gundersausage

The Death Of Raid

Recommended Posts

Game Load Times! That's why people stripe these days! Not because of results on "application" benchmarks. Nobody gives a turd how many Word documents/second they can save, it's yanking in the megs and megs of textures and meshes and sounds and crud that you need to play ONS-Torlan in from disk that people care about. If you get in the game 0.2 seconds before someone else, maybe you get the killer vehicle while the other guy is hoof'n it. That's why people buy RAID-0 rigs.

I have a RAID-0 rig and I always assumed it made a big difference but I never tested it well. I know I can sling giant files arround fast, that's a good thing, but I can't really say I'm amazed at the disk speed. I'll do some side-by-side comparisons with other machines and see how they come out. I think my RAID-0 is helping game load times (I expect it's mostly sequential transfers of large blocks of data which is what RAID-0 is good at) but I'm not sure.

Share this post


Link to post
Share on other sites

Some people will never get it:

I just can't believe Storage Review is making such bold claims on such limited data.

Limited data? SR's hard drive tests are more in depth then anything on the web - you point me to a different site and prove me wrong. There is nothing out there that can compare to the IPEAK tests for accurate relevant data.

The SR results only prove there is a limited performance improvent from striping in very light desktop workloads on a legacy PCI bus.

So even though STR is a non factor in desktop performance, you feel the 133mb/sec PCI bus was the bottleneck? Very light workloads? Why don't you put up some evidence, tell us what your doing, and in performance monitor, tell us what queue depths you are generating.

The average computer enthusiast can create some pretty heavy disk I/O. Many people are heavy into p2p sharing, digital photography or video-editing. They move large amounts of data around their hard drives. These people will greatly benefit from striping (in combination with mirroring or parity). Your hard drive won't stall when you're making a system backup over a gigabit network or when the weekly full system virus scan is running. File copy times will be greatly reduced.

no, no, no and no. You think P2P applications benefit from striping? I don't know what to say about that, other then its insane. Digital photography and video editing will benefit much more from running your disks independently (Rather then using silly striping). I do video editing. I use 3 fast drives to do it. It doesn't take much to see that slapping your drives together in a raid is putting all your eggs in one basket. The first thing you do puts your entire raid under load, rather then taxing a single disk, leaving the rest free and clear to do whatever other multi tasking you want

A few other points on this quote: Don't run your damn virus scanning while your working on something disk intensive. Read the damn manual and figure out how to schedule it late at night. I would think this wouldn't even have to be stated. As for file copying, I regularly move 2+ gig files around, and moving from one disk to the other requires only a small amount of forethought, and will copy faster then a raid.

Oh yeah - this goes along with the background processes thing - don't schedule your full system backups while your going to be working on editing the next Star Wars.

Why do people buy cars with six seats when most of the times only one or two seats are occupied? Because sometimes they take their family on a trip. Why use striping? Not to improve MS Office load times but to improve performance in scenarios with heavy I/O. This may happen frequently or rarely, but when it happens you will be happy to have multiple drives working in parallel.

Guh – again, what kind of heavy I/O loads are you referring too? What you are doing, and average queue depths again would be appreciated. I put my system into overload all the time, and when I do, I’m happy that I have independent disks to service the requests. It’s faster then RAID.

I use a RAID 5 array with five Maxtor Atlas 15K drives on a LSI MegaRAID SCSI 320-2X with 512MB cache. Performance: 240MB/s sequential read, 180MB/s sequantial write, desktop I/O performance three times higher than a single Raptor WD360GD. It is virtually impossible to get this array to stall. It doesn't matter if you're pumping 35MB/s over the network or have a virus scanner or file defragmentation running in the background. The only way you will notice it is because of the blinking LEDs on the front panel or the crispy seek noises making its way out of the case.

Well – now I understand why you are so desperate to justify striping – that’s a big investment. Unfortunately, we already know that STR doesn’t matter except for loading large files, and file copying – and those are better served for the most part with independent disks. Your claim of desktop I/O performance three times higher then a single Raptor is ridiculous, and based on STR? Come on! Even Anandtechs article showed that where STR should matter (FarCry level loads) it still DOESN’T!

Even with all this information available – your denial is impressive, at least you are sticking to your guns. Unfortunately your long post is nothing more then a glorified:

I don’t care what all the evidence shows, my RAID is still faster, I know it is!

Share this post


Link to post
Share on other sites

RAID-0 will always be an added blessing to my eyes. Of course I'm talking cause I've never had a drive failure in my life and can't say how it feels like to lose an array from one drive. <_<

Other than that, even the minimal writing/reading advantage it gives is welcome.

Share this post


Link to post
Share on other sites
People got the idea that RAID-0 helps performance in all cases.  People got the idea that STR matters.  Unfortunately, people have a hard time getting rid of those ideas.

And... don't forget to add SCSI to this list.

Just wanted to post this! :D

I started a thread about WD740GD in a popular forum here in Hungary, it ended up in a terrible flame with some SCSI fans :rolleyes::rolleyes:

Share this post


Link to post
Share on other sites

Assuming you have a decent RAID controller, RAID1 can be faster than RAID0 for _reads_.

For both you can do read interleaving. But the difference is with RAID1/disk mirroring, the system can read from EITHER drive - the data is the same on both drives, so a smart controller can pick the most convenient drive to read from. Whereas with RAID0, if the data is on a particular stripe, the controller has no choice but to read from that particular drive (this is assuming a simple dual drive stripe), and so it has to wait till the drive is ready.

This is why stripe size is important for RAID0 and why sometimes you don't see as much performance gain.

Of course, there are plenty of crappy RAID controllers- for RAID1 they do stupid stuff like just read from one drive, or read the same thing from both drives. A smart controller will keep track of "where" the heads are on the drives and pick the right drive to read from, and also interleave reads for bulk transfers.

Summary: with a decent RAID controller, RAID1 gives you faster (2X) read speeds, same write speeds, 100% data redundancy, fast rebuild. But zero storage gain.

RAID5 is overrated - only should be used if you need redundancy AND lots of storage - the disadvantages are lower sequential write speeds, much slower random write speeds, and slow rebuilds.

RAID1+0 gives you 4x read, 2x write. But less storage than RAID5.

The sad thing is I can't find a cheap ATA RAID controller that does read interleaving for RAID1 (3ware isn't that cheap :) ). Any suggestions welcome.

Share this post


Link to post
Share on other sites

I never understood when anyone would want to double the chance that their computer data would be unavailable? Nope never understood that.

Share this post


Link to post
Share on other sites
The sad thing is I can't find a cheap ATA RAID controller that does read interleaving for RAID1 (3ware isn't that cheap :) ).

I think you should read the SR article lyeoh.

All the RAID controllers featured in the article are balancing IOs on their mirrored disk arrays. You can observe the RAID 1 results here and then contrast them with the single drive results here. The technique you describe is designed to improve positional performance by allowing disks in an array to respond to read requests independently. It is therefore only an advantage in very specific situations which I have detailed on this forum several times. One of these conditions is the queue depth must be greater than 1 for any benefit to be observed. SR's IOMeter tests satisfy all the conditions required for independent responses from disks in arrays to some extent. At queue depths >1 you can observe that load balancing is occuring.

It should go without saying, but I'm going to say it anyway: Despite bringing load balancing to the table, these RAID controllers were unable to deliver significant performance gains in single-user tasks.

Share this post


Link to post
Share on other sites
Why can't we all just get along? :D

Are you a RAID-0 sympathizer? :) How many card carying members are there for gosh sakes?

Free

Share this post


Link to post
Share on other sites
Are you a RAID-0 sympathizer?  How many card carying members are there for gosh sakes?

No actually, I have used RAID 0 in the past (4 Cheetahs & 2 7200.7's) and my new 74GB Raptor is still faster. Actually, in a recent test my Raptor is still faster than a pair of 8MB cache SATA 7200.7's with DVD encoding, game map loading, and image editing with Photoshop.

I just find it interesting how passionate people get over having or not having a RAID 0 array. Personally, after seeing the lack of benefits I would not waste my money on RAID 0 again. Been there...done that...got the Raptor. :D

Share this post


Link to post
Share on other sites
Game Load Times!  That's why people stripe these days!  Not because of results on "application" benchmarks.  Nobody gives a turd how many Word documents/second they can save, it's yanking in the megs and megs of textures and meshes and sounds and crud that you need to play ONS-Torlan in from disk that people care about.  If you get in the game 0.2 seconds before someone else, maybe you get the killer vehicle while the other guy is hoof'n it.  That's why people buy RAID-0 rigs. 

And this is important enough to justify putting the rest of your data at twice the risk???? :blink: And to do a little math here, for the (undeniably) faster STR of striping (I won't even dignify using the term RAID here for something that is inherently not redundant) to give you an 0.2 second advantage means you're loading about 24 MB of data (assuming the drives each have a 60 MB/sec transfer rate). Most games don't load that amount of data that often for STR to even be much of factor, and many of those that do offer to precache if you have enough RAM available (and any decent machine these days should have at least 768 MB of RAM). Indeed, MS Train Simulator is one of the heaviest disk users in that regard (routes can be upwards of 1 GB, and the game won't precache), and even here seek times have been found to be way more important than STRs. In fact, one person running MSTS on a RAM drive all but eliminated the jerkiness caused by loading shapes and textures.

I have a RAID-0 rig and I always assumed it made a big difference but I never tested it well. ..... I think my RAID-0 is helping game load times (I expect it's mostly sequential transfers of large blocks of data which is what RAID-0 is good at) but I'm not sure.

Notice the words in bold-your words, not mine. Ever heard of the placebo effect? If you're made to think something is helpful even when it really isn't then your perceptions can be grossly distorted. Too many out there assume striping is faster even though real-world tests show it to be of very limited use. As a result, many enthusiasts are incapable of impartially analyzing the benefits (or lack thereof) of striping. I'm not blaming you or anyone else here for what amounts to an illogical defense of striping. I'm just saying you're all so biased by what so-called experts have said that you're now incapable of seeing striping for what it really is. It is of little or no benefit unless all you do is transfer vast amounts of data, and even then it is not necessarily the best solution since it puts your data at twice the risk. If transfer speeds are really that important, then it's cheaper to buy a rig with 3 or 4 GB of RAM, use most of it for a RAM drive, and start your "STR-intensive" games from the RAM drive. Of course, you'll initially take some time copying from hard disk to RAM drive, but after that you'll have STRs far exceeding anything striping can do, and all without putting any of your data at more risk.

Share this post


Link to post
Share on other sites

One might also question the ongoing value of loading a game .2 seconds (or any number of seconds) faster than others. It seems to me that it is in the best interests of game developers to figure out ways to make sure that everyone has an equal crack at selecting "class" (or position, or weapons, or whatever you want to call it). At least in the one game I play often online (America's Army), load time has zero effect on selections.

Share this post


Link to post
Share on other sites

While I agree with the lack of RAID0 performance gains, I'd like to point out that RAID0 does NOT put your data at "twice the risk".

By far, the greatest risk to your data is a virus/trojan, not hard drive failure. Backups protect your data, not hard drives. Running RAID0 or spanning does not significantly increase the chances of lost data.

You may say, "but people don't run backups". Well, that only proves my point, because if someone's too lazy to run backups, they're probably too lazy to update their antivirus software, scan attachments before opening, stay current with Windows Update, etc.

System builders -- amateur or professional -- are partly to blame here, because every new system budget should include a backup strategy. Since Ghost requires a reboot, it isn't schedulable. That's why I use Acronis True Image (www.acronis.com), ideally backing up images to a hidden partition on a second hard drive, which protects both against virus/trojan attacks and HD failure.

KC

Share this post


Link to post
Share on other sites
I use Acronis True Image (www.acronis.com), ideally backing up images to a hidden partition on a second hard drive, which protects both against virus/trojan attacks and HD failure.

Seconded.

Also, there is a disadvantage to RAID-0 other than increased risk of data loss, for those who are asking what the disadvantages are. Mars said it well already, but I'd like to also put in my two cents also. Well-thought-out independent-drive setups are faster in nearly all real-world uses than any RAID setup. This is theory, but backed by empirical data which is readily available throughout this forum. So, if you have two or more drives already, you could be getting better performance by setting them up independently.

Share this post


Link to post
Share on other sites

Never say never.

The claims that RAID0 never has any place on any desktop computer ever are just as false as claims that RAID0 is always better. Both seem to be totalitarian points of view that try to force their perspects on everyone. The true answer is that it depends.

I use a RAID0 array for my Photoshop scratch disk. Do I put my valuable data, programs, OS, email, etc. on that RAID0? Of course not. Does this make a huge difference to the productivity of my Photoshop time? You bet! I edit large files that max out the full 2GB RAM capacity of Photoshop, so the speed of the scratch disk determines the speed of my edits after a while. Smart use of RAID0!

And, since I've got a RAID0 array for Photoshop scratch, and since the scratch partition uses only a tiny portion of the size of the array, why not use the remainder of the space for other non-critical uses where -- dare I say it? -- STR makes a difference. For example, I have copied several gigs worth of TOPO! topographical map data onto another partition on that RAID0 array. These consist of many files in the 2MB - 10MB range. Now I can scroll around multiple US States worth of detailed topo maps as fast as I please because all this big fat juicy detailed ata is coming off the RAID0 array. Is this a stupid use of RAID0? Heck no! It's smart. Oh my oh my, what if the array crashes? Well, I spend 20 minutes re-copying the CDs. And the space is on that array anyway, regardless of whether I use it. And the maps do, in fact, load faster than from a single disk.

Oh wait, here is another use for my handy RAID0 array: a spool directory for my Epson photo printer. When I print a large picture, say a 13" x 40" panorama at 300dpi, the Epson printer driver generates an intermediate file that is a few hundred megabytes in size. Writing a file this large on my C drive takes a few 10s of seconds, whereas it takes a handful of seconds on yet another small partion on My Friend The RAID0 array. Is this stupid use of RAID0? Heck no!

Is my computer not a "desktop computer" because I use Photoshop? Because I often browse through large topo maps? Because I print large pictures to a photo printer? It depends on who is doing the defining, I suppose.

RAID0 is just like any other tool: it can be used properly or not. But to say it is never appropriate? Never say never.

Share this post


Link to post
Share on other sites

As for the commentary about striping being for servers, can we please put a bullet in that one? I deal daily with IT types who run RAID-0 stripes at home for their gaming system then come to work and advocate scatter storage across 10, 20, even 30 disk volumes with heavy transactional access patterns. They don't realize that by striping small amounts of data across a massive number of disks, they are artificially generating an enormous amount of aggregate I/O and seeking, and the problem gets worth with an increasing number of concurrent transactions.

This is something I've benchmarked and seen firsthand: 4-disk RAID array, 16kb stripe size, 64kb read requests vs. four non-RAIDed disks, OS and application tuned for 64kb I/Os. The latter almost always wins under heavy load: more TPS and less CPU load. Plus the fact that you can meaningfully tune by approportioning more physical spindles to the data that you know is needed more often.

Here's desktop/server RAID striping that makes sense: the controller/OS takes two physical disks, stripes part of each disk together to present two non-striped volumes and one stripled volume, and the controller/OS then places files on these three volumes based on a STR-vs-I/O ratio that is calculated based on the size of the data set, its frequency of access, and level of parallelism with which it is accessed. Unfortunately, no controller/OS/filesystem that I know of is that intelligent.

I have not found RAID striping to be terribly useful in most corporate environments save document management repositories and data warehouse / data mining systems. The vast majority work better with mirroring and as many different mount points as possible.

Share this post


Link to post
Share on other sites

The RAID king has no clothes...

Get yourself some cheap 7200RPM ATAs and if that doesn't make a big enough dent in your finances, buy yourself a lot of ram.

Share this post


Link to post
Share on other sites

I don't give a stinker what anyone says.

I have 2 Raptors in a Raid 0 array on a P4 3ghz.

I _definately_ notice a difference on my box.

The files are defragged monthly on my machine and I only keep 20gb of data on the 74gb partition - I (beleive) that the WINNT dir is towards the start of the disk also.

My 800mb swap file is on D: (PATA controller, not ICH5 SATA controller) and that's a 200gb WD Caviar JB.

My Raptors have a "problem" where I can only get about 25mb sustained real world throughput from them in ANY application, including apps which simply read but do nothing with the data (CRC checking apps -etc)

It's a shame about the problem - but thanks to the raid I can now write files quite quickly to C:.

Share this post


Link to post
Share on other sites
I don't give a stinker what anyone says.

This nicely somes up the opinion of the RAID-0 enthusiast. Here's your card...

Free

Share this post


Link to post
Share on other sites
I don't give a stinker what anyone says.

This nicely somes up the opinion of the RAID-0 enthusiast. Here's your card...

Free

Maybe he's using the banana card.... I definitely notice a difference when I'm using those. Everything is so snappy and responsive.

Share this post


Link to post
Share on other sites
RAID-0 arrays will win you just about any benchmark, but they'll deliver virtually nothing more than that for real world desktop performance.

That's from Anandtech. The first part is indisputable, the second the whole point of this discussion.

...unbelievably fast, but it's nothing we haven't seen before.

That's from Maximum PC (June 2004, page 76). They are talking about a top-of-the-line system for which any hardcore geek would trade his girlfriend (or boyfriend, let's not discriminate).

Regardless of its merits, you will continue to see RAID-0 in top-end computers reviewed by enthusiast magazines, and therefore you will see it in enthusiast computers, because for system integrators that negligible 2% means the difference between a pat on the head and an Editor's Choice / Kiss^H^Hck Ass / Golden Baboon Award.

Enough of that. I really have to question Anand's methods, and the inclusiveness of his conclusion. He tested a single RAID array on a single RAID controller. Raptors are the fastest hard drives available for a desktop system. What about a pair of inexpensive SATA drives? How would 2 Seagate 7200.7 drives do compared to 1? How would a pair of old 5400 RPM drives do? What about other RAID controllers? ICH5 is Intel's first try at a consumer RAID controller. What about VIA and their VT8237 south bridge? What about ye olde Promise S378?

And (and this is just a niggling annoyance that won't go away) should we give credence to hand-timed benchmarks that display results to the nearest tenth of a second?

-- Craig

Share this post


Link to post
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now