35 replies to this topic

[MedEvil]

Inspired by this comparison of ramdisk drivers and the huge differences in the Random 4k performance, i was wondering, if anyone already tested the two drivers for performance, in this most important category for a system drive?

[Wonko the Sane]

NO , but someone already asked about this, even without inspiration.
http://reboot.pro/8804/page__st__70



Wonko

[Sha0]

Inspired by this comparison of ramdisk drivers and the huge differences in the Random 4k performance, i was wondering, if anyone already tested the two drivers for performance, in this most important category for a system drive?

If you mean the RAM disk performance (as opposed to the AoE, HTTPDisk, sector-mapped disk, file-backed disk performances), I thought that someone had posted screen-shots of their favourite timing software in the main WinVBlock thread.

[MedEvil]

NO , but someone already asked about this, even without inspiration.

Actually, if you look into the previous posts, the other question was: "If someone had made a test on their machine with firadisk and also another ramdisk as reference. Cause both testers claimed, that firadisk was way slow, without giving a reference to what kind of performance their machines were capable of."

Also that subtopic dealt with read speed of sequential files or top read speed, while this topic deals explicitly with performance of the Random 4k accesses.

[Wonko the Sane]

Actually, if you look into the previous posts, the other question was: "If someone had made a test on their machine with firadisk and also another ramdisk as reference. Cause both testers claimed, that firadisk was way slow, without giving a reference to what kind of performance their machines were capable of."

Also that subtopic dealt with read speed of sequential files or top read speed, while this topic deals explicitly with performance of the Random 4k accesses.

Sure .
The answer(s) is/are still (and AFAIK):
NO. (to the generic question about people having posted performance comparison)
and
NO. (to the specific question about reading/writing 4kb chunks)

You also have to take into consideration that, several different versions of BOTH Winvblock and Firadisk have been published lately (so ANY test not performed with the very latest of both may give incorrect or anyway meaningless results).

The understated part was (now explicited ):
Why - since both drivers are freely available and at least three of the ramdisks listed in the linked to page are also freely available, if you are so curious about the performance comparison you don't do it yourself? (and possibly publish the results )

Personally I find - without having EVER tested them - their speed "sufficient".
But you know how we don't agree on the "good enough" computing approach.


Wonko

[MedEvil]

if you are so curious about the performance comparison you don't do it yourself? (and possibly publish the results )

I actually intend to do that, just wanted to know, if anyone already done that.

But you know how we don't agree on the "good enough" computing approach.

I know, that i don't agree with Peters definition of "good enough", can't remember, that we ever had a fall out over that topic.

[Wonko the Sane]

I know, that i don't agree with Peters definition of "good enough", can't remember, that we ever had a fall out over that topic.

We did.
http://reboot.pro/10906/page__st__15


Wonko

[MedEvil]

That was so minor, not even worth remembering.

[L A M A]

Inspired by this comparison of ramdisk drivers and the huge differences in the Random 4k performance....

wow , so when properly optimized, ram disk can reach this level. Take a look at this monster Desktop PC, benchmarks would report much higher with these?

[Wonko the Sane]

That was so minor, not even worth remembering.

Hmmm.




Wonko

[MedEvil]

And heeeere they are! The long awaited test results.

Don't be confused about the low values shown in the screenshots. I ran the tests in a VM on my WinbuilderPC.

Let's start off with ImDisk.
 ImDisk.gif   25.95KB   58 downloads
It shows good performance with big files and then performance slowly drops as the files get smaller and smaller.
Would have loved to show QSoft Enterprise Ramdisk too, to show the max possible values in this setup. However all scripts for QSoft ramdisks have myteriously disappeared from this forum.
So we can only assume based on this test that max. possible 4k performance is about 2 times ImDisks.


Next stop Win2k3 ramdisk.sys.
 MS-RamDisk.gif   26.31KB   47 downloads
It's performance is seriously worst than ImDisks. However the smaller the files get, the better the M$ driver gets in comparison. At 4k it finally outperforms ImDisk and is almost twice as fast.
Being optimized for small files is the behaviour i'd expected from a driver, meant to power systemdrives.


Now to our two homegrown candidates.

Let's start with Firadisk. (v0.0.1.16)
 FiraDisk.gif   25.82KB   37 downloads
It's performance with big files is slightly worst, than that of the M$ one, but not that much that it would be noticable.
It's not until we get to the 4k category, that things start looking realy bad. Just half the performance of the M$ driver and still worst than ImDisk.


Last but not least WInVBlock. (v 0.0.1.7)
 WinVBlock.gif   29.38KB   29 downloads
Performace with big files is actually a bit better than with the two competitioners. However performance takes a huge blow in the 4k category. It drops again by about 50% compared to FiraDisk and delivers only a quater of the performace of the M$ driver.


So now of course everyone wants to know, what do those numbers exactly mean for my boot time?
I timed the boot process of a LIveXP with a 133MB ISO from end of ramdisk loading till PENetwork startup.

with ramdisk - 37 seconds
with FiraDisk - 44 seconds
with WinVBlock - 46 seconds

Interestingly, the differences are not that big, as one might though based on the test results.
With WinVBlock just being 24% slower than ramdisk.sys.

So in conclusion, yes there is room for improvement and not little, but if it is actually worth to invest work into improving the performance is still open for discussion.

[karyonix]

Please try latest version, in a real machine.
Let's see if we get any improvement.

ImDisk RAM Disk CPU: Core2 Duo E8400 RAM: DDR2-800

-----------------------------------------------------------------------
CrystalDiskMark 3.0 x64 © 2007-2010 hiyohiyo
Crystal Dew World : http://crystalmark.info/
-----------------------------------------------------------------------
* MB/s = 1,000,000 byte/s [SATA/300 = 300,000,000 byte/s]

Sequential Read : 2402.629 MB/s
Sequential Write : 4602.827 MB/s
Random Read 512KB : 2352.485 MB/s
Random Write 512KB : 4374.599 MB/s
Random Read 4KB (QD=1) : 520.924 MB/s [127178.8 IOPS]
Random Write 4KB (QD=1) : 242.948 MB/s [ 59313.6 IOPS]
Random Read 4KB (QD=32) : 884.080 MB/s [215839.8 IOPS]
Random Write 4KB (QD=32) : 726.321 MB/s [177324.6 IOPS]

Test : 500 MB [H: 2.0% (40.8/2048.0 MB)] (x5)
Date : 2011/02/18 9:20:07
OS : Windows 7 Ultimate Edition [6.1 Build 7600] (x64)

FiraDisk 0.0.1.30 RAM Disk CPU: Core2 Duo E8400 RAM: DDR2-800

-----------------------------------------------------------------------
CrystalDiskMark 3.0 x64 © 2007-2010 hiyohiyo
Crystal Dew World : http://crystalmark.info/
-----------------------------------------------------------------------
* MB/s = 1,000,000 byte/s [SATA/300 = 300,000,000 byte/s]

Sequential Read : 2172.224 MB/s
Sequential Write : 3108.427 MB/s
Random Read 512KB : 2053.840 MB/s
Random Write 512KB : 2875.944 MB/s
Random Read 4KB (QD=1) : 168.044 MB/s [ 41026.3 IOPS]
Random Write 4KB (QD=1) : 159.666 MB/s [ 38980.8 IOPS]
Random Read 4KB (QD=32) : 243.468 MB/s [ 59440.4 IOPS]
Random Write 4KB (QD=32) : 224.352 MB/s [ 54773.4 IOPS]

Test : 500 MB [H: 2.0% (40.8/2045.0 MB)] (x5)
Date : 2011/02/18 9:04:22
OS : Windows 7 Ultimate Edition [6.1 Build 7600] (x64)

[MedEvil]

Please try latest version, in a real machine.
Let's see if we get any improvement.

In my test, FiraDisk showed ~75% of ImDisks performance in the Random Read/Write 4KB category. In your test it shows only ~30% of ImDisks performance.
Unless you hope, that ramdrive.sys will bomb even more, i don't understand your request.

[Sha0]

...i don't understand your request...

Different versions. Your test was with Firadisk 0.0.1.16. Firadisk is now at 0.0.1.30.

[MedEvil]

Yes, i understand, that there's a new version out.
But karyonix intention to propose another test, the way i read it, was to show that the newer driver had improved performance.
Yet, the test posted by him, clearly shows a way worst performance compared to ImDisk, than my test. So i don't understand, what he hopes to gain from another test done by me.

[Wonko the Sane]

....what he hopes to gain from another test done by me.

Maybe something that is NOT meaningless but actually representative of CURRENT status of the drivers?

You also have to take into consideration that, several different versions of BOTH Winvblock and Firadisk have been published lately (so ANY test not performed with the very latest of both may give incorrect or anyway meaningless results).

Wonko

[dog]

So i don't understand, what he hopes to gain from another test done by me.

Apples are different to oranges
But there is a new orange, which may be better than the old orange
Hope that helps

[MedEvil]

Maybe something that is NOT meaningless but actually representative of CURRENT status of the drivers?

But there is a new orange, which may be better than the old orange

Would you two do me the favour and at least try to read the tests done by karyonix and me!
Karyonix tests, with the new driver, show a way worst performance, than my test with the old one.
So a retest by my can only give worst results, than the one before.

Or to put it more clearly: "Karyonix tests should have shown at least an improved level of performance, in the 4k category, for a retest by someone else to make sense!"

[Sha0]

Ok, so with:

...i don't understand your request...

You actually meant:

...I don't understand why this test is needed when you've already shown some test results...

Is that right? If so, is it worth considering the difference(s) between karyonix' testing hardware and your testing hardware?

Please try latest version, in a real machine.
Let's see if we get any improvement.

[MedEvil]

Ok, ok, i will do another test, once i have figured out, how to actually use a newer driver in the firadisk script. it seems to be especialy made to not allow for easy swapping of driver versions.

[karyonix]

I think the change in 0.0.1.18 should improve performance, especially in Windows XP x86.
I post the two line first before I run the Win7 x64 test.
To test for improvement, you should compare old version with new version in similar environment.
And the environment should be real machine.

Later, I run Crystal DiskMark in FiraDisk in Windows 7 x64.
FiraDisk runs slower in Windows 7 x64 than in Windows XP x86.
Yesterday morning, I did not have time to continue the test with different version of FiraDisk and WinVBlock before I go to work, so I just post the result.
At least it give some information to users who will vote "Are FiraDisk and WinVBlock fast enought or should they get optimized?".
I think 12-86 MB/s is unacceptable. It is surely not fast enough.
But 160-3100 MB/s maybe acceptable even if it is slower than ImDisk.

FiraDisk before 0.0.1.18 use memcpy function to copy memory between RAM drive and buffer.
FiraDisk 0.0.1.18 use SSE2 to copy memory between RAM drive and buffer.
It results in faster sequential read/write in x86 build.
However, in x64 build, memcpy already utilize SSE2, so there is no speed improvement.

Edited by karyonix, 19 February 2011 - 05:14 AM.

[Wonko the Sane]

So a retest by my can only give worst results, than the one before.

Sure, but since hopefully a new version has "better" capabilities, and it is advised as "version to use", which is the sense of testing an obsolete version ONLY?

A test has sense if performed on what you are using or should use, otherwise someone may have a false impression.

Or - if you really want to test an old one - have BOTH the old and new version tested on SAME hardware and OS to see which speed differences there are.

As always I may be wrong, but I don't see this kind of tests as a contest to see which is faster, but rather to see whether the "current" one (and NOT a previous version) is "fast enough" for the intended use.

As you have correctly evidenced in your booting test timing , what may seem an abyssal difference in theoretical speed actually turns out as a minor difference once applied to "real life".

Performing the tests on "real life" versions would be even more representative.


Wonko

[MedEvil]

At least it give some information to users who will vote "Are FiraDisk and WinVBlock fast enought or should they get optimized?".
I think 12-86 MB/s is unacceptable. It is surely not fast enough.
But 160-3100 MB/s maybe acceptable even if it is slower than ImDisk.

Since ramdisk performance depends havily on CPU Power and Speed of used RAM, everybody will get different results, no matter what test system is used.
However by running a simple test on the imdisk drive, which all PE include, on own hardware, everyone can calculate, about what to expect from each of the 3 drivers.

New test, with latest driver versions, will come in a day or two, once i have figured out, why todays builds froze on start.

[MedEvil]

which is the sense of testing an obsolete version ONLY?

The FiraDisk and WinVBlock scripts contain lots of hardcoded entries, which makes it very hard to extend/update them.
So i used the included versions of the drivers.

A test has sense if performed on what you are using or should use, otherwise someone may have a false impression.

From what i know about your computers, i guess the values of the first test will give you a more realistic picture then the new one.

As always I may be wrong, but I don't see this kind of tests as a contest to see which is faster, but rather to see whether the "current" one (and NOT a previous version) is "fast enough" for the intended use.

Well, this is where we clearly contradict. To just know if a driver is "fast enough" there is no need for a comparison test.
Those are done to see, which one is better/faster/...

As you have correctly evidenced in your booting test timing , what may seem an abyssal difference in theoretical speed actually turns out as a minor difference once applied to "real life".

I don't know about minor. 25% difference is clearly perceivable to a human being without test equipment.

[Wonko the Sane]

I don't know about minor. 25% difference is clearly perceivable to a human being without test equipment.

I would say "minor" for the same reason you yourself used "just" and "not that big" :

So now of course everyone wants to know, what do those numbers exactly mean for my boot time?
I timed the boot process of a LIveXP with a 133MB ISO from end of ramdisk loading till PENetwork startup.

with ramdisk - 37 seconds
with FiraDisk - 44 seconds
with WinVBlock - 46 seconds

Interestingly, the differences are not that big, as one might though based on the test results.
With WinVBlock just being 24% slower than ramdisk.sys.

But as I see it, you are somehow mixing apple with oranges (which is by now re-known as being a very possible thing: http://improbable.co...1-3-apples.html )

I mean in this "real-life" test you apparently miss the "structural" dfference of the booting methods.
With 2k3 ramdisk.sys it's ramdisk.sys itself (or SETUPLDR.BIN, or both) that load the image to RAM and then start the OS from it.
With BOTH firadisk and winvblock it is grub4dos (or memdisk - though I believe you used grub4dos) that loads the image to RAM and only later the actual driver is used to load the OS from it.

So, what you are timing is not "pure" firadisk/WinVblock performance, but rather the combined effect of grub4dos+firadisk or that of grub4dos+WinVblock.

You can test the drivers also with memdisk mapping, maybe the actual time will be different...


Wonko