RAIDing the SSD: Intel's 32GB SLC SSD
THE OTHER DAY, we saw that these new shiny SSDs, with ten channel flash arrays aided by efficient controllers and DRAM buffers for spread out writes, do achieve some nasty benchmark results. What happens when we put two or four of them in parallel as a RAID0 stripe?
I took two of the Intel 32GB SLC high speed "enterprise class" drives and set up the RAID using the inbuilt Intel chipset controller. Ran the Vista disk bandwidth and removable device iops benchmarks. How about the results?
With two drives, each on its own SATA channel, the average read speed climbs to an astonishing 468 MB/s, actually nearing the half a gigabyte per second number almost half of the time - not bad at all and, still at the same below 1ms access time.
So, two drives in this case really do seem to double the total performance. How about, say... four? After all, we've got six SATA channels in there and, after this array plus the OS HDD and the SATA DVD, there are still two ports left.
OK, I added two more drives and recreated a fresh four-disk RAID0 array. Ran the benchmarks again...
Oh well, the read speed remained nearly the same, at 476 MB/s average, while the 74K ops/minute number did go lower than either the two or one drive configurations. Why? The south bridge SATA controller's total bandwidth could be an issue here, as well as the DMI bandwidth to the North Bridge and memory, which all could limit the net transfer rate. There's no reason why at least the read bandwidth shouldn't near double again when jumping from dual to quad parallel drive RAID0 setup.
Frankly, I'd rather stick with a two-disk RAID0 here, and consider RAID5 or RAID10 for anything more - the incremental performance benefits from two to four are non-existent on the Intel south bridge RAID controller at least, and - for safety sake - some fault tolerance would be welcome. It may be, in any case, useful to try this four-disk RAID0 configuration speed-wise on a true hardware RAID controller, which I'm about to do as well. µ
Share this:
Business Intelligence Software
Customer Relationship Management (CRM)
Enterprise Accounting Software
Enterprise Resource Planning (ERP)
Enterprise Systems Management
You have $3000 worth of SSDs but no hardware RAID controller to benchmark?
This is really helpful stuff. I have been wondering where the raid sweet spot would be. I am anxious to see both your "real raid controller" test, AND some comparisons against other SSD drives.
INTELS Wins Magic Leather Pouch Today. Obiviously BIGgangbandEDUpe' Improvement. Still, Wheres RAID 5 come In with No Spinning Disk, Is It Some New Lib & Controller function or IS it even SSD? Like Incorporation of New & Faster to Workstation thats Needs Price/Performance Edge, performance seems sugarDll.icious.drashek 13 Fingers.
Any chance you can do a review with an Intel SRCSASJV RAID controller ( http://www.intel.com/products/server/raid-controllers/srcsasjv/srcsasjv-overview.htm ) and the drives configured in RAID-5 and RAID-6 configurations? Redundancy is the enterprise is a must, so while having the RAID-0 benchmark is useful when compared against other drives, it's not really practical for most users -- especially at their price point.
You are comparing performance of expensive "enterprise grade" SSDs and economical SATA drives ?
How about something a bit more meaningful for enterprise performance comparisons, and use Cheetah 15K SAS drives, and a decent raid controller with some cache.
I'm pretty sure I could also easily trounce the SATA drives without needing to resort to blowing the budget (GB/$) on SSDs.
As much as I drool over these benchmarks I can't help but remember that RAID actually stands for Redundant Array or INEXPENSIVE DISKS.
The latest news from SanDisk talks about a possibility of 100x speed up of current SSD's. This mean that S-ATA can't offer enough speed to transfer this high amount of data/sec. Therefore the real soulution for SSD's will be the PCI-Express connection.
Now it might be the time for some manufacturer to make an internal RAID-O style setup, so that a lets say 512 GB Flash drive would have 4 SATA connectors and the user could attach 1, 2 or 4 connectors and get 1x, 2x or 4x SATA speed.
I might myself prefer using a separate RAID controller but I assume there might be users that would prefer a drive that can be used with one single cable or with 4 SATA cables.
What was the test system used for this?
(Mobo and CPU)
Oh, and btw...RAID now stands for Redundant Array of INDEPENDENT Disks...
Inexpensive was the original term.
SATA-II tops out at 3Gb/s or 375MB/sec max (although the Serial ATA International Organization states 300MB/sec), so putting more than 3 x SSD's in a RAID-5 is effectively a mute point and will not give any further performance increases.
For this reason, SATA-III or SAS 6.0Gb/s is needed and even then is limiting to a maximum of 600MB/sec. This is the main reason why Fusion-IO makes sense: http://www.fusionio.com/
FusionIO is a PCI x4 card with 80GB to 320GB per card and offering random IO write speeds in the excess of 600MB/s and sequential write speeds 1GB/s.
I'll buy an SSD RAID when they can do that.
With all that theoretical power of ssd why would anyone use an on board mobo raid to give accurate scaling of ssd's? You need an actual RAID controller such as an Adaptec 5805 or Areca 1231 where the bottleneck appears around 800 mb/s or higher. I see this benchmark as a wake up call for people to start using controller cards instead of underpowered Mobo RAID
With SSD's now taking the limelight, I'm surprised there isnt some sort of internal parity on the drives themselves.
Considering the possibility that a bit can go bad, wouldnt it make sense to use the same raid5 stripind and algorithms inside the SSD itself?
This could be a great selling point to the upper end of the market. Since the drive is unlikely to actually FAIL outright, it will almost certainly keep a drive running near it's EOL.
Add it on to SMART and you can get nice handy warnings when the drive has finally outdone its write cycles.
- Adaptec 1430SA SATA controller
- 2 Samsung Spinpoint HD103UJ (1TB) in software raid 0 (linux)
Reading 200 GB dataset from xfs filesystem: 242 MB/s on average
Dirt cheap.
Pretty dumb assumption. Duh...like yah hardware raid you need and this was figured out a long time ago...in internet years...Read Below.
http://www.nextlevelhardware.com/storage/battleship/
Another winning artilce by the Inq.
Maybe if the author of this segment had some common knowledge, they'd know that the south bridge to north bridge cap is 500 MB/s. They'd also know that the sata controller is on the south bridge. That's just me though...
Where do I find this built in vista drive benchmark program? I tried looking in device manager as I found a post on a web forum insinuating as much or should I be in disk management or perhaps in reliablity and performance monitor? Do I need a disk that is devoid of any volumes?
It isn't critical that I know but I am just curious. Thanks in advance.
Tom Kolbe, you are wrong on a few points. SATA II was the name of the committee that developed the SATA 3.0Gb/s spec and not the name of the spec. SATA 3.0Gb/s tops out at 3GB/s since 8bit = 1byte plus 8b/10b encoding. Finally, SATA is a point-to-point link. Each individual SATA drives gets a 3GB/s link. Theortically, if you have 5 drives capable of achieving 3GB/s and ran them in RAID0, you could hit 15GB/s (assuming no overhead and large enough interconnect bandwidth).
There are several problems with your testing, the least of which is the purpose behind it. Your hardware was so far off from what should have been required, that this almost wasn't worth reading. Also, you should probably be aware of what to use for testing SSD performance. There was a recent article that went in-depth on the topic of SSD performance benchmark testing, but I can't seem to recall the website.
The real speed of SATA 1.5 Gbit/s and SATA 3 Gbit/s is 150MegaBytes and 300MegaBytes per second signaling layer. SATA 2 is 3 Gigabits per second, not Bytes my friend.
SATA-IO is a consortium of vendors to draft SATA specifications:
http://www.sata-io.org/
SATA-IO determines the specifications of of the Serial Interface and has done such since SATA I, which again runs at 1.5Gigabits or 150MegaBytes per second.
I will agree that this speed is per device however and concede to having not remembered that, but the limiting factor of the technology still exists.
Finally a good candidate to stress test different RAID cards.
It's now obvious that non-enterprise storage has to make it's way up on the NB like intel did with LAN when they introduced the CSA interface. That would put to good use the free space they must have on the NB now that the memory controller is on die with i7.
Increasing the NB to SB bus wouldn't be fixing the problem, just patching it. We'll always need faster storage, i'm not buying a san for home.
Simon from canada.
467MB/s from an onboard RAID is great. This looks like the sort of tech that will help to increase hardware sales, which is good news during a recession.
Selling point: This hard drive pair is approx 10x faster than your old IDE hard drive.
Pairing these in a laptop could mean that laptops are properly quick, at last.
Just a note on the other benchmarks. A recently purchased western digital WD6400AAKS has an average read of around 94Mb/s.
I find it odd that INQ would test such high end products against such low end ones. Raptors, 15k rpm drives, other ssd's and high end 7.2k drives should be in there. Not 2 drives that can't even manage 120mb/s in raid0
Some very nice speeds there would love to know the boot time of XP/Vista with a quad core cpu. also game loads would be interesting.
A few problems I have with the benchmarks though, and thats that they are a little irrelivent. What I would like to see compared are say 1x none raid modern (fast/dense platters) 7200 hdd and then 2x and 4x of that drive in raid0, 1x 10,000 VelociRaptor and then 2x and 4x in raid0.
I think the raptors may manage 300+MBs with 4x raid0 but giving 10x the storage space and probably costing less?