Re: [tlug] Making better use of SSDs?

[Kalin KOZHUHAROV <me.kalin@example.com>]

Hello Satoshi,


On Mon, May 28, 2012 at 6:15 PM,  <satoshi.nagayasu@example.com> wrote:
> A few days ago, I just published a technical report that examine
> SSD performance compared with HDD to execute DWH (or analytics)
> workloads on PostgreSQL. SSD was used for its data directory.
>
Will the report be available somewhere to read?

> According to the experiments, it's been determined that SSD is able to
> improve the query performance extremely if it has massive random
> acccess I/O operations. As a results, more than 20 times faster
> performance has been observed with some workloads.
>
Yep, that is to be expected, more or less. Actually the more random
and more processes try to read a HDD, the better improvement you'll
see with databases. 10-100 times is not uncommon.
No idea what is the size of your tables, but have you tried to fit
them in memory?
It will be a good indication of how much wait for IO is there, better
than measuring it.
If you don't want to change anything, just mount a tmpfs and put the
data directory there and `mount -o bind` it to the old place (when db
is down).

> On the other hand, by replacing HDD with SSD, sequential access (read)
> was not improved as much as I expected (almost x1.3 faster than HDD),
> because sequential access on HDD is enough fast.
>
Or your process is stuffing the CPU already, so it cannot take more IO.

> Generally speaking, it's safe to say SSD would increase the performance.
> But practically, you need to determine a type of workload (or a bottleneck
> of workload) that you're running. And it's really fun! :)
>
++1!

Cheers,
Kalin.