but it wasn't very consistent and it varied
heavily depending on the prefetch depth.
I don't know what to make of this test. I don't know how to reproduce the same
data distribution, I have no idea what raid configuration it's been run on,
what version of what OS it's on, etc. It's quite possible posix_fadvise isn't
working on it, I don't know.
It's also possible the overhead of the extra buffer lookups and syscalls
outweight any benefit of overlapping i/o and cpu on a single spindle.
Trying to contrive a situation where a single spindle sees a significant
benefit is going to be very tricky. Avoiding caching effects and the
confounding effect of any overhead will make it hard to see a consistent
benefit without a raid array.
--
Gregory Stark
EnterpriseDB http://www.enterprisedb.com
Ask me about EnterpriseDB's 24x7 Postgres support!
From | Date | Subject | |
---|---|---|---|
Next Message | Greg Stark | 2009-01-03 04:26:41 | Re: posix_fadvise v22 |
Previous Message | Robert Haas | 2009-01-03 04:13:10 | Re: posix_fadvise v22 |