Eric Sosman wrote:
> It might be a good idea to learn more about the
> anecdote before drawing too many conclusions ... Just
> a few quick observations:
>
> 1) The anecdote does not reveal what was measured. This
> five-to-one difference might have been in latency,
> throughput, capacity, license price, or debugging time.
Transactions per minute in a Java Enterprise application hitting an Oracle
back end. It was a saturation test, so exact boundaries aren't available
and
I don't have access to the hard data anyway.
> 2) The anecdote does not reveal which configuration had
> the better result on the measured quantity, only that
> a difference existed. The final score may have been
> ten to two, but which team won?
Oracle with its direct disk management vs. Oracle working with the file
system. Direct disk management sped up the database throughput by a
factor
that they could tell was above five to one, but not by how much.
> 3) The anecdote tells us only that the test was "large-
> scale," but nothing else. A few weeks ago near where
> I live, a "large-scale" test called the Boston Marathon
> demonstrated conclusively that wheelchairs are faster
> than motorcycles (even though the motorcycles had a
> head start, the wheelchairs were first to the finish).
Terabyte-order data, on the order of hundreds of thousands of do***ents
per
hour processed through an enterprise application with relevant data
extracted
and stored in the database.
> 4) The fact (if we assume its existence) that some database
> performed better without a file system than with one does
> not prove that the file system performs poorly. It might
> well be that the database in question does things dumbly
> and forces the file system to do a lot of needless work.
In this case, no. If you read up on Oracle's direct-disk management it is
not
dumb, nor is its file-system interaction. None of the big players are
stupid
with file-system access, be they Oracle, DB2, Postgres or whomever. You
can
pose it as a theoretical risk, but if you pick a reputable product it
won't be
an issue at all in real life. No need to panic.
> I'd suggest that you not dismiss Lew's anecdote, but that
> you examine its actual information content before forming firm
> opinions about file systems vs. raw devices.
Which is exactly why I presented it as an anecdote. Still, seeing it work
like that even in one case is very indicative, wouldn't you agree?
--
Lew
|
