RAID0 + SATA + NCQ?

qbker

Distinguished
Dec 27, 2004
29
0
18,530
0
Hello, I just want to know if anyone had tried a RAID0 + SATA + NCQ setup? (the hdd, the controller and the software have all got to support NCQ)

I'm just wondering, could NCQ and RAID0 be used at the same time? I'm just a bit worried about the following scenario:

If you have 2 SATA NCQ HDD, you can run them in RAID0 but not with NCQ. To use NCQ, you cannot run them in RAID0.

Or am I worrying too much? Thanks! :p
 

supremelawfirm

Distinguished
Jan 9, 2005
47
0
18,530
0
Greetings,

Here's a well written and empirical analysis
of RAIDs in various configurations:

http://www.storagereview.com/articles/200406/20040625TCQ_1.html

This article may answer your question directly:
they have concluded that NCQ will only benefit
file servers and not single-user workstations.

We recently assigned C: to a single WD 74GB Raptor
on an ASUS P4C800-E Deluxe, and never looked back.

It works marvelously fast with a 2.8GHz P4 512K L2 cache
and 800MHz FSB (Northwood core).

We got hit pretty hard by a virus last year
on an aging Windows 98/SE machine.

We now depend a LOT on Drive Image 7 to create and
restore "image" files of our C: partition on the
new ASUS motherboard with Windows XP/Pro.

This is the fastest way we know of recovering from
a destructive virus or worm.

This software (now acquired by Symantec and re-named
"GHOST") does not appear to work if C: is on a RAID 0,
however.

We're planning right now to build an experimental
machine which will also have a single HD for C:,
plus a RAID 0 with 2 x SATA HDs @ 40GB each
(80GB total "striped").

On this special-purpose RAID 0, we plan to
store ONLY the Internet Explorer cache, and
possibly also the Windows swap file.

Because the IE cache tends to get large,
the more so as we browse the Internet,
Drive Images of C: grow larger accordingly.

By moving the IE cache to different drives,
C: stays quite static.

Moreover, the Windows swap file is volatile and
does not need to be saved between shutdown
and startup. So, it too can be assigned
to such a RAID 0.

And, for our database, we will go with a
single large 300GB PATA Maxtor with 16MB cache
(which we just bought at Office Depot at 50% discount)
and possibly add future SATA drives of similar size.

Another way of insuring "snappy" program launch
speeds is to make sure you have extra RAM,
which reduces the need for swap file I/O
in the first place.

I hope this helps.


Sincerely yours,
/s/ Paul Andrew Mitchell
Webmaster, Supreme Law Library
 

qbker

Distinguished
Dec 27, 2004
29
0
18,530
0
Yes, that has answered my question, thanks.
So SATA+RAID0+NCQ do all work together, it is just a matter of how much performance gained depending on how my system is run (server or single user).
Even if it is only marginal, I would still like to think that SATA+RAID0+NCQ would run faster than just SATA+RAID0.
 

supremelawfirm

Distinguished
Jan 9, 2005
47
0
18,530
0
Yes, and that's the main reason why it has already
been developed and implemented for file servers
using SCSI subsystems, e.g. Ultra SCSI 160 & 320.

Head movement can be very time-consuming, particularly
if it happens a lot. If a controller can keep a
read/write head positioned over the same or nearby
tracks, the cumulative amount of time being spent
moving the head from one track position to another
can be reduced considerably.

Several years ago, we prepared presentation to a
minicomputer user group, in which we documented
these important statistics:

(1) rotational speed and rotational latency
(latter is 1/2 the time required for one rotation)

(2) time to read or write one sector (which is
derived from sector density and rotational speed);

(3) access time, or the time required to re-position
the read/write head in worst case, average, and
best case situations.

As CPU speeds have climbed into the stratosphere,
the microelectronics used in IDE controllers have
improved apace.

You can see how Native Command Queuing is ONLY
going to help if there are a LOT of disk I/O
requests pending, and those requests are spread
across disparate cylinders that require a
lot of head movement to access completely,
i.e. process the entire set of such I/O requests.

Now, we can appreciate how server environments would see
this condition quite often, but workstation settings
would not see it as frequently.

What we want to test here is the benefit, if any,
that may be obtained from assigning the Internet
Explorer cache to a RAID 0 of some kind, most
probably using 2 x SATA drives.

The only way to evaluate RAID is to do a scientific
comparison of a single Raptor, for example, with
multiple identical Raptors in a RAID 0 setup.

When a single Raptor is compared to 2 x 7,200 SATA
drives, there are two factors which explain the
outcome (at least): differing rotational speeds, and
RAID v. no-RAID.

This is not good scientific method.

Don't forget: P-ATA/133 also has a lower maximum
of 133MBps as compared to S-ATA's higher maximum
of 150MBps.

It is better to design experiments which
measure the effect that a single factor has on
the outcome variable, i.e. data throughput.

Clearly, the Raptor is superior not only because
of its fast rotation speed (10,000 rpm v. 7,200 rpm);
also, Western Digital found that the load on the
bearings was much greater if it used the same
platter size as they were using on their 7,200
IDE ATA/100 drives.

So, they reduced the Raptor's overall capacity
by significantly reducing the platter diameter.

This, in turn, reduced the average time required
to re-position the read/write head from one track
to the other: obviously, the worst case -- from
inside track to outside track -- would automatically
be much lower solely because the platter diameter
is smaller.

Now, setting aside NCQ for the moment, one 74GB Raptor
is going to perform differently from two identical
74GB Raptors in a RAID 0 setup: the latter is capable
of a much higher throughput, necessarily.

Of course, other factors will also affect the outcome
variable, such as whether the controller is doing
computations in hardware, firmware, or software.

Nevertheless, CPU's have become so fast, the penalty
to be paid by doing these calculations in software
appears to be negligible now.

Our hypothesis (as yet untested) is that the Internet
Explorer cache, on average, exhibits the very same
or similar conditions which benefit greatly from
NCQ on file servers with multiple I/O requests
pending.

If you look at complex web pages which display a lot
of separate graphic images, each subregion of that
web page will map into a different Windows data file.
And, a web "page" is often much larger than a single
CRT "screen".

Even if a single data file is allocated sequentially
on a hard disk, it is very UNlikely that multiple
data files will be allocated sequentially on that
same hard disk. In this instance, we are referring
specifically to the disk subsystem which hosts the
Internet Explorer cache, and possibly also the
Windows swap file.

We'll let you know what we find: I am not completely
convinced that NCQ will NEVER benefit a high-performance
workstation which accesses the Internet frequently
and which is storing upwards of 20GB of files
at any given moment of time in its browser cache.


Sincerely yours,
/s/ Paul Andrew Mitchell
Webmaster, Supreme Law Library
 

Similar threads


ASK THE COMMUNITY