Intel 311 SSD as ZIL/SLOG

Beta4Me

Weaksauce
Joined
Nov 3, 2011
Messages
103
What are people's thoughts on using an Intel 311 SSD as a ZFS write-cache (ZIL/SLOG)?

It has fairly similar specs to the outgoing X-25E SSDs and is also SLC but it lacks a supercapicator (can be mitigated with dual UPSs etc.)

As a big plus, it is really cheap at around AU$140 which is about half the cost/GB of of the X-25E and, as a smaller size (which is more relevant for a ZIL) it makes it even cheaper! :)
 
Get an Intel SSD 320 instead.. The reason people rave about SLC SSD's for ZIL drives is because of write cycles, but what people forget is that the parameter that really matters is total amount of writes to the drive before the cells start to die, and the extra NAND on the 320 will probably (impossible to know for sure) make up for the difference in MLC vs. SLC.

For the price of the 311 you could get an 80GB 320, or even a 120GB 320 for a bit more...

This way you also get the supercapacitor.
 
Get an Intel SSD 320 instead.. The reason people rave about SLC SSD's for ZIL drives is because of write cycles, but what people forget is that the parameter that really matters is total amount of writes to the drive before the cells start to die, and the extra NAND on the 320 will probably (impossible to know for sure) make up for the difference in MLC vs. SLC.

For the price of the 311 you could get an 80GB 320, or even a 120GB 320 for a bit more...

This way you also get the supercapacitor.

What about IOPS performance? That is mostly the reason Intel went SSD for the 311 (designed for Z68 caching).
 
Intel SSD 320 4K write IOPS: 23000
Intel SSD 311 4K write IOPS: 3300 (not a typo)

Intel SSD 320 4K read IOPS: 39500
Intel SSD 311 4K read IOPS: 37000
 
I was thinking of getting 2 mirrored pairs of (i.e. 4) 311s for ZIL but maybe I should be getting 80GB 320s and over-provisioning them down to 20GB... :confused:
 
Really? :O

According to Intels official specs, yes.

And don't bother partitioning the drives, any unused space will be used for wear leveling, the drive doesn't care if it's partitioned or not, and since the ZIL will never contain more than relatively few GBs of data, an 80GB drive will have lots and lots of space for wear leveling.
 
And don't bother partitioning the drives, any unused space will be used for wear leveling, the drive doesn't care if it's partitioned or not, and since the ZIL will never contain more than relatively few GBs of data, an 80GB drive will have lots and lots of space for wear leveling.
Actually, that's not entirely true. I'll just be setting an HPA so that only 16GB is accessible or something...after all, I'm going to have a stripe of mirrors so there'll be 2 pairs to go round for double the space and IOPS :) 16GB would be great as there'd be 4 spare spots for every 1 used/data spot and I'd therefore have 5 times the life.
 
No supercapacitors = bad idea for a ZLOG device, even if you can somehow reattach a log device with the new zpool versions from Oracle.
I heard that a new version of the 311 should come with the Cherryville refresh this month introducing 2x nm SLC chips, name's Hawley Creek, perhaps we will get supercapacitors with it.
 
Actually, that's not entirely true. I'll just be setting an HPA so that only 16GB is accessible or something...after all, I'm going to have a stripe of mirrors so there'll be 2 pairs to go round for double the space and IOPS :) 16GB would be great as there'd be 4 spare spots for every 1 used/data spot and I'd therefore have 5 times the life.

What's wrong about my statement? :)

I'm sure you can add log devices as mirrors but I haven't seen anyone striping their ZILs. I'm not sure it's possible or recommended even if you do have both of them mirrored. I'm not sure you would gain much from it either, but that depends on your specific situation of course..

Do let us know what your experiences are if you do this!
 
What's wrong about my statement? :)

What's wrong is that it's only true with trim or an eventual file system aware controller (which will never be ZFS aware).

Even if ZFS only uses a few GB per drive, pretty soon all the flash will be "dirty". AFAIK ZFS doesn't support trim.
 
TRIM is not wear leveling. Wear leveling goes on at the controller level, TRIM is the collaboration between OS and controller for garbage collection. Wear leveling is something all (most newer anyway) SSD's do, no matter what OS it's running under.
 
any unused space will be used for wear leveling, the drive doesn't care if it's partitioned or not
This part isn't correct and there are several reports from users that it isn't the case. You need to actually partition the drive or HPA it for the wear levelling etc. to work properly with the over-provision.
 
I'm fairly sure that your statement is incorrect, and Wiki seems to agree with me:
Wikipedia said:
Free user space

The SSD controller will use any free blocks on the SSD for garbage collection and wear leveling. The portion of the user capacity which is free from user data (either already TRIMed or never written in the first place) will look the same as over-provisioning space (until the user saves new data to the SSD). If the user only saves data consuming 1/2 of the total user capacity of the drive, the other half of the user capacity will look like additional over-provisioning (as long as the TRIM command is supported in the system).

I know this is Wikipedia, but there are 2 independent references that both seem to agree and none that do not.
 
I'm fairly sure that your statement is incorrect, and Wiki seems to agree with me:


I know this is Wikipedia, but there are 2 independent references that both seem to agree and none that do not.
I've read both of the references and they don't seem to address this at all.

My understanding (which is in part based on an Intel white paper than I remember reading) is that the SSD controller views free space as different to unpartioned/HPA'd space. Free space is used for wear levelling but once written it needs to be TRIM'd. The latter doesn't need to be because the controller knows it is unallocated/over-provisioned.
 
I found this...
http://ubuntuforums.org/showpost.php?p=9106293&postcount=27 said:
Free space: This should be interpreted as NAND-pages/blocks which are interpreted by the SSD-controler as free space. And since the only way to occupy that space is when the OS writes to a LBA and send that command to the drive the following situations are in play:

If you wipe your drive clean and you apply an HPA (Host Protected Area) that space is free.
If you wipe your drive clean and you under-partition your drive (a.k.a. over-provisioning) that space is free.
If you wipe your drive clean and have native TRIM running on your file-system (and the trim-implemantation is to do everything), all your free space in your file-system is free space.
If you wipe your drive clean and have native TRIM running on your file-system (and the trim-implemantation is only best effort), a large part of your free space in your file-system is free space, but not nessecarely all.
If you wipe your drive clean and don't have TRIM running, not all free space in your file-system is free space. You're dependent on the LBA-reuse implementation of your OS and the min amount of free space in your filesystem you had since your last cleaned your drive.
 
I see your point.

I guess it's up to the controller. If the controller can effectively clean dirty blocks without TRIM it still shouldn't matter, but HPA is a way to make sure that the controller has plenty of free space to use for wear leveling, and seeing as the drive will not be used for anything else, it will not matter.
 
I see your point.

I guess it's up to the controller. If the controller can effectively clean dirty blocks without TRIM it still shouldn't matter, but HPA is a way to make sure that the controller has plenty of free space to use for wear leveling, and seeing as the drive will not be used for anything else, it will not matter.
At the end of the day I don't really care how I have to set my drive up, especially as creating an HPA is easy anyway, but it'd be interesting to know what the truth is--could be controller-dependent...
 
Intel SSD 320 4K write IOPS: 23000
Intel SSD 311 4K write IOPS: 3300 (not a typo)

Intel SSD 320 4K read IOPS: 39500
Intel SSD 311 4K read IOPS: 37000

Just to clarify - those are 4k RANDOM writes quoted for the 311 (320 is 6600 for the 40 GB model, should be higher for the 80 GB). I am not sure if ZFS's ZIL writes are random or not.
 
I believe in normal operation the ZIL should be entirely sequential writes.
 
Yeah? That is good news!

For sequential writes the Intel 311 does 100 MB/sec on my system. I did not put them (I have 2) into my server yet as I am not convinced it would give me any benefit as most of my data is accessed sequentially anyways.

I only bought 2x 311s because they were on sale for $60 each a few weeks ago. One day I will figure out a better use for them! (Maybe for a small PfSense install).
 
==============
Correct Intel Specs
==============

Intel SSD 320 40GB 4K Random Write (100% Span) IOPS: 300
Intel SSD 320 80GB 4K Random Write (100% Span) IOPS: 300

Intel SSD 320 40GB 4K Random Write (8GB Span) IOPS: 3,700
Intel SSD 320 80GB 4K Random Write (8GB Span) IOPS: 10,000
Intel SSD 311 20GB 4K Random Write (8GB Span) IOPS: 3,300

------------------------------------------------------------------------------------------------

Intel SSD 320 40GB 4K Random Read (8GB/100% Span) IOPS: 30,000
Intel SSD 320 80GB 4K Random Read (8GB/100% Span) IOPS: 38,000
Intel SSD 311 20GB 4K Random Read (8GB Span) IOPS: 37,000

------------------------------------------------------------------------------------------------

Intel SSD 320 40GB 128K Sequential Write MB/s: 45
Intel SSD 320 80GB 128K Sequential Write MB/s: 90
Intel SSD 311 20GB ?K Sequential Write MB/s: 100

------------------------------------------------------------------------------------------------

Intel SSD 320 40GB 128K Sequential Read MB/s: 200
Intel SSD 320 80GB 128K Sequential Read MB/s: 270
Intel SSD 311 20GB ?K Sequential Read MB/s: 190

------------------------------------------------------------------------------------------------

Intel SSD 320 Series Write Latency µs: 65
Intel SSD 311 Series Write Latency µs: 75

------------------------------------------------------------------------------------------------

Intel SSD 320 Series Read Latency µs: 75
Intel SSD 311 Series Read Latency µs: 90
 
Last edited:
==============
Correct Intel Specs
==============

....
------------------------------------------------------------------------------------------------

Intel SSD 320 Series Write Latency µs: 65
Intel SSD 311 Series Write Latency µs: 75

------------------------------------------------------------------------------------------------

Intel SSD 320 Series Read Latency µs: 75
Intel SSD 311 Series Read Latency µs: 90

Hi,
is it possible, that you interchanged the specs?
The intel spec brief:

311 read: 65µs
311 write: 75µs

320 read: 75µs
320 write: 90µs

Has someone new experiences about the intel 320 or 311 sdd for slog device?
Some wirte tests would be great.

Greets,
ghandalf
 
My mirror intel 311 stops right around 3100 iops to it. and stays steady there for hours, while I was doing some large nfs transfers.

During those writes, it only managed to get to about 40MB/sec, but I also had lots of other write operations happening also.
 
Back
Top