considering SSD raid for new SB server

[ThePiston]

I have a SB server which is 5 years old so I need to build a new one before the inevitable happens. I like to stay on the edge of technology usually so I really want this next build to use SSDs - if it makes sense.

The server is used for an SQL business app. We connect to the app using remote desktop and also the office uses the server for a basic file server. About 10 people use it at a time and usage is spotty of those people - so there's not really that much I/O. Nightly backups of the data and database are written to an external spinning drive.

So here's what I'm considering for new build:

For the SQL/data drive I'm considering using raid 5 SSD.

Someone please talk me out of this - there's seems to be no good reason for using SSDs in raid for a SB server except for an increase in speed and possibly reliability.

- SSDs are prone to die (especially if you use the multiple drives from the same batch in raid)
- they can corrupt data if power is lost in rid config (even 840Pro is susceptible)
- need a really expensive enterprise SSD if you are serious
- enterprise SSDs are super expensive

So can anyone think of a good reason why I shouldn't just get 3 reliable spinning drives in raid 5?

I might mirror raid the OS drive with consumer SSDs (proven reliable such as M4 or 830) since that would increase speeds of opening the app in RD.

Any input is appreciated.

 

[sub.mesa]

If you require random write-like performance, you should opt for RAID1 aka mirroring - not parity. You should pick SSDs with power-safe capacitors; i.e. Intel 320. You certainly don't want corruption on your SQL databases; which can seriously damage them.

 

[kbp]

^ +1 and a ups battery backup.

 

[Phynaz]

You should pick SSDs with power-safe capacitors; i.e. Intel 320.

You don't put desktop storage in servers. Didn't we already have this discussion?

OP - Stick to business class storage.

 

[Hellhammer]

You don't put desktop storage in servers. Didn't we already have this discussion?

OP - Stick to business class storage.

Our servers are powered by Intel's X-25M G2s. Besides, Intel SSD 320 isn't a consumer SSD anymore, even Intel lists it as a datacenter/IT drive nowadays.

OP, have you thought of getting an UPS to protect against power loss?

 

[ThePiston]

we have a 1500va UPS on this server which should last an hour without power.

I'm not sold on SSD just yet. I might just go with WD RE4s. I'll check out the 320. When I looked at enterprise SSDs the only real one I found was the Intel 710 which is $400 for 100GB. 4 of those in R5 would be $1600 for data storage alone. 4 WD RE4s would be $320 and they have 5 year warranty.

 

[ThePiston]

a lof of bad reviews for the Intel 320 - mostly bricked complaints: http://www.amazon.com/Intel-Series-2...owViewpoints=1

Intel 520 on the other hand has really good reviews: http://www.amazon.com/Intel-Solid-St...ords=intel+520

 

[Cerb]

For the SQL/data drive I'm considering using raid 5 SSD.

No. 1, or 10.

So can anyone think of a good reason why I shouldn't just get 3 reliable spinning drives in raid 5?

http://storagemojo.com/2007/09/19/cerns-data-corruption-research/
https://docs.google.com/viewer?a=v&...Rhapkj&sig=AHIEtbR5gqIJQzYXSBTJUV4rydWn-jXf8A

There's your start. Spinners or not, try to avoid RAID 5 whenever possible, these days.

I might mirror raid the OS drive with consumer SSDs (proven reliable such as M4 or 830) since that would increase speeds of opening the app in RD.

Provided it's not write-heavy, the M4 has a real use history in that scenario, and without any major problems, so with a properly set up UPS, that should work well.

a lof of bad reviews for the Intel 320 - mostly bricked complaints: http://www.amazon.com/Intel-Series-2...owViewpoints=1

Intel 520 on the other hand has really good reviews: http://www.amazon.com/Intel-Solid-St...ords=intel+520

Not in RAID w/o TRIM, it doesn't. You can easily find Linux users with unbearable performance drops over time due to SF controller (in large part because software RAID TRIM has been slow in gaining support, while it would fix their problem). I wish they'd just kept the 510 around, and tuned its random performance a bit more. The Intel 320 and Crucial M4 have been successfully used in servers with light writing duty for some time now.

 

[Phynaz]

Our servers are powered by Intel's X-25M G2s.

And what current generation consumer SSD has the same reliability as those?

 

[Hellhammer]

And what current generation consumer SSD has the same reliability as those?

Since you brought this up, do you have any statistics showing that the X-25M is more reliable than the current generation consumer-grade SSDs?

Reliability is, and has always been, a moot point because there are no sufficient statistics to show long-term reliability.

 

[dave_the_nerd]

I have a SB server which is 5 years old so I need to build a new one before the inevitable happens. I like to stay on the edge of technology usually so I really want this next build to use SSDs - if it makes sense.

The server is used for an SQL business app. We connect to the app using remote desktop and also the office uses the server for a basic file server. About 10 people use it at a time and usage is spotty of those people - so there's not really that much I/O. Nightly backups of the data and database are written to an external spinning drive.

So here's what I'm considering for new build:

For the SQL/data drive I'm considering using raid 5 SSD.

Someone please talk me out of this - there's seems to be no good reason for using SSDs in raid for a SB server except for an increase in speed and possibly reliability.

- SSDs are prone to die (especially if you use the multiple drives from the same batch in raid)
- they can corrupt data if power is lost in rid config (even 840Pro is susceptible)
- need a really expensive enterprise SSD if you are serious
- enterprise SSDs are super expensive

So can anyone think of a good reason why I shouldn't just get 3 reliable spinning drives in raid 5?

I might mirror raid the OS drive with consumer SSDs (proven reliable such as M4 or 830) since that would increase speeds of opening the app in RD.

Any input is appreciated.

If you don't need the extra performance of an SSD (no complaints about current setup) then don't get an SSD.

But go with RAID 1 or 10, as has been said.

Hell, if it's SQL - shouldn't the table be cached into RAM most of the time anyway? (Or is it too big for that?)

You didn't describe the nature of the work done, how important it is, and how much DT costs you, but if it's not a super-critical thing and stoage cost is a factor, could you consider a single drive? I know that's blasphemy in server circles, but a single SSD will be embarrassingly faster than even four platters in RAID-10.

 

[ThePiston]

this is a small medical office. DT has to be minimized, but cost is also a factor. We use an EMR on this server.

I do not mind staying with R10 spinning drives, but I only like to build the server for the office every 4-5 years since it eats up my weekend and you never really know what type of issues may pop up.

I have time to wait since I'll be installing Server 2012 on this server and the program doesn't even support it yet. Soon, they say.

The speed jump from SSD will be so dramatic though I might just do it. I think a lot of people want to do same thing so I'll report back if I do.

So, what do you think - should I just Raid 10 the OS + SQL/Data all on one drive? Maybe partition out the data and SQL database onto their own partitions?

I do not think the I/O is so dramatic in our situation to require dedicated drives with the kind of throughput raid 10 SSDs will provide.

 

[dave_the_nerd]

No reason to RAID-10 SSDs. RAID-1 is adequate.

If you want SSDs for OS/Applications, that's fine. Those should be mirrored too, though.

That leaves the question of what you store your DB on. RAID-10 HDDs or RAID-1 SSDs would be the 2 options I'd lean towards.

(Partitioning the same physical array is possible, but it's not common/best practice.)

What's your budget for storage? This may be a moot point. Four enterprise-grade SSDs (RAID-1 OS + RAID-1 data) would start around $2k. Six SAS spinners (RAID-1 OS + RAID-10 data) would be more like $800.

$1200 spent on MOAR RAM and a (cheap, disposable) SSD cache for the OS/Applications would probably be money better spent.

SQL should load as much as possible of the DB into RAM, making disk I/O less important, and Intel's SRT doesn't lose data if the SSD fails, so it wouldn't be an availability issue if the drive died. (It would just be slower.)

The frequently-accessed applications would be cached to the SSD first and give you most of the speed boost you're looking for.

 

[Cerb]

Small office EMR? Hmmmm. If it's anything like eClinicalWorks, disk IO will be negligible, after the initial setup and data import, except during upgrades and such.

 

[Phynaz]

Since you brought this up, do you have any statistics showing that the X-25M is more reliable than the current generation consumer-grade SSDs?

Reliability is, and has always been, a moot point because there are no sufficient statistics to show long-term reliability.

Outside of anecdotal evidence? No.

But I think everyone agrees that the G2's are very reliable.

Which website is it that does the write endurance testing? I recall their G2's still writing after multiple petabytes, where current drives aren't getting near that.

 

[Zxian]

Outside of anecdotal evidence? No.

But I think everyone agrees that the G2's are very reliable.

Which website is it that does the write endurance testing? I recall their G2's still writing after multiple petabytes, where current drives aren't getting near that.

Write endurance is not directly correlated to reliability. I have heard of several SSDs dying before the apparent maximum as tested by those on xtremesystems. Return rates are also not failure rates. Current SSDs have not been on the market long enough and do not have enough of a userbase to determine a valid reliability metric.

 

[Phynaz]

Current SSDs have not been on the market long enough and do not have enough of a userbase to determine a valid reliability metric.

Which would make it a dumb idea to put them in a server, don't you agree?

 

[Zxian]

Which would make it a dumb idea to put them in a server, don't you agree?

Not at all. RAID is intended to ensure uptime, which is the key factor we're asking about here. RAID1 and RAID10 provide sufficient redundancy to account for a single drive failure (multiples in RAID10 if you're lucky with with which drives fail). The performance benefit of SSDs is self explainitory.

New hard drives are also unknowns when it comes to reliability. Why do you choose the brand you do in your servers? It's usually because their previous reliability records are good.

 

[ThePiston]

just looking at the prices on 15K SAS drives. I'd need 4 of them in Raid 10 which would be about $1400. It would still be way slower than mirror SSDs.

I'm going to give the mirror SSDs a shot. I'll use Intel 520s. 2 mirrored for OS and 2 mirrored for SQL/data. Less power, heat, cheaper than SAS, and faster. Good backups and good UPS.

 

[sub.mesa]

The Intel 520 uses a Sandforce controller, which is usually not equipped with power-safe capacitors. Since you will be using the SSD for a medical facility, I would argue that you would want the most reliable SSD series available for an affordable price.

The Intel 320 is the only affordable consumer-grade SSD that is inherently reliable. All other SSDs can corrupt themselves on unexpected power loss. Your alternative would be enterprise-grade SSDs which are far more expensive. The Intel 320 should be the perfect choice for your job, in my opinion.

The Intel 320 had a firmware bug in the past - popularly known as the '8MB bug'. However, the 8MB is not a bug; it is the result of corruption that occurred. The bug was that the power-safe capacitors in the Intel 320 that should have protected against corruption, didn't work properly in relation to the firmware. After Intel fixed this problem by issuing newer firmware, its failure rate should be lower than the Intel X25-M series, which is seen as a traditionally reliable SSD. The Intel 320 uses the exact same controller as the X25-M does, but with different firmware, newer NAND memory and power-safe capacitors to protect against corruption. Furthermore, the Intel 320 uses RAID4 on the NAND chips to protect against bit rot, commonly known as bad sectors. Despite it having lower-quality 25nm NAND versus the 34/55nm NAND used in the X25-M, the Intel 320 with modern firmware should be the most reliable consumer-grade SSD on the market.

So when people refer to the '8MB' bug, I would argue that all other consumer SSDs which are unprotected would have this 'bug' - the only exception being Intel 320 with modern firmware. Another contender is the Crucial M500, but it is not yet released and it uses a fairly new controller and firmware which still needs to be proven. So for multiple reasons, I think the Intel 320 should deserve your attention as a serious option. Use two in RAID1 if you like.

 

[Cerb]

Which would make it a dumb idea to put them in a server, don't you agree?

The same can be said for HDDs, to lesser extent. HDDs with data-corrupting firmware bugs are not unheard-of. The WD REs had such a bug, several years back, FI. Hardware moves too fast to let others beta test for more than a year or two. Better to be slightly behind the curve, and have crash plans.

The Intel 520 uses a Sandforce controller, which is usually not equipped with power-safe capacitors.

It can also get dog-slow over time, with no TRIM to pep it back up. I doubt it would malfunction in a way that would hose anything up. Neither 520 nor 330 drives have the spotty behavioral history of most other SF drives, that rely on SF for their firmwares. But, I wouldn't want to take the chance of performance issues down the road, when Crucial's offering is well-priced, performs well, and has a solid history. Intel's 320 would likely be great, too, but the price never came down.

 

[ThePiston]

I thought the 520 did have those capacitors, but apparently not. This is what is says on the Intel site:

Additionally, the Intel SSD 520 Series contains “End-to-End Data Protection” ensuring integrity of stored data from the computer to the SSD and back.

Not sure what that means but it sounds like it would work just as the capacitors would work, no?

 

[sub.mesa]

No. End to end data integrity protection means that once the harddrive reads something, it guarantees that in between sending it to the host, the data will not be corrupted. If it is, it will read again before sending it back to the host. Harddrives have the same feature. Its fairly simple, but it sounds good for the marketing folks. They always concentrate on things that matter the least, to distract you from the awful truth:

The only consumer-grade SSDs that are safe, are Intel 320 and soon the new Crucial M500.

 

[ThePiston]

ok, then I'll use the 320s for the database. I will probably use mirrored M4s for OS. Now I need to find a server mobo that will support trim in raid.

 

[sub.mesa]

Simple consumer Intel chipsets support TRIM in RAID. Nevertheless, you should apply overprovisioning by only partitioning the SSDs up to 75% of their maximum capacity. If you pick 50%, you won't need TRIM altogether. But 75%+TRIM is excellent.

I'm not certain whether the TRIM feature works on older chipsets - they may have build a 'political' protection in the drivers - but it should work on Ivy Bridge Z77 chipsets and others that support RAID.