Meausuring QD in win7

[Ourasi]

The write latency goes up without NCQ but it is the 4k reads at qd64 that really get hit. Without NCQ the read performance at qd64 stays around the same as qd1. (As can be seen here, which I found randomly during a search)

I still can't work out why reads get so much better at higher qd's. (Not to say they are good at qd1) It must be something to do with NCQ but how that works I have no idea.

Took a stab at some explanation here: http://communities.intel.com/thread/9838?tstart=0
The reason the Intel shines at higher load/QD, is down to NCQ and the performance of their controller/layout...

[Computurd]

also numbers of channels on the controller. 10 of them right? you would think that sure helps a bit
nice link ourasi, ty!

[Anvil]

SATA 6Gb/s includes a few enhancements wrt NCQ

(source wikipedia) Link
The new specification contains the following changes:
A new Native Command Queuing (NCQ) streaming command to enable Isochronous data transfers for bandwidth-hungry audio and video applications.
An NCQ Management feature that helps optimize performance by enabling host processing and management of outstanding NCQ commands.
The NCQ Management feature looks interesting.

[Computurd]

i asked at another forum about how QD is assigned and received this answer...emphasis added

the queue depth is not "assigned" in any way, it's just how many commands are outstanding on your storage subsystem
meaning -> the more random IO -> the higher the queue
the slower your Storage -> the higher your queue
now you got a ssd which is fast as hell on random IOs (because of low access times) and a workload which is very light -> so you got no queued commands for your storage because your storage is not the bottleneck
some ssd benchmarks do even use a queue depth of 64, take a look at the AS SSD Benchmark

this is making more sense to me tbh. during my own testing i have come to this conclusion as well...with my HDD caviar black during gaming and different usages my QD goes much higher than with ssd. i think this is where you can see where there is benefits to raid and SSD...it keeps the QD low because there is such high throughput at the lower QD, tons of channels, etc.
for instance, at a QD of 1 a raid array will have the throughput of a single ssd at a queue depth of 32. so the queue depth doesnt go higher, thus lower access times...
for instance sequential read at QD of 1 with my controller...
4k- 136
8k- 200
16k 355
32k- 519
64k- 692

that is at a QD of one, and helps explain why my QD never goes over two on my array. on a caviar black during same usage it goes 6+

[Levish]

they don't "keep the qd low", their latency is so low (.1ms or less compared to 12ms) that they can find and retrieve anything small before the normal drives could even find them. This is also why if you work with mostly large files SSD's will not be a major benefit.

At a QD1 of 1 with small files (below stripe size) you will have single SSD speed. At a QD of 2 with a large file you will have however many drives worth of performance that the file is striped across (i.e. a 128k stripe will go across 2 drives for a 256kb file, 4 drives for a 512kb file etc). But this isn't entirely realistic since its possible that the 2nd 4k file might be on one of the other disks in real useage.

But I believe you pay for using too low a stripe by incuring extra IOPs on the controller/CPU and drive so there are no free lunches.

*side note*

ideally if you want to look at disk information that windows provides you want to use performance monitor (aka system monitor) and load up the respective counters or creat a new counter log for physical or logical disk (whichever you are interested in measuring).

[Computurd]

i agree levish... they do find the files extremely fast and thus "keep qd low". if they were slower the qd would be higher. so the throughput and access time is what 'keeps it lower' is what i am trying to say

mainly what i am trying to say is that the QD is not 'coded' into an app as was started above. your QD is relative to the speed of your storage system.

[One_Hertz]

mainly what i am trying to say is that the QD is not 'coded' into an app as was started above. your QD is relative to the speed of your storage system.

Do you really think all apps que up commands (even if the storage device is slow)?

[Computurd]

i dont understand what you are saying hertz...i dont think the app queues it up....the QD is not coded into the app as stated above.



i think :if the program needs info, it asks for it. if the disk is able at that very millisecond to supply that info, it does. if not, it has to wait for whatever is taking place at that second to finish, thus a queue forms...
whether or not that is 100 percent accurate I am not sure, but it is close. i dont believe that a program would say "ask for information at a QD of 32" i think it just says "ask for information:"
well a program doesnt say anything of course

[One_Hertz]

i dont understand what you are saying hertz...i dont think the app queues it up....the QD is not coded into the app as stated above.



i think :if the program needs info, it asks for it. if the disk is able at that very millisecond to supply that info, it does. if not, it has to wait for whatever is taking place at that second to finish, thus a queue forms...
whether or not that is 100 percent accurate I am not sure, but it is close. i dont believe that a program would say "ask for information at a QD of 32" i think it just says "ask for information:"
well a program doesnt say anything of course

What I was meaning to say is that I bet some apps will not make another request until their previous one has been met, thereby forcing the queue to be 1 regardless of anything. Other apps may be more flexible and try to do more than one thing at a time (send more requests).

[Computurd]

OH i see..hmm that is an interesting thought, and one that would take some testing to answer

i have however done some rudimentary testing with the QD and several configurations and these are my results...
what i did is load a full run of PCMarkVantage and ran it on three setups...one uber SSD array, one HDD raid 0 (two cav blacks) and then a single HDD... monitoring with performance monitor on another screen.

on the ssd array the QD was nil most of the time. when QD was raised it would shoot to 1 or rarrely 2, but only for a spike, a very short amount of time. avg QD was very low, usually nonexistent until spikes.

the raid 0 HDD array the QD shot up more frequently, and higher, QD of 2 and 3 were the norm, and the average QD was much higher, constantly. also the spikes lasted longer.

and as you can imagin, QD on the single HDD were much higher, regularly hitting 4 and 5, 6 8...the spikes were much longer in duration, and most importantly the average QD was high....always there no matter what.

of course as you say it is probably relative to the complexity and sophistication of the program. if you have multiple windows open and multiple programs the qd does go higher of course, they all want attention!
i think that if some programs would only wait for thier request to be answered before asking for more, that you would run into big problems when running multiple apps. they would have to wait a ton of time, thus running slower...

[Ao1]

Comp, try this…
Run the Anandtech heavy use Iometer config @ qd1 on a single hhd.
Run the Anandtech heavy use Iometer config @ qd1 on a single ssd
Run the Anandtech heavy use Iometer config @ qd1 on a soft raid 0 ssd array
Run the Anandtech heavy use Iometer config @ qd1 on a hard raid 0 ssd array

Repeat test runs with incremental increases of qd up to 7.

I think you will find that read/ write speeds will not vary significantly on any of the set ups once access times are taken out of the equation.

[Anvil]

Computurd,

You may test your "thoughts" on this but you have already been given the answer to your question.
There is only one way for an applicaiton to produce QD>1.
The answer is Multi Threading.

The reason to why an HDD shows higher QD is because it is slow compared to an SSD, it has nothing to to with the application unless the application is multi threaded.
The faster the storage the lower the QD. (for the same given task)

[Ao1]

Comp, if you run the benchmarks I propose I think you will find out why SSD raid 0 is not faster than a single drive in non enterprise applications. Access time is much more important than read/write speeds for the vast majority of desktop tasks at low qd’s and that does not improve with raid 0, in fact in can make it worse.

With raid 0 all you are doing is improving sequential read/ write speeds and that is not much use for the vast majority of desktop tasks.

I’m going to bet that at qd 1 to 7 (using a heavy desktop usage pattern) that read/ write speeds, once access time has been subtracted, is not going to be that different between hdd/ ssd & raid 0 ssd, but I stand to be corrected.

[Computurd]

i can see what you arte saying audieneofone, but the thing is that you are asking for benches to prove real world performance difference.
NO one SSD will load games faster than an array of like disks, provided that you have a sophisticated controller that can handle the load. look at napalm in raid 0. ridiculous fast. i think steve-ro and others with arrays could also chime in here.
i hate to quote myself, but here i am going to do it anyway, i posted this in another thread :

I agree with much of what audienceofone states, it is truly up to the end user.
alot of the links provided by audienceofone that show no benefit in raid usage, however, focus on load times. i think sometimes load times is kinda like 4k. people just think it is a definite statement of performance, and you definitely want to be careful of making assessments on a few factors only.
for instance, the photshop loading times. what about when you are manipulating images and such when you are actually using the program? there is no data for that. same as the autocad loading times assessment. both of those programs use VERY large files that benefit tremendously from enhanced throughput, however, it is hard to bench such things. when you are manipulating images, is there hangs and lags? how does actual in-program performance handle. how many people just sit around and load autocad repeatedly? unless they are some sort of weirdo, no one. they load the program to USE it. with very large files that are acceseed and manipulated constantly.
one group of people who can definitely attest to this is anyone who has manipulated video files and things of that nature, especially professionals. that is why they go for the z-drives and the pcie solutions for these types of usages. they know that the load time is the most IRRELEVANT thing that they do when they use a program. a developer loads a video. takes 15 seconds. then manipulates/edits that image for HOURS> that is where the performance advantages of the raid arrays really shine., and we arent even speaking of the writing of these types of large documents and files, etc, after they are done working with them.

about the only benchmark that is truly indicative in alot of storage capacities anymore is pcmarkvantage. and that is because it doesn't load synthetics when it does the assessment, it actually performs tasks that end users do on their computers every single day.
such as watching a high def video and also performing other tasks in the background. such as checking mail or working on a project.
importing digital photos into libraries and media players. cataloging those libraries?
what about when you transcode a highdef video to a portable media player? have you ever done that? i have with my HD zune player. takes time.
what if you are streaming to a xbox 360 or ps3 off your rig while you are also recording video from a television source on that same computer? time shifting? what if you are even watching HD video on that computer yourself while these other things are going on as well? do you think that is going to play smoothly?
how many times have you went through levels in your game and noticed lags and hangs when you are going through to other areas of the map? in alot of the newer games streaming data from the hdd comes into play alot, especially MMO and RTS games.
encryption and compression of ANYTHING is also taxing on your storage subsystem.
the computer is becoming more and more of a hub for all activities that are media related in the home. anyone who watched steve ballmers keynote at CES could tell you that. they are showing the direction of computing that it is going for the average end user, and it is multiple things at once and tons of HD vid. what happens if your weekly virus scan kicks in while any of the above things are happening??
pcmark vantage tests 26 different things of this nature in the default benchmark. of the 26 tasks performed only one is application loading. it is important, but not that important. performance DURING use is key.
you can look at pcmark vantage top users and see the real truth. not one person near the top of that is using a single ssd. probably not in the top hundred. they simply do not have the performance. it is all RAID SSD and acards in RAID.
of course this is a heavy heavy use of your computer that is being simulated, but it is USAGE testing. not loading testing. and it is the type of usage that everyday users do now, and will do increasingly in the future.
-------------------------------------------------------------------------------------
@anvil-i agree totally with that you are saying. i was trying to find the answer my own way, because i was getting some information that i felt was incorrect, which it was. i just do what i do when i am not sure, i find out for my damn self

[SteveRo]

Comp, if you run the benchmarks I propose I think you will find out why SSD raid 0 is not faster than a single drive in non enterprise applications. Access time is much more important than read/write speeds for the vast majority of desktop tasks at low qd’s and that does not improve with raid 0, in fact in can make it worse.

With raid 0 all you are doing is improving sequential read/ write speeds and that is not much use for the vast majority of desktop tasks.

I’m going to bet that at qd 1 to 7 (using a heavy desktop usage pattern) that read/ write speeds, once access time has been subtracted, is not going to be that different between hdd/ ssd & raid 0 ssd, but I stand to be corrected.

+1 with Comp - looks like all the top pcmark vantage scores are with controller cards and either SSDs or acards.

[Ao1]

+1 with Comp - looks like all the top pcmark vantage scores are with controller cards and either SSDs or acards.

That assumes that pcmark vantage represents real life

I’m not saying I’m right or wrong I’m trying to put theory to what I have physically observed. Would you agree that the Iometer test I propose would settle the discussion?

[SteveRo]

That assumes that pcmark vantage represents real life

I’m not saying I’m right or wrong I’m trying to put theory to what I have physically observed. Would you agree that the Iometer test I propose would settle the discussion?

It would provide additional evidence yes.

[SteveRo]

That assumes that pcmark vantage represents real life

I’m not saying I’m right or wrong I’m trying to put theory to what I have physically observed. Would you agree that the Iometer test I propose would settle the discussion?

I will say this also - for me - regular use being very simple - email and alot of browsing - the "snappiest" set up - quickest response from a mouse click - is a pair of SSDs in raid 0 on ich10r - actually acard 9010 raid 0 on ich.

[Ao1]

It would provide additional evidence yes.

Let’s go for it then. I could run the Iometer config on a single hdd and a single ssd. At a push I could pull out an ssd from my spare pc for soft raid, but unfortunately I sold my hard raid card some time ago.

[Tiltevros]

did u remember when napalmV5 made a program that opens 100+ applications???
just open the resource monitor and see the queue depth that it use.... every day applications and disk real usage...
the answer is there... just try it...
for me the queue depth of 69 application like word excel power point internet explorer and some other stuff.. is from 5 - 41 queues... now for mechanicals drives like a WD 320AAKS is like 50-161 queue depth

[Computurd]

ok i ran the 4k on single drive. the results are going to be shocking!!
the latency is higher on the single ssd with increasing Queue Depth because the single SSD simply cannot deliver as fast as eight. the single ssd wins at QD of 1, by .01 of a second and a slight win in mb/s. after that it is no contest. the right side is raid 0 eight vertex. the left hand side is single ssd.

NOW if you take into the fact that the queue depth on a single ssd is going to raise much more quickly, and that the successive access time of EACH queue depth on the ssd is going to get higher than on the array, the array is way way faster. the fact that the ssd has higher access time with qd, and also higher qd also, it is like a multiplier. that is why the raids far far exceed speeds of single ssd.

[Ao1]

Comp did you run the single ssd via the mobo or via the 9211-8i? If you ran it on the 9211-8i the results would have included unnecessary latency.

I've tried to run the same io config to compare Tilts results with 7 x X25-M's and I am generally faster at qd's below 4. (assuming it was an apples for apples test config). Over and above qd 4 and raid starts to clean up, but as I don't go above qd 2 for the vast majority of times that is irrelevant to me.

Raid 0 is much faster, that is not what I try to discuss. What I try to discuss is what is fastest for desktop users for whom speeds that occur above qd 7 are irrelevant.

From what you have said you can't get above qd 2 for most of the time. I think you will find that does not change if you use a single ssd on ich.

Again, I'm not trying to say raid 0 is not faster. It is, however it is only faster when you get to high enterprise qd's. At low qds it can be slower as you have seen for yourself.

Your results show that scaling is bad at low qd. At qd 8 you are getting 148% extra read speeds by multiplying your ssd's by 8. That is roughly 18% scaling for every ssd that you add and at qd 1 you are actually losing by 7%.

I'm more than happy to run and compare any io config tests that reflects desktop usage patterns (reads and writes at qd's 1 to 7.) if you want to do an apples for apples comparison.

[Tiltevros]

i have to buy the 8th intel computardddddddddddddddddddddd i will get u :P
Now for qd i told u something before just make a bat file to run like 20-30 applications in the same time and u will see ur QD to raise up just try it.. its fun every application handles its own QD like a perfect example the crystaldiskmark version 3... it runs @ 1 queue depth and @ 32 . OK now just make a program in access and try to search or even that just use the search from windows and u will see the QD.

[Anvil]

audienceofone

I've prepared a graph comparing QD on a single X25-M vs 2 x X25-M's in raid-0.

iops_qd_single_vs_raid0.png

Up to QD3 there is virtually no difference but from then on...

iometer, 4K 100% random read.

edit:
I compiled a new graph integrating Computurd's 9211-8i

iops_qd_single_vs_raid0_9211.png

[Ao1]

^

Nice. Thanks. That shows around the same as I am finding. Was your raid array hard or soft?

It seems Tilts 7 x X25-M's are quite a bit slower at low qd's than Comps 8 x vertex on the same controller.