Michael S (already5chosen@yahoo.com) on 8/21/08 wrote:
>
>Actually, nowadays TPC-C benchmarks use 73GB disks only
>because 36GB disks becoming rare and cost about the same as
>73GB.

You miss the point.

TPC-C doesn't use small disks because people want small
disks.

TPC-C uses small disks because with rotating media, you
need a metric sh*tload of actuators to get the IOPS. So
using big disks is a waste of everybody's time and money,
because you still want about a million disks.

But there are actually fairly high costs from using a
million disks - and not just in money. There's a huge
complexity cost in attaching that many disks, and it's
only done because you have to. And it's actually seldom
done in real life, because in any non-benchmark setting,
the complexity cost simply isn't worth it!

What SSD's could do is to give essentially the
same number of IOPS, but using a much smaller set of disks.
You could literally get the same number of IOPS using
probably something like 20 SSD's instead of 2000 rotating
disks!

And if you only connect a small number of disks, it gives
you a lot of other advantages. For example, you could now
do it without any fancy fiberchannel setup or other exotic
controllers. In fact, you could do it in a box that you
actually expect people to use. Yes, it would still
be expensive, but it wouldn't be insanely fragile like
having several thousand disks connected is.

In other words, you could have a realistic box that people
would actually buy to run a benchmark on.

And yes, I may be odd, but I think that the TPC rules
really should say that the box you test is not only
something you can buy, but that people actually
do buy them in that configuration (with some minimum
number of real sales).

But realistically, you cannot do that right now, simply
because of the SSD size issue. That will change,
but it will take quite a few more years.

So I agree - you could connect just a thousand SSD's, and
you'd actually have enough disk. But connecting that many
disks is what you want to get away from in the first place,
and with that many disks you probably have more latency in
the storage subsystem than you have in any of the disks!

See what I'm trying to say? A thousand rotating disks is
better than ten rotating disks, because it actually does
improve performance through improving IOPS. But a thousand
SSD's are actually likely to be inferior to ten
SSD's because performance will probably not even go up!

Yes, the IOPS may still scale, but it now scales at the
cost of latency for any individual IO due to the much more
complex interconnect, which means that it's no longer a
good trade-off.

Most people would take ten thousand IOPS with .1ms latency
per IO over twice as many IOPS but with twice the latency
per IO. The second version may get twice the throughput,
but it needs four times as many IO's outstanding to get
there!

With rotating media, the latency of the more complex
disk subsystem is not nearly as noticeable, because the
per-IO latency is still dominated by the disk itself.

So IOPS has been the limiting factor with rotating media,
with individual IO latency being fairly flat (and even
improving, because with bigger disks you can choose to only
use a small portion of the disk an get somewhat better seek
latencies).

However, with SSD's an individual SSD can already
do tons and tons of IOPS, and the limiting factor is more
likely to be the individual IO latency and all the overhead
to try to queue the IO in the first place.

See?

(And no, I have no numbers to back this up.)

Linus