Chipset Serial ATA RAID Performance Exposed

TheRaindog writes "Serial ATA RAID has become a common check-box feature for new motherboards, but The Tech Report's chipset Serial ATA and RAID comparison reveals that the performance of Intel, NVIDIA, SiS, and VIA's SATA RAID implementations can be anything but common. There are distinct and sometimes alarming performance differences between each chipset's Serial ATA and RAID implementations. It's also interesting to see performance scale from single-drive configurations to multi-disk arrays, which don't offer as much of a performance gain in day-to-day applications as one might expect."

Comments

Surprised much?

[Anonymous Coward]

SiS, nVidia and Via are hardly world renowned for their RAID controllers, so why should we all act surprised that a consumer level product from low-cost manufacturers with very little experience designing these types of device doesn't exactly have screaming fast performance?

Best Upgrade

[swordboy]

I think that the hard drive is the most overlooked upgrade for a "power user". If at all possible, go out and pick up a 15krpm Ultra SCSI hard drive and controller for the boot partition. Use that slow ATA crap for storage of non-performance type stuff.

18 or 36 gig drives aren't exactly too expensive given the performance that they offer.

Re:Best Upgrade

[Lead Butthead]

Good deal, ONLY if the manufacturors are being honest with drive spec. Several coworkers that used to work for Quantum have indicate that the actual drive mechanism used in SCSI and ATA drives frequently shares common mechanical parts (platters, spindle motors, etc.) Their differences are ENTIRELY artifical. SCSI drive spec at times looked "better" in part because of the firmware difference and cheating on SPEC; for instance seek time for SCSI drive are computed differently to create the illusion that somehow SCSI drives seek faster...

Re:Best Upgrade

[leerpm]

Yes, and that it means it is stored in 2 locations. So when you want to look up 2 different parts of the game data, you can delegate each task to a separate drive, instead of one drive doing the read for both.

Re:Best Upgrade

[tuxlove]

How does RAID 1 help loading times? RAID 1 is all about mirroring.

Mirroring generally improves performance, which most users and most inexperienced engineers don't realize. Because you have the exact same data on at least two different spindles, you can transfer data with twice the concurrency, and at times approaching twice as fast. When reading a large file, for instance, if each disk can transfer, say, 10 MB/second and the file is 20 MB in size, the file can be loaded in one second with mirroring and two seconds without.

In addition, concurrency allows you to load two different files simultaneously on different disks. Not only do you get faster transfer times, you don't suffer from disk head seeks back and forth as you read the files. This can actually improve "load time" by much more than twice.

Since most filesystem operations are reads, the concurrency gained by mirroring usually helps immensely. However, writes do not suffer significantly either. When you write to a file on a mirrored filesystem, it obviously must be written out to both sides of the mirror. But, it doesn't take twice as long, as one might immediately think. Data can be written simultaneously to both drives, at a cost which is only marginally slower than writing to a single disk (assuming they are attached to different disk controllers/buses, as best practices dictate).

All-around, mirroring is very good for performance.

Re:Best Upgrade

[Jim_Maryland]

Shamelessly presenting a link to define the RAID levels.

Seriously though, the proper RAID level all depends on how much money your willing to spend for the speed and/or performance you require. Consideration for the types of operations (mixed read/write, read-only, or write-only) and reliability (can you afford to lose the filesystem or do you need fault protection) along with your budget usually determine the RAID level for a given system. Also throw in that you can use hardware and software RAID and the choice becomes even more difficult.

Personally, I tend to mirror the OS and application filesystems and use RAID5 for data, but these are systems we deploy and need a high degree of reliability and performance (pretty even mix of read/write data transactions).

Raid 5 is a combination of the 2 in some ways, but it requires 3 hard disks.

This is the minimum configuration, but RAID5 really just requires a disk to maintain parity. You lose capacity for the sake of reliability (example: 5 disk setup could use 4 for data while the other disk maintains parity). Optionally you could add "spare" pool disk(s) to provide failover to automatically take the place of the failed disk until it is replaced (to ensure availability - wouldn't want a two disk failure, rare but possible).

Re:Best Upgrade

[stilwebm]

Another reason that SCSI drives perform better in RAID arrays is that SCSI permits out-of-order I/O request execution.

It also has great command queuing as part of the out of ourder command execution. Serial ATA supports Native Command Queuing, providing these features plus First Party DMA and Interrupt Aggregation. Hardware support is relatively new. Seagate was the first to make a drive that supported it. My understanding is that the majority of Serial ATA drives out there essentially have parallel IDE controllers with a Serial ATA converter.

Here is a great article from Intel on NCQ: PDF HTML.

IDE performs blocking I/O, so everything would have to wait until drive 3's read was complete. I don't know if this also applies to SATA.

Interrupt Aggregation and First Party DMA were designed to limit the effects of this. SCSI still has an advantage with its offloading controller though. I also understand that the maximum queue depth for commands on the SATA is 32, while it is 256 for SCSI.

Question about striped/mirrored raid

[biscuit67]

Does the raid driver typically allow two independent seeks on the seperate drives with mirroring enabled? I would expect this to significantly improve things like boot times as most of the time is spent seeking for new data. I would have expected a 50% drop in seek. If they don't do independent seeks, why the hell not?

Re:Question about striped/mirrored raid

[cduffy]

It's not as big of a boost as you might think, because not infrequently you'll be reading enough data to require two consecutive stripes to be read (anything that crosses a typically 64k boundary).

Then you can be penalized for seeking your heads independently, because you need to pay your seek time separately for the second 64k of a given read.

Re:Question about striped/mirrored raid

[Tranzig]

Don't forget that those RAID controllers are just toys for the kiddies. Industrial grade RAID controllers have on board processor and memory, and they do optimize the read access for RAID 1 arrays. Though they don't halve seek time on two disk arrays, they still provide noticeable speedup for reading.

Re:Too Many Checkbox feature

[mattkime]

Holyshit, at this rate I will have 1 new input per year. Why can't we wait a couple years and all agree on 1 super format.

Yeah, then they'll come out with double sided bluetooth and the upgrade cycle will start again!

Built-in RAID chipset performance has always...

[Dagny Taggert]

..been a sore spot with me. Most users do a RAID 0 setup so their cool rig is hottie fast. Truth is, I can't see much real world performance difference. For my money, a large SATA drive and an external FireWire for backups is the way to go. Simple setup, no worries about drive failure and losing data, and still fast enough for UT2004.

Seemed to miss an important implementation

[SilentChris]

I recently put together a rig with a K8V SE Deluxe. The chipset includes two SATA RAID chipsets: the standard VIA one and a Promise one. I've been absolutely floored by the Promise's performance (easily the fastest desktop RAID I've ever tested) and I don't see it anywhere in this review.

For those hankering for another opinion, setting up the SATA RAID was a breeze. It was literally set it up and forget about it. The servers at work were much more difficult to set up. If you have the extra money for a spare drive (mine is two WD 10,000 RPM HDs :) ), it's worth it. Nearly double the speed.

Too...Many...Graphs

[dangerweasel]

Must...Stop...Reading.

Re:Too Many Checkbox feature

[Anonymous Coward]

Remember the days when IDE/SCSI was all you needed.

Yeah, when all you had to worry about was MFM or RLL? ST506, IDE, E-IDE, Western Digital IDE, ATA, ATA-2, ATA-3, ATA-4, ATA-5, ATA-6, SCSI, SCSI-2, SCSI-3, Wide SCSI, Fast SCSI, Fast SCSI-2, UltraWide SCSI, Ultra SCSI-160? Connectors were just as simple; 40pin, 44pin or 80pin? 25pin D conector for external SCSI, male or female? How about a dense 50 pin D connector, or wait, maybe 64pin? 50 or 64 pin cable for internal drives; your choice.

Don't forget to setup your SCSI bus and wave that chicken. Does your SCSI controller boot from SCSI ID 0 or 7? Maybe 6 or 4? Did you set your master and slave jumpers on those IDE devices properly? Your IDE performance sucks; you didn't put a PIO device as a slave on the same channel as your screaming-fast UDMA166 120Gb hard drive now did you? By the way, does your BIOS support 48bit LBA for that drive? Got SCSI terminators. Need a terminator block or is it an internal jumper perhaps?

Oh boy, things were so much simplier back then..

Re:Too Many Checkbox feature

[Anonymous Coward]

Got SCSI terminators. Need a terminator block or is it an internal jumper perhaps?

Dear god I'm having sysadmin flashbacks now. Gonna be thinking of sendmail.cf all day...

Bastard.

RAID vs. single drive performance

[fugu]

storage review did a writeup a while ago comparing RAID 0 performace to that of a single drive. more often that not you're better off getting a single, faster drive if you're looking for desktop performance.

Wrong

[sczimme]

I normally don't respond to ACs, but this one is just incorrect.

Yes, RAID {1|5|10} are generally used for their redundancy purposes, but RAID 0 is used because it offers improved I/O performance. It is certainly not used for redundancy because - guess what - it doesn't offer any on its own*. Go read this before you provide more misinformation.

* it can be used in combination with other levels - e.g. RAID 0+1 - to provide performance and redundancy.

storage market sluggishness?

[davejenkins]

From what I can see in the market right now,
1. Everyone says they need more storage, so the market for it should be huge
2. SAN or NAS configurations are always more expensive than people think (even though they are radically more cheap than they were two-three years ago).
3. Because of the sticker-shock, a lot of people actually spend their first swipe at the problem cleaning out the cruft and streamlining their business processes and data management rather than drop coinage on storage kit
4. Storage companies are having a very hard time here in Japan, probably from the influx of vendors (see #1 above).

Re:Well lets see here

[LookSharp]

So why make a RAID?

They are not talking about mirroring (RAID 1) exclusively. They are talking about RAID 0 so people can stripe drives and achieve considerable performance increases.

As for me, I have an 8-channel IDE raid card with 8 x 120GB drives, hardware RAID 5, and in 24 months have blown oh about 4 drives (3 on the same channel til I found a faulty cable)... I have really appreciated the 860GB array having fault tolerance. And yes, I do some backups of critical data, but I can't afford the storage required for regular full backups.

Quick question

[sczimme]

You are a hardware vendor. Would you rather sell a) 10,000 units that are broadly compatible but offer [arbitrary number] 80% performance or b) 3,000 narrowly-focused units that offer 100% performance at a slight price premium?

I believe the revenue generated by selling 10,000 units would outweigh that of the 3,000 higher-priced units, even if the technology in a) is inferior.

I'm not saying this is the best/worst/right/wrong way of looking at the situation; I'm saying this is probably the compromise the vendor has to make when offering such items.

Surprise, surprise, surprise

[Anonymous Coward]

Why do folks act shocked when commodity hardware behaves like commodity hardware?

Why should computer hardware be exempt from the "You get what you pay for?" dictum which dominates other markets.

And when you make millions and millions of any one thing, a "couple of pennies a chipset" adds up. Once again, that's what you get when you buy a commodity.

RAID Perfomance

[Berylium]

For the past 3 years I've had a RAID array set up on my home computer. It is a RAID 5 array with four 18GB Seagate X15 hard drives on an AcceleRAID 170 PCI card. I'm on the computer several hours a day during which time I play various video games, program in visual studio, and transfer a bunch of MP3 sized files and very large video files (~2GB). From my experience, the RAID 5 is definitely faster in some tasks than a high-performance ATA drive (like game loads) but for the types of activities I'm doing the expense of the SCSI drives and the noise they generate is more costly to me than the (perceived) slight speed disadvantage of a single disk serial ATA drive.

Don't get me wrong, the RAID 5 array is sweet and certainly amps up geek appeal, but I don't have enough friends who know what the hell a RAID array is to really impressive them.

-Berylium

Driver software is probably key

[Anonymous Coward]

Since every SATA raid controller (bar the i960 based one from adaptec) is done using software, I reckon that what is actually benchmarked here is how optimum the drivers are, not the hardware performance. Besides (I'm guessing, as I only read the conclusions page) that each of these interfaces is connected off of a crummy 32 bit 33MHz PCI interface... That's the real killer right there.

I have a Dell PowerEdge in the back room with 2 15k scsi drives running linux and raid 0 - with hdparm -t this thing gets 125-128 mb/sec! The HD interface on that machine is definitely hung off of a PCI-E interface or something better; as the maximum theoretical transfer rate of PCI is about 33*32 million bits per second or 132 megabytes per second.

What would be really nice is if the filesystem was put on the i960 based adaptec card...

we've opened scsi drives and ide

[john_uy]

there are physical differences in the manufacturing of the drives.

1. the surface disk are different from ide and in scsi. the scsi drives are much reflective than the ide drives. though i am not sure if this affects reliability.

2. the size of the platter (diameter) is much smaller in scsi than in ide. probably this will help them achieve a higher rpm than the ide counterparts.

3. the head movement is much sturdier in scsi (probably attributed to more better magnets.) i find it much difficult to move the heads in scsi than in ide.

4. there are more chips underneath the scsi drive than in ide. however, this does not tell much. but in fc drives, there are 2 dsp chips, one that handle internal drive functions like motor and head, while the other handle io host requests making them much faster!

5. scsi drives have higher mtbf. though this may not be much the only guage for quality but scsi drives are much better in quality.

Re:we've opened scsi drives and ide

[Jeff DeMaagd]

While SCSI and IDE are just interfaces, that often isn't the only differences, because they are sold to different markets.

The IDE drives are sold to a consumer market where they don't need to be tested as vigorously. SCSI drives are often tested more vigorously from a mechanical, electrical and firmware aspects. Because the SCSI drives are often sold for heavy server use, they must be able to withstand constant use, around the clock for years.

While it is possible to get the same mechanicals in both SCSI and IDE formats, I don't think that is done for any of the cheapest drives, IIRC,
WD Raptor is one. So far that I know, there aren't any 15k RPM SATA or IDE drives. It could be done, but it wouldn't be that much cheaper.

10k and 15k RPM drives also have different platters, cases and mechanicals - the platters are more like 2" in diameter than 3".

Generally a SCSI drive is expected to last for five years, and I suspect that there really is an improved build quality to make it worth putting the 5yr warranties that drive makers put on SCSI drives, in a day when a typical IDE drive gets 1 yr, or if you are lucky, three.

I know it isn't much, to say, but I've yet to have any of my SCSI drives fail on my, something I can't say for the IDEs.

SATA RAID? I pass that

[kill $(pidof explore]

Let's talk about profermance. Most SATA drives are, still low end IDE drive, 8ms seeking is not a hit. the one SATA fanboys talking all the time is Western Digital Raptor, but, hey, they are the same price as 10K SCSI U320, what the point?

I agree Raptor are great disks, 2 of them will out run PCI bus bandwidth, would you go PCI-X for SATA raid? a good PCI-X RAID card will cost $300+ for 4 ports, no thanks, I will stay my SCSI solution.

The bottom line is SATA don't even have a BUS.

Intel RAID crashing under load

[jeffmeden]

"Of course, for all its prowess, I'm still a little troubled that the ICH5R's RAID 1 arrays crashed out of IOMeter under our highest load level. A load of 256 outstanding IOs is quite a bit beyond what most desktops and workstations will encounter, but it's well within the realm of possibility for servers" Can anyone confirm or deny that this occurs in real world settings? Its definitely troubling that the crash condition was consistent, but I am suspicious that it was simply an incompatibility between the benchmarking tool and the raid controller. Does someone know more? Jeff

RAID 0,1,5

[mr_rizla]

Raid 0 = striped disks for improved performance. No redundancy. In fact, increasing your chances of losing data because if one goes down, no chance of data recovery. (total storage = total of disks)

Raid 1 = Mirrored disks, writing same data to all disks so if one fails you simply replace it and no loss of data. (Total storage = 1/2 of disks)

Raid 5 = Redundant striped disks. One of the disks is used to store a XOR bit, so that basically any one of the disks can go down and once it is replaced the RAID system will rebuild the data on to that disk. (Total storage = total storage of (all disks minus one))

In RAID 1 and RAID 5, which is used in business servers, you really need hotswappable drives so any drives going kaka will not impact the server in any way, just replace the hard drive under warranty without even rebooting the server and the RAID system will rebuild the drive.

RAID 5 is most effective in a business situation, offering a good compromise of speed, capacity and redundancy.

Re:RAID 0,1,5

[bobv-pillars-net]

RAID 5 is most effective in a business situation, offering a good compromise of speed, capacity and redundancy.

Nope. In a real business situation, i.e. data-warehousing or ISP hosting environment, nobody trusts RAID 5. It's slow and fragile. Instead, everybody I know goes with RAID 10 (striped mirrors). Here's a typical 8-drive configuration:

Stripe:

1. Disk 1 mirrored with Disk 2
2. Disk 3 mirrored with Disk 4
3. Disk 5 mirrored with Disk 6
4. Disk 7 mirrored with Disk 8

Total storage equals the same as a 4-drive RAID-0 system. Performance should be slightly better, on a high-end dedicated controller, as the mirrors should be able to seek to different files independently for concurrent read requests (thus lowering latency), while the stripes should be able to operate simultaneously for large-block i/o (thus raising the streaming i/o rate).

Reliability is better than Raid-5, for two reasons:

1. When a drive fails and is replaced, only that particular stripe is rebuilt. That means that until the rebuild is done, one drive will be doing streaming-reads, and the other will be doing streaming writes. None of the other drives are affected. Contrast this with Raid-5, where one drive is doing block-writes and all the others are doing block-reads, interspersed with CPU checksum calculations, until the entire drive array is rebuilt. The result is that RAID-10 has much shorter disaster recovery times.
2. In a RAID-10 system, up to half the drives can fail simultaneously without data loss, as long as one drive in each stripe remains functional. In a RAID-5 system, the loss of two drives guarantees loss of all your data.

Re:RAID 0,1,5

[dfghjk]

Tecnically, the only justification for hot-swap is a zero-downtime requirement. If downtime can be scheduled, then an online spare is all you need and spares are good in any case. The need for hotplug is consistently overstated.

RAID 5 is increasingly marginalized by the low cost of drives and high capacity they offer. RAID 1 *should* increasingly replace RAID 5 in the minds of people who understand the issues but sadly it does not. Many people believe that RAID 5 is simply "four better". Those same people also like hot-swap.

Re:not about transfer rate

[Gr8Apes]

For me, personally, transfer rates are #1, seek times are #2.

And for the record, this article only cements in my mind that SATA is seriously no better than IDE, just a faster version of IDE, with all its inherent problems.

Firewire is another one that's basically DOA for HDs. SCSI really is your only solution, especially if you're looking for RAID performance. (Of course, that's not your normal consumer purchase, but I have 3 SCSI RAIDed systems, so I'm not your normal consumer;)

For hard drives, read/write & reliability.

[khasim]

Sustainable speeds. From the pictures in the article, you can see that drives that may have the highest max speed don't always have the highest average speed.

But then you have to hook your drives to a controller. And controllers have the read/write & reliability factors that hard drives do AND they also have CPU utilization.

Ideally, you'll want hard drives with fast read/writes and high reliability hooked to a controller that does fast read/writes and has high reliability AND very low CPU utilization.

But if you're just looking at hard drives, you're correct in your statement.

But for best utilization of the hard drive, you at least have to look at the controller, also.

And cables. :)

They are not Hardware RAID!

[farrellj]

I found out a few months back some interesting things about the state of SATA RAID...most of the SATA chipset RAIDS are not hardware RAID controllers.

If you check Linux Mafia's web page on SATA controllers, you will find that very few of the SATA RAID controllers are actually hardware RAID. What their "Drivers" really are is proprietory software RAID pretending to be Hardware RAID. I think of all the SATA RAID controllers and chipsets being offered, there are only three that are really hardware RAID. And 3Ware's offering is the least expensive of the real hardware RAID.

ttyl
Farrell

3Ware s-ata hardware RAID

[Axello]

I've been using a couple of 3Ware hardware RAID cards in my FreeBSD servers. More expensive than the onboard crap, but Very Nice. Full hardware RAID 0,1,10,5,50, remote control, hot swap, hot spare, email notification on failure, the works.
You can configure your RAID remotely while your server is running. (But always be careful with your boot disc ;-) Or you can install your OS while the RAID is building in the background. Works with Linux & Windows as well, unfortunately not with MacOS X.
But for MacOS X (& linux) geeks, the XRaid RuleZ!

Exposed? Everybody knows it's software RAID.

[jgarzik]

Being the person implementing Serial ATA for Linux...

Most "SATA RAID" is a bunch of marketing malarkey. It is provided by the BIOS and OS, not the hardware.

There are a few "true" hardware RAID controllers, such as 3ware or some of the more advanced Adaptec controllers.

In the middle is Promise, which produces controllers what I call "RAID offload" features -- not true RAID, but faster than non-RAID if you use Promise-specific features.

Finally, the third group of SATA controllers is vast majority -- no RAID support whatsoever, but they are being sold as RAID.

Any benchmark of SATA RAID simply benchmarks the OS- or vendor-provided software RAID driver.