NCQ usage/support in linux

NCQ usage/support in linux

[xerces8]

Hello!

I have a few questions about SATA NCQ.

Background story:
Many users noticed that concurrent (by two programs) read access to a disk
under Windows goes very slow with modern systems, using SATA AHCI mode.

References:
 - "Slower concurrent disk access with NCQ ?"
<http://forums.hexus.net/hexus-hardware/138640-slower-concurrent-disk-access-ncq.html>
 - "Slower with NCQ ?, Concurent access" <http://forums.storagereview.net/index.php?showtopic=26864>
 - "NCQ: Best Upgrade For a Power User!" <http://forums.storagereview.net/index.php?showtopic=26965>

(for short summary you can read the 3rd page of the last forum topic linked above)

Here is the gist of the issue:
--quote from a forum post--
320GB WD Caviar:

Windows:
63 MB/s: 1 instance
28 MB/s: 2 instances (both at 0% position)
9 MB/s: 2 instances (0% position and 90% position)
15 MB/s: 10 instances (10% gap between each)

Linux:
63 MB/s: 1 instance
63 MB/s: 2 instances (both at 0% position)
45 MB/s: 2 instances (0% position and 90% position)
48 MB/s: 10 instances (10% gap between each)

Linux, stock install of Fedora 9 x64 (2.4.25 kernel).

Linux commands issued:
dd if=/dev/sdb of=/dev/null bs=256K skip=200K (skip was incremented by 200K each instance which
corresponds to about 50GB forward in this 320GB drive)

all dd invocations started at the same time (batch).

throughput measured with:
iostat -m 1 /dev/sdb (m shows in megabytes, 1 is the update interval of graph every second)

Windows commands:
dd if=\\.\PHYSICALDRIVE1 of=NUL bs=256K skip=200K (as above)
--end quote--

My question is: How does linux deal with NCQ ?
Is it used by default ? Any known problems with it ?

In the forum the Seagate 7200.11 harddrive family is mentioned as the only one
not having the performance issue. Does those drives also behave exceptionally
in linux ?

Thanks for any answers.
Sincerely,
David

Re: NCQ usage/support in linux

[Jeff Garzik]

xerces8 wrote:
> Hello!
> 
> I have a few questions about SATA NCQ.
> 
> Background story:
> Many users noticed that concurrent (by two programs) read access to a disk
> under Windows goes very slow with modern systems, using SATA AHCI mode.
> 
> References:
>  - "Slower concurrent disk access with NCQ ?"
> <http://forums.hexus.net/hexus-hardware/138640-slower-concurrent-disk-access-ncq.html>
>  - "Slower with NCQ ?, Concurent access" <http://forums.storagereview.net/index.php?showtopic=26864>
>  - "NCQ: Best Upgrade For a Power User!" <http://forums.storagereview.net/index.php?showtopic=26965>
> 
> (for short summary you can read the 3rd page of the last forum topic linked above)
> 
> Here is the gist of the issue:
> --quote from a forum post--
> 320GB WD Caviar:
> 
> Windows:
> 63 MB/s: 1 instance
> 28 MB/s: 2 instances (both at 0% position)
> 9 MB/s: 2 instances (0% position and 90% position)
> 15 MB/s: 10 instances (10% gap between each)
> 
> Linux:
> 63 MB/s: 1 instance
> 63 MB/s: 2 instances (both at 0% position)
> 45 MB/s: 2 instances (0% position and 90% position)
> 48 MB/s: 10 instances (10% gap between each)
> 
> Linux, stock install of Fedora 9 x64 (2.4.25 kernel).
> 
> Linux commands issued:
> dd if=/dev/sdb of=/dev/null bs=256K skip=200K (skip was incremented by 200K each instance which
> corresponds to about 50GB forward in this 320GB drive)
> 
> all dd invocations started at the same time (batch).
> 
> throughput measured with:
> iostat -m 1 /dev/sdb (m shows in megabytes, 1 is the update interval of graph every second)
> 
> Windows commands:
> dd if=\\.\PHYSICALDRIVE1 of=NUL bs=256K skip=200K (as above)
> --end quote--
> 
> My question is: How does linux deal with NCQ ?
> Is it used by default ?

We turn on NCQ if its available on both disk and controller.  In 
general, we program your hardware to go as fast as possible while still 
reliably accessing data.

	Jeff