SATA RAID 5 for games?

[MIRTBelGeran]

hmm what are your thoughts?

Considering Read performance is good on RAID 5 and it only incurs a penalty for wrights. If I can afford the extra drives is it worth it and why? BTW I can not justify the cost for 0+1.

Please back up your opinion with at least technical speculation. no Yo RAID 1 all the way replies or stripe that biatch who needs a backup!!!

Also what SATA drives would be best and why? Again Technical speculation based on features and stats and would be apreciated.

Thanks in advance

 

[pj-schmidt]

I wouldn't recommend RAID 5 for speed, only for reliability. If you just want raw speed get a 150GB Raptor, or maybe go with RAID 0.

My 4x 250GB Seagate SATA RAID 5 Array on a 3ware 9550s-8 (hardware parity calculations) only gets 55MB/s for both sustained and burst, as reported by HDtune.

Good Luck,

~PJ

 

[venm11]

hmm what are your thoughts?
...
no Yo RAID 1 all the way replies or stripe that biatch who needs a backup!!!

Raid DOES NOT remove the need for backups. All raid does (besides performance) is prevent one disk failure from bringing down the system. There are MANY reasons for doing backup/restore, and drive failures are the less common incidents.

I wouldn't bother with raid 5 or 0+1 for games... too many headaches. Stick with 0 or 1.

 

[chanchan]

Raptor Raptor Raptor!

With the money invested in a RAID 5, just buy Raptor's!!!

If the 150 GB is too small, then Raptor in RAID 0!!

I'm not too happy with my setup I think I need to Raptor it myself ..

Either way RAID 1 will be faster anyways, your read speeds will be ub3r fast!

 

[venm11]

Either way RAID 1 will be faster anyways, your read speeds will be ub3r fast!

It really depends on your controller. If you're shelling out for lots of raptors, and want raid 5, the 3ware or Areca controllers are vastly better than the vanilla nvidia/promise/highpoint crap.

 

[MIRTBelGeran]

Raid DOES NOT remove the need for backups. All raid does (besides performance) is prevent one disk failure from bringing down the system. There are MANY reasons for doing backup/restore, and drive failures are the less common incidents.

I wouldn't bother with raid 5 or 0+1 for games... too many headaches. Stick with 0 or 1.

LOL I should have expected this type of responce. Was my sarcasm in that statement not evedent and the desire for performance and reliability. Never was I saying that a system backup was not nessesary or that RAID provieds this. That was an example of a responce I did not want. Also "too many headaches" is not technical speculation based on facts. That statement is a feeling. Can you back up this with examples? Generally speaking headaches are caused by the PPP effect.

Also can you quantify "vastly better" and provide linkage to a reviews or roundups on controllers. I have not been able to find decent reviews of like disks in differant array configurations For SATA drives plenty of review for SCSI and SAS.

Theoretically RAID 5 read performance should be almost the same as RAID 0. The Parity disk is not read unless there is a drive failure. However there is a write penalty for 5 over RAID 0 becuse of writting the parity bit.

Lets run the cost numbers, to determin the $ value. I know the numbers are not exactly scientific but the point is there

WD Raptor 150GB $200 * 2 = $400
Storage
RAID 0 = 300GB = $1.33 Per GB
RAID 1 = 150GB = $2.66 Per GB

WD Cavier 250GB $70 * 4 = $280
Storage
RAID 0 = 1000GB = $0.28 Per GB
RAID 0 + 1 = 500GB = $0.56 Per GB
RAID 5 = 750GB = $0.37 Per GB

What do you RAPTOR fan boyz think of that? 0+1 with more than double the storage at almost half the cost.

 

[venm11]

LOL I should have expected this type of responce. Was my sarcasm in that statement not evedent and the desire for performance and reliability. Never was I saying that a system backup was not nessesary or that RAID provieds this. That was an example of a responce I did not want. Also "too many headaches" is not technical speculation based on facts. That statement is a feeling. Can you back up this with examples? Generally speaking headaches are caused by the PPP effect.

Also can you quantify "vastly better" and provide linkage to a reviews or roundups on controllers. I have not been able to find decent reviews of like disks in differant array configurations For SATA drives plenty of review for SCSI and SAS.

Theoretically RAID 5 read performance should be almost the same as RAID 0. The Parity disk is not read unless there is a drive failure. However there is a write penalty for 5 over RAID 0 becuse of writting the parity bit.

Lets run the cost numbers, to determin the $ value. I know the numbers are not exactly scientific but the point is there

WD Raptor 150GB $200 * 2 = $400
Storage
RAID 0 = 300GB = $1.33 Per GB
RAID 1 = 150GB = $2.66 Per GB

WD Cavier 250GB $70 * 4 = $280
Storage
RAID 0 = 1000GB = $0.28 Per GB
RAID 0 + 1 = 500GB = $0.56 Per GB
RAID 5 = 750GB = $0.37 Per GB

What do you RAPTOR fan boyz think of that? 0+1 with more than double the storage at almost half the cost.

Sorry man... I couldn't figure out exactly how you meant that. RAID is for maximizing uptime to beyond 99.9% (or so), and it's not at all clear why you want to go through that trouble and hassle for a gaming box.

I just said headaches for the sake of brevity. More drives = more complication, quite simply. It increases the chance that you will have a drive failure. Then, you need to make sure the drives aren't from the same batch because those drives will all tend to fail at the same time (given same conditions). Then you need to worry about what your controller actually supports --- does it actually support a "hot" failure (not guaranteed, despite what they say- you need to test it), or does it require that you replace the drive and rebuild before proceeding? Does it support rebuilding on the fly (hot swap or hot spare)?

At work, raid 5 is necessary, but I'd never bother with that at home. For a gaming box.... like I said before, 2x in raid 0, plus a decent backup solution should be fine. Personally I don't do much net gaming, so the worst scenario for me is a couple of lost game saves (since last incremental backup) if there is a disaster.

Random other stuff - Raid 5 doesn't use a parity drive (raid 3/4 might), the parity is spread across all drives. There isn't a hit for parity writes *if* you're using a good controller (> $200?) that have all the bells and whistles (another topic).

I hope this has helped....

 

[Farva]

Raptor Raptor Raptor!

With the money invested in a RAID 5, just buy Raptor's!!!

If the 150 GB is too small, then Raptor in RAID 0!!

I'm not too happy with my setup I think I need to Raptor it myself ..

Either way RAID 1 will be faster anyways, your read speeds will be ub3r fast!

Raid 0 does not necessarily help for games. In fact, it can be quite detrimental to load times and other things.

Raid 1 is not faster than a single disk, nor is it faster than Raid 0. About the only thing I agree with your post is to use a Raptor 150GB.

 

[venm11]

raid 1 can be faster, depending on the controller; it can issue non-parallel reads in many situations. Not raid 0 fast, but it should reduce the bogging under a high-transaction load.

 

[unhappy_mage]

Raid 1 is not faster than a single disk, nor is it faster than Raid 0. About the only thing I agree with your post is to use a Raptor 150GB.

It depends what you're doing Many small reads, raid 1 can do substantially better than a single disk, or in some cases better than one 10k disk.

Theoretically RAID 5 read performance should be almost the same as RAID 0. The Parity disk is not read unless there is a drive failure. However there is a write penalty for 5 over RAID 0 becuse of writting the parity bit.

And with a good adapter, it is. However, latency is never helped by striped raid (that I can think of - but if you can name a counterexample I'll retract my statement), and for many uses that latency is the primary limitation.

What do you RAPTOR fan boyz think of that? 0+1 with more than double the storage at almost half the cost.

And what happens when you do a 65k write to your new 64k-striped array? It doesn't finish until all four disks are done seeking, that's what. The Raptor seeks once (and almost twice as fast as one of those disks, I might add!) and is done. 65k reads aren't as godawful; you have to wait for one disk on each side of the mirror to finish seeking. So you're only waiting on two disks.

No, while I'm not a huge fan of the Raptor, they have their market niche, and 7200 RPM raid 0 simply doesn't do the same things as Raptors. As for capacity - well, duh, buying 10k disks gets you less capacity than 7k disks. Amazing! It's almost as if they cost more to make or had higher quality than consumer drives...

 

[rodsfree]

Raid 0 does not necessarily help for games. In fact, it can be quite detrimental to load times and other things.

Raid 1 is not faster than a single disk, nor is it faster than Raid 0. About the only thing I agree with your post is to use a Raptor 150GB.

Acutally, it depends on the game.
2 Raptors in RAID 0 is the preferred drive setup for CAL in all of their FPS leagues.

The big money guys use 2 or more U320 SCSI 15k rpm Fujisitus in RAID 0 with really good 8X PCIe controllers

And Oracle, the HUGE database company, recommends a RAID 10 or RAID 0+1 setup for the computer running one of their databases. They like 10 better than 0+1 - it has to do with the way the mirrors are arranged, but not all controllers can do a 10.

And you should see the HDTach numbers on 4 Raptors in RAID 0.

 

[unhappy_mage]

Acutally, it depends on the game.
2 Raptors in RAID 0 is the preferred drive setup for CAL in all of their FPS leagues.

The big money guys use 2 or more U320 SCSI 15k rpm Fujisitus in RAID 0 with really good 8X PCIe controllers

But that doesn't mean it's efficient to do so 8x pci express is way, way overkill for 2 or 3 disks - one lane is 250 MB/s, and even those fast expensive disks only do 100 MB/s or so flat out sequential transfer. Then take into account that many of those cards are probably bridged designs from PCI-X to express, and one wonders what they're doing with those.... Looking at how poorly most games do in doing transfers in a logical order, they might be better off with a PCI-X controller in a PCI slot than the bridge. The loss of possible sequential transfer rate could be overcome by the lack of a bridge. I haven't tested it, but I bet the CAL folks haven't done double-blind tests either.

And Oracle, the HUGE database company, recommends a RAID 10 or RAID 0+1 setup for the computer running one of their databases. They like 10 better than 0+1 - it has to do with the way the mirrors are arranged, but not all controllers can do a 10.

Yep. But just because database systems work better (in general) on 0+1 doesn't mean anything else will. For example, applications with a lot of sequential writes (like video capturing - no guarantees on editing, though ) raid 5 can have better performance. Think about a 6-disk setup - with 0+1 or 10 you've got three copies of the same data, mirrored in two places. So when you capture video you get to stripe to three disks, and write it to two arrays. With raid 5, you'd be essentially striping to five disks instead of three, and calculating parity for the last disk. Since you're doing full-stripe writes, you don't have the raid 5 write penalty to worry about. So you can get a significantly higher write transfer rate out of 5 than 0+1, in controlled circumstances.

In short, my point is that storage performance isn't very intuitive. Adding more disks can help or hurt your application performance. All you can do is test a few ways and see what works best.

 

[rodsfree]

But that doesn't mean it's efficient to do so 8x pci express is way, way overkill for 2 or 3 disks - one lane is 250 MB/s, and even those fast expensive disks only do 100 MB/s or so flat out sequential transfer. Then take into account that many of those cards are probably bridged designs from PCI-X to express, and one wonders what they're doing with those.... Looking at how poorly most games do in doing transfers in a logical order, they might be better off with a PCI-X controller in a PCI slot than the bridge. The loss of possible sequential transfer rate could be overcome by the lack of a bridge. I haven't tested it, but I bet the CAL folks haven't done double-blind tests either.

Yep. But just because database systems work better (in general) on 0+1 doesn't mean anything else will. For example, applications with a lot of sequential writes (like video capturing - no guarantees on editing, though ) raid 5 can have better performance. Think about a 6-disk setup - with 0+1 or 10 you've got three copies of the same data, mirrored in two places. So when you capture video you get to stripe to three disks, and write it to two arrays. With raid 5, you'd be essentially striping to five disks instead of three, and calculating parity for the last disk. Since you're doing full-stripe writes, you don't have the raid 5 write penalty to worry about. So you can get a significantly higher write transfer rate out of 5 than 0+1, in controlled circumstances.

In short, my point is that storage performance isn't very intuitive. Adding more disks can help or hurt your application performance. All you can do is test a few ways and see what works best.

u_m,
Not disputing anything that you've said, but games, most db apps, and most video applications are focused on reads - not writes. Heck, even most OSes focus on reads.

And the problem with putting most PCI-X cards in a PCI slot is that most of them aren't notched to allow for the shorter slot. And using a dremal tool on a $300.00+ card is a quick way to lose a lot of money.

Most of the good PCIe, SCSI or SATA cards that I've seen are 8X (i.e Adaptec & Areca if I'm not mistaken), have you found any 4X cards that are any good?

And I agree that the CAL guys probably haven't done any serious testing.
But, they are all about frame rates and load times, and if anything can give them a perceived advantage - they'll do it. (Keyword = PERCEIVED)

 

[unhappy_mage]

u_m,
Not disputing anything that you've said, but games, most db apps, and most video applications are focused on reads - not writes. Heck, even most OSes focus on reads.

Yes, but that's by no means exclusive. Systems like DVRs for security purposes may do many, many more writes than reads during the life of the system.

And the problem with putting most PCI-X cards in a PCI slot is that most of them aren't notched to allow for the shorter slot. And using a dremal tool on a $300.00+ card is a quick way to lose a lot of money.

Really? The only cards I've seen with this misdesign recently are the 3ware cards.

Most of the good PCIe, SCSI or SATA cards that I've seen are 8X (i.e Adaptec & Areca if I'm not mistaken), have you found any 4X cards that are any good?

If all you're doing is a few ports with simple raid, these are pretty good. LSI has like 80% of the SAS market, and as far as I can tell it's for a reason.

And I agree that the CAL guys probably haven't done any serious testing.
But, they are all about frame rates and load times, and if anything can give them a perceived advantage - they'll do it. (Keyword = PERCEIVED)

Yeah. And it's probably not worth arguing with them; if they want to spend the money that's their business.