Re: separate hard drive for scratch disks for two different programs?

dnichols@xxxxxxxxxxx (DoN. Nichols) wrote:
According to Floyd L. Davidson <floyd@xxxxxxxxxx>:

[ ... ]

It's worse than it looks too. While the 300GB SCSI
measures out at a higher speed than the 750GB SATA
drive, that happens only when testing the outer
cylinders. When the two devices have 270GB of data on
them, not only does the SATA drive still have that much
more (while the SCSI drive is full), the SATA drive is
much faster using the middle cylinders than the SCSI
drive is using the inner most cylinders.

This is a function of the way the OS deals with the disk, If it
insists on seeking to the early tracks frequently, yes. If it keeps the
heads fairly localized, this is not a problem.

That is only true for consideration of head seek times.
It does not address i/o speed variation due to higher
rotational speeds for outside cylinders.

Modern disks effectively provide localized head
operation by prioritizing i/o operations to minimize
head seeks. The OS of course still has finer
granularity, which will work just as well under a lower
load condition than will the drive's built in
optimization. Which is to say, for a server the OS
probably has little effect; but for a workstation with
significantly less disk activity, the OS's optimizations
are much more significant.

It is a problem too, because the figures given by i/o
speed tests are not useful for comparing different sized
disks in actual use, in any case other than possibly on
a nearly empty disk.

Hence if you compare two 300GB disks, the tested i/o
speeds are useful. But comparing tested i/o speeds
between 300GB disk and 600GB disks is not a valid
indication of actual in use performance.

This would argue for not
putting the swap partition (unix) or the swapfile (Windows) on the same
disk where your heavy data activity is located.

If a system actually works from swap space, it makes
very little difference where it goes... any place will
be best described as "slow"! ;-)

But yes, it can be optimized in a number of ways, the
best of which is to be on a totally separate set of
drives (and to have them all at an equal priority, which
is similar to striping with a RAID array).
Implementation is problematical: the fastest disks are
all so much larger than any possible requirement for swap
space that spreading swap across 4 disks would be an
enormous waste of disk space!

If a very active swap area is on a disk with other data
that is also very actively being used, the swap area
should perhaps be in the middle of the disk to reduce
head seek times to/from other data and back to the swap
area. The other option is to put swap on the outer
cylinders to optimize r/w rates, which would be suitable
if the swap area is used much more often than the other
data areas, indicating that head seek times have less
effect.

By far the best solution is more RAM though, don't work
from swap, and put swap anywhere convenient. :-)

And the same argument can be used against the SATA drives, if
they happen to get near full too. And there, you don't have the option
of sharing the load between multiple drives, because you are wanting the
maximum capacity per drive.

Why would that preclude sharing with other drives? And
no the same argument doesn't work against SATA, because
there is no way to buy a two times larger disk for the
same price that will be faster if it is only half full.

That *is* the point. If the need is for 300GB of disk,
a 600GB SATA is faster and less expensive than a 300GB
SCSI system. If the need is for a 600GB disk... the
same is true of 1200GB SATA vs. 600GB SCSI systems.

But SCSI is a better option when the absolute largest
and fastest possible system is required.

Given the use of hot swappable RAID arrays, I don't see
the increased reliability of SCSI drives as having much
value for most systems.

Of course there are situations where even that is not
true. But exceptions in this case simply prove the
rule. For example there is a satellite imaging earth
station here in Barrow. They of course use SCSI RAID
arrays. There essentially is no such thing as local
skilled maintenance available (that they are willing to
pay for...), and the cost of servicing equipment even
once is higher than the cost of acquisition. They do
need high reliability SCSI disks, even if they are
slower and smaller.

That means if the actual need is for anything more than
perhaps half of the storage space available on the SCSI
drive, the SATA drive will not only be 6:1 less
expensive, but have a higher performance in *all* ways.

I don't accept that if a good OS is used which keeps seeks
localized.

Localizing seeks does not change the fact that inner
cylinders have a slower data rate than outer cylinders.
We aren't talking head seek times, we are talking about
rotational speed. The outer cylinders have more data
per revolution.

And -- if you are using two SCSI disks in place of a single SATA
disk, you have the option of striping (with RAID) or simply partitioning
the two disks with each project allocated a partition, which will keep
the seeks while working on that project localized.

That works equally well with SATA or SCSI, and is not a
SCSI advantage.

One reason that the SATA disks are larger is that they are being
used as a less critical testbed to determine the failure rate until
someone decides whether they are reliable enough to be sold as SCSI
disks. All it takes is a swap of logic card to turn a SATA disk into a
SCSI disk -- or vice versa. The higher (new) cost of the SCSI disks
relates to the smaller customer base, and the extra testing which goes
into producing reliable disks for the high-end users.

If that is true, the same platters being used for high
reliability SCSI can also be used for high reliability
SATA.

The cost of SCSI is not just the smaller customer base,
which would not be smaller if SCSI were the same cost.
The cost reflects the complexity of that logic card
which gets swapped. SCSI has an onboard CPU of
significant complexity; SATA does not and instead uses
the computer's CPU instead.

Note that off loading the computation to a peripheral
CPU is part of what makes SCSI faster, and it also
unloads the computer's CPU, which is an added advantage
of SCSI that is not reflected in the usual benchmark
tests.

So -- yes, the SATA disks will probably continue to be larger
than the largest SCSI disk -- simply because the platters and heads
don't have a long enough track record yet to be trusted as serious SCSI
drives.

If that were true, they would indeed be showing up as
"less serious" SCSI drives. Are they?

(On the other hand, I don't know what exactly is the
current limiting factor in the size of SCSI disks. I
doubt that is is just a matter of price tag, and expect
it is actually technical. But I have no idea why the
same platter/head system cannot be controlled by SCSI
logic.)

--
Floyd L. Davidson http://www.apaflo.com/floyd_davidson
Ukpeagvik (Barrow, Alaska) floyd@xxxxxxxxxx
.

Relevant Pages

  • Asus A7N8X-E w/SATA and SCSI, 4.10 sys install snafu
    ... Old buslogic SCSI controller with 2 drives, a cd, and a tape. ... and nt going off the sata drives. ... I configured the mobo to boot from cd first, ...
    (freebsd-questions)
  • Re: Hot Swapping a SATA drive in Windows 2000 and XP.
    ... Things showing up as SCSI is just odd. ... > put it in my SATA enclosure. ... > I even took my USB memory stick, and was able to format it as NTFS, ... > Do your SATA adapters and associated drives show up as SCSI devices? ...
    (microsoft.public.windowsxp.hardware)
  • Re: Using REP with INx and OUTx?
    ... separate SCSI controller running MFM or ESDI drives), ... even though I don't know of any hard drives using ... it's comforting to know that my memory isn't ... but smaller/faster/more expensive than disk. ...
    (comp.lang.asm.x86)
  • Re: New system using AHA 2940U2W SCSI???
    ... I would like to build a new system centered around an Asus/Intel chipset and possibly a Q6600 CPU but want to keep my my old AHA 2940u2w SCSI. ... I hate to lose my Plextor SCSI drives or my SCSI scanner at this time. ... Know of any benchmarks pitting UW SCSI2 against SATA, and possibly even Firewire in the mix?? ... The cache on the disk controller board ...
    (alt.comp.periphs.mainboard.asus)
  • Re: What do you use?
    ... I haven't used any SATA stuff yet. ... >>I can guarantee modern SCSI throughput is superior to any of the SATA ... you've got drives like the WD 'Raptor' ... so effectively need one channel per disk. ...
    (freebsd-questions)