What We Learned About SSDs In 2015

Megalith

24-bit/48kHz
Staff member
Joined
Aug 20, 2006
Messages
13,000
This guy thinks that SSDs are a poor choice for data centers, but I’m not sure his argument is entirely legit…

…the unpredictable latency of SSD-based arrays - often called all-flash arrays - is gaining mind share. The problem: if there are too many writes for an SSD to keep up with, reads have to wait for writes to complete - which can be many milliseconds. Reads taking as long as writes? That's not the performance customers think they are buying.
 
Now that was interesting...

It started out interesting, but then devolved quickly into buzzword soup.

This makes no sense.

SSDs have always been a transitional or bridge technology. There's no way that we'd be using SSDs today if we'd had flash technology in 1957 instead of IBM's RAMAC.

But billions of open SATA ports made for a ready market. And now that the early fears of low endurance have passed, we're ready to move on. NVMe, 3D, TLC and all-flash/no SSD arrays will drive the market in 2016.

NVMe is still an SSD.
TLC = bridge technology and less endurance
All flash array is still an SSD array

And this paragraph

Today's SSDs run on NAND flash, which is far from the ideal storage medium. Clunky addressing. Very slow writes. Poor endurance. And it's an analog medium, driving vendors to 3D architectures.

NAND is the fastest storage medium ever mass produced....period. And it's a analog medium? Is this like some metaphysical reference?

Also, let's not lose our shit without taking into consideration that NAND SSD's have only been mainstream for about 8 years now. HDD sucked well into the late 90's. NAND has done exceptionally well for maturing.
 
"1912 was the beginning of the end for automobiles on the road. Why? Because researchers have delved deep into their actual behavior and found multiple problems. Here's what you need to know..."
 
Nice!

"1912 was the beginning of the end for automobiles on the road. Why? Because researchers have delved deep into their actual behavior and found multiple problems. Here's what you need to know..."

Sponsored***** Enjoy the 60 page slideshow!! ;)
 
i knew that. i heard a time ago that hdds are faster when it comes to write and reading at the same time, at massive data "crunching" (if that's the english word i remember).
 
I think he's generally correct. SSD's lifespans are still too short for my tastes. When they die, there's no chance to recover data. And the speed concerns for enterprise environments are real. Thumb drives die for no reason, even tho they can sometimes survive going thru the wash. Normal use kills them. It's same technology as SSD's. Don't get me wrong, they are making progress and improvements, but I'm still running mechanicals for the reliability.
 
Actually most SSD when they come to end of life they stop allowing writes so you can still read your data.....unlike mechanical drives that normally just up and die or get bad sectors.

Reliability comes in proper backups, raid arrays and other storage arrays, no one tech is flawless. We use Nimble storage arrays in many clients and to my knowledge we have not had to replace a single flash drive, but we have replaced mechanical drives many times...
 
eh -- just some neckbeard who saw a chance to ruffle some feathers, get some clicks, and what have you.

I'll take an SSD array any day of the week. Dude can keep his spinning pieces of metal.
 
Actually most SSD when they come to end of life they stop allowing writes so you can still read your data.....unlike mechanical drives that normally just up and die or get bad sectors.

Reliability comes in proper backups, raid arrays and other storage arrays, no one tech is flawless. We use Nimble storage arrays in many clients and to my knowledge we have not had to replace a single flash drive, but we have replaced mechanical drives many times...

Now that's only for enterprise class ssd. They go read only. Consumer drives suicide. Which I feel is kinda bs since that was the major advantage of an ssd drive.
 
If you have a petabyte a month writes, then yes. Mechanical hdds make sense. There has to be a way to reduce the writes on a ssd drive. Maybe a huge ram cache with a universal power backup. Heck, maybe a raid setup that would notify you to swap drives because the writes have reduced the storage to become too small.
 
This guy is a genius. He seems to think that its somehow better to wait 200 milliseconds for hdd instead of randomly wait 2 milliseconds more on SSD lol.
 
If you have a petabyte a month writes, then yes. Mechanical hdds make sense. There has to be a way to reduce the writes on a ssd drive. Maybe a huge ram cache with a universal power backup. Heck, maybe a raid setup that would notify you to swap drives because the writes have reduced the storage to become too small.

How about just using the 6TB Optane DIMMs that are due out next year?
 
I remember the night we moved all our machines in the datacenter from SSD's back to HDD's.

the SSD's at the time (a year and ah half ago) had great i/o in bursts, but didn't have really great 24/7 read/writes from our benchmarks compared to enterprise hard drives. they couldn't handle all the writes we were making to SQL.
 
SSDs actually have far superior reliability to mechanical drives, the idea to the contrary is valid, however, based on what metrics are used. If I'm writing data to a drive 24/7 at a decent rate then it's quite possible a mechanical will outlive a SSD. In the vast majority of cases, however, a SSD will greatly outlive the mechanical. I didn't read the article but it seems it's talking about data centers and performance, in which case, da fuq? You will always have better I/O performance per dollar with a SSD array unless your controller is garbage.
 
the SSD's at the time (a year and ah half ago) had great i/o in bursts, but didn't have really great 24/7 read/writes from our benchmarks compared to enterprise hard drives. they couldn't handle all the writes we were making to SQL.

I have a SSD array at home on my workstation and I run a server (HTTP + Media) off a SSD as well. You're right, the main issue is consistent writes. A SSD array will have a bottleneck where the fast reads aren't realized because of this. It's a problem based on the technology itself (you will always see flash memory with higher reads than writes) which reminds me a lot of broadband (e.g., DSL back in the day); the ideal way around it is to purchase 'symmetric' drives.
 
I think he's generally correct. SSD's lifespans are still too short for my tastes. When they die, there's no chance to recover data. And the speed concerns for enterprise environments are real. Thumb drives die for no reason, even tho they can sometimes survive going thru the wash. Normal use kills them. It's same technology as SSD's. Don't get me wrong, they are making progress and improvements, but I'm still running mechanicals for the reliability.


Comparing SSD's to thumb drives is like comparing 1980's HDDs to current ones.

As for the short lifespans:
http://techreport.com/review/27909/the-ssd-endurance-experiment-theyre-all-dead

The fact that they can't be read is indeed a concern, but the rest not so much.
 
In practice SSD's are mindbogglingly faster than conventional drives...
 
I remember the night we moved all our machines in the datacenter from SSD's back to HDD's.

the SSD's at the time (a year and ah half ago) had great i/o in bursts, but didn't have really great 24/7 read/writes from our benchmarks compared to enterprise hard drives. they couldn't handle all the writes we were making to SQL.


But was this because you were making writes that were less then 4K? SSD's do have an Achilles heal, but it's not a secret or anything.
 
Actually most SSD when they come to end of life they stop allowing writes so you can still read your data.....unlike mechanical drives that normally just up and die or get bad sectors.

I have not seen this failure mode. At home, every SSD death I've seen is just abrupt, won't ever power on again death. Professionally, I've seen a lot of wear percentage on SSD utilized with LSI CacheCade that you most definitely cannot read the data off of, but the drive still functions well enough to only zero it out, and then act like data was never written to it every time the drive powers down or drops out. All of the SSD's I've seen fail in vSphere SAN arrays regarding my professional life have also met the same fate as my home experience. I'm convinced it is a myth created on the Internet, unless you're willing to desolder the NAND and put it on a new controller like a data recovery center for a hail mary.
 
First of all can we confirm if this guy is getting kickbacks from a company heavily vested in spinning rust. My money's on dell/emc "sponsoring his research"
Some things to chew on:
What are the rebuild times on SSD arrays like?
In a 1U server you can have 10 DAS SSD; samsung has a 16TB drive. 128 TB with 8 active 2 standby.
Now what do rebuild times look like.
New storage tech in Server 2016, storage spaces direct looks like it could kill off the big expensive SAN. Everything can now be direct attached storage.
SSD continue to get cheaper; a drive fails after 6 months you replace with same drive that's now 20 cents a gig versus 30 cents a gig @ original purchase date.
In large data centers drive failures are no big deal. You roll down the aisle pull out failed ( or false failure) insert new.
The fear of failure is why we have raid, hourly backups offsite, secondary and tertiary data replication,
megawats of batteries for outages that may never happen.
 
This article is shit. As soon as they started talking about URE rates, without taking about volume or time period, I knew it was time to stop reading.
 
In practice SSD's are mindbogglingly faster than conventional drives...

On reads - yes
On heavy writes - no

An SSD is only fast on writes, until it runs out of free blocks (blocks of data in the flash that are erased)
Once they run out free blocks, you have to wait while the drive erases more unused blocks. This is a slow process compared to normal access. Under normal desktop use, this usually isn't a problem unless you happen fill the drive to almost full capacity, and you are replacing more that that the available free blocks.

Try this on a consumer level SSD.
1. Fill the drive with large files.
2. Once the drive is almost full, delete the files.
3. Try copying large files to fill the drive again and watch how slow the copy gets after a few files.
 
Some things to chew on:
What are the rebuild times on SSD arrays like?
In a 1U server you can have 10 DAS SSD; samsung has a 16TB drive. 128 TB with 8 active 2 standby.
Now what do rebuild times look like.
New storage tech in Server 2016, storage spaces direct looks like it could kill off the big expensive SAN. Everything can now be direct attached storage.
SSD continue to get cheaper; a drive fails after 6 months you replace with same drive that's now 20 cents a gig versus 30 cents a gig @ original purchase date.


Except Enterprise SSD's are still way more expansive.
I can buy a 4TB enterprise SATA drive for around $200
A 2TB Enterprise drive (Intel p3600 or p3700) runs around $3,500 or $4,500

It would speed up my database server and my exchange server if they where moved to SSD's, but the price is still too much. I can't justify 20 times the price for a drive, even if it is a lot faster. Instead it's cheaper to just use a larger number of mechanical drives and optimize the server with multiple spindles, etc.

I'd agree that mechanical drives are more likely to fail form a hardware problem, however, an SSD is more likely to wear out in a heavily used server.

As for failures, with my existing servers, I'm more likely to replace a drive due to needing more space (upgrading with larger drives), than I am to replace a drive because it has failed. I have a stack of old 80GB and a growing stack of 250GB drives that still work fine, but are just too small to be used any more.
 
Now that's only for enterprise class ssd. They go read only. Consumer drives suicide. Which I feel is kinda bs since that was the major advantage of an ssd drive.

Intel drives (even consumer) die on purpose at 750TB of writes.
 
The only part of this article that isn't stupid is the line about SSDs not being that great for databases... Well no, it would be faster to have your idea database in RAM and write out changes asynchronous... Oh wait that's how they work. Unless you have crazy high volume you don't need SSDs behind your databases, but then again it's not going to hurt. This guy sounds like a Seagate spokesman or something.
 
i guess the samsung Pro 840 SSD is still a good buy (MLC for the Win), Writing 2PB of data is nothing to sniff at (with 0 bytes of loss data until the drive suddenly stopped working, but they was throwing 24/7 constant Random Writes at all of them) you never really do this normally tho, most consumer SSDs should last 10 years unless your using them in workstation/server loads that use lots of writes

unreliability wise i only had Sandisk SSDs that have failed on me in odd ways (1 stops working after 5 minutes, another one it stops reads from happening when trying to copy files off it had to use a unstoppable copier and manually skip files that would not copy as it did not error the copy a file would pause indefinitely until i stopped skipped the file manually and then continued)

why most SSDs take there own life when they fail is to prevent them from been resold (but its Very bad on a recovery standpoint)

i some enterprise SSDs have QOS where they guarantee response times of less than 1ms (something like 95% load) and then another at less than 10ms (99.9% load) they have lower hard capped Max Write IOPS but are consistent (something like 30K IOPS when the drive can do 90K so gives the drive time to do GC and TRIM without it affecting latency and gives you a Solid and Predictable response time for reads and Writes)

also enterprise SSDs typically have what 25% or more spare space, compared to under 10% on consumer SSDs
 
Intel drives (even consumer) die on purpose at 750TB of writes.

That's interesting considering the Intel 910 800Gb that I bought for 360 dollars is rated for 14PB writes and is claimed to have an another 800Gb of provisioning to make that happen.
 
That's interesting considering the Intel 910 800Gb that I bought for 360 dollars is rated for 14PB writes and is claimed to have an another 800Gb of provisioning to make that happen.

has just under 1900GB of total flash on the 800GB version (seems Intel are using 2x amount of flash on enterprise stuff) i would imagine them things would do 100PB of data or more (if the drive allowed that thought, as intel drives tend to suicide at set limits)
 
Don't lie.

The Intel DC P3700 has endurance up to 62 PB (that's 62,000 TB).
The Intel DC P3608 has endurance up to 21.9 PB.

I wanted to make a quick post to say thank you very much for correcting w00t69's post. I took his word for the truth and almost entirely wrote off Intel SSDs because of it. That's why I love the [H]ard so much! Because of honest posters like you.
 
Back
Top