Slashdot Mirror
MS Researchers Call Moving Server Storage To SSDs a Bad Idea
An anonymous reader writes "As an IT administrator did you ever think of replacing disks by SSDs? Or using SSDs as an intermediate caching layer? A recent paper by Microsoft researchers provides detailed cost/benefit analysis for several real workloads. The conclusion is that, for a range of typical enterprise workloads, using SSDs makes no sense in the short to medium future. Their price needs to decrease by 3-3000 times for them to make sense. Note that this paper has nothing to do with laptop workloads, for which SSDs probably make more sense (due to SSDs' ruggedness)."
News at 11!
This is an ACM article behind a paywall.
How about a slashdot policy of not linking to articles behind paywalls?
Test your net with Netalyzr
Hm. I was thinking the same thing about the ACM subscription.
I'm a 2000 man.
SSD is already cheaper per gig than some SAS drives. Also, 3-3000 times? What the hell sort of estimate is that?
they don't use NTFS?
One of our competitors trademarked the term "hypothesis". From now on, we will call them "boneheaded ideas".
they don't use NTFS?
I cast thee OUT!
It's called "pulling numbers out of your ass".
My goodness! They have really done their research in order to produce data as accurate as that!
The fact was, they said the same thing when it came to magnetic tape versus magnetic disks. These days, hard drives are cheaper than tapes and will hold their data longer and more compatibly.
Microsoft fears change that they do not control. If they don't control the changes, someone might write them out of the story.
Geez, that was hard. Also has a PDF version if you don't like the colors.
I bet this has absolutely nothing to do with the fact that Apple updated their Xserve yesterday with an SSD option.
From what I understand, the SSD is for the OS itself and not the data/storage.
What do you mean, an african or european ass?
This report brought to you by the same people that said Linux was a more expensive platform than Windows. Seriously, why do we even listen to these nutbags?
seriously? "we don't have enough people here. we need between 2-2000 times as many people in the configuration department." Does that sound like I have ANY idea how many people we need?
Sorry, that is a *ridiculous* range to give.
I can only guess they're referring to differently priced SSDs. Some cost in the thousands, but provide top-teir performance. Their price would be justified at approximately 1/3rd the current price, as that's what would be necessary to provide similar cost/performance to a raid array of rotational drives.
On the other hand, the low cost MLC ssds typically provide lower performance than a single rotational drive at a cost premium in the range of about 100x the cost of a rotational drive. These lower cost drives are frequently seen as needing an improvement to capacity to be worth considering, you see them in 64gb sizes and the like. They would have to be available at a price lower than rotational drives in similar capacities to be considered, which would a 100+ scale price reduction.
I'm not seeing the 3000x reduction necessary, I guess a small exaggeration for effect may be the source of that.
"And we have seen and do testify that the Father sent the Son to be the Savior of the World"
1 John 4:14
SSD gives phenomenal random read performance, equally good serial read performance, and average write and random write performance (at least if you get a good SSD, the low end ones using the crap IO chips are worse than budget HDs). The only way to beat the read performance of a good SSD is a really expensive SAS RAID, and even then it's not going to be by much. Yes you can take a hit on serial write performance, but not much of one (it's on par with most medium to high end HDs, with surprisingly few high end HDs able to outperform it). If you're primarily going to be doing reads, particular random reads, or even if you're going to be doing mostly random writes rather than serial writes, an SSD is probably a good idea.
Curiosity was framed, Ignorance killed the cat.
Of course, SAS drives are also often too expensive to survive a purely cost/benefit driven analysis. For many real-world loads you're better off adding more spindles which can give you similar iops per dollar but with the added benefit of vastly more storage space.
There's a lot of snake oil and very little quality analysis in enterprise storage these days, so it's good to see at least some do attempt to do actual real-world cost/benefit calculations before jumping onto the marketing train.
It depends on the application which to select right now, but in the long run the SSD drives will have an advantage.
So soon we may no longer need those noisy hard disks at all.
And when a storage is built on flash memories it may be possible to work with it in segments where parts of the disk isn't powered in order to save power and generate less heat. The latter is a huge advantage in datacenters where cooling is expensive.
The ruggedness is also an advantage, but not in datacenters. What you usually want in a datacenter is good performance to a low cost, but you also want reliability.
If builders built buildings the way programmers wrote programs, then the first woodpecker would destroy civilization.
I lolled. Thanks.
That wasn't hard.
Price point depends on the server workload pattern. Non-paywalled article http://research.microsoft.com/pubs/76522/tr-2008-169.pdf
What a misleading term - I know of companies using Enterprise SSD in production precisely because it's financially sound for them to utilise the ridiculous speed improvement it provides.
Sure, it's not a lot of companies that are using this yet, but as longevity increases with better garbage collection and write-spreading algorithms as well as stabilty and feature set through maturing software and firmware it's closer than you think.
For clarity, the product wasn't SSD behind SATAII, it was FusionIO's PCI devices.
No statement is true, not even this one.
Not ones you'd use in an enterprise! X-25e is the only SLC based flash with a decent controller under $1k and it's still $24/GB. Unless you have a WORM application that needs fast seeks (pretty rare) MLC based flash isn't a good fit for most enterprise applications. The only areas we've found for them are log drives for high transation database servers where the insane IOPS per $ make sense and cache for a BI system which still sees enough writes to rule out MLC. Oh and their analysis is based on a rather small set of data, I'm at a midsized shop and I have more storage online and a more varied workload than that.
There are 4 boxes to use in the defense of liberty: soap, ballot, jury, ammo. Use in that order. Starting now.
For many real-world loads you're better off adding more spindles which can give you similar iops per dollar but with the added benefit of vastly more storage space.
I'm not going to dispute the half of adding more spindles for better performance, but additional storage space is not always an added benefit. Do you have any idea how frustrating it can be to explain to users that "yes, we have 4TB of unused space" and "no, you can't put another database there without killing your existing one"?
I'd say Microsoft does not have any "solution" in which SSD would help.
OTOH Sun does ... http://www.sun.com/storage/flash/. YMMV.
Microsoft researchers provides detailed cost/benefit analysis for several real workloads.
If Microsoft researchers report that SSD's are not cost effective storage, it means that Microsoft is not getting any revenue from SSD storage. Or that they're behind on incorporating SSD's into the server stack. Or they caught blind-sided by the trend like they did with netbooks and are now scrambling to explain why they didn't see it coming. Oh, we found that wasn't cost effective, so we didn't incorporate it.
I really miss the days Microsoft had it together. There was a time they were great to work with. Now they seem like the Three Stooges Do IT. SSD, eh? Oh, a wise guy! SMACK! Wo-wo-wo-wo!
That's our life, the big wheel of shit. - The Fat Man, Blue Tango Salvage
What is this, an SSD for ants???
Whale
Of course, SAS drives are also often too expensive to survive a purely cost/benefit driven analysis. For many real-world loads you're better off adding more spindles which can give you similar iops per dollar but with the added benefit of vastly more storage space.
You need 2-3x as many SATA drives as 15k SAS/FC drives to get equivalent IOPS. That means 2-3x as much physical space required, probably around 1.5-2x as power usage and decreased reliability overall.
Storage volume is rarely a concern when the primary objective is performance.
I don't know!
waaarrggghhhh
Knowledge is power. Knowledge shared is power lost.
First, they don't say 3-3000 percent (nor do they say 3-3000 times, which is what the original post above says). They say 1-3 orders of magnitude.
re: tape vs. disk:
cheaper per byte, sure.
more compatibly, possibly. i'm not sure whether you'd have an easier time reading a 1985 disk or a 1985 tape; in either case you'd need to do some digging to find the appropriate hardware to read the media.
longer? Nope. That 1985 disk might be readable by ontrack in the lab, but odds are pretty good that it won't work right once you dig up that MFM controller. The tape will probably be fine.
re: TFA:
If you RTFA they do some reasonable analysis.
They ignore the possibility that you might drop the RAID1 on your boot disk and go with a single SSD; I'm certainly considering that and if you think hardware raid controller + a pair of disks vs. 1 SSD the cost is very favorable.
They don't appear to take read latency into account; IOPS and latency are not the same thing.
Also, prices have already fallen significantly since their published data. They list a 32G SSD at $739; the same one is $449 today at http://rocketdisk.com/index.php?cPath=8&gclid=CJbK8OH14ZkCFSQeDQodaikyXA
Windows 2020 will have the same features as Open Solaris 10, just wait and see. They will be able to use a SSD as a cache reader I swear!
They could call it... ReadyBoost.
High-end HDDs still edge out SSDs for serial reads in many setups.
Keep in mind that write performance degrades over time (goes from great to very good) as the pages get full.
When you're out of free pages, you have to read an entire block of pages to cache, erase the entire block, then write back the new block.
Current OSs and controllers do not yet support the "yes, actually delete it" command, and current controllers do not yet support any sort of automatic drive-level page consolidation.
If money is no object, then SSD is a good choice. A better choice is a RAM drive, though prices for these make SSDs look like a bargain.
$-for-$ there is no contest. HDDs win. I think they've got about 2 years left before they start to be marginalized to bulk storage devices.
MS isn't making any observations or claims that aren't already well understood by the corporate IT world wide. You don't adopt new technology into a production environment until 1) its failure modes are infrequent, well understood, and able to be mitigated and 2) it is cost effective and/or necessary from a functional or business standpoint to do so. Last time I checked, SSDs were significantly more expensive and significantly lower capacity than conventional HDDs.
They list the write IOPS of their "Enterprise SSD" drive as only ~350. That number seems like it's an order of magnitude too low, which would obviously skew the conclusions.
http://blogs.sun.com/brendan/entry/test has some background information, and http://blogs.sun.com/brendan/entry/l2arc_screenshots and http://blogs.sun.com/brendan/entry/my_sun_storage_7410_perf has some performance numbers.
Basically, what Sun is claiming is that by adding a SSD cache layer you can improve IOPS by about 5x, for what amounts to a really small amount of money for say a 100tb system. This is being marketed quite heavily by Sun as well. (The numbers look convincing, and the prices for the Sun Storage servers are certainly very competitive, well, compared to say NetApp.)
IMHO this is just a repeat of the well known Microsoft tactic of spreading massive amounts of FUD about any competing technology that you can't reproduce yourself - you'll have to wait until Windows Server 2013 for this.
More like 13$/GB. HTH! HAND!
http://www.provantage.com/intel-ssdsa2sh064g101~7ITE90J5.htm
X25E SLC 64GB 2.5INCH SATA SSD $827
And Provantage is rarely a price-leader.
-- Have you ever imagined a world with no hypothetical situations?
Dismissing using SSD because it's only cost effective for the boot partition is a mistake. Anyone who's put together servers before knows the boot partition is critical to the system, and the hardest part to backup. Once you get a system booted, there's a million things you can do to fix it or restore the relevant data. Getting it bootable if the boot partition is toast is much harder.
AccountKiller
Since this paper is focused on solid-state storage, and wear is a novel, SSD-specific phenomenon, we include it in our device models. Currently we do not model other failures, such as mechanical failures in disks.
The correct approach to incomplete data is, of course, to gather complete data, and they have no excuse here, because there is PLENTY of data on mechanical drive failure rates. However, if you are not willing to do that, the least you can do is ignore the data equally on both sides. The authors' failure to treat both sides equally leads to a hopelessly biased and skewed analysis.
I'd say Microsoft does not have any "solution" in which SSD would help.
ReadyBoost in Vista is doing basically the same thing (at least for reads). So the foundation is already there (and has been for a couple of years now).
SSD is already cheaper per gig than some SAS drives. Also, 3-3000 times? What the hell sort of estimate is that?
I remember an article a while ago talking about how Windows disk drivers are not optimized for SSD. Now there is a white paper showing how SSD is not practical by Microsoft. So to answer your question, it is a PR estimate.
I just finished the reading the paper.
The paper boils down to this:
SSD disk when measured against IOPS, Watts, and Capacity in relation to cost based on several different server types is not cost effective yet. Depending on the type of server costs need to come down at least 3 fold, and under some scenarios as much as 3000 times. Hosting MP3s that are largely sequental, low write storage SSDs are 3000 times over priced. For insaine random IO scenarios that need to come down 3 fold to make it worth it compared to conventional drives.
Depending on the type of server they can perform worse then standard mechanical disks.
They found no advantage to 15k RPM drives versus 10k RPM drives when cost is factored in.
SSD drives pay for themselves in power saving in about 5 years, well past their expected longevity.
Mechanical disks wear out more or less independant of their data load, SSDs wear out proportional to their data load.
SSDs do not handle tiny files very well due to how data is written.
I see nothing in the paper that is pro-microsoft, rather straight dealing on the drives themselves.
I would suggest MOD-TROLL any evanglest on any side of the OS wars as this paper doesn't seem to deal with OS touting.
It was a boring but informative read.
-=[ Who Is John Galt? ]=-
I can only guess they're referring to differently priced SSDs. Some cost in the thousands, but provide top-teir performance. Their price would be justified at approximately 1/3rd the current price, as that's what would be necessary to provide similar cost/performance to a raid array of rotational drives.
The interesting thing is, according to the performance table on page 6, the SSD they used only had write performance of ~350 IOPS. Either that number is missing a zero, or something is _seriously_ wrong with their SSD.
If it's the latter, then clearly any conclusions drawn from that write performance are completely invalid.
If you're primarily going to be doing reads, particular random reads, or even if you're going to be doing mostly random writes rather than serial writes, an SSD is probably a good idea.
Which, as you undoubtedly know, is why enterprise data centers use SANs (Storage Attached Networks) with multiple tiers of storage. You use big slow drives for archival storage (old emails, for instance), and smaller faster drives for day-to-day use (databases, etc). Flash drives get used when performance really matters, such as database indexes, not the actual data.
Nothing for 6-digit uids?
Cool, when I bought mine the 64GB wasn't available and the 32GB was almost that much. That's why SSD's are so cool right now, even at the high end the $/GB is falling rapidly, much more so then enterprise HDD's. 450GB 15K FC drives are about three times that much so about 1/3rds the $/GB but MUCH higher $/IOP.
There are 4 boxes to use in the defense of liberty: soap, ballot, jury, ammo. Use in that order. Starting now.
I'm not going to dispute the half of adding more spindles for better performance, but additional storage space is not always an added benefit. Do you have any idea how frustrating it can be to explain to users that "yes, we have 4TB of unused space" and "no, you can't put another database there without killing your existing one"?
That's called short stroking. No, really, you only use the innermost tracks of your drives to avoid long seeks, turning a 144 GB drive into 72 or even 36 GB. If you do it right, the unused capacity doesn't show up on any of your reports, so no one ever realizes that the extra space is there.
Nothing for 6-digit uids?
Have you ever heard of thing called RAID? It is actually designed for inexpensive crappy of the shelf disks.
More disks means less reliability due to more points of failure. A 16-drive array is more likely to have a failure than a 6-drive array by virtue of simple statistics.
And when a storage is built on flash memories it may be possible to work with it in segments where parts of the disk isn't powered in order to save power and generate less heat.
Flash drives all do this right now automatically.
Unless the controller is actively accessing a particular block, there is no power to that part of the flash, as there is no need to power it in any way. This is the underlying concept behind all non-volatile memory (like flash)...it doesn't have to be continuously powered to maintain its state.
I wonder how that donkey feels working for MS.
Actually the write performance degradation problem is a bit more involved than that, but the way you describe it is essentially correct. There are some applications out there that can consolidate an SSD to get back to the original performance (think of it as SSD defrag), but having support for the new erase command in both hardware and the OS will ultimately be the best solution. I've heard rumor that Windows 7 is going to have support for it when it ships, and I assume Linux either has, or will have soon support for it as well. Interestingly enough, attempting to defragment a SSD drive using traditional utilities will only make the problem worse.
As for the $-for-$ argument, if you're talking low to mid range performance, then yes, once you start talking high performance SSD edges out HDD because you start having to do all kinds of expensive RAID setups in order to match performance. Now, in terms of dollar per gigabyte costs, yes HDD is still going to smoke SSD for a few more years.
Curiosity was framed, Ignorance killed the cat.
For read data, it makes more economic sense to cache to RAM instead of SSD and just read everything into RAM at startup. Fpr writes, I'm not that sure -- is a write to SSD really that much faster than a write to disk? It might make sense to use SSD for journaling in those cases where a transaction can't complete until you are certain the results have been saved. But in that case, your network latencies are probably much greater than you disk write latency anyway.
I've abandoned my search for truth; now I'm just looking for some useful delusions.
The chance of a single disk failing is irrelevant when the array is capable of handling disk failures without failing as a whole. I was actually kidding when I asked if you hadn't heard of RAID, but you obviously haven't; so read up on it, it will explain the details for you.
I can only guess they're referring to differently priced SSDs. Some cost in the thousands, but provide top-teir performance. Their price would be justified at approximately 1/3rd the current price, as that's what would be necessary to provide similar cost/performance to a raid array of rotational drives.
The interesting thing is, according to the performance table on page 6, the SSD they used only had write performance of ~350 IOPS. Either that number is missing a zero, or something is _seriously_ wrong with their SSD.
If it's the latter, then clearly any conclusions drawn from that write performance are completely invalid.
No 350 IOPS is pretty standard for SSD in real world conditions. The problem is manufacturers can claim speeds of 45,000 IOPS which while technically true, those speeds due not hold up in typical conditions. I don't fully understand the reasons but feel free to research my claims.
The chance of a single disk failing is irrelevant when the array is capable of handling disk failures without failing as a whole.
No, it's not.
I was actually kidding when I asked if you hadn't heard of RAID, but you obviously haven't; so read up on it, it will explain the details for you.
I've been working with RAID for ~15 years now. I have a rough idea about how it works. You, OTOH, seem to think an array that has lost a disk suffers no change in ongoing reliability or performance, which strongly suggests that *you* are the one who needs to do the reading.
First, they don't say 3-3000 percent (nor do they say 3-3000 times, which is what the original post above says). They say 1-3 orders of magnitude.
Also, prices have already fallen significantly since their published data. They list a 32G SSD at $739; the same one is $449 today at http://rocketdisk.com/index.php?cPath=8&gclid=CJbK8OH14ZkCFSQeDQodaikyXA
I'm not sure how a 3 order of magnitude drop could possible happen, as that would imply a 32GB drive at $0.44, which is 1.375 cents/GB.
At that price, a 1TB drive would be just about the same price as lunch for two at a fast food place. Two orders of magnitude is just barely possible, but by then, I think spinning platter hard drives will be given away with Happy Meals.
Especially if they're EXT3rnal...
They could grip it by the husk!
A slashdotter who didn't build his own computer is like a Jedi who didn't build his own lightsaber.
Microsoft Windows, and their vendor have a lot of work to do to make their offerings compatible for SSD.
This appears to be the same thing that Intel did, there is no advantage to 64-bit--until we do it.
https://www.youtube.com/c/BrendaEM
No 350 IOPS is pretty standard for SSD in real world conditions.
Intel specs their X25-E at 3300 IOPS for random 4K writes. I'm willing to consider there might be a bit of fudge factor in that number (although all the benchmarks I've seen suggest it is conservative, if anything), but certainly not an order of magnitude.
> I'm not sure how a 3 order of magnitude drop could possible happen, as that would imply a 32GB drive at $0.44, which is 1.375 cents/GB.
I'd rather say that would mean a 32TB drive for $449 (or more likely 3TB for $50). Not so unplausible, considering that the 20MB HDD my dad once bought cost five times the 1TB drive I bought last week. That's about 5 orders of magnitude difference for the cost per capacity, and in only 20 years or so.
but additional storage space is not always an added benefit
True, you need a matching use case for it too. However, in my experience there's a huge pent up demand for cheap low-performance storage in enterprise environments (further indicated by in-vogue pitches about cost reduction through 'deduplication'), so while putting another database on the spindles with free space is a no-no, you might be able to match the space up with backups, archived user data or similar low priority data that won't interfere with peak load performance.
It may not be trivial to accomplish such pooling with storage vendor tools, but hey, the whole sales pitch for SANs was that it was more manageable, and supposed to somehow be cost effective (which is obviously not what's happening when companies are worrying about storage cost for power point presentations and email archives at a time when you can get storage enough for every bit of data most employees generate over a lifetime for less than the cost of their office chair).
HP has this technology in a Mezzanine adapter for blades. It's called "IO Accelerator".
Help stamp out iliturcy.
A well designed RAID in a robust SAN can survive not just the death of a drive but often the death of an entire enclosure (10-16 drives depending on age and enclosure design). Most of the time a small enterprise class SAN has 8-12 enclosures worth of drives. Big ones can span half a dozen or more racks. I don't think this article is talking about a couple drives thrown into a box with a hardware RAID controller here. When a player like Microsoft starts talking about "storage" they are talking 100 TB or more. Last place I worked had ONLY 25 TB of storage, made up of older storage tech that only gave us 300 GB FC drives. We had 8x14 disk enclosures and could loose and entire enclosure without data loss. The disks were striped in such a way as to ensure that none of our RIAD5s had more than one disk in any one enclosure, and 4 spares made sure that up to 4 disks could die before we even had a chance of any long term performance issues. If you're really paranoid you can build a RAID5+1 to make sure that up to two drives per RAID could die without data loss. I've heard of, but not seen companies so paranoid that they use RAID5+2.
The storage system at my current place is even fancier and dynamically handles the RAIDs. We've got about 100TB spread across two racks worth of enclosures and any 20 or so disks could die at one time before we lost data.
I don't need a million points of light, just two points of multi-mode fiber and a 10 Gig-E router.
Outer sides of the disk. The disk is spinning faster on the outsides.
I spazzed and accidentally marked it +funny ;-P
I'm very happy that Bill's company is so worried about the little money they have left us after all those years, but for my company I'm considering an Intel X25-E for my new server, hoping that the MySQL database will run faster. Currently that is on a Raid 1 disk. I like speed, but don't want to lose my data in the first place. So, I'm not that interested in cost. I'm interested in whether my data stays OK, whether I still should put those SSD in a Raid 1 configuration, and whether it will work OK for a couple of years (even with small files). (Mac OS X will continue to run from a regular HD). Experiences, anyone?
Bert
How many terabytes to these "tapes" hold? The Wikipedia article must be really old, because they're talking about storage of 80GB. That's like not even two high def movies. If I can get 60 TB on a tape that would be sweet. I don't think it would be practical at less than 1TB. After all, does it matter how long the data stays on the tape if it won't fit in the first place?
Help stamp out iliturcy.
The chance of a single disk failing is irrelevant when the array is capable of handling disk failures without failing as a whole.
umm no it isn't.
If your talking about a raid 5 "Not what you would use for a database server" then once you loose that drive your in big trouble. You are running with no back-up and x time the chance of a drive falure taking out the entire array until the rebuild is done. That is if you have a hot spare in the system.
If you are running RAID 10 then which is a far better solution your better off but still not a happy camper.
What most people don't understand is that all too often all the drives in a raid are from the same litter. You tend to buy them all at the same time from the same source and they tend to all be from the same batch and usually are the age when one dies.
I have seen way too many raids start to have one drive after another die once they start having failures.
See my blog http://ilovecookes.blogspot.com/ for light hearted technical information.
You need 2-3x as many SATA drives
Yep, the more extensive calculations I've seen come up with approximately equivalent prices per IOPS when all the factors you mention are accounted for. At least one of which then extended the analysis into really bizarre arguments about watt per IOPS being of supreme importance while ignoring the 3-5 times as large storage volume.
You're right that volume isn't a concern in the cases when performance is the objective, but in my experience (and yours may differ), in consolidated storage environments the complaints about low performance storage being expensive outnumber the complaints about high-performance storage being slow by several times. Adding massive amounts of very low-cost but near zero IOPS storage would reduce problems and costs associated with everything from employees storing data on USB drives to the man hours spent sorting out data and mail that they're certain they can throw away. So if you can triple the volume by doing a near-zero cost benefit substitution of fast disks for more spindles with equivalent IOPS, you can get a significant benefit in another segment of storage.
One place I worked at did something slightly fancier a couple years ago before RAID 6 went into common use:
They had three hardware RAID enclosures which did the RAID calculations on the enclosure and appeared to the host as a single hard disk volume (each had a number of drives including one as a hot spare) plugged into a midrange UNIX server via fiber channel. Then the three volumes from the enclosures were configured as a RAID volume in software using Symantec's volume manager (formerly Veritas). This way if a drive fails while one enclosure is in degraded mode, the data is still accessible. It would take multiple controller failures, multiple drive failures in different enclosures, a combination of the above, or a sitewide disaster to lose the data contained.
With RAID 6, this is less of an issue because it would take three drive failures in the time it takes for one drive to rebuild to kill the stored data, but at the time RAID-ing the RAID arrays seemed to provide some decent redundancy.
This page has a price history graph for the 32GB model (scroll down a bit). The price dropped quite a bit a couple months after the introduction. Hopefully the 64GB model will follow the same pattern (pushing down the 32GB model along with it ... $200 range maybe?).
I was thinking along the same lines; when I converted my laptop to an SSD I found that I had to do a number of things to my Vista install to get it to match my Mac OS X install, stutters, pauses and other problems continue to plague the windows boot on this machine while the OS X side just runs (and fast) the biggest gains I saw with this weren't in the sustained MB/s or IOPs, it was its ability to service multiple IO streams with none of the effects I typically would see with a single spinning disk, (i.e. running multiple VM's with multiple virtual HD's) I can see from a storage perspective why MS might claim this is a non-starter, but from a server perspective (laptop, desktop, server) using them as a boot device and application device is brilliant. What I take from this is a concession that SSD's don't work so well as a M$ boot, and don't make a lot of sense for storage that would typically be served via a large disk array with very large ram caches.
Unix, an obscure operating system developed by bored researchers in an attempt to get a better game playing experience.
One thing about this research paper is that they used only one model MemoRight GT MR25.2 in 8/16/32 GB capacities to do their testing before 2008-11-11 publication of the paper in the United Kingdom.
I'm concerned that the research test and results are largely skewed against SSDs because they used only that one model to do all their testing with based on only one price point for the SSDs.
There is a very large difference in performance between many various SSD drives based on the original flawed JMicron JMF602 chipset (stuttering/freezing on write), newer JMF602B (smaller stuttering), Samsung's chipset, Intel's chipset (fastest random writes by 4x), and the newest Indilnix Barefoot chipset (balanced sequential/random read/write). Additionally the huge drops in prices in the last 6-12 months ($1,500->$400) is a big change in the SSD arena. These price, capacity, and performance changes are going to continue fluctuating for the next few years yielding much better drives for the consumers.
I believe that the research in the paper will be shortly obsolete, if it isn't already, given the latest products on the market and price points and the Q3/Q4 new upcoming products from Intel and others.
I'm helping a friend of mine build an all-in-one HTPC / Desktop / Gaming system and I've been doing research into SSDs for the past few weeks based on reviews and benchmarks so I wanted to share my info.
Basically there are only two drives to consider and I list them below. A good alternative at this time is to purchase smaller SSDs and create RAID-0 (stripping) sets to effectively double their performance instead of buying a single large SSD. The RAID-0 article below shows great benchmark results to this effect.
Intel X25-M
The Intel X25-M series of drives is the top performance leader right now, and the 80GB drive is barely affordable for a desktop system build if you consider the increased performance of the drive.
Intel X25-M SSDSA2MH080G1 80GB SATA Internal Solid state disk (SSD) - Retail - $383.00 USD ($ 4.7875 / per GB)
OCZ Vertex
The new OCZ Vertex series of drives with the newer 1275 firmware is the price/performance leader and they are much more affordable than the Intel drives. When you combine two of these smaller 30/60 GB drives into RAID-0 (stripping) you get double the performance at still acceptable prices.
OCZ Vertex Series OCZSSD2-1VTX30G 2.5" 30GB SATA II MLC Internal Solid state disk (SSD) - Retail - $129.00 USD ($ 4.3 / per GB)
OCZ Vertex Series OCZSSD2-1VTX60G 2.5" 60GB SATA II MLC Internal Solid state disk (SSD) - Retail - $209.00 USD ( $ 3.483 / per GB)
Reviews
Required Reading:
AnandTech - The SSD Anthology: Understanding SSDs and New Drives from OCZ
AnandTech - Intel X25-M SSD: Intel Delivers One of the World's Fastest Drives
AnandTech - The SSD Update: Vertex Gets Faster, New Indilinx Drives and Intel/MacBook Problems Resolved
RAID-0 Performance:
ExtremeTech - Intel X25 80GB Solid-State Drive Review - PCMark Vantage Disk Tests
BenchmarkReviews - OCZ Vertex SSD RAID-0 Performance
(Be Warned about BenchmarkReviews! Synthetic benchmark results only, no real-life benchmarks such as PCMark Vantage.)
I wonder how that Monkey feels working for MS.
Ballmer's well paid for his, er, efforts.
Actually you typically use the outermost, not the innermost. Data rate at the OD is ~2x that of the ID.
More data, damnit!
eeeewwwww! You get to pick that one up!
SAIT holds 500 gigs native, latest rev of LTO holds 800 gigs native.
This is why enterprises use large tape libraries and RAIT.
I would absolutely do RAID-1 for any system where the data has more value than a second drive. As far as them running for years, sure the X-25E is rated at 1PB of random writes. Even our busiest SAN volumes have only seen about 14TB of writes in 2.5 years, this is for a 1 million row per month OLTP and reporting system. If you have a single RAID1 today there is no way you are pushing enough IOPS to fail the drive in less than 5 years. SSD's have another advantage which is that even when a cell has failed they are still readable so you shouldn't lose data if you manage to wear it out.
There are 4 boxes to use in the defense of liberty: soap, ballot, jury, ammo. Use in that order. Starting now.
"As an IT administrator did you ever think of replacing disks by SSDs? Or using SSDs as an intermediate caching layer?
SSDs aren't big enough for some uses as mass storage but they could speed up things if used as a cache.
Note that this paper has nothing to do with laptop workloads, for which SSDs probably make more sense (due to SSDs' ruggedness)."
I think laptops are where SSDs can come into their own. There shouldn't as much need to large mass storage and SSDs extend battery life. Having said that, I replaced the 160GB HDD in my 1 1/2 year old laptop with a 320GB drive, the biggest I could find.
Falcon
Should there be a Law?
Anyone know the real-world performance numbers of IoDrives?
I realize most SSDs perform worse than advertised, but they have quite a bit going against them.
Lots of them are MLC; with sustained write workloads they run out of erased clusters. Suddenly their write speed drops to half.
Most SSDs are also limited by the SATA controller, which is a massive limiting factor in IOps and latency.
And the SSDs themselves have shoddy controllers with hardly any cache, that are too slow to maintain high performance under demanding erase-rewrite workloads.
But the IoDrive's specs are so absurdly high, that even if they aren't real-world, it's surely faster than a meagre 350 IOps, or even a meagre 20,000 IOps. They benchmark at almost 200,000 IOps, so I'd assume real-world performance would be at least a third to a half of that.
Using six of them to get over a million IOps seems to indicate they scale well, and real-world performance may actually be pretty close to the benchmarks.
What you have to remember is some companies like to push technology. Others will try to sell you crap better than your current crap, but crappy enough that in a year they can put out better crap for you to upgrade to.
*whoops*. Misread. 11 drives - Five 320GB, and six 160GB IoDrives. Still, it's impressive - 8GB/sec throughput!
Hm. I was thinking the same thing about the ACM subscription.
Try Google.
Falcon
Should there be a Law?
Today you just buy 2.5" drives with 2" platters, much more cost effective =) As an example HP 146GB 15K 2.5" has a full stoke latency of 4.85ms, nearly as fast as the average (short stroke) latency of a 144GB 10K 3.5" drive (3.9ms). No way the 2.5" cost 2x more =)
There are 4 boxes to use in the defense of liberty: soap, ballot, jury, ammo. Use in that order. Starting now.
LTO4 Tapes are 800G
Mostly people using these tapes have a robot, so they write until a tape is full, then start on the next tape. Units holding 10 tapes are pretty cheap for a company to buy.
Well, given that all modern domestic asses are derived from the African ass, and the European ass is extinct, it should be kind of obvious which.
Oh, wait.
I just got myself a OCZ VERTEX 250GB drive, and I've seen SQL Server do ~6000 random IOPS on it, sustained, in a practical case (complex UPDATE statement for a huge table). This is on a laptop, mind you -- that UPDATE command was CPU limited!
I ran it head-to-head with the same ~60GB database against two production servers, one with a 20-something spindle SAN volume, and the other with a 3-drive 15K RPM SCSI RAID. It won against both for all cases where IO was a significant bottleneck in the query. Obviously, my laptop lost out against the 8-CPU server with 32GB of memory for 'small' queries, but for un-cached data sets, it was usually faster.
I've seen a number of enterprise environments where SSDs make perfect sense:
* System disks for Citrix XenApp servers - if they're hosting 100+ users on a modern 64-bit Nehelem platform, disk I/O will be the bottleneck.
* ARC cache disks for ZFS volumes - Solaris now lets you use a "fast" disk as a kind of cache for "slow" disks. Whack in a few SSDs in a mirror or a stripe if you're brave, and you could see faster speed than even SAN storage.
* Data disks for business intelligence or integration servers. Both tend to do horrible things to disks, like run multiple parallel read and write streams at the same time.
Personally, I'd stick with a RAID-5 of standard HDD's for now. I'm waiting a couple years for SSD's to get cheaper and to have more reliable statistics on failure rates.
Databases has always been an area where I stick with old and proven before going with new and shiny. I've been called into projects where new and shiny ended up costing more than if they had gone with old and proven in the first place.
That being said...I'd never be getting an Xserve to start with. Failure rates and hardware issues in the field were a lot higher than they should have been and other than the Xgrid or Qmaster controller, I never saw anything that couldn't be done cheaper with other hardware and FreeBSD.
"The problem with socialism is eventually you run out of other people's money" - Thatcher.
Yes, Intel's SSDs completely invalidate their whitepaper. The X25-Ms were first released in September 2008, this paper was published in November 2008. Their research likely took longer than that. Furthermore, from page 2 it seems clear their servers are mostly file/print servers (which typically aren't bottlenecked on disk IO). Also, what are they doing running a 7 TB Exchange server when Microsoft PSS told me in no uncertain terms not less than a year ago to keep each information store below 50GB?
Not for the whole dam SAN. Just for the prefetch or other apps the user loads. You cannot add FIVE SSD drives and have all of them prefetch and on a server you cannot have it prefetch the data that users are requesting.
Mac OS X might have some of this ZFS SSD Cache drive madness in the next release, when they fully support ZFS for the boot volume.
Thanks for your input. Any comments on possible small files problems?
Bert
If MS says its bad, whoe gives a shit! We all know how M$ like to buy studies that say exactly waht they want them to say.
Quote from Steve Ballmer "Facts?! FACTS!?!> We don't need no stinking facts!!!!!"
Their price needs to decrease by 3-3000 times for them to make sense.
How do you decrease something 0.5 times ?
Then how do you decrease it 3000 times ?
echo '[q]sa[ln0=aln80~Psnlbx]16isb572CCB9AE9DB03273snlbxq' |dc
Actually, SATA protocol is reasonably low overhead, compared to SAS/SCSI, I wouldn't say it's a massive limiter.
As long as you plan on allocating 1 PCIe bitlane per SSD attached, whether SATA or anything else, you should be fine. The problem is those PCIe x1 devices with four drives attached. They can't keep any of the SATA busses saturated.
However, get a good x4 controller to attach 4 devices, or x8 for 8 devices, and you'll be able to keep everything humming along at their peak rates.
Remember, a single modern rotating drive (130-140MB/s sequential at the OD) can saturate any regular PCI controller, and almost saturate an x1 PCIe controller by itself. It gets even worse when your NIC and your HBA share a PCI bus, because you can lose half your bandwidth going in-and-out of memory.
Problem with getting six IoDrives is having six PCI-e slots available that are all x4 or better. However, you can get that same IO rate in IOPS with a pair of high-end 16-port PCI-e SATA/SAS boards, and connect 32 Intel X25-E drives, and likely still have expansion slots left over in your system. That whole setup will burn about 5W per controller, and ~2W per drive under load.
More data, damnit!
The only real difference between the two in the SSD world is the 'enterprise' and 'extreme' tend to be SLC rather than MLC. It'll be a matter of time before the performance difference between the two will be so minor that it'll be difficult to justify the higher price tag for performance alone.
Decreasing something by 0.5 is equivalent to multiplying it by 2.
It depends on the workload. It looks like they got the least benefit from the "test web server" class of devices. So if you define you metrics as IOPS/$, or MB/s/$ and the difference between SSD an HDD is neglible your measurements are going to blow up. Obviously, if the results are correct, it wouldn't make sense to use SSDs in the test web server area.
Decreasing something by 0.5 is equivalent to multiplying it by 2.
So if I ask you to decrease my wealth by half, you would actually double it ?
Sir, I would love to do some business with you :)
echo '[q]sa[ln0=aln80~Psnlbx]16isb572CCB9AE9DB03273snlbxq' |dc
Just as an academic exercise, we recently built out a server running SSD's. It was certainly interesting.
Basically, my testing found that generally when running a Windows operating system, the system booted quite quickly and the "feel" of the desktop was pretty good. However, I was hard pressed to say that it was really significantly better than the 10K 146GB SAS drives we were already using. Granted this was using SATA drives on a SAS/SATA combined controller (from HP), so we were able to test both configurations just by switching out drives.
Now, I will say though that once booted and running, synthetic benchmarks (Atto) actually resulted in some surprising results. When configured in a RAID 1, the read performance was generally pretty good, but only in the realms of 5-10% faster than the SAS drives in the same configuration. Write performance was the achilles heel because it plateaued early in the test and as you got up to larger write sizes the SAS drives actually pulled ahead by up to 20%.
Note that this wasn't terribly scientific, but a quick test hitting the servers across the network from a workstation attached to gig showed no discernible difference between application or file performance in either case. We also tested some SQL in this configuration and the results were about the same.
We blew both configurations away at that point and then created a VMware ESXi server. Again, the results were rather disappointing, with even a virtual desktop hosted on the ESXi box running with almost identical performance on the SAS drives to the SSD's. The virtual desktops tested were XP and Vista. A Linux guest also performed about the same.
Atto on the virtual desktops showed results pretty consistent with the raw Windows Server on the SSD's... about 5-10% max increase in performance on reads, but writes falling behind... though VMware's aggressive caching meant that performance was better as a VMware guest than bare metal (though only slightly)
I then ran a torture test with Atto running for a week on the SSD's, using VMware guests all running Atto at the same time. Note that during this week of constant read/write activity we noted a marked decline in performance on the SSD's, which my research says would be expected. In fact the read performance was almost a 1:1 match at the end of that week to the SAS drives, and write performance had suffered also by 5-10%.
It was decided that the problems associated with SSD's (performance degradation over time) coupled with cost meant that in reality it was not likely we were going to start using SSD's any time soon.
Granted, we didn't do any power monitoring; that was not a focal point of the test.
Now, having said all that, I put an SSD (Samsung) in my laptop recently with Vista and saw a visible and marked improvement in performance. I've been incredibly impressed with the SSD in a laptop... but if I'm totally honest with myself the performance of Vista on a server-class controller with SAS drives is pretty much identical. So the conclusion I drew from that (without a significant amount of benchmarking... just "seat of the pants") is that while SSD's DO bring server-class I/O performance to laptops and desktops, they really do nothing for a well optimized and properly configured server storage array.
the primary merit of SSDs is the miniscule inter-track seek time. The primary demerits are the cost per byte and the paucity of appropriate buses.
There are server tasks for which they are completely appropriate. There are very simple loads where the alternative to reducing seek latency is divide the load many times and then replicate hardware and networking interconnect to match.
By "paucity" I just mean that dividing the SSD BW (1.76Gbps?) into the bus streaming BW _should_ give you a number as big as one. And of course you may want several SSDs...
Serial being the keyword here. You probably know that random reads are several orders of magnitude slower on a HDD than serial reads, but the difference is much smaler for a SSD. Lots of people reading your post won't know it tough, so I think it is better to put this advice here.
Rethinking email
I don't get all the fuzz about the small file problem. Yeah, SSDs behave better when you have big files, but so do HDDs. More yet, the troughtput difference between a HDD when reading big files and small files can have 6 orders of magnitude (9 if you go to the theoretical worst case), for a SSD, it has 2 orders of magnitude at most (a few models keep that bellow 10).
Rethinking email
Ah, but it should have this qualifier, to round it out (so to speak): A cocunut laden African or European ass?
"$-for-$ there is no contest. HDDs win."
It's really not that simple. I can deliver 1,000,000 IOPS with an HP DL785 and 6 ioDrives. It would take approximately 4,000 2.5" 15K RPM SAS drives to deliver the same number of IOPS. And that's in RAID0. In RAID10 it would take 8,000 drives. Do you have any idea how much 8,000 drives would cost? Let's go with one the smallest commercially available enterprise 2.5" SAS drive, HP 36GB 2.5" SAS at $350. Even in RAID0 with 4,000 drives you're talking $1.4 million JUST IN DRIVES! In RAID10 we're talking about $2.8 million. You can get 6 640GB ioDrive DUO's for $60,000 and put them in a single $30,000 server, and deliver those IOPS out of the PCI-e BUS (32GB/s IIRC, thats a big B, as in bytes).
Now let's talk enclosures, how about an HP MSA50 with 10 drives per U. Now we need 400U, or 9 and a half racks worth of disks. Now we need controllers, cabling, power and cooling.
Now, obviously we're talking 1-2TB worth of ioDrive storage vs 144TB worth of spinning disk storage. Of course we're also talking about $4 million vs $100k, even before power/cooling.
Point being, workload is a real consideration here.
We have a name for that, it's called "online" and "nearline" storage. What I think interesting now is that some SAN vendors are using SSD's and calling it a "tier 0" and sticking it between the controller cache and the online storage.
You really can't compare the cost per gigabyte of individual disks vs the cost per gigabyte of a SAN, it's just not that simple, you have to look at the entire solution. SATA storage space is around $0.10 per gigabyte, but a very fast high end SAN (fiber channel, redundant controllers, FC switches, HBA's, etc) will run you easily $30/GB.
But when I can consolidate an entire datacenter into a single rack with a blade enclosure, vmware and a 3par SAN, all the sudden the entire solution becomes dramatically cheaper, easier to manage, easier to cool, etc, etc, etc.
I have never, ever seen a fiber channel san that wasn't running some kind of multi-level RAID (eg - RAID10, RAID50 or RAID60). The only case I can think of where you'd spend all that money on fiber channel but put it in RAID5 would be on some OLAP workload that was 99% reads. (Obviously because of the RAID5 write penalty)
You clearly don't understand the cost per square foot of datacenter space, in terms of lease/ownership, cooling and power. I think I'll stick with 5 racks of fiber channel san(s) vs 15 racks of SATA disks failing all day.
And compressed LTO4 is 1.6TB and you can get a tape for $50 or $60.
There is supposed to be a devastating fragmentation problem if you have many small files on a SSD. Had hope the guy could report his experiences on that.
Bert
loves it's swap...
I know tobacco is bad for you, so I smoke weed with crack.
Is there any real benefit from SAS/SCSI drives over Infiniband/FC connections, or is it just snake oil?
I mean, SATA roughly matches SCSI drives in features and capability, and whatever is lacking can be compensated by vastly cheaper CPU/RAM bolt-ons.
Nobody is worried about controllers hitting their limits in rotating rust drives, they are wondering whether or not they are going to physically fall apart.
The specialized SAN backbones seem even more pointless. Much more expensive and hard to replace electronics, far from commodity, and with the exception of Infiniband, bring nothing to network topology.
So seriously, tell me, is there any point in Fiber Channel & Co., or should people just get some ATAoE bridges, and simplify their lives?
I know tobacco is bad for you, so I smoke weed with crack.
The internal quality of the drive is (usually) unrelated to the interface technology. The SAN backbone even less so.
I know tobacco is bad for you, so I smoke weed with crack.
Looks like you need to go over your math. The only way it makes sense is if IOPS/watt is the only thing that matters to you and you have very small amounts of data.
The rest of us that actually need to store things on our SAN will find we get a vastly higher storage density in GB/watt or GB/sq foot with SATA disks.
My math is simple, if SATA is 2-3x slower than fiber channel, then to deliver the same IOPS you need 2-3x more sata disks. I'm not sure what you're not understanding. Some (all) of us need to consider performance, not just total disk space. This becomes not only GB/watt, GB/sq ft, but also IOPS/sq ft and eventually $/sq ft to deliver storage AND IOPS.
You really can't compare the cost per gigabyte
Yes, you can compare. You need to set the parameters for the comparison and decide how to measure, etc, but yes, you really can compare.
The figure you mention is certainly in the ballpark of SAN storage I've seen. I content that most enterprise SAN storage is both grossly misapplied and often a huge waste of money, and further the actual criteria necessary can often be reached with far cheaper hardware.
Of course, the biggest waste is probably the misapplication, which is perhaps a bit beside the point, but still; in the name of that 'easy to manage' theory, everything from word documents to high-performance databases gets stored on high-performance FC disks, reached over redundant FC SAN, connected through a redundant FC fabric.
At a ballpark, I'd say about 80% or more of the stored data in a large corporation could be stuck on USB disks and run through the metal shredder once they're filled. Much of the data is simply unimportant junk that could be lost and nobody would care. ISO images of installations, temporary backups, web caches, etc, etc. None of it has performance requirements, and none of has availability requirements beyond getting it back within a week if service is disrupted. When you could hire another employee for the same cost of storing Linux ISOS and RPMs locally, and there is no other alternative available, there's a problem.
$30 per gigabyte is justifiable when we're talking the most important applications, but it simply isn't when we're talking the junk data. And like most high-end hardware, high-end SAN equipment is not subject to economies of scale; use more of it and it gets _more_ expensive, not less. The more of it you use, the faster fabric you need, the bigger switches you need to serve the high-end case appropriately because that expensive fabric is clogged by crap.
all the sudden the entire solution becomes dramatically cheaper, easier to manage, easier to cool
I'm certainly not arguing against SAN when used appropriately. Heck, I consolidated my storage at home several years ago and run most my stuff off iSCSI, with exactly the advantages you name.
Correctly implemented and highly tiered SAN storage, even with expensive disks in some places, can certainly be cost effective. But don't mistake the solutions the vendors will try to sell you for being the cheapest or most appropriate way to do things.
And don't say you can't compare; it's data of varying importance on some type of media. Any storage technology can be trunked, raided, copied, distributed, to obtain any specific criteria (look at google as an example). The only interesting part is what price you pay to store at the criteria you have, and price per unit may not correlate with anything but how gullible the marks are. Sifting through the smoke you'll get blown to actual analysis like the article here is necessary to avoid paying far more than you have to for your storage.
Uhhhhhh, $30/GB fiber channel SAN is not for your fileserver. It's for your massive high volume transactional databases, your VMFS for your vmware clusters, etc etc.
All the crap you're talking about should be on $3/GB NetApp iSCSI SAN full of SATA disks.
I didn't guess it would be cheaper, I did the math. By using a FC san to back our vmware cluster, we're seeing as many as 15 guests on dual socket X5550 equipped machines with 32-64GB of RAM (depending). Trust me, the SAN pays for itself.
Numerous benefits: http://www.storusint.com/pdf/SCSI_vs_FC.pdf
More bandwidth, more nodes on the link, blah blah.
Thanks for the link, but I still don't see anything that fiber channel has to offer that Ethernet doesn't have.
High speed link - Gigabit Ethernet - check
CRC-32 hash-summing - built into Ethernet, check
Switched fabric - trickier, but also check here you go
SCSI command transport? - ATAPI specification and ATAoE cover that.
Am I missing something?
Seriously, I'm not trying to be flippant.
I know tobacco is bad for you, so I smoke weed with crack.
Depends on what we're comparing exactly. Are we comparing SCSI disks vs FC disks or are we trying to compare FC arbitrated loop against ethernet?
The second one. Though, AFAIK the good parts about SCSI/SAS disks is about the drive controller, which has advantages, but I don't see how they can't be duplicated with a secondary contoler, which you will probably need anyway for RAID and the like. Cheers!
I know tobacco is bad for you, so I smoke weed with crack.
Well that's a little tougher comparison. The short answer is ethernet is probably the better choice. Just look at FCoE and cisco's unified fabric solution using converged network adapters over 10GbE (copper at that). You definitely won't convince me that we should keep building two networks, one for ethernet and one for storage.
I'm not even trying. I was implying that Ethernet was the way to go, and you seem to agree.
I know tobacco is bad for you, so I smoke weed with crack.
Yes, exactly, I'm agreeing with you. Maybe it's a colloquial thing, but I wasn't referring to you specifically when I said "You definitely won't convince me", it's meant in a general sense, as in, anyone else who may take that position. Sorry for the confusion :)
No problem.
I know tobacco is bad for you, so I smoke weed with crack.
2 phoronix studies combined prove that there is no real world advantage of SSD. Advantage is only in synthetic benchmarks. Note that they used one of the fastest SSDs available (Intel X25-E), and a 7200 rpm laptop hard disk. For a slower SSD, difference might be even lower or hard disk might beat SSD in performance. This SSD was not even available at the time of this MS study.
1. http://www.phoronix.com/scan.php?page=article&item=intel_x25e_ssd_linux&num=1 : Uses ext3 filesystem and compares a hard disk and an SSD . Negligible performance difference in real world tests, huge difference in synthetic tests.
2. http://www.phoronix.com/scan.php?page=article&item=intel_x25e_filesystems&num=1: Using SSD, compared different file systems (JFS, ext4, ext3, XFS). Again, negligible difference in real world tests, but different filesystems did well in different synthetic tests.
Bingo Dictionary - Pragmatist, n. A myopic idealist.
Slight problem with your statement. If data is more valuable than a second drive, do periodic backups to a second drive. RAID-1 is for situations when continuous uptime is more valuable than a second drive.
Bingo Dictionary - Pragmatist, n. A myopic idealist.
Yes, I'm aware of that. That comes with some lost performance, that could be as severe as 2 orders of magnitude increase on writting times.
My point was just that disks also have devastating fragmentation and performance problems when dealing with many small files, and those are way more severe than the ones solid state devices show.
Rethinking email
then to deliver the same IOPS you need 2-3x more sata disks.
Which also gives you 6-9x more storage space. Storage space, which, unless you have a very performance oriented load, you'd have to buy anyway. Combining appropriate access patterns with each other and you can get both high-IOPS disk and low-IOPS disk out of the same pile, while if you get the expensive disks you're locked into those disks being used only for expensive high-IOPS use, and you still have to buy the cheap disks for storage, ie, over the whole storage need you don't save any power or space.
There are, of course, exceptions. If your load is heavily IOPS oriented so you don't actually need the space from the 3x spindles anyway, then you're certainly better off using faster disks. But like I said, I hear many more complaints about enterprise storage being expensive than it being slow, and when concepts like storage de-duplication are bandied about it's certainly an indication that storage costs are far out of a reasonable range.