In theory there should be no benefit on a single spindle system. There
could be a slight benefit due to reordering of I/o but only on a raid
array would you see a significant speedup -- which should be about
equal to the number of spindles.
What would be interesting is whether you see a noticable speed
*decrease* from having prefetching enabled when it isn't helping.
Either due to having everything fit in shared buffers or everything
fit in the filesystem cache (the latter should be more of a hit)
Even if there is it doesn't really worry me. By default the feature is
disabled and you should only really turn it on if ulu do have a raid
array and want an individual query to make use if it.
Now that there's an actual run-time sysconf check for the buggy glibc
called by the guc function we arguably don't need the autoconf
check_run check anymore anyways.
--
Greg
On 1 Jan 2009, at 21:43, "Robert Haas" <robertmhaas@gmail.com> wrote:
> On Thu, Jan 1, 2009 at 3:55 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> "Robert Haas" <robertmhaas@gmail.com> writes:
>>> Am I correct in thinking that the only thing we're really checking
>>> for
>>> here is whether a trivial posix_fadvise() call returns success? If
>>> so, is this test really worth doing?
>>
>> Runtime tests performed during configure are generally a bad idea to
>> start with --- it's impossible to do any such thing in a
>> cross-compilation scenario, for example.
>
> OK, here's an update of Greg's patch with the runtime configure test
> ripped out, some minor documentation tweaks, and a few unnecessary
> whitespace diff hunks quashed. I think this is about ready for
> committer review. The only thing I haven't been able to do is
> demonstrate that this change actually produces a performance
> improvement. Either I'm testing the wrong thing, or it just doesn't
> provide any benefit on a single-spindle system. However, I believe
> that Greg has previously posted some fairly impressive performance
> results, so I'm not sure that my shortcomings in this area should be a
> bar to having a committer pick this one up. If more testing is
> needed, it would at least be helpful to have a committer specify what
> areas they are concerned about.
>
> ...Robert
> <posix_fadvise_v23_rh1.diff.gz>