Top Solid State Disks and TB Drives Reviewed

Lucas123 writes "Computerworld has reviewed six of the latest hard disk drives, including 32GB and 64GB solid state disks, a low-energy consumption 'green' drive and several terabyte-size drives. With the exception of capacity, the solid state disk drives appear to beat spinning disk in every category, from CPU utilization, energy consumption and read/writes. The Samsung SSD drive was the most impressive, with a read speed of 100MB/sec and write speed of 80 MB/sec, compared to an average 59MB/sec and 60MB/sec read/write speed for a traditional hard drive."

Comments

Longevity of NAND flash

[ASkGNet]

NAND flash deteriorates with use. When used in a high-I/O situations like hard drives, just how much time will it be able to work correctly? If I recall correctly, NAND blocks are guaranteed to the order of 100000 writes.

Re:Longevity of NAND flash

[peragrin]

Yes, new on disk topology mappings, and new tech give you roughly a million r/w and the mappings help to evenly distribute the load.

Re:Longevity of NAND flash

[plague3106]

This is always claimed as the solution, "evening" writes. But I think the question about how long will the drive last is still relevent; all it takes is a mostly full disk, which has a high I/O load. Even with evening, it seems that at least part of the disk can fail before the rest of the disk.

Do traditional drives fail if the same sector is written to over and over again as well?

Re:Longevity of NAND flash

[goofy183]

Will this ever die? The write cycle counts in modern flash is in the millions now. Doing the math you very easily get 20+ years before write cycle wear is a concern: http://www.storagesearch.com/ssdmyths-endurance.html

How many heavily used spinning drives do you know that last even 10+ years?

Re:Longevity of NAND flash

[baywulf]

Actually the endurance on NAND has been going lower over the years as they switched to smaller cell geometry, larger capacity and MLC technology. Some are as low as 5000 cycle endurance. These MLC(multi-level cell) NAND tend also to be much slower than SLC(single-level cell) NAND. Most SLC NAND have around 50K or 100K endurance.

Re:Longevity of NAND flash

[Khyber]

And this is why we're moving away from NAND, so get that damned term out of your head already! OUM/OVM is coming, uses a nearly identical manufacturing process (It's the same thing found in RW optical media, except you use electricity instead of a laser to change it's state) as CMOS does, and it has FAR more read/write cycles than anything NAND could have ever hoped to achieve, in the range of 10^8 as opposed to NAND 10^5-10^6

Re:Longevity of NAND flash

[ComputerSlicer23]

You ever actually done this? I work on embedded systems that use flash drives... Even with write levelling, we've had failures. It's lots of fun, when your 512MB flash isn't 512MB, and will suddenly lose ~41MB suddenly. As a work around, we've had to start partitioning with extra space left lying around at the end of a disk. This isn't even a heavy workload system.

Some friends of mine at another company that were using them in a I/O laden system that wanted to replace laptop drives to make the machinews lower power and more reliable can blow out a flash drive in about 4 weeks.

Kirby

Re:Longevity of NAND flash

[ComputerSlicer23]

Yes I have. However, I've never had one magically get smaller on me in such a way that fsck decides that your done fixing the filesystem. With SSD, YES, I've had exactly that happen to me.

In my life, I've lost a total of about 42Kb be completely unrecoverable with spinning media (yes, I mean that number literally). I use RAID extensively, I was the DBA/SA/Developer at a place that had ~10TB of disk online for 5 years. In all that time, 42KB is all I lost. Oh, that was in the off-line, tertiary backup of the production database (it was one of 5 copies that could be used as a starting point for recovery, we also had the redo logs for 5 days, each DB was a snapshot from one of the previous 5 days). It was stored on bleeding edge IDE drives put in a RAID 5 array. We used it as a cheap staging area before pushing the data over Firewire/USB to a removable drive that an officer of the company took home as part of the disaster recovery system (it had only the most recent DB and redo logs). The guy didn't RMA the hot spare, and we had two drives fail in about 3 days while the hot spare was waiting for the RMA paper work to fill out. In that one particular case, using ddrescue, I recovered all of the data off of the RAID5 array but 42KB (even though it was an ext3 filesystem on LVM, on a RAID5 array, which made the recovery even more complex). Every other bit and byte of data in my life from spinning media that I cared about, I've recovered (I've had a number of drives die with data I didn't care about, but I could have recovered from if need be). Trust me, I know about reliability, backups, and how to manage media to ensure that failure doesn't happen. I know about failure modes of drives. I've hot swapped my fair share of drives, and done the RMA paperwork. I've been in charge of drives that losing any one of the ~200 drives would have cost 10 times as much as I made in a year if I couldn't reproduce the data on it within hours.

If it had been worth $10K, I'd have sent off the drive to get that 42KB of data recovered. But it wasn't. It's well understood how the failure mode of spinning media. People know exactly how to do things like erase drives securely. People know who to call that has a clean room that can remove the magentic media to and put it under a microscope to get the data recovered. SSD isn't nearly as mature in that sense.

All of that is really to say: Yes, I know something about disks and drives. My point is to say that SSD's aren't magic pixie dust in terms of reliablabilty. I've had exactly what he's saying I shouldn't worry about happen to me on a regular basis. Enough, that our engineering department has developed specific procedures to deal with them in the field. We've changed our release procedures to accout for them. If your going to use an SSD or flash drive, go kick the crap out of it. Don't believe on faith anything you read on Slashdot (including this post, which is anecdotal). We order lots of 5,000 flash disk, and you can bet that at least 100 of them has serious flaws within being fielded. The ones the developers and testing uses regularly develop problems in terms of months, not years. The manufacturer tells us essentially, it's not worth it to find those, so deal with it.

The whole point of replacing the laptop drive was to make the silly thing more reliable. But making it uber-reliable for 4 weeks until the write leveling crapped out wasn't the idea.

Kirby

Re:Longevity of NAND flash

[TooMuchToDo]

I can understand your reluctance to trust flash media. Indeed, it hasn't been proven like spinning media has. Let's take another example. An in-car radio. I want a 100GB hard drive in my car, solid state, that is for all intents and purposes write once. I should be able to dump 10s of GBs of MP3s onto it, and the index should be stored on a replaceable CF card (as the index would be changed often). But why would I remove music from the drive? I can just add more music.

For the above example, a flash drive works very well. If you need the benefits of flash storage mediums (vs spinning media) you should be prepared to engineer around the situation. Run temporary data out of RAM with battery backup, and only commit the data to flash between reboots and power outages.

Re:Longevity of NAND flash

[vertinox]

Do traditional drives fail if the same sector is written to over and over again as well?

No, but they'll fail either reading or writing over time regardless if you are writing or just reading just because the drive is moving. Even if you cool your standard drive, eventually it could just fail because it was left on for 10 years (since an active drive is constantly spinning).

Now its not guaranteed to fail, but the chances of a standard HDD failing that you only read from and don't write it is far greater than a SSD that you put files on it one time and don't write further.

I think SSD shine in archival types of things that you don't plan on trashing and rewriting that often such as image collections, movies, and MP3s. That said, swap disks, scratch disk, and cache file directories would logically still have better performance on your spinning platter drives and if that drive goes belly up you haven't lost much.

Re:Longevity of NAND flash

[s_p_oneil]

"all it takes is a mostly full disk, which has a high I/O load"

It is a relevant question, but this wouldn't kill your hard drive, it would simply reduce the amount of free disk space. And it's not difficult to imagine a file system smart enough to move files around when this happens. When a sector gets written to too many times, it can simply look for and move a really old file onto that sector to free up some of the rarely used sectors of the drive. With the increased performance of SSD, you probably wouldn't even notice it.

Aside from the re-write issue, flash memory drives should be WAY more reliable than a mechanical HD. It should never just completely die or start getting bad sectors so fast you don't have time to retrieve your data. It should also be a lot easier to replace when it starts to degrade. It shouldn't be as susceptible to damage when you drop it from a height of 3-5 feet, or due to heat, cold, vibration, dust, humidity, etc. I'm not sure whether a magnetic field could erase it like a hard drive, but if not, that's another plus for SSD. I imagine SSD's are more susceptible to static electricity, but so is almost everything else plugged into your motherboard, so I'm not sure if that could be considered a minus.

I'm sure if you ever tried an SSD on a laptop, you'd never want to go back to an old HD. The improved performance and battery life would make going back to an old laptop HD seem like going from broadband back to an old 56K modem.

Re:Longevity of NAND flash

[psydeshow]

Just remember to mount these drives noatime to avoid a write every time you read a file.

For that matter, noatime is a sensible default for any desktop OS. When was the last time you actually searched for files you hadn't accessed in six months?

Re:Longevity of NAND flash

[ComputerSlicer23]

When I was young and stupid about drives and media, I lost a 1.2GB WD drive and lost everything on it. I couldn't spell "mkfs" or "fsck" and had no idea how to recover the drive at the time (I also didn't have the money to have a second drive to recover too, and no credit card so I could hold onto the first while having the second during the RMA). I was just young and ignorant. I lost a 1-2GB laptop drive that I literally just rode into the ground, I could have copied everything off and moved along. I knew the drive was going bad, but it was just a knock around system that I didn't care about. In the end, had I been thinking, I'd have saved the e-mail on it. I lost the first ~5-6 years of e-mail I had, but who wants e-mail from when they were 18-24? That was probably a couple of hundred MB that I might regret, but of nothing more then sentimental value. I'd never read it, and only be amused that I could prove I'm getting the same chain letters 15 years later.

I believe I had 4-5 drives I lost due to a virus or pilot error, but not a mechanical/media problem.

I've RMA'ed probably 100-200 drives due to some type of failure. I've had lots of of drives fail that were in a RAID array, that the mirror saved me. I've had lots of drives fail that were stand alone that had a section of bad sectors. All of that I recovered every byte of data from. Normally a drive that is going bad, you can still recover from for a very limited amount of time. Normally you have plenty of lead time, especially with SMART drive monitoring that your drive is going south. As long as you pay attention, spinning media isn't that hard to keep in good shape.

As a professional IT person, 42KB is it. On machines where production work is done for money at a company. 42KB is it, and in that case I was bound and determined to recover absolutely everything, and I invested a week into that project. I gave up on the 42KB once I proved that it was in a backup for the database that was at that point 15 days old (and thus of no use). Had it been necessary or cost effective, I'd have spent the $1-3K to get that drive images recovered by a professional data recovery shop. I think I've lost a drive or two on my personal machines at work, but the drive was fine, the laptop SATA controller was overheating. Using FSCK, I recovered the entire FS once the RAID controller was replaced. I think I had to re-rip some music from CD, because I failed to back it up prior to sending the laptop in for repair. I re-imaged the drive just to be safe in case the RAID controller had corrupted something important on the OS drive, which was the only reason I actually lost the music.

Again, it's the fact that the flash drives we have decided the drives are smaller at the interface level. Using fsck just scragged the system pretty much start to finish. I don't have a clue where the missing blocks are from. I have no idea what happened, upon reboot it decided that the block devices was smaller. Filesystem recover tools haven't had a chance to mature to understand those types of failures. Flash makers haven't yet decided that access to diagnostics and re-mapping logs might be of value to data recovery tools (at least none that I'm aware of). Access to the raw data (in case they are holding blocks in reserve). All of these things are reasons to be concerned about write leveling.

Kirby

Re:Longevity of NAND flash

[ComputerSlicer23]

I think that makes perfect sense, but then I'd think that all the money they'd spent in making the thing perform faster then say 1MB/sec read rate is totally wasted. I'd assume folks are trying to push these are replacements for enterprise server machines, which I'd be extremely relucant to do.

Folks talk about these things in the theoretical (the original poster linked to a story that crunched numbers to show it should be safe). My question is does anyone have solid experience they can point to show that it has actually been safe for say 6-18 months under some well known duty cycle (A database, a file server, an e-mail server).

I have actual experience, with crappy flash made by a low end manufacturer that shows me, it's not terrible reliable. It is my understanding that we've had better luck with other makers, but their parts were too expensive (but software development is free *sigh*).

There are other threads in here that make me want to cram a CF-IDE converter into my machines and try putting my journal onto a Flash drive. Sounds like the performance boost and power consumption is a big win, but the fact that every byte of data pushed to the journal might be an issue. On a home machine, it might be worth playing with for giggles for performance testing.

Other folks I know who have tried to do things with flash have also been disappointed over the past 12-24 months, despite assurances from various experts that "it should work"... I'm looking for, "I'd done it, here it is, go play with it.". Now obviously MP3 players have been doing it for a while. I'm more interested in general purpose usage of a Flash drive. Those are the types of things I'm currently working on, cramming a flash into a machine that runs an ext2/ext3/xfs/reiserfs/jfs or some other read/write heavy usage ready FS on it.

Kirby

Re:Longevity of NAND flash

[TooMuchToDo]

Do RAID1 on the flash drives. You may lose data on one drive, but not both. And you'll stick get huge power savings (no spinning disks). As reliability improves, you could do away with RAID.

Reliability

[RonnyJ]

It's not mentioned in the summary, but added reliability might make these types of disks more appealling too:

The no-moving-parts characteristic is, in part, what protects your data longer, since accidentally bumping your laptop won't scramble your stored files. Samsung says the drive can withstand an operating shock of 1,500Gs at .5 miliseconds (versus 300Gs at 2 miliseconds for a traditional hard drive). The drive is heartier in one other important way: Mean time between failure is rated at over 2 million hours, versus under 500,000 hours for the company's other drives.

Hmm

[orclevegam]

I'm really interested in the SSD drives as high performance replacements (particularly for holding OS images where boot times should be nicely reduced), but I've got to wonder how the mean time to failure of one of these compares to a traditional magnetic disk. I know they use write leveling, but that just means everything will have a tendency to fail around the some time later, rather than a spot or two now and then. Anyone have any actual reports on these? I can usually make it 2 or 3 years before I start to see errors crop up on magnetic disks (sometimes more or less depending on how much thrashing the disk is subjected to). Might it be cheaper to simply buy a decent sized CF or SD card and an ide/sata adapter rather then paying for an actual disk, or is there some inherit advantage to one of these you'd be missing out on?

Is it just me?

[crymeph0]

Or does the linked article say nothing about TB sized drives, only the flash drive?

MTBF/Write Cycles

[Lookin4Trouble]

Since I've seen this plenty of times, I'll address it.

Write Cycles: Even at the lowest estimate, 100,000 write cycles to failure

Meaning on a 32GB Drive, before you start seeing failures, you would have to (thanks to wear-leveling) write 32*100,000 GB, or 3.2Petabytes

at 60MB/sec write speed of the Samsung drives, you would need to write (and never, ever read) for 3,200,000,000/60, or ~53Million seconds straight.

53Million divided by 86,400 means you would need to be writing (and never ever reading) for ~617 Days straight (That's roughly 20 months of just writing, no reading, no downtime, etc...

So... the sky is not falling, these drives are slated to last longer than I've ever gotten a traditional drive to last in my laptop(s)

Almost forgot to mention, standard NAND of late has been more in the 500k-1M write cycle between failures range. 100k was earlier technology, so multiply numbers accordingly.

Re:MTBF/Write Cycles

[goofy183]

Except most wear-leveling MOVES data around on the drive. Since random access is 'free' shuffling mainly read-only data around on the disk periodically is perfectly reasonable.

Re:MTBF/Write Cycles

[NMerriam]

You obviously wouldn't be overwriting any data already stored on the drive

No, but the wear-leveling routines in the drive will happily move around your existing data so that rarely written sectors are available for heavy writing operations.

Seriously, this "issue" comes up in every discussion about SSDs, and it seems like people are just unwilling or unable to accept that what was once a huge problem with the technology is now not even remotely an issue. Any SSD you buy today should outlive a spinning disk, regardless of the operating conditions or use pattern. It is no longer 1989, engineers have solved these problems.

What about real performance

[B5_geek]

How do these SSD compare to a real high-end disk like a 15k rpm Ultra320 SCSI drive?
Of course SSD will beat an IDE disk hands-down, but that is not why you buy IDE drives.
I have always used SCSI for my OS/system and IDE for my storage, this combination (in addition to SMP rigs when available) has allowed me to out-live 3 generations of processors. Therefore saving me money on upgrades.

SSD seems best marketed to 'gamers' so why is it always connected to a very limited IO bus?

Re:What about real performance

[KonoWatakushi]

No need to compare with 15k rpm drives; flash disks lose spectacularly to low rpm laptop drives for random write performance. For obvious reasons though, no one ever tests random write performance. Manufacturers also rarely report random write IOPS.

Flash is great, if your disk is basically read-only.

Re:What about real performance

[DDumitru]

Random write performance to bare drives is usually quite bad. Most "reputable" vendors do publish random write figures. SanDisk quotes 13 IOPS in the spec sheets. Mtron quotes 120 IOPS. I have not seen quotes from Samsung, but have tested their old drives at 27 IOPS. I even tested one drive at 3.3 write IOPS.

On the other hand, random writes issues are "fixable". My company just published tests for various Raid-5 Flash SSDs setups. For 4 drives testing with 10 threads on Linux 2.6.22 using our MFT "driver", we get:

4K random reads 39,689 IOPS 155 MB/sec
4K random writes 29,618 IOPS 115 MB/sec

These are real numbers and the application does see the performance improvement.

For full details on drive performance see:

    http://managedflash.com/news/papers/index.htm

Re:Number of writes?

[Planesdragon]

Why is the ultimate number of writes never taken into account in these comparison reviews? Why are solid state drives tested so that their weaknesses are not probed? Because it's a measure best reflected by Baysean Data, and they don't have enough time to test them.

If you want, buy an HDD and a Flash-Drive of the same cost, hook them up to a program that runs each at equal data-transfer rates, and see how much data you can read and write to each before they fail. Report back to us in the six months it'll take you.

Oh, and you need to do the trial over a wide sample, so get, oh, at least ten of each.

It gets better

[WindBourne]

My home server has a terabyte of disk, but I added a CF-IDE adaptor card, along with 4G CF card. I loaded Linux kernel on it, and then mapped a few dirs to partitions on the HD. After about 6 months at it, I noticed that the temp in the case dropped. It appears to be about 5-10C lower (depending on load). The disk spend the bulk of their time sleeping. I have been pleased enough with this server, that I am going to do the same to my small shoe box computer. Rip out the HD, add CF for /, and then mount my home dir from the server.

Waiting for low-end drives

[alegrepublic]

I am still waiting for a reasonably priced low-end drive. An 8GB usb drive can be found for about $50. Packing 4 of them and replacing the usb circuitry with SATA would make for a 32MB for $200. Granted, it may not be the fastest drive around, but sometimes speed is not the most important factor. A 32MB would be enough for installing any current OS and still have some room for personal files to carry along on a trip. So, I think the current trend of providing high-end drives only is just an attempt to milk users to the maximum without much concern for what we actually need.

Re:Waiting for low-end drives

[skolima]

A 32MB would be enough for installing any current OS and still have some room for personal files to carry along on a trip. We're still in 2007, riiight?

Speed for a mech. HD is burst, not track-to-track?

[Futurepower(R)]

Quote from the Computerworld article and the Slashdot summary:

"Samsung rates the drive with a read speed of 100MB/sec and write speed of 80 MB/sec, compared to 59MB/sec and 60MB/sec (respectively) for a traditional 2.5" hard drive."

The speed quoted for a mechanical hard drive is a burst speed, accurate for reading only one track, and doesn't include the time it takes for a conventional rotating hard drive to change tracks. Isn't that correct?

Then don't fill the drive

[tepples]

This is always claimed as the solution, "evening" writes. But I think the question about how long will the drive last is still relevent; all it takes is a mostly full disk, which has a high I/O load. Easy: don't let the drive become mostly full. This means heavy-duty drives will be a 64 GB chip reformatted for 48 GB with the rest designated as spare sectors for wear leveling, but the power consumption and seeking speed benefits can still make it worthwhile.