Hi!

In theory NCQ should speed up things when multiple threads/taks/processes/programs are accessing the hard drive, right ?

But I see the opposite.

My mainboard (details below) can run its ICH9R SATA adapter in legacy or AHCI mode.
I noticed, that in AHCI mode (it allows NCQ, as opposed to legacy mode) I get huge (reproducible) slowdowns.

I did the following test procedure:

Preparations
- create a test directory , like : md C:\src
- put two large (gigabyte) files in it ([1])
- create a destination directory, like : md c:\dest

Test run:
- reboot (to make same conditions for each test run)
- open a command prompt (cmd.exe)
- run the command : copy c:\src c:\dest
- while the copy is running, open the start menu and start Wordpad
- after Wordpad starts, close it
- from the start menu start (and again immediately close) also: Notepad, Calculator, Paint
(They are all under "All Programs"/"Accessories")
(note: do this while the copy is running)

Measure the time it takes to start each program (no need for a clock, counting seconds in your head is more than enough).

What happens on my system is this :
- in legacy mode, each program starts in a seconds (Wordpad, being larger, takes 3-5 seconds)
- in AHCI mode, it sometimes it takes 2 seconds, sometimes it takes a minute ! Often, it takes 10 seconds just for the Start submenus to open.


I tried several operating systems: Windows 2003 Enterprise, Vista Enterprise, Windows XP Professional SP2, Windows 2008 x64
In all cases: legacy mode works smooth, AHCI mode have the "minute to start wordpad" problem.

Driver used:
- for legacy mode : Windows builtin Standard IDE driver
- for AHCI mode :
-- Windows XP and 2003 : Intel Matrix Storage Manager v7.8
-- Windows 2008 and Vista : msahci.sys (Windows builtin AHCI driver)
-- Vista : also tried Intel Matrix Storage Manager v7.8

There was only one exception to the "AHCI is much slower" result: Windows XP Pro SP2, updated to SP3, then cloned to another partition and finally changed from legacy mode to AHCI (by: force install the intel driver, reboot, change setting in BIOS, boot Windows)
In this case, I did not see the problem.

I contacted MS newsgroups, Asus support, Intel support. But no useful response. (Intel said : We have not replicated such issue nor have other customers reported it.) Talks with Asus are ongoing.

Did anyone experience this ?
Solution ? Just run in legacy mode ?


Regards,
David

[1] - I used the two freely downloadable iso images from this page : Windows Server 2008 Trial Software

My system details:
mainboard : Asus P3K-E WiFi AP , BIOS version 1013, Intel P35 & ICH9R chipset
CPU: Intel Q6600 (not overclocked)
RAM: 4x2GB DDR2-800 (Transcend JM800QLU-2G)
gfx: Gigabyte GV-NX88T512HP/typeA
HD: Western Digital WD7500AAKS - connected to one of the ICH9R SATA ports
optical: Samsung SH-S203P - connected to one of the ICH9R SATA ports
(problem same even if I disconnect the optical unit)

Your I/O queue depths aren't enough for NCQ to do anything beneficial. Your tasks are far more likely to be bound by something else in your computer, including driver issues.

Well NCQ has been a very mixed bag from the start. Back in the early days I think pretty much everyone was saying turn it off for desktop use, full stop - what we think of as multiple tasks (ie file copy and launch notepad) isn't the sort of multi-tasking servers with databases etc have to do. More recently I saw a thread here on the 7200.11 Seagate drive, where the poster used a little program that tested read speeds I believe it was at different segments of the drive, and the results looked like even when trying to access two places far apart (in the physical sense) on the drive the results were very promising. Personally I don't have the variety of drives to be able to comment much, however the conclusion I took away and am still sticking with is that when in doubt turn off NCQ, and that at the very best as the saying goes "YMMV". There are settings for the controllers and/or drives in Device Manager in XP, you can give toggling those a try (tagged command queuing, etc - I'm not sure but it's possible I think that the OS tried to optimise commands via it's command queuing system and then the drive does it's own as well and together they are destructive).

On a complete tangent, it seems support at even the largest "respectable" tech companies these days is a total waste of time and I find that ridiculous. Like for example when my LSI RAID card didn't work on my Gigabyte mobo, Gigabyte said oh well it's LSI's responsibility to take care of compatibility and LSI said it's Gigabyte's. Very helpful for the customer, and that's just one example of MANY.

In theory, NCQ starts being relevant for queue depth >2 (intensive concurrent usage requires at least 32 queue depth)
In theory, NCQ has a very small execution cost, about 0.1 ms, that you may see on single access like benchmark.
==> You "may" explain a 1s delay with 1000 io which is really above your test usage so it looks more like a driver or OS parameter issue

In device manager there are no options for the Intel AHCI driver.

The only choice I see is AHCI versus legacy mode.

So I should just switch over to legacy and forget about AHCI ?
Are there any important differences ?
(the only one I know is hot-plug support in AHCI)

David

PS: Yes, support sucks big time...

You need an AHCI controller to support NCQ as well.

Of course. So besides NCQ and hot-plug, is there anything else ?

I thought the other big aspect of ACPI was power management. So I don't know about you, but that doesn't seem too important to me as far as the drives are concerned.

I didn't mean options for the ACPI driver. I meant options for the harddrives themselves (always have at least a couple options) and/or the Intel storage controller.

In general it sounds like you should stick to legacy, unless you find some options you can play around that help.

Man I musta been tired or something, don't know how I went from AHCI to ACPI. I think long story short you should go with legacy mode if AHCI hinders performance, plain and simple.

Given that nobody else has seen this in Windows on an ICH*R, consider other faults.

What hard drive do you have? It could have broken SATA2 firmware.

See my first post.

Nope, not seeing this problem, and I've exclusively run my personal systems in AHCI mode since ICH7.

In addition the majority of Workstation we build are either run in AHCI or RAID mode on ICH7, ICH8 & ICH9, and have not seen this

We tend to exclusively use WD 10000rpm Raptors and Seagate 7200rpm Barracudas

I suspect the problem is really related to your system, most likely some software/driver/firmware is misbehaving or you a faulty component

NCQ vs. NO NCQ benchmarks:
http://home.comcast.net/~jpiszcz/ncq_vs_noncq/index.html

How do you turn NCQ on and off ?

I tried with a new HD (Samsung F1 640GB - HD642JJ). It seems a bit better (less slowdown in AHCI mode), but still noticeable difference compared to IDE mode.

I also saw Intel released a new version of their driver (Intel Matrix Storage ... v8.2), but it does not help.

personally I think the intel ncq implementation isnt optimal, I have found jmicron to have better results on NCQ drives, in addition some drives have very bad NCQ implementation so its a shot in the dark how well it performs, my advice is only enable it for the os drive and all data storage drives and especially media drives use legacy ide mode.

Another problem is all ICH ide drivers earlier than 8.0 dont allow write cache in ahci mode which will have a significant impact on performance.

I repeat my question : How ?

(... to turn on and off NCQ)

That drive does not seem to support NCQ properly as one user posted on my thread. Put a Seagate 7200.11 and try with that.

Also... Windows AHCI driver does not work with NCQ properly. Use only Intel.

I am sorry, this all is misleading. It all comes down to blocking. I don't really care if it takes 5 mins for a copy to finish on non-ICQ drive while I cannot use anything else until that copy finishes when i can have an NCQ drive that takes the same copy to finish in 10 mins but allows me instant access to anything else i might throw at it while doing that copy.

The results are very pronounced in real-world as opposed to minute percent performance differences in benchmarks. I also do have 15K U320 Cheetahs and i would take a SATA drive with properly working NCQ any time of the day.

cant turn it on and off per drive on ICH chipsets, if bios is in RAID or AHCI mode NCQ will be used, if in IDE mode it wont be used.

I moved my samsung f1 to my jmicron 2 days ago, the jmicron is in ide mode and the ICH is in RAID mode.

On the ICH7 port playing videos had stutters and the drive seemed to generally struggle to play the file, delays on seeking etc. waiting for drive.
On the jmicron the hd light is on way less and seeking is instant, not a single stutter even when machine is under heavy load.

So the results are clear on this drive.



yeah depends on the drive, on the samsung f1 in NCQ mode the entire system pretty much freezes when copying large files. The iastor process locks up which is part of the ICH drivers.

I now plan to do large file copying tests on all the other drives on the ICH controller to see if they are actually compatible.

In Windows or Linux?
Not sure in Windows.

Windows.

Why don't you just run hd_speed for 1, 2 and 9 instances and post the results? If NCQ is working properly, your throughput will remain the same. If it is not, there will be a large drop in throughput for anything more than a single instance. I wrote a whole thread on NCQ based on czr's initial post. Why to waste all this time if you can just run a bench for 5 mins and find out for sure if it is working or not?

I spent months on my config with Samsung 500GB... Apparently, the drive did not implement NCQ properly even though NCQ was advertised all over. Save yourself the hassle.

OK, I started HD_Speed v1.5.4.72, selected the F: drive (a 300 GB partition in the test disk) and Block Size 64KB.
A single instance shows 110 MBytes/s. If I start another one, both show around 33 MB/S.
If I start a third one, all 3 show 16 MB/s.
This is a Samsung HD642JJ (F1 640GB) disk.

Controller is in IDE legacy mode.

I get similar results if I select PHYSICALDRIVE1 instead of F:.

The results for the WD7500AAKS (PHYSICALDRIVE0 , 64KB block size):
- 1 instance : 95 MB/s
- 2 instances : each 34 MB/s
- 3 instances : each 19 MB/s

I will now switch to AHCI mode and post results in a few minutes (hopefully).

I now did the AHCI mode.
On the WD7500AAKS it behaves very strange. I again started 3 instances with 64KB block size.
One instance runs at full speed (about 90 MB/s) for several seconds (like 10), during which the other two are at 0.00 KB/s), then another instance gets the full speed and the other two again have zero.
See picture.


The HD642JJ on the other hand run at 110 MB/s in one instance. When 2 are running, both measure 3.6 MB/S (10 times less than in IDE mode) and 3 instances each measure 2.5 MB/s.


In each case AHCI sucks compared to IDE mode.

This was again on Windows 2008, with MS AHCI drivers ("Standard AHCI 1.0 Serial ATA Controller" msahci.sys , version 6.0.6001.18000)

I will try Intel drivers next.

RAID/AHCI Software - Intel® Matrix Storage Manager , version 8.2

Same behavior as MSAHCI.

The picture I mentioned two post back.

Please see your throughput not on each HD_speed instance. You cannot reliably get averages for more than 2 instances since NCQ alternates the reads.

TASK MANAGER > PERFORMANCE > RESOURCE MONITOR > DISK

See the throughput there. Just have a look at my thread for testing methods.

http://forums.storagereview.net/index.php?showtopic=26965

And please submit your results there too so that we can have an idea.

PS: Do not use Windows Storage Controllers. Use Intel's. I tested long ago and default windows AHCI controller did not perform at all with NCQ.

Well, the Resource Monitor value more or less matches the sum of hd_speed values.
(just tried with the 2 HDs in a raid 0 setup)

Why do you say the averages in hd_speed are not accurate ?
They are (data transferred) divided thru (time past). Why should it not be accurate ?

Example:
- thread 1 transfers 5 MB in 5 seconds
- thread 2 transfers 15 MB in 5 seconds
- thread 3 transfers 10 MB in 5 seconds

So:
- thread 1 shows average 1 MB/s
- thread 2 shows average 3 MB/s
- thread 3 shows average 2 MB/s

total throughput by adding the averages : 1+2+3 = 6 MB/s
total throughput by counting data and time : (5+15+10 MB) / 5 s = 30 / 5 = 6 MB/s

Same result.

Listen, we are interested in THROUGHPUT... not what each ones value represent. And res monitor is the proper way to diagnose that (provided that nothing else is reading / writing to your system).

But yes, you can get averages on hd_speed as well. Just make sure 1 min duration is chosen below (1:00 or 60). Be careful not to make it 1 second (1). This way you can get the averages.

Now do this:

Run a single instance of hd_speed (read, 256k blocks, 0% position, 60 secs duration) ... and record the average from hd_speed average section. compare if you wish from res monitor. This you will use as reference.

After, run 2 instances of hd_speed:
1st instance: read, 256k block, 0% position, 60 secs
2nd instance: read, 256k block, 50% position, 60 secs

When they finish, get their averages from hd_speed and sum them up to find your total throughput. Make sure START is pressed almost at the same time in both instances. You can monitor their values from res monitor while they are running to have an idea of total disk throuhput.

After this, run 9 instances of hd_speed. Same data above... only each instance will be seperated by 5% increments... For example, 1st instance at 0%, 2nd at 5%, 3rd at 10%, and so on.

Now see your total throughput from res monitor (better than summing up 9 instances from hd_speed).

And post your results here. All these i explained in the other thread... so did czr. This is just redundant info now that i had to type.

No need to get upset.

I will do a bunch of tests and then post them in the other thread.

Hi Xerces8, thx for the link to this thread, and hello to the others!

I have the same problem with two RAID 0 (one with 2x WD2500YD, one with 2x WD6400AAKS).
What you describe here, Xerces8, is exactly what I "feel" in every kind of application I run: one process gets the full bandwidth, and other processes can't access the drives at all for long periods of time, so they hang completely, including GUIs. Movies stutter, etc.

And I've got one problem more: since I'm using RAIDs, I can't disable NCQ at all, because the controller has to stay in RAID mode.


No, 6_6_6, we're not. Throughput is what we get, although we don't want it, we want Responsiveness.

I would seriously prefer this behaviour described by xerces8:

This is what I'm used to. It's possible to watch a video while doing other stuff (like copying, comparing, whatever), while it is not possible in all cases with our problem.

Oh, and even setting the drives to SATA 1 with a jumper doesn't help, because they still support NCQ, so it's still used.

You get responsiveness with throughput. For example, this drive posts 25 mbps throughput with 9 instances while 7200.11 posts 80-90 mbps with 9 instances. This drive does not support NCQ... 7200.11 does support NCQ as evident from these throughput figures and my system is responsive. But unfortunately, I had to suffer for so long (with the same probs you are suffering, see my thread) with drives that were supposed to have NCQ.

Do yourselves a favor, and grab a 7200.11 next time... And thank czr for it.

Not in our case.
Throughput is fantastic (e.g. >120 MB/s), while one application hangs completely nevertheless. I guess you have to see it to understand our problem. The problem is that one process gets all available throughput for some time and all others get none at all.

I will certainly not throw out my four (!) Western Digital drives which are simply great and buy new ones that I don't even want to have, only because this Intel crap does something weird with these drives. I had no trouble with my old nVidia nForce 4 SLI board, so it's not only the drives fault.

I'll try to get some serious SAS/SATA RAID controller and see if it helps.

Not in our case.
Throughput is fantastic (e.g. >120 MB/s), while one application hangs completely nevertheless. I guess you have to see it to understand our problem. The problem is that one process gets all available throughput for some time and all others get none at all.

I will certainly not throw out my four (!) Western Digital drives which are simply great and buy new ones that I don't even want to have, only because this Intel crap does something weird with these drives. I had no trouble with my old nVidia nForce 4 SLI board, so it's not only the drives fault.

I'll try to get some serious SAS/SATA RAID controller and see if it helps.

No it is not. At 9 instances of the program, the throughput is fantastic at 7200.11 at 90 mbps... while it is 26 mbps on the WD.

That is why my system is responsive when i do multiple disk operations at the same time... and yours is NOT.

That exactly is how a properly working drive with NCQ behaves -- alternates the disk operations between processes so a single program does not have blocking disk access while all the rest waits. WD does NOT have proper NCQ implementation (reference: xerces8 and kreatif benchmarks).

Shit.

I've read more threads here at SR about NCQ, and I fear you're right, my drives suck. I don't find appropriate words to describe my frustration...

It will be quite a financial loss if I try to sell my WD6400AAKS's at eBay, and buy 7200.11 instead.
Do you know if the 640 GB 7200.11s have the "working" NCQ as well? I would prefer them because of their price and the two 320 GB platters.

But what do I do with my two 250 GB WD2500YD...

SHIT!


- The trouble is: NCQ like you describe it in "NCQ: Best Upgrade For a Power User!" is exactly what I am expecting from NCQ, and exactly what I need! I hate my PC becoming faster and faster with every upgrade, except the ******* hard drives.

Note that it is not all the disks fault.

For example here is how it behaves under Linux (RIPLinux 6.1 - 64bit kernel 2.6.26):

the HP nw8440 laptop (AHCI mode)
Disk : ST980825AS

1 instance : 42 MB/s
2 instances : 36 MB/s
9 instances : 35 MB/s

Windows values were : 42/11/15

--
HW: P5K-E WiFi
BIOS setting : AHCI

Disk : WD7500AAKS
1 instance : 96 MB/s
2 instances : 78 MB/s
9 instances : 75 MB/s

Disk : Samsung HD642JJ
1 instance : 114 MB/s
2 instances : 92 MB/s
9 instances : 89 MB/s

In Windows it was: 99/90/26 and 116/14/25

The test were instances of this command:
dd if=/dev/sda of=/dev/null bs=256K skip=xxxx # values for offset
measurement with : iostat -c sda

xerces8, thanks a bunch. this is very interesting find. I will test my Samsung on linux when i have time.

Conlusion would be: Drives implement NCQ properly...but:

a ) Microsoft/Intel conspired to support only Seagate 7200.11 drives on NCQ front.
b ) Seagate is the only hd manufacturer to find the problem in Intel/Microsoft NCQ implementation and fixed it at their firmware (since 7200.10 does not work properly) or altered their NCQ implementation.

Whatever the scenario / configuration / drivers / controllers are, nobody reported NCQ working properly on windows except 7200.11s thus far. Hence, there is no choice.

FAT punisher, I have no idea. I specifically hunted the model czr reported as working instead of trying my luck with a bazillion of hard drives.

If someone will have that drive, I hope they will be kind enough to benchmark and post the results for rest of us.

I have a question..

Does this mean that with NCQ enabled (7200.11s) as a main bootup OS drive it would beat using any RAPTOR model or specifically at least a 10k 150 gb?

It sounds like everyone been getting ripped off for years with NCQ drives.. as I must be on my 4th or 5th by now..

NCQ's main benefit is in high-I/O-queue depths... which, is NOT a typical desktop...

Considering Velociraptors and whatnot have a working NCQ implementation (heck regardless if they do), in typical desktop loads they have a significant speed advantage over their 7200rpm brethen, and NCQ is not a consideration.