What Makes Parity RAID Safe on SSDs
-
Maybe you can clear this up for me because I'm not the biggest hardware/raid buff and don't have that extremely deep level of hardware knowledge to the atom.
I typically go RAID 10 with spinning rust and RAID 5 in most SSD cases.... unless where the performance gain will actually make a difference and it's just that damn important to have the extra level of protection vs RAID5.
Anyways, to get to what I'm not clear on...
The biggest argument I've seen everywhere on RAID 5 + SSD's vs HDDs is that URE doesn't matter because SSD's work so fast (basically speaking, and maybe URE isn't the correct one to use, I can't remember at the moment). But doesn't the same exact amount of bits get processed?... Just faster? Everyone says a RAID 5 rebuild of high-terabyte drives will take FOREVER (with HDDs) and the changes of a failure go up. But I'm thinking to myself... wait, who cares about the time... isn't it about the amount of bits moving? Isn't it because of the massive number of bits moving that increases the chance of URE or whatever... not necessarily the time it takes to move them?
That said, wouldn't it mean that SSDs and HDDs with equal capacity experience the same number of bits moving during a rebuild... but the SSDs do it faster? Therefore, the chance of a bit flipping or <insert issue here> is the same in both. Oh, I'm not talking about mechanical failure, that's obvious. Let's keep that out of the equation in this particular thought experiment.
Thoughts on that SAM? Or am I just drifting off the road?
-
OK I had to do some digging before posting.
I recall seeing @scottalanmiller post many times that SSDs don't suffer UREs, but perhaps he was meaning that they are so unlikely in current use patterns we can ignore them.
http://www.theregister.co.uk/2015/05/07/flash_banishes_the_spectre_of_the_unrecoverable_data_error/
This post talks about this topic exactly. But I'll shorten it for this post.
https://i.imgur.com/JfTT7E0.pngYou asked - is it all about the bits, yep, it sure is. As you can see in that graphic, even consumer SSDs are 10x less likely to hit a URE than an enterprise HDD. And there are two levels greater than that. You can see where Scott got the 12 TB basically means 100% likeliness that a resilver will fail. You'll notice that a 600 TB array has about a 50% chance of failure, 300 TB has 25% 100 TB is around 8%.
I think the new question we need to ask ourselves, what level of risk are we willing to accept? Which isn't new at all, it's something we should have been (and continue) to ask ourselves.
-
@Dashrender said in RAID for Four Disks:
OK I had to do some digging before posting.
I recall seeing @scottalanmiller post many times that SSDs don't suffer UREs, but perhaps he was meaning that they are so unlikely in current use patterns we can ignore them.
http://www.theregister.co.uk/2015/05/07/flash_banishes_the_spectre_of_the_unrecoverable_data_error/
This post talks about this topic exactly. But I'll shorten it for this post.
https://i.imgur.com/JfTT7E0.pngYou asked - is it all about the bits, yep, it sure is. As you can see in that graphic, even consumer SSDs are 10x less likely to hit a URE than an enterprise HDD. And there are two levels greater than that. You can see where Scott got the 12 TB basically means 100% likeliness that a resilver will fail. You'll notice that a 600 TB array has about a 50% chance of failure, 300 TB has 25% 100 TB is around 8%.
I think the new question we need to ask ourselves, what level of risk are we willing to accept?
Oh man, that clears it right up. Thanks for that! So SSDs just straight up are much more resilient and can pump through way more bits before a URE or whatever. (I keep thinking it was something else, not URE, but I guess it is URE...)
Anyways, good info here, and summed up real nicely.
-
@Tim_G said in RAID for Four Disks:
The biggest argument I've seen everywhere on RAID 5 + SSD's vs HDDs is that URE doesn't matter because SSD's work so fast (basically speaking, and maybe URE isn't the correct one to use, I can't remember at the moment).
A bunch of SW people keep combining unrelated things and making some really weird quotes. The issue stems from a continuous repeating of "half" of the info with a few details reversed. The end result is right, RAID 5 on SSD is normally safe. But the reasons get crazy.
- Fast SSD rebuilds are important because they eliminate downtime caused by the rebuild process
- SSDs are not subject to URE (they are we think, but the numbers are so large that it is the same as not having them right now)
Those two things are unrelated, other than that they both make RAID 5 safe for SSDs. Notice that the speed and the URE don't connect. UREs are never affected by speed, but by access. So 5400 RPM drives and 15,000 RPM drives with the same URE rate are equally risky for UREs, speed isn't a factor. That's why no speed and/or time is part of the URE equation.
-
@Tim_G said in What Makes Parity RAID Safe on SSDs:
Oh man, that clears it right up. Thanks for that! So SSDs just straight up are much more resilient and can pump through way more bits before a URE or whatever. (I keep thinking it was something else, not URE, but I guess it is URE...)
Correct, AND they are typically smaller (capacity) so a URE that happens every 1.25PB happens far less than on spinning disks. A typical SSD is about 250GB in size, maybe 512GB for servers. An average spinning disk is about 2TB in size. And spinning arrays are normally larger.
So we might decide that a typical SSD array is 256GB - 1TB in size. And a typical spinning disk array might be 4TB - 100TB in size. That's a dramatic different.
100TB with a failure every 125TB is terrifying on a rebuild. That's almost certainly going to hit a URE.
1TB with a failure every 1,250TB isn't scary at all. It's literally 1,000 times different in those numbers!
-
@Tim_G said in What Makes Parity RAID Safe on SSDs:
Everyone says a RAID 5 rebuild of high-terabyte drives will take FOREVER (with HDDs) and the changes of a failure go up. But I'm thinking to myself... wait, who cares about the time... isn't it about the amount of bits moving? Isn't it because of the massive number of bits moving that increases the chance of URE or whatever... not necessarily the time it takes to move them?
It's both. This is where you can tell that so many Spiceheads are just repeating things without taking time to understand them, even when posting so called answers. I see this almost daily.
This is people, no matter how many times this is corrected, conflating UREs with disk failures. They are different things, a URE does not imply a failed disk, the disk is fine and still running. Winchester disks are always spinning and will fail even when not being read. So the time to recovery a failed array matters because:
- Winchester drives die over time, even when not being read
- HDD array rebuilds can grow from hours to literally months
- Rebuilding puts extra strain on the array
So in the Winchester world, time to recovery does matter because the times are long and there are risks from the extended risk period. SSDs don't fail over time when idle, only significantly from write operations and very trivially from read operations, so speed doesn't matter, only the amount and type of access happening during the rebuild and since the rebuilds are fast, that access is fractional compared to Winchester disks.
So speed matters, but not for the reasons that they keep stating.
-
This is also why I have a couple of older servers with RAID 5 that I will be decommissioning this year but haven't removed the RAID 5 from. They are 10K 146GB SAS drives, so the chances of hittiting a URE during a rebuild are low enough.
-
@stacksofplates said in What Makes Parity RAID Safe on SSDs:
This is also why I have a couple of older servers with RAID 5 that I will be decommissioning this year but haven't removed the RAID 5 from. They are 10K 146GB SAS drives, so the chances of hittiting a URE during a rebuild are low enough.
LOL - I'm in this same boat. I removed an old server from service with 300 GB drives about a month ago. I wasn't worried about it for the same reason - two RAID 5 arrays, each with three drives (yeah don't ask I didn't build it.).
-
@Dashrender said in What Makes Parity RAID Safe on SSDs:
OK I had to do some digging before posting.
I recall seeing @scottalanmiller post many times that SSDs don't suffer UREs, but perhaps he was meaning that they are so unlikely in current use patterns we can ignore them.
http://www.theregister.co.uk/2015/05/07/flash_banishes_the_spectre_of_the_unrecoverable_data_error/
This post talks about this topic exactly. But I'll shorten it for this post.
https://i.imgur.com/JfTT7E0.pngYou asked - is it all about the bits, yep, it sure is. As you can see in that graphic, even consumer SSDs are 10x less likely to hit a URE than an enterprise HDD. And there are two levels greater than that. You can see where Scott got the 12 TB basically means 100% likeliness that a resilver will fail. You'll notice that a 600 TB array has about a 50% chance of failure, 300 TB has 25% 100 TB is around 8%.
I think the new question we need to ask ourselves, what level of risk are we willing to accept? Which isn't new at all, it's something we should have been (and continue) to ask ourselves.
Don't take those numbers for granted, there are plenty of consumer SSD drives with 10^14 URE. So if you happen to put 6x 2TB SSDs in RAID 5, you can still run into it.
Now another thing is most misunderstand what URE is and assume it's a set number and you always hit URE after reading 12.5 TB. It's not, it's just a probability or running into sector on disk that cannot be read from. Similar to playing Russian Roulette, if you're lucky, you can keep pulling the trigger indefinitely without blowing your brains out, you can have RAID 5 rebuilding PBs of data without ever hitting URE. -
@marcinozga said in What Makes Parity RAID Safe on SSDs:
Now another thing is most misunderstand what URE is and assume it's a set number and you always hit URE after reading 12.5 TB. It's not, it's just a probability or running into sector on disk that cannot be read from. Similar to playing Russian Roulette, if you're lucky, you can keep pulling the trigger indefinitely without blowing your brains out, you can have RAID 5 rebuilding PBs of data without ever hitting URE.
That's why it's a percentage risk number. Otherwise an array of 12.5TB would have a 100% failure rate, but that never quite happens. There is never 100% success nor 100% failure no matter how big or small the array. But it gets really high, really quickly.
-
@scottalanmiller said in What Makes Parity RAID Safe on SSDs:
@marcinozga said in What Makes Parity RAID Safe on SSDs:
Now another thing is most misunderstand what URE is and assume it's a set number and you always hit URE after reading 12.5 TB. It's not, it's just a probability or running into sector on disk that cannot be read from. Similar to playing Russian Roulette, if you're lucky, you can keep pulling the trigger indefinitely without blowing your brains out, you can have RAID 5 rebuilding PBs of data without ever hitting URE.
That's why it's a percentage risk number. Otherwise an array of 12.5TB would have a 100% failure rate, but that never quite happens. There is never 100% success nor 100% failure no matter how big or small the array. But it gets really high, really quickly.
I can honestly say in the dozen or so RAID-5 rebuilds that I've seen in the last 6 years, literally half of them were successful, and the other half were not. This was in arrays of 6 x 250GB drives to 8 x 1TB drives (all Winchester / spinning rust).
-
@dafyre How many of the failed ones were actually caused by URE? I'd be interested to see the %.
-
@marcinozga said in What Makes Parity RAID Safe on SSDs:
@dafyre How many of the failed ones were actually caused by URE? I'd be interested to see the %.
That I can't tell you. Drive "help me" lights came on, and we replaced drives.
At one point, with one of our LeftHand SANs, I was sent 8 brand new drives, as the techs saw severe errors on all 8 of them, lol. I had to re-install the LeftHand OS from scratch and let that thing sync with our other unit across the way. Only took a week to sync 7TB... The upshot is that nothing went down or slowed to a crawl.
-
If I'm thinking about this correctly, a help me light is a failed/failing drive, not usually a URE. Single UREs happen all the time but don't take the drive down/offline.
This kinda tells me that the lights indicate drive failures.
-
@dafyre said in What Makes Parity RAID Safe on SSDs:
@scottalanmiller said in What Makes Parity RAID Safe on SSDs:
@marcinozga said in What Makes Parity RAID Safe on SSDs:
Now another thing is most misunderstand what URE is and assume it's a set number and you always hit URE after reading 12.5 TB. It's not, it's just a probability or running into sector on disk that cannot be read from. Similar to playing Russian Roulette, if you're lucky, you can keep pulling the trigger indefinitely without blowing your brains out, you can have RAID 5 rebuilding PBs of data without ever hitting URE.
That's why it's a percentage risk number. Otherwise an array of 12.5TB would have a 100% failure rate, but that never quite happens. There is never 100% success nor 100% failure no matter how big or small the array. But it gets really high, really quickly.
I can honestly say in the dozen or so RAID-5 rebuilds that I've seen in the last 6 years, literally half of them were successful, and the other half were not. This was in arrays of 6 x 250GB drives to 8 x 1TB drives (all Winchester / spinning rust).
That's a rough fail rate.
-
I've been lucky I guess - I've never lost an array when resilvering a RAID 5 array.
-
@marcinozga said in What Makes Parity RAID Safe on SSDs:
@dafyre How many of the failed ones were actually caused by URE? I'd be interested to see the %.
Yeah. RAID fails from so many factors. Often no one records what affected what. We were hit with URE losses on an array that lost zero disks! We never used RAID 5 again
-
@Dashrender said in What Makes Parity RAID Safe on SSDs:
I've been lucky I guess - I've never lost an array when resilvering a RAID 5 array.
Ha, I've been lucky to never use RAID 5
-
@Dashrender said in What Makes Parity RAID Safe on SSDs:
If I'm thinking about this correctly, a help me light is a failed/failing drive, not usually a URE. Single UREs happen all the time but don't take the drive down/offline.
This kinda tells me that the lights indicate drive failures.
URE would never cause a drive failed light as the drive is fine. Light indicators are always failed or failing drives. But a single drive loss would not be what killed the machine, it's a drive loss PLUS a URE that would do it.
-
@scottalanmiller said in What Makes Parity RAID Safe on SSDs:
@dafyre said in What Makes Parity RAID Safe on SSDs:
@scottalanmiller said in What Makes Parity RAID Safe on SSDs:
@marcinozga said in What Makes Parity RAID Safe on SSDs:
Now another thing is most misunderstand what URE is and assume it's a set number and you always hit URE after reading 12.5 TB. It's not, it's just a probability or running into sector on disk that cannot be read from. Similar to playing Russian Roulette, if you're lucky, you can keep pulling the trigger indefinitely without blowing your brains out, you can have RAID 5 rebuilding PBs of data without ever hitting URE.
That's why it's a percentage risk number. Otherwise an array of 12.5TB would have a 100% failure rate, but that never quite happens. There is never 100% success nor 100% failure no matter how big or small the array. But it gets really high, really quickly.
I can honestly say in the dozen or so RAID-5 rebuilds that I've seen in the last 6 years, literally half of them were successful, and the other half were not. This was in arrays of 6 x 250GB drives to 8 x 1TB drives (all Winchester / spinning rust).
That's a rough fail rate.
There for a few years, we had a crazy time with stuff crapping out. Even a new drive array went crazy. After the LeftHand, the RAID 5 rebuilds got better, but still lost far more of them than we should have.
