Re: Scaling further up

On Wed, 3 Mar 2004, Paul Thomas wrote:

>
> On 02/03/2004 23:25 johnnnnnn wrote:
> > [snip]
> > random_page_cost should be set with the following things taken into
> > account:
> > - seek speed
>
> Which is not exactly the same thing as spindle speed as it's a combination
> of spindle speed and track-to-track speed. I think you'll find that a 15K
> rpm disk, whilst it will probably have a lower seek time than a 10K rpm
> disk, won't have a proportionately (i.e., 2/3rds) lower seek time.

There are three factors that affect how fast you can get to the next
sector:

seek time
settle time
rotational latency

Most drives only list the first, and don't bother to mention the other
two.

On many modern drives, the seek times are around 5 to 10 milliseconds.
The settle time varies as well. the longer the seek, the longer the
settle, generally. This is the time it takes for the head to stop shaking
and rest quietly over a particular track.
Rotational Latency is the amount of time you have to wait, on average, for
the sector you want to come under the heads.

Assuming an 8 ms seek, and 2 ms settle (typical numbers), and that the
rotational latency on average is 1/2 of a rotation: At 10k rpm, a
rotation takes 1/166.667 of a second, or 6 mS. So, a half a rotation is
approximately 3 mS. By going to a 15k rpm drive, the latency drops to 2
mS. So, if we add them up, on the same basic drive, one being 10k and one
being 15k, we get:

10krpm: 8+2+3 = 13 mS
15krpm: 8+2+2 = 12 mS

So, based on the decrease in rotational latency being the only advantage
the 15krpm drive has over the 10krpm drive, we get an decrease in access
time of only 1 mS, or only about an 8% decrease in actual seek time.

So, if you're random page cost on 10krpm drives was 1.7, you'd need to
drop it to 1.57 or so to reflect the speed increase from 15krpm drives.

I.e. it's much more likely that going from 1 gig to 2 gigs of ram will
make a noticeable difference than going from 10k to 15k drives.