10K SAS or OCZ Vertex

Rurki

Limp Gawd
Joined
Dec 18, 2006
Messages
150
I'm considering a couple drive configurations.

Raid 0
4 x 10K/rpm 2.5 SAS 74GB Fujitsu (already own)
LSI MegaRaid SAS Controller

or

Raid 0
2 x OCZ Vertex 60GB SSD
Onboard Jmicron363 (GigaByte) Raid support

Either direction will cost me the same. I have to replace my existing GA-965P-DS3 to support the LSI MegaRaid.

Which way to go?
 
I'm planning on going with a couple of sas drives in raid myself. though I'm unsure of which config I plan to go with. I'm hoping that all the reliability I've heard about from enterprise drives is true. so I will most likely go with striped. but again I'm not certain.

since you have the sas drives already I would say make use of them. you will gain overall more capacity than the SSD for one thing. I also suspect they should be almost as fast. I'm sure others will chime in with more facts than my opinions though :)
 
I'm leaning toward the SAS drives, mainly due to SSD price. The only other concern I have is the noise level of the 10k SAS Drives. Anyone know what the db range is of a 2.5 10k drive?
 
SAS. It's cheaper, and it will be faster. You're doing a LOT of writing with the encoding - which is where SSD fails badly still.

Sorry, missed the question about 2.5" SAS noise. Well, let me put it this way - it's a distinct noise. It depends ENTIRELY on your mounting assembly as to how it presents. The cheapie SAS hotswap setups that go in a 5.25" bay? I don't recommend them if you're worried about noise. Imagine 10kRPM of vibrational noise, amplified through that, and then through your chassis. It's not necessarily loud, but it is grating upon the nerves. A well isolated mounting setup, they're actually quieter than a Seagate 7200.10 or 7200.11 typically, but they're certainly not Samsung SpinPoint quiet.

For an Apples to Apples on noise, we'll compare a Hitachi UltraStar C10K300 to a Hitachi DeskStar E7K1000.
C10K300 - 2.9 bels idle - http://www.hitachigst.com/portal/site/en/products/ultrastar/c10k300/
E7K1000 - 2.7 or 2.4 bels idle - http://www.hitachigst.com/portal/site/en/menuitem.06d8ccb10579a6bb10441762eac4f0a0/
So, you're talking about a difference of .2-.5 bels, which will be almost completely masked with them mounted inside a case, meaning for purposes of comparison they're effectively the same when mounted inside a chassis.
 
Last edited:
I definitely would strongly recommend against the I-Star. It has the double whammy of nasty vibration and those fans in back are extremely loud, annoying, and cheap junk. (When I work with those, the first thing I do is replace the fans.) You'll get more than adequate cooling in, for example, an Antec P180 with the middle 120 mounted and the SAS drives in the bottom drive cage.
 
Desk space is a valuable asset in my office. I am using a Lian-li A05B with 3 internal 3.5 bays.

Would you recommend another unit besides the iStar?
 
I highly recommend Enhance's S35, provided you can fit in your A05B.
http://www.scsi4me.com/images/S35_Side_Big.gif
You'll notice in the back image, there's no fan. I believe there's a squirrel cage on the side, going from memory. You shouldn't have <5 drives in a RAID5 ever. Otherwise, performance will be way off what it could be.
If you don't have the room for that, I would recommend the Enhance Quadrapack
http://www.scsi4me.com/enhance-q14ss.html - but be warned that the fan is rather loud
... or to instead use slightly cheaper but no less respectable 3.5" SAS drives (which are actually usually faster!) and an external enclosure that matches the LiLi stuff, like, say, this.
http://www.scsi4me.com/enhancebox-e...-sataii-drives-multi-lane-host-interface.html
 
I have 2 I-stars and you can change the fan speed with the jumper on the back.
I had 7 Fijitsu 76gb 10k drives in a Raid0 in mine before I upgraded to some vr's. And the enclosure does not rattle or make any noise.

Honestly I found those 10k drives to be lacking for Encoding which is whtat I do with mine.
I got pretty much identical speed from just a couple vr's.

Considering you already have those drives, that might be the way to go, but how much is it going to cost you to replace the mobo?
And buy a card?

Heres what I would do:
Sell the sas drives.
Take that money plus the money you planned on spending on a new mobo, and raid card, and buy 2-4 OCZ vertex drives.
Put them in RAID0 using the onboard controller.

Since you are doing Raid0 you can get away with the onboard raid, you dont need a true raid controller since there is no parity calculation.
 
nitro, you have your RAID configured entirely wrong. That's a plain old fact. You do not do 7 drive RAID0, because that's not how RAID0 works. It is not a magic N-drive performance wand, and never will be. You should have been running a 6+1 RAID5 or a 5+1+HS RAID5.

As far as the I-Stars, they cook the drives at low. They do rattle, as well. I've handled more than a few. Yes, some were fine - and more were not. The quality is so spotty, that I can't begin to consider recommending them. I don't know why the hell they're so spotty, but they are. Especially compared to the Enhance stuff, which is frankly, amazing. It's better than IBM and HP in terms of fit and finish, even.

An LSI 8480 controller to 6x 7200RPM SATA drives in RAID5 (4+1+HS) will do >350MB/sec write. To two arrays doing software striping, >500MB/sec. (I ran out of source at 500MB/sec, 4x LTO4.)
An LSI 8480 controller to 5x 10kRPM SAS in RAID5 (4+1, no HS) will therefore, do >400MB/sec. At this point, your bottleneck is going to be the encode itself. However, SAS will permit you to have the source and destination on the same array without taking a significant performance penalty, due to the much much lower seek time of SAS versus SATA.

Is the LSI 8480 a hot controller? Oh GODS yes. That thing is a monster. It chews up data and spits it out so fast, you'd think it was powered by quad Xeon 3.6's.
So, everyone starts screaming about the cost. It must cost an absolute fortune and a half, right?
How about $352 before shipping?
http://www.techforless.com/cgi-bin/tech4less/39R8850?mv_pc=cnet-pg&tts=20090402210144
And unlike others, IBM leaves everything exactly as LSI makes it. You will need the BBU, which is about $100.

STILL not happy? Okay.
http://www.newegg.com/Product/Product.aspx?Item=N82E16816118099
$350. 4 ports. 128MB cache. True hardware RAID. Rock on. Only 4 ports though.

BUT WAIT!
http://www.newegg.com/Product/Product.aspx?Item=N82E16816101205
http://www.newegg.com/Product/Product.aspx?Item=N82E16816151038
http://www.newegg.com/Product/Product.aspx?Item=N82E16816401015
(All available elsewhere for less, of course.)

Point made!
 
nitro, you have your RAID configured entirely wrong. That's a plain old fact. You do not do 7 drive RAID0, because that's not how RAID0 works. It is not a magic N-drive performance wand, and never will be. You should have been running a 6+1 RAID5 or a 5+1+HS RAID5.

...

Point made!

Just wondering, why you are saying he should have done RAID5 over RAID0? You really think that even HW RAID5 is gonna be faster?
 
nitro, you have your RAID configured entirely wrong. That's a plain old fact. You do not do 7 drive RAID0, because that's not how RAID0 works. It is not a magic N-drive performance wand, and never will be. You should have been running a 6+1 RAID5 or a 5+1+HS RAID5.

...

Point made!

AreEss...

Even with the best of controllers RAID0 writes will be faster than RAID5/6 because there is no parity calculation.

I fail to see how he is doing it wrong. Of course RAID0 doesn't scale perfectly well, I know this as 4 of my Raptors on my controller are almost as fast as 8 Raptors, but I went with more drives for storage needs.

Nitro is fine with his RAID0, and what does having 7 drives have to do with it either, you can do a RAID0 with whatever the hell you want to.
 
Heres what I would do:
Sell the sas drives.
Take that money plus the money you planned on spending on a new mobo, and raid card, and buy 2-4 OCZ vertex drives.
Put them in RAID0 using the onboard controller.

Since you are doing Raid0 you can get away with the onboard raid, you dont need a true raid controller since there is no parity calculation.

You bring up a good point. Put the existing drives on ebay and grab 2 vertex 120s or 4 60s.

I have a GA-965p-ds3 motherboard with two SATA raid ports. Wouldn't I need to upgrade the board to support 4 drives in raid 0?
 
*sigh*
Why does nobody here actually do their homework? You all just read what one or two people say, and presume they're the gurus of something, without bothering to check other sources. (Which is why I back what I say up, every single time.)

Firstly, RAID5 is faster than RAID0 with parity calculations, period, past a certain point. RAID0 does not magically scale infinitely, period. Don't like it? Tough. You are not entitled to your own facts. I can build RAID5 that will blow RAID0 out of the water till the cows come home.

If we presume onboard RAID, that means a stripe size of 64KB. If we presume NTFS, that means a block size of 4KB. If we presume that it's a sequential layout, that means each block is written in order from drive 0 through 6. (e.g. write of 4MB gets written to 0, 1, 2, 3, 4, 5, 6 in that order, next write goes 0, 1, 2, 3, next write 4, 5, 6, etcetera.) The maximum combined disk operation on NTFS is typically 16KB.
You never ever get a full stripe read. You max out at 4 drives ever being used.
That is why Nitro's configuration was completely wrong. He sacrificed all data integrity for no performance gain, and in fact, performance losses due to partial stripe writes and reads.

That's why 4 drives isn't faster than 8 drives, too. You're maxing out at 4 drives on any read or write operation, because each operation is 4KB at the file system level, with a maximum of 16KB typical.

RAID is about finding the balance between the stripe and the file system laying down atop it. Almost no truly high performance array uses a 64KB stripe - it's too hard to do full stripe operations, and only when you do full stripe will you see the actual throughput numbers. That's why a 3 disk RAID5 can seem faster than a 5 disk RAID5, too. Not because of parity calculations - they are not anywhere near as intensive as people seem to think. (Especially not on cards like the LSI 8704! It's a 500MHz AltiVec! That's faster than the processors on some "Enterprise" arrays - by a lot.)

Right now, I really wish I could do ASCII illustrations. Because it's so much easier for people to understand when it's actually drawn. But I can't. Suck.

So, why wouldn't I - for example - use a stripe size of 32KB and to do source and destination on the same array. Because this gives you two operations of 16KB attempted on the same array, adding up to 32KB, and making a full stripe operation more likely, right?
Wrong. Full stripe, means full stripe in a single operation. Otherwise, you're only ever using part of the array.
For an encoding setup, I would still recommend that source and destination be on the same disks here. There's no reason to have two arrays for a home setup, it's not going to give you a performance gain - your bottleneck is already going to be your encoding processing.
Here's where Windows gives me fits; NTFS maxes out at a 4KB block. On FreeBSD, UFS2 maxes out at 64KB blocks, making full-stripe possible. NTFS? Not so possible. On any NTFS configuration, you want the smallest possible stripe on the array - if you can get 4KB, go for it. But most will not go below 16KB (4x4KB or 4 disk operations in NTFS land.) Yes. This is also why 8+1 isn't so great; you'll struggle to use half the array.

Optimally, what I would recommend is having the array on a FreeBSD system, formatting the array and UFS2 at 64KB (larger blocks are better for large sequential operations like encoding!) and using 802.3ad trunked Gigabit between the systems. Presuming the encoder can crank out that much. Gigabit might be just fine, provided it's good NICs on both ends with TOEs. But I don't think that's realistic here.

Another option would be RAID3, which uses a dedicated parity disk. It's more viable on SAS than other interfaces, because you don't have a shared bus for the disks. However, you're limited by single disk performance still. So it really only works with solid state disks or where your parity disk is solid state.

Hopefully this explains things in a way that makes sense now.
 
Last edited:
Actually with fast SSD's RAID0 does scale nearly perfectly. Add a drive, get almost twice the speeds. Up to whatever the bus can handle. This has been proven many times with the large Vertex drives, the Intel drives, and pretty much any SLC drive. The issue is not the drives. The issue is maxing out the bandwidth between the controller and the system.

It should be noted though that I know next to nothing about the inner workings of RAID and what you are saying does make perfect sense on a more general level, especially with spindle based drives.

For the question about drive mounting - I'm using this: http://www.startech.com/item/SATABAY425BK-4-x-25-SAS-SATA-RAID-Backplane-for-525-Bay.aspx with 4x SSD and it is perfectly quiet :)
 
I have a GA-965p-ds3 motherboard with two SATA raid ports. Wouldn't I need to upgrade the board to support 4 drives in raid 0?
Yes, you will need a controller, or a better mobo to support 4 drives.
 
Who said anything about a 64kb stripe?

Raid0 doesn't scale as well as raid 5?
Are you telling bad jokes?
An array maxing out at any given speed is going to due the controller used.
x# drives in raid0 vs x# drives in raid5. Raid 0 will win everytime assuming all other conditions are the same?(Ie mobo proc controller etc.)

I don't know what you are smoking but I want some of it.

BTW I have seen 4x 30 gb OCZ do 600mbs read and 400 write on raid 0 with Intel raid.
Show me 4 disk 2.5in setup that will do that.

Also you don't need raid 5 for encoding.
You need pure speed.
I know this because it is what I do everyday.

The array that you encode on is typically not where it is stored.

Yes Arees is right that the Enhance hotswaps are better,but its not worth the price IMO(I would rather have better/more drives) and I think he is exaggerating the shortcomings of the i-star.
And if you have SSDs then you don't have any vibrations, and you really don't need the fan at all.
 
nitro, it's pretty clear to me that you have no experience with RAID except on motherboards. You seem interested in personal insults, rather than facts. And all of your statements so far, excepting the Enhance, are false - and can be proven false. Also, to clarify my stand on the I-Star; the issue is not with the design itself. The issue is with the very erratic quality of them. I cannot in good faith recommend a piece of hardware unless I am certain that quality will be consistent, regardless of what the quality is. The I-Star is so wildly inconsistent, that I'm not comfortable recommending or using it. I can't say for certain whether it will even function, or if it will be the best fitted part you've ever seen. It could be either - it's a roll of the dice as to which.

Again, I do this for a living. I have seen more RAID controllers than you have seen computers in your life. I have stunned product engineers by getting more out of their arrays than they were able to (IBM DS3200, IBM DS3400, IBM DS4200, EMC CX500 this year alone.)

http://docs.hp.com/en/B7961-90018/ch08s12.html
Down toward the bottom of the page, explains Full-Stripe Writes (if you search on that exact phrase, it will take you right to it.)
There is also some good information on what Full-Stripe Writes are, and why they are good, located here.
http://sunsite.uakom.sk/sunworldonline/swol-07-1999/swol-07-raid2.html

http://www.xssist.com/blog.htm
The link showing RAID0 scaling demonstrates what I have been saying - RAID0 does not scale infinitely. It stops scaling generally around 4 drives, after which point, additional drives only add additional space and increase failure chance - they do not increase performance. At 4 drives, their IOPS are double a single drive and then add <10% each additional drive. Throughput does not double past two drives - it increases by <25% for each additional drive.
With the LSI Elite1650 with 256MB + BBU (to ensure cache functions) I performed some testing, which I lost the complete notes for, sadly. Spanning two controllers and 4 channels, I saw identical behavior. RAID0 does not scale anywhere near as well as people claim. You see IOPS scaling to 4 - in my case, I got scaling to 6 due to multiple controllers and channels - but percentage scaling rather than exponential scaling (percent of drive rather than throughput = drive throughput * drives) past 4.
 
Again, how much of that is due to the overhead (and whatnot) of using spindle drives? How much of scaling issues are due to the PCI-e bus getting overloaded? Four 120g vertex drives will completely saturate something like a 2404/5404/ICH9r/ICH10r (ie, general "consumer" level RAID), and the PCI-e bus it is on. Adding more drives doesn't scale because the underlying hardware can't handle it. Even going to multiple controllers won't resolve this because the vast majority of systems on the market today share bandwidth between PCI-e slots. Even those that don't do this have a single link between the south bridge and the rest of the system. And even if your controller isn't on the south bridge (like if you use a pci-e 16x GPU slot [generally the only ones not on the south bridge]) you still have a hard bandwidth limit between the system and the controller.

RAID0 in and of itself is not the problem, the problem is the system under it can't handle the scaling.
 
Again, I do this for a living. I have seen more RAID controllers than you have seen computers in your life.

That is a pretty damn arrogant statement.

How do you know what nitro/me/others do for a living, how do you know you've seen more of this and that than we have. That alone makes me laugh that you automatically think you know more.

I and the others appreciate what you are trying to say but calling us noobs etc, psh... whatever man, I totally lost any respect for anything you have to say.

I've been working with computers for over 10 years, in a corporate environment for over 5. Maybe I don't know as much as you but don't treat us like crap because maybe in our environments and our testing we get different results from you and you don't agree.

/rant
 
Again, how much of that is due to the overhead (and whatnot) of using spindle drives? How much of scaling issues are due to the PCI-e bus getting overloaded? Four 120g vertex drives will completely saturate something like a 2404/5404/ICH9r/ICH10r (ie, general "consumer" level RAID), and the PCI-e bus it is on. Adding more drives doesn't scale because the underlying hardware can't handle it. Even going to multiple controllers won't resolve this because the vast majority of systems on the market today share bandwidth between PCI-e slots. Even those that don't do this have a single link between the south bridge and the rest of the system. And even if your controller isn't on the south bridge (like if you use a pci-e 16x GPU slot [generally the only ones not on the south bridge]) you still have a hard bandwidth limit between the system and the controller.

RAID0 in and of itself is not the problem, the problem is the system under it can't handle the scaling.

That's a good question, though your answer to it is flat out wrong, I'm afraid.

RAID0 will not scale past 4 drives because of the full stripe issue, as outlined before. It also doesn't scale past 4 drives because the bus craps out - you are not trashing a PCIe bus with SATA, period. Not even 4 lanes of it. You're also not doing it with SSDs.

RAID0 is absolutely NOT zero cost. Learn that simple fact now, please, everyone. NO RAID is zero cost. RAID0's cost is in stripe calculations and table lookups. It has to break apart or piece together file system blocks into a segment for the RAID, and it has to look up what drives they are on.
RAID0 does not scale past four drives for the same reason RAID5 does not scale past four data drives and RAID6 does not scale past four data drives and RAID1 does not scale past four drives.
Striping.

Seriously. Learn that word, love that word, because it is the one true dictator of performance, period. Less than full stripe, uses less than full drive count. More than full stripe, stacks IOs onto multiple drives. Stripes must be divisible by 512 bytes or 4096 bytes (4KB) depending on implementation. Why 512 bytes? That is what the physical block size on a hard drive is, and that is what they work in - blocks. So if we presume a stripe size of 4KB, and a file system block size of 4KB, we have a full stripe write. This actually works out to 8 physical blocks, so 8+1 means 1 block operation, 4+1 means 2 block operations, 2+1 means 4 block operations, and so on. To determine block operations, take your segment size in bytes and divide by 512. If it doesn't divide cleanly by your drive count, that means you will have hot spots - drives that get hit harder than others.
MORE RELEVANT EXAMPLE: 64kb stripe, 4k file system block. 65536 bytes per stripe segment, 512 bytes, 128 blocks. 16 per drive in 8+P, 32 in 4+P, 64 in 2+P. 4k file system means parity reads cannot be bypassed, adding 4-12KB read operations for each file system write, from each drive. These are random seek activity, which is what cripples SATA.

One of these days, I'll write a book or something, and then just give you all an Amazon link. :p
 
That's a good question, though your answer to it is flat out wrong, I'm afraid.
I'm willing to accept that. Reading the rest of your post I do have even more questions though. At what point does the file system on top of the RAID become a factor? How do things like journaling, and true log fs systems affect it? What about a FS designed for SSD's that ignore the stupid 512k block size crap? What about the fact that you can (fairly easily) have a file system that does not use 4k blocks? Hell even NTFS can be configured to use different sized blocks.

As to the Anandtech - factually incorrect aside, are you saying that they are lying about the numbers they posting for the benchmarks?
 
... look above...
I hate to shatter your little world, but 4 vertexes will shatter anything he is doing with SAS.

You are proud that your little 4 drive SAS array does 350 mb/sec.

ONE Intel X25-M does 250+ mb/sec.

One SSD = THIRTEEN 15k Sas drives in Random IO per Anandtech.
10k 2.5 in sas drives do like 90 mb/sec each, similar to the max transfer of this seagate...

3x 30gb vertexes on an ICH10r do 600+ mb/sec write and 400+ mb/sec read

lolssd.jpg
 
Last edited:
"Firstly, RAID5 is faster than RAID0 with parity calculations, period, past a certain point. RAID0 does not magically scale infinitely, period. Don't like it? Tough. You are not entitled to your own facts. I can build RAID5 that will blow RAID0 out of the water till the cows come home."

......
 
nitro, it's pretty clear to me that you have no experience with RAID except on motherboards. You seem interested in personal insults, rather than facts. And all of your statements so far, excepting the Enhance, are false - and can be proven false. Also, to clarify my stand on the I-Star; the issue is not with the design itself. The issue is with the very erratic quality of them. I cannot in good faith recommend a piece of hardware unless I am certain that quality will be consistent, regardless of what the quality is. The I-Star is so wildly inconsistent, that I'm not comfortable recommending or using it. I can't say for certain whether it will even function, or if it will be the best fitted part you've ever seen. It could be either - it's a roll of the dice as to which.
How do you know my experience? Sounds like you are guessing?
All of my statements are false? Like what the fact that a 4 drive R0 array will be faster than the current SAS drives he has, because it will.
Or the fact that if he uses SSD he doesnt have to worry about noise, vibration, because they dont have moving parts?

Also you opinion on the I-Star is just that an OPINION.
My opinion was to go with the I-Star.

Again, I do this for a living. I have seen more RAID controllers than you have seen computers in your life.
Congratulations

I have stunned product engineers by getting more out of their arrays than they were able to (IBM DS3200, IBM DS3400, IBM DS4200, EMC CX500 this year alone.)

http://docs.hp.com/en/B7961-90018/ch08s12.html
Down toward the bottom of the page, explains Full-Stripe Writes (if you search on that exact phrase, it will take you right to it.)
There is also some good information on what Full-Stripe Writes are, and why they are good, located here.
http://sunsite.uakom.sk/sunworldonline/swol-07-1999/swol-07-raid2.html

http://www.xssist.com/blog.htm
The link showing RAID0 scaling demonstrates what I have been saying - RAID0 does not scale infinitely. It stops scaling generally around 4 drives, after which point, additional drives only add additional space and increase failure chance - they do not increase performance. At 4 drives, their IOPS are double a single drive and then add <10% each additional drive. Throughput does not double past two drives - it increases by <25% for each additional drive.
With the LSI Elite1650 with 256MB + BBU (to ensure cache functions) I performed some testing, which I lost the complete notes for, sadly. Spanning two controllers and 4 channels, I saw identical behavior. RAID0 does not scale anywhere near as well as people claim. You see IOPS scaling to 4 - in my case, I got scaling to 6 due to multiple controllers and channels - but percentage scaling rather than exponential scaling (percent of drive rather than throughput = drive throughput * drives) past 4.

Nobody is talking about Enterpirse SAN systems here.
How does that help the OP with his decsion?

Again, how much of that is due to the overhead (and whatnot) of using spindle drives? How much of scaling issues are due to the PCI-e bus getting overloaded? Four 120g vertex drives will completely saturate something like a 2404/5404/ICH9r/ICH10r (ie, general "consumer" level RAID), and the PCI-e bus it is on. Adding more drives doesn't scale because the underlying hardware can't handle it. Even going to multiple controllers won't resolve this because the vast majority of systems on the market today share bandwidth between PCI-e slots. Even those that don't do this have a single link between the south bridge and the rest of the system. And even if your controller isn't on the south bridge (like if you use a pci-e 16x GPU slot [generally the only ones not on the south bridge]) you still have a hard bandwidth limit between the system and the controller.

RAID0 in and of itself is not the problem, the problem is the system under it can't handle the scaling.

I completely agree.

I hate to shatter your little world, but 4 vertexes will shatter anything he is doing with SAS.

You are proud that your little 4 drive SAS array does 350 mb/sec.

ONE Intel X25-M does 250+ mb/sec.

One SSD = THIRTEEN 15k Sas drives in Random IO per Anandtech.
10k 2.5 in sas drives do like 90 mb/sec each, similar to the max transfer of this seagate...

3x 30gb vertexes on an ICH10r do 600+ mb/sec write and 400+ mb/sec read

lolssd.jpg

Thank you for proving my point/recomendation.

_____________________________________

To the OP:
My original reccomendation still stands.
For a consumer board, and on a consumer budget.
An SSD array will work better for what you are trying to do.
It's just a simple fact.
 
What I like here is how at the end of page one, Mr "I know raid" starts using IOPS to show us how 4x 2.5 10k sas drives are faster than two modern SSDs. I think he even mentioned power consumption somewhere... LOL....

An LSI 8480 controller to 6x 7200RPM SATA drives in RAID5 (4+1+HS) will do >350MB/sec write. To two arrays doing software striping, >500MB/sec. (I ran out of source at 500MB/sec, 4x LTO4.)
An LSI 8480 controller to 5x 10kRPM SAS in RAID5 (4+1, no HS) will therefore, do >400MB/sec. At this point, your bottleneck is going to be the encode itself. However, SAS will permit you to have the source and destination on the same array without taking a significant performance penalty, due to the much much lower seek time of SAS versus SATA.

So your twelve 7200 RPM SATA drive array with software striping can do 500 mb/sec

What about a Highpoint 3520 + 4*30gb Vertex 1275 Raid 0

attachment.php


Power Consumption? Seek Time? IOPS? Sas WHAT? Get owned?

I do not think I will see you here again.....
 
WOW!

My apologies for creating an environment of hostility :)

I have to agree with both sides. On an enterprise level Raid0 does reach a cap. However, I am not looking for more than 4 drives.

The cost of a decent capacity Vertex SSD is a bit pricey at the moment. Hopefully prices will begin to fall now that OCZ has shown the way beyond the JMicron stutters of other SSDs.

I most likely will wait till summer to see if prices drop on SSD.

Then sell my SAS Drives and various parts to purchase a new motherboard, raid controller and 4 SSDs.

I just hate waiting!
 
That is a pretty damn arrogant statement.

How do you know what nitro/me/others do for a living, how do you know you've seen more of this and that than we have. That alone makes me laugh that you automatically think you know more.

I and the others appreciate what you are trying to say but calling us noobs etc, psh... whatever man, I totally lost any respect for anything you have to say.

I've been working with computers for over 10 years, in a corporate environment for over 5. Maybe I don't know as much as you but don't treat us like crap because maybe in our environments and our testing we get different results from you and you don't agree.

/rant

QFT.
 
*sigh*
Why does nobody here actually do their homework? You all just read what one or two people say, and presume they're the gurus of something, without bothering to check other sources. (Which is why I back what I say up, every single time.)

Firstly, RAID5 is faster than RAID0 with parity calculations, period, past a certain point. RAID0 does not magically scale infinitely, period. Don't like it? Tough. You are not entitled to your own facts. I can build RAID5 that will blow RAID0 out of the water till the cows come home.

If we presume onboard RAID, that means a stripe size of 64KB. If we presume NTFS, that means a block size of 4KB. If we presume that it's a sequential layout, that means each block is written in order from drive 0 through 6. (e.g. write of 4MB gets written to 0, 1, 2, 3, 4, 5, 6 in that order, next write goes 0, 1, 2, 3, next write 4, 5, 6, etcetera.) The maximum combined disk operation on NTFS is typically 16KB.
You never ever get a full stripe read. You max out at 4 drives ever being used.
That is why Nitro's configuration was completely wrong. He sacrificed all data integrity for no performance gain, and in fact, performance losses due to partial stripe writes and reads.

That's why 4 drives isn't faster than 8 drives, too. You're maxing out at 4 drives on any read or write operation, because each operation is 4KB at the file system level, with a maximum of 16KB typical.

RAID is about finding the balance between the stripe and the file system laying down atop it. Almost no truly high performance array uses a 64KB stripe - it's too hard to do full stripe operations, and only when you do full stripe will you see the actual throughput numbers. That's why a 3 disk RAID5 can seem faster than a 5 disk RAID5, too. Not because of parity calculations - they are not anywhere near as intensive as people seem to think. (Especially not on cards like the LSI 8704! It's a 500MHz AltiVec! That's faster than the processors on some "Enterprise" arrays - by a lot.)

Right now, I really wish I could do ASCII illustrations. Because it's so much easier for people to understand when it's actually drawn. But I can't. Suck.

So, why wouldn't I - for example - use a stripe size of 32KB and to do source and destination on the same array. Because this gives you two operations of 16KB attempted on the same array, adding up to 32KB, and making a full stripe operation more likely, right?
Wrong. Full stripe, means full stripe in a single operation. Otherwise, you're only ever using part of the array.
For an encoding setup, I would still recommend that source and destination be on the same disks here. There's no reason to have two arrays for a home setup, it's not going to give you a performance gain - your bottleneck is already going to be your encoding processing.
Here's where Windows gives me fits; NTFS maxes out at a 4KB block. On FreeBSD, UFS2 maxes out at 64KB blocks, making full-stripe possible. NTFS? Not so possible. On any NTFS configuration, you want the smallest possible stripe on the array - if you can get 4KB, go for it. But most will not go below 16KB (4x4KB or 4 disk operations in NTFS land.) Yes. This is also why 8+1 isn't so great; you'll struggle to use half the array.

Optimally, what I would recommend is having the array on a FreeBSD system, formatting the array and UFS2 at 64KB (larger blocks are better for large sequential operations like encoding!) and using 802.3ad trunked Gigabit between the systems. Presuming the encoder can crank out that much. Gigabit might be just fine, provided it's good NICs on both ends with TOEs. But I don't think that's realistic here.

Another option would be RAID3, which uses a dedicated parity disk. It's more viable on SAS than other interfaces, because you don't have a shared bus for the disks. However, you're limited by single disk performance still. So it really only works with solid state disks or where your parity disk is solid state.

Hopefully this explains things in a way that makes sense now.

SSD's don't scale? You should tell that to these guys:

http://www.youtube.com/watch?v=wInHYpXRRf0

Also, I tend to PLENTY of servers at work (that you can bet your bottom dollar are set up correctly), that hd speed testing software is used to monitor speed and degredation. In certain instances, sometimes a raid5 can beat a raid0 by a little, but raid5's blowing away raid0's??!!! I agree with the other guy, you're telling bad jokes.

Don't bother with the book lol, it won't be a best seller. You wouldn't believe what kind of reads and writes I can get off of 3.5 year old HP Proliant 380dl g3's, and it certainly doesn't back up what you're saying.

OP - you're in a tough spot. SSD is still new, and there is a lot of unanswered questions. I'm currently in the market for new drives and I want to buy the OCZ drives and raid tehm oh so badly, but there are just too many variables and question marks.

I'd hold off for a year. Think That's what I'm going to do anyway.
 
SSD's don't scale? You should tell that to these guys:

http://www.youtube.com/watch?v=wInHYpXRRf0

Also, I tend to PLENTY of servers at work (that you can bet your bottom dollar are set up correctly), that hd speed testing software is used to monitor speed and degredation. In certain instances, sometimes a raid5 can beat a raid0 by a little, but raid5's blowing away raid0's??!!! I agree with the other guy, you're telling bad jokes.

Don't bother with the book lol, it won't be a best seller. You wouldn't believe what kind of reads and writes I can get off of 3.5 year old HP Proliant 380dl g3's, and it certainly doesn't back up what you're saying.

OP - you're in a tough spot. SSD is still new, and there is a lot of unanswered questions. I'm currently in the market for new drives and I want to buy the OCZ drives and raid tehm oh so badly, but there are just too many variables and question marks.

I'd hold off for a year. Think That's what I'm going to do anyway.

QFT.

i'm in the same boat as you. I really want to try SSD but not cry if my drive goes bottom up or prices drop significantly.

For this reason I bought a 30GB vertex. $100 or so after MIR, works very well. Even random write performance is pretty good and 30GB is plenty big enough for OS & all my apps.
 
I started off with 2x 30GB Vertex drives in RAID0 using onboard ICH controller. Total cost was $250 w/tax&shipping, $230 if I send in the rebate-- but I probably won't cause I'm lazy.

That is pretty cheap considering the performance. Controller seems to be limiting me to 300MB/s seq read. Plus I'm not stuck with anything. I want better random writes? $100 more and 3x RAID0. Need more space? $100 extra for another 30GB plus better performance. SSDs scale so well you're not really stuck with anything. Onboard controller is limiting you too much? Buy an addin card. Just seems like you can do whatever you want and get crazy performance. You don't need the upfront cost of buying everything at once. You can start small and slowly upgrade your setup as you feel it necessary.

Plus the other reason I got these, I'm not really worried about something else coming out or prices dropping. I don't ever see these drives getting wasted. Something awesome comes out, I'll put these Vertex drives in my laptop, HTPC, and/or my GF's system. All setups will benefit. No noise, no vibration, fast, physically small, and standard SATA.

All the talk about what hotswap unit to use for SAS drives... I got my Vertex drives sitting on the bottom of the case cause I was too lazy to actually mount them. How could that be any more simple :p.

I just don't see regretting not waiting.
 
I started off with 2x 30GB Vertex drives in RAID0 using onboard ICH controller. Total cost was $250 w/tax&shipping, $230 if I send in the rebate-- but I probably won't cause I'm lazy.

That is pretty cheap considering the performance. Controller seems to be limiting me to 300MB/s seq read. Plus I'm not stuck with anything. I want better random writes? $100 more and 3x RAID0. Need more space? $100 extra for another 30GB plus better performance. SSDs scale so well you're not really stuck with anything. Onboard controller is limiting you too much? Buy an addin card. Just seems like you can do whatever you want and get crazy performance. You don't need the upfront cost of buying everything at once. You can start small and slowly upgrade your setup as you feel it necessary.

Plus the other reason I got these, I'm not really worried about something else coming out or prices dropping. I don't ever see these drives getting wasted. Something awesome comes out, I'll put these Vertex drives in my laptop, HTPC, and/or my GF's system. All setups will benefit. No noise, no vibration, fast, physically small, and standard SATA.

All the talk about what hotswap unit to use for SAS drives... I got my Vertex drives sitting on the bottom of the case cause I was too lazy to actually mount them. How could that be any more simple :p.

I just don't see regretting not waiting.

This is one of my lines of thinking, but I still have questions about the life of the drive in terms of actual mileage (1 year, 2 years, et cetera), and on top of that, how long will they be "blazing" fast over their life span? I heard they slow significantly with age, but how fast and how much I can't seem to get a precise answer to.

I want soooo badly though....lol
 
I think it's better to wait for 60GB or 120GB drives.

30GB capacity is very small. You might be able to put it in a laptop but even now, in 2009, how many of us are using 30GB or smaller drives in our single-drive-bay laptops as it is? Where do you think things will be in 2010?

I do see some value in putting these 30GB Vertex SSDs into cheap eSATA enclosures down the road, if such a thing exists. Do that, and you have the equivalent of a 32GB USB key, but with 3-4x the bandwidth.... the Vertex becomes a very attractive "shuttle" drive for moving files between PCs and I could see myself using those in 2010. But, no matter what scenario I play out in my mind, 30GB SSD disks strike me as closet / trash can devices 24 months out, whereas a 120GB drive I'd feel comfortable tossing into a late 2010 netbook.
 
I hate to shatter your little world, but 4 vertexes will shatter anything he is doing with SAS.

Come back when you have legitimate, real world testing. Not an SSD-oriented benchmarketing toy.

As a matter of fact, I have yet to see anyone present any actual evidence of anything but low to mid SSDs with toy "benchmarks," and absolutely zero representation of, or experience with SAS. I also see a lot of people demonstrating illiteracy or simply a desire not to read the fact that I said SATA does >350MB/sec and SAS does well over 500MB/sec for a six drive RAID5 (4+P+HS) configuration - one that is 1.2TB, I'll add. And I have yet to see any SSD benchmarks here that have gotten near that performance reliably, and the only ones I've ever seen get near it, were attached to $1200+ Arecas. Test rig is not real world - the real world has stuff to do besides benchmarketing.

In fact, I price checked myself this evening. I do that sometimes, because I don't rely on sites with little to no experience in the SCSI or SAS realm, or people who don't have any specialized training or experience in storage, to tell me things I can test for myself. I also don't trust vendors not to play pricing games on me.

Using 5 Hitachi 15K300's at $320 each, and an LSI 8480 (previously mentioned at $352,) I can do 1.2TB at over 600MB/s random read and over 500MB/s sequential write regardless of "alignment fixes" or other crap, for $1952.
Here's your challenge; match capacity and performance with SSD at or below that price. No, you aren't allowed to mix SSD and SATA.

STANDARD DISCLAIMER: You can possibly do better than me on pricing. Or you can do worse. I go by best price from reputable vendors (as determined by Google Shopping or ResellerRatings.)

That's what we call putting up or shutting up.
 
Back
Top