SSD Write Endurance Considered Sufficient | Hacker News Reader

SSD Write Endurance Considered Sufficient

ef.gy

41 pointsforbes32 comments

This is a pretty poor article. Consumers haven't bought SLC in years and MLC is much less reliable (although the actual reliability of flash is still higher than the rating). It doesn't take write amplification or any kind of real workload into account. And it may contribute to the alarmism that it claims to dispel; at worst your SSD will last 172 days? Yikes!

Not like magnetic is necessarily immune. WD greens used to have astoundingly bad spin down logic. I killed a bunch of them with default Linux settings and near idle servers in about that time frame.

Greens are not enterprise drives and shouldn't be used in RAID arrays.

Their logic makes perfect sense for single disk and "green" (power saving.) Its just incredible what people will do to save a little cash on servers. Your company's data depends on them and you put in the cheapest Best Buy inventory you can find?

I'm not sure who you are replying to. Your fictitious account does not match my reality.

A drive that is accessed a couple times a day seems like the perfect use case for a green. Sadly, it head parks after 8 seconds of idle time, and Linux was doing one IO every 30 seconds, I believe as part of the S.M.A.R.T. monitoring (its been a while), the drive is rated for 300000 head parks in its lifetime, so in 105 days the drive has thrashed away its specified lifetime because you were monitoring the drives' health, and head parks wasn't one of the parameters tracked, so it didn't even help to monitor!

This is not because the Green is not Black. It is not the grade of bearing, or the balance tolerance of the platters. This is because WD wrote ridiculously silly firmware for their drives and they self destructed on one of the most common computing platforms.

Back to the article topic: The similar workload machines I deployed 14 months ago have an SSD for their root drive and newer greens (tweaked to not load cycle) for their bulk. None of the SSDs have even reach 1% of their media lifetime count in their S.M.A.R.T. data. I'm not worried about wearing out SSDs.

AFAIK the RE-GP series have the same issue and are advertised as server/enterprise grade.

I have a Scorpio Blue that parked its heads incessantly until I deployed a utility to force better ACPI settings.

Not write endurance, per se. But you only get to park the heads so many times, on average...

Yeah, statistics work. I had greens fail at 30k parks and survive over 1M parks. 300k was the published number for parks. I think they screwed up by allowing the firmware to thrash them to death, but I think they got the 300k number right.

Do you have any concrete/verified descriptions of the firmware problems? I ran across plenty of people commenting on the problem, and my own intuition (or deduction, based on listening to the "click... click... click..." and eventually noting the escalating number of parks reported) plus the fix that worked for me strengthened this intuition. But I never saw a documented analysis let alone an admission on WD's part. (Admittedly, I may have missed one or both.)

While I completely agree, I think it is important to point out that SLC is still used pretty widely in Enterprise-grade SSDs. Standard configuration on these drives is 100% over-provisioning, so lifespan is likely 20+ years.

People go completely irrational about avoiding writes to SSD's, I've seen researchers at an EE dept spending evenings tuning their Windows installations and going as far as stuffing frequently used stuff onto a RAM disk.

I guess the early off brand SSD's randomly eating your data didn't help the popular perception (even though that wasn't because of flash wear). But still...

What's wrong with locating frequently used stuff on a RAM disk?

It's work, but it's unlikely to have any benefit.

I remember when SSD's first started hitting the market years ago, everyone's utmost concern was about wear leveling.

As it turns out, if you ship a drive with buggy firmware and it refuses to be recognized after the 50th cold boot, write leveling ends up not being so important.

tl;dr

I'm guilty of the SSD fear. My first Intel (320) drive was meticulously maintained in fear of shortened life span and performance losses. I went so far as to question every file copied to the drive vs. mounting an external HDD. The thing bricked itself due to a firmware bug in less than 12 months.

Using a Samsung drive now. I don't think about read/writes anymore. I just use the darn thing.

> The thing bricked itself due to a firmware bug in less than 12 months.

Was it an OCZ, by any chance? The exact same thing happened to me. I scrupulously avoided unnecessary wear, disabled swap (though I've been disabling swap for years), mounted /tmp on a RAM disk, and the accurs't thing killed itself stone dead waking from sleep due to a firmware bug.

I also replaced it with a Samsung, as it happens.

> My first Intel (320) drive

Oh, that's embarrassing! I read that, but somehow convinced myself that it was part of another comment.

I don't worry much about my ssd, however I do occasionally look at the wear with smartools.

The 320 had a brickable fw bug? I have 5-10 or so in production and have no issues. Maybe mine have a newer FW. I always update them when I give them out.

It's pretty well documented. The supposed firmware fix doesn't fix the issue [1]

This drive came with my x220 (160gb), shipped March 2012. I have a second Intel 320 (120gb) which is older, and hasn't exhibited issue. Lenovo replaced the 160gb 320 under warranty. I opted to go with a 256gb Samsung 840 pro. The 320's are now external drives.

[1] http://www.anandtech.com/show/4646/intel-ssd-320-firmware-po...

c't has a lang duration test where it became obvious many SSD's were dying long before they should according to the theoretical numbers. Unfortunately it doesn't seem to be available for free: https://www.heise.de/artikel-archiv/ct/2012/03/066_SSD-Zerst...

This technology is too dependent on software to make these kind of models even remotely approximate reality.

Here's a forum doing write tests on a variety of mlc and slc drives...ongoing:

http://www.xtremesystems.org/forums/showthread.php?271063-SS...

Also, most MLC has between 10k and 1k writes, not 100k. TLC is even worse!

It looks like he's only considering ideal writes with no amplification. This can make a big difference: http://lwn.net/Articles/428584/

Considering he's modeling the drive writing at the full SATA 3.0 interface speed for the entire lifespan of the drive, continuously, amplification shouldn't affect his numbers, only the perceived throughput the user receives under this workload. He mentions this is already faster than the drive can actually write: "the maximum link speed for SATA 3.0, which happens to be larger than any other relevant interface or drive speed so it's a solid upper bound."

I thought write amplification happened inside the drive (due to the drive internally working in bigger blocks than the 512-byte sectors it presents to the outside world), so the amount of data written to flash would be greater than the amount sent over the SATA link?

I've always wondered what these theoretical numbers, which has been quoted since long before consumers could ever afford an SSD, should tell you... They should not, in any way, make you trust your SSD. They are beyond useless. Also, this article assumes no write amplification and assumes perfect wear level algorithms. He/She even throws in usb-thumb drives in the mix, so, how do you know if your thumb drive has a perfect (TM) wear level algorithm? What about that flash card? Would you recommend storing swap and log files on that too? Sigh.

Compared to this article the advice of not ever storing swap/log files on a flash drive would be sound advice.

But yes, you are probably not going to hit the limits of your consumer drive SSD. But whether you are going to or not you will not, under any circumstances, be able to find out by making some numbers up. If anyone knows it is the manufacturer, and you can bet you won't get access to that data.

You can use smartctl or equivalent to see the media wear indicator.

But SSD firmware bugs have been a larger problem in practice than wearout, and MWI won't help you with them; keep backups.

SMART says the years-old, low-capacity MLC SSD in this computer is 5% worn out. SSDs in RAID arrays at work are getting a much heavier workout and handling it fine.

You can use smartctl or the equivalent to check the media wearout indicator to see how much the actual workload has worn out a drive. If your computer isn't constantly writing, you're probably fine--that is, your writes will last until you'd be upgrading anyway.

But, keep backups. Besides the many practical problems you can have (laptop lost! fire!), SSDs have firmware bugs every so often.

SSD write fear is kinda nuts, IMO. It's a device subject to wear, tear, and catastrophic failure -- just like HDDs. Keep backups (three for important data; a cloud copy is only one copy), enjoy the performance, and don't sweat it. If you aren't keeping backups, then go ahead and cower under the sheets at night.

- Smartctl has a media wearout indicator if you want to check. Most likely, you'll check it after months of use, see that you've used like 3% of the write capacity, and then not sweat it.

- Firmware reliability sadly has been an issue. So, yes, keep backups.

- It's fun to check out the ongoing SSD write exhaustion derby here: http://www.xtremesystems.org/forums/showthread.php?271063-SS...

Since when do SSD's have a 100,000 write limit? My 256gb Samsung 840 has something like 2-3k, not 100k.

SLC. This article is a blast from the past.

Reply on news.ycombinator.com