SSD vs RAID

[Killamus]

So, my friend and I are having an argument. Which is better, 2x 150g 10k RPM drives, or one 128 SSD? My current 10k drive gets about 60mb/s read, 42mb/s write. In RAID0, that's doubled.

For about 200$ (The cost of 2x 10k 150g drives), this is the best I could find in SSDs: http://www.newegg.com/Product/Product.a ... 6820227395

Half the drive space, same sustained write speeds, but twice the read speeds. Assuming that I do a lot of read/write access on this machine, which would be better? Also, are there any benefits (Aside from the lack of spinning parts) from the SSD that I'm missing?

Thanks.

[Endless Mike]

Well, they're smaller than a desktop hard drive, quieter, and can be put anywhere in a case there's room since vibrations aren't an issue.

[enk]

Assuming that I do a lot of read/write access on this machine

I find this statement hilarious

Anyway, assuming you mean IOPS, an SSD will win by two orders of magnitude.

[Killamus]

Assuming that I do a lot of read/write access on this machine

I find this statement hilarious

Anyway, assuming you mean IOPS, an SSD will win by two orders of magnitude.

Why's it funny? Anyways, according to Wikipedia, a 10k drive gets ~200iops (mean of 1 queue depth/24 queue depth). Considering that RAID0 gives 2(n) performance, wouldn't the iops be equal?

[Bhelliom]

Are you actually getting double the performance in RAID 0? Or are you talking "theoretical" performance?

Will Smith in MiB2 says it the best - RAID=Old and Busted SSD = New Hotness

Although RAID still has the better data security.

When SSDs become cheap and support TRIM in RAID arrays the war will be over and SSDs will be the one drive to rule them all. *


*At least until magic lightspeed crystal data cubes come out.

[Killamus]

Are you actually getting double the performance in RAID 0? Or are you talking "theoretical" performance?

Theoretical is 2x, but the motherboard I am getting supports hardware raid, so it'll be close.

[Bhelliom]

If you are talking motherboard RAID then you won't see much of a performance gain. You really need a dedicated RAID card with its own CPU to see these types of gains. These cards are typically $200.00 or more.

[enk]

Why's it funny?

Because read/write is all harddisks do. And if you just meant "use the harddisk a lot", you'd need more criteria to decide which would be better. And if it was meant to mean something specific, it wasn't very specific. Does it mean a lot of sequential read/write? Or a lot of read/write as opposed to simply a lot of read? Or a lot of alternating read/write?

Anyways, according to Wikipedia, a 10k drive gets ~200iops (mean of 1 queue depth/24 queue depth). Considering that RAID0 gives 2(n) performance, wouldn't the iops be equal?

But according to random stuff I read about SSDs on the intertubes, it doesn't matter if your hard disks do 200 or 400 IOPS, because the SSDs do a hundred times more anyway.

Are you actually getting double the performance in RAID 0? Or are you talking "theoretical" performance?

When I did a test with high end consumer hardware and Intel Matrix RAID a year ago, RAID1 real-world read speeds were less than 1% lower than theoretical speeds, so I guess something similar would apply for RAID0.

[Killamus]

Why's it funny?

Because read/write is all harddisks do. And if you just meant "use the harddisk a lot", you'd need more criteria to decide which would be better. And if it was meant to mean something specific, it wasn't very specific. Does it mean a lot of sequential read/write? Or a lot of read/write as opposed to simply a lot of read? Or a lot of alternating read/write?

I meant a lot of read and write computations. For instance, I do a lot of gaming on my home machine (Read, mostly) and a lot of self programming to write iterations of things to a disk (e.x. a permutations creator). Moreso write then read, lately, to be honest. Just checking out my home server, my read-write ratio is almost 1:1.5 on it.

Anyways, according to Wikipedia, a 10k drive gets ~200iops (mean of 1 queue depth/24 queue depth). Considering that RAID0 gives 2(n) performance, wouldn't the iops be equal?

But according to random stuff I read about SSDs on the intertubes, it doesn't matter if your hard disks do 200 or 400 IOPS, because the SSDs do a hundred times more anyway.

Simple SLC SSD ~400 IOPS[citation needed]
10k RPM SATA drives, queue depth 24 ~290 IOPS
10k RPM SATA drives, queue depth 1 ~130 IOPS

Those are the stats I was looking at, over the IOPS page on Wikipedia, if you were curious. They may be completely off, but I find wikipedia to be at least reasonably within range on things like this.

[GeorgeH]

SSDs are all about the access latency. Hard drives are in the 10ms range; SSDs are in the 0.1ms range.

Compare it to gaming ping. Even if you have the fastest internet connection money can buy, if your ping sucks you're screwed; a slow connection with minimal latency will be much better. Then you go download a massive BluRay rip with that slow connection and see "5 days, 23 hours remaining"; with the fast connection it would've been done in an hour. So which is better, the fast connection or the slow connection?

Moving back to storage, if you've got a lot of large files that are accessed and written in a predictable fashion, hard drives are the way to go. If you've got a lot of small files that are accessed and written in an unpredictable fashion, SSDs are the way to go. Despite hundreds of thousands (if not millions) of hours of work done by incredibly smart people to hide hard drive latency penalties with hardware and software trickses, the average computer user is still significantly bound by their hard drive latency, meaning SSDs are the way to go.

Clear as mud?

[Kromix]

maybe you should mix both and make a RAID SSD and be done.

[Bhelliom]

for now if you do that TRIM does not work and the drives degrade over time unless you run a cleanup program.

[GeorgeH]

I wrote this somewhere else as a response to an "Gee Willikers, HDD makers are doomed!" argument; as such it's a little off-topic, but still has enough relevant information that I thought it'd be worth sharing:

In the typical computer, you have 5 levels of memory. On the CPU, you have L1, L2, and L3 cache, followed by RAM, followed by your hard drive. The access latency of each goes something like this:

L1 – 5ns
L2 – 10ns
L3 – 50ns
RAM – 100ns
HDD – 10,000,000ns

Smoothing out L1->L2->L3->RAM transitions is largely a solved problem. That’s why super fast RAM is worthless to most people; you can expect 1-5% performance increases going from low end sticks to insanely expensive high end RAM. There are some tasks that significantly benefit from quick RAM, but in general the amount of RAM is the only thing that you need to worry about.

The RAM->HDD transition, by contrast, is far from a solved problem; as you might expect from the access latency differences, performance falls off a cliff when making that transition. That’s where flash memory comes in:

L1 – 5ns
L2 – 10ns
L3 – 50ns
RAM – 100ns
Flash – 100,000ns
HDD – 10,000,000ns

There’s still a cliff there, but it’s much less severe. Once we get flash closer to the 1000-10,000ns range, the RAM->Mass Storage transition will be a solved problem in the same way that the L3->RAM transition is. Speed will be irrelevant to most people, they’ll just need sufficient capacity.

Some tentative stabs at diminishing the RAM->HDD cliff have been made by Intel with Turbo Memory, Microsoft with ReadyBoost, and Seagate with the Momentus XT Hybrid HDD, but so far none have managed to get it done. The gap between RAM and mass storage remains as large as it has ever been.

SSD makers have stepped into that gap by offering their fast, flash based mass storage devices. The problem is that SSDs are a duct tape and string solution. They completely replace hard drives instead of augmenting them, they’re very expensive, they have reliability issues (although those have been mostly quashed), and they’re still too far away from RAM’s performance to be no-brainer “must have” devices.

That’s a big reason why HDD manufacturers aren’t flooding the market with their own SSD lines; SSDs for mass storage are transient gap-filling devices, not permanent fixtures of the computing landscape. Enthusiasts who demand the best possible performance today will happily suffer with installing their OS on an expensive SSD and managing slower HDDs for their files, but the average user won’t be using flash devices until they become seamlessly integrated into their PC’s cache hierarchy.

The problem for SSD makers is that you don’t need a massive amount of flash for the RAM->Mass Storage transition; the 500GB Momentus XT gets very good results with only 4GB of flash. Based on that number, the PC of the future will probably end up looking something like this:

CPU – Intel iOMG
RAM – 100GB DDR5
TurboFlash – 400GB Intel integrated iCache
Storage – 50TB Western Digital

The only place OCZ, Corsair, and the like have in that PC is making their own “TurboFlash” module. Unfortunately for them, at best I’d lay even odds that “TurboFlash” will be a user replaceable component instead of ending up soldered onto motherboards somewhere (probably as part of Intel’s ICH20R chipset.)

So don’t worry for the Seagates and Western Digitals of the world; so long as they can keep pushing mass storage capacity upwards, they’ll be just fine. The SSD makers “obsoleting” HDDs are the ones that need to be worried about the future.

[enk]

Simple SLC SSD ~400 IOPS[citation needed]
10k RPM SATA drives, queue depth 24 ~290 IOPS
10k RPM SATA drives, queue depth 1 ~130 IOPS

Those are the stats I was looking at, over the IOPS page on Wikipedia, if you were curious. They may be completely off, but I find wikipedia to be at least reasonably within range on things like this.

But if you look just below that, you also see numbers like 8k, 5k and 20k IOPS for other SSDs. Not exactly a 100 times bigger (that's only when you trust the claims of the manufacturers, apparently), but still bigger than harddisks.

[zmatt]

This is just a sample of tests, but modern TRIM supporting SSDs kick hard drives to the curb. It's no contest. I have also seen benchmarks run with worst case scenario SSDs that were first gen, had no TRIM and were severely slowed down and they still outpaced 10k rpm hard drives in most cases. You just cant get around the latency issue. Anandtech has been pretty active in reviewing SSDs and they have a few very informative writeups on how they work, why they are so good and their shortcomings. I highly recommend everyone read them, they are extremely education. They reached the conclusion that assuming you can afford them they are the single best upgrade you can get for your pc.

EDIT:
So you dont have to look them up:
http://www.anandtech.com/show/2738
http://www.anandtech.com/show/2829

They are old and the recommendations no longer hold, however the information is still very valuable in understanding SSDs. It's actually quite hard to keep up with the market as someone has a new SSD just about every week and their performance has skyrocketed since 2007.

[Moose Hole]

This is from like a year ago. SSD RAID with TRIM.
http://www.bit-tech.net/news/hardware/2 ... for-raid/1