Samsung sm863a as slog?

[Martin Maisey]

Hi,

I’m looking to replace an aging Intel 320 80GB drive as a slog device with something more capable and less likely to give up the ghost. I’m running home lab VMs locally requiring synchronous writes, and even the 320 makes quite a difference to performance. Budget not high - hopefully around the £250 mark (or ideally even less, if possible).

I know the standard advice is to go NVMe given latency benefits, but the sole accessible PCIe slot on the riser is taken up with my HBA and my budget precludes that in any case. I also doubt my pool would keep up as it currently stands, or even when extended with another mirrored vdev, so it would probably be a waste even if I could do it. As a result, I've been looking at cheaper enterprise SATA SSDs. I’ve been looking out for a UK/European 200GB Intel DC3700 (preferably new, or little used) on eBay, but there seem to be limited options that aren’t secondhand with no information on how hard they’ve been hammered or how long they've been in use, which is a bit of a worry given how old these drives are. I’ve already bought one that was claimed to be a hot spare with no write workload, but it was DOA so I’m returning - will hopefully get my money back without too much trouble … at least it’s a UK seller so easy to post.

While researching alternatives, I’ve come across the Samsung SM863a (http://www.samsung.com/semiconductor/minisite/ssd/downloads/document/PM863a_and_SM863a_Brochure.pdf), which looks at least superficially attractive. I can get a 240GB drive new for £235 from a reputable UK VAR (https://www.ballicom.co.uk/samsung-...WT20J69hfD_8kDUkLbJINaJPPRQgiABR28aAmp48P8HAQ). It’s a write-orientated enterprise SSD with PLP, rated for 3.6 DWPD over 5 years, and claims 450MB/s sequential write performance and ‘extremely low, consistent latency’ - although I can’t, unfortunately, find any actual figures on the latter. Or any posts regarding use as a slog device.

That looks promising enough that I'm tempted to abandon my search for a DC3700, even though the latter is frequently recommended for low-end slog. Any thoughts/experience would be very much appreciated.

Martin

—

Chenbro RM23212-O12C 12-bay 2u chassis/backplane
Tyan S7012 motherboard
2 x E5645 6-core Xeon
128GB RAM
LSI 9211-16i HBA
4 x WD RED 4TB, mirrored/striped zpool with Intel 320 80GB slog
2 x Kingston SSDNow UV400 120GB mirrored boot zpool

 

[norbs]

I think the big requirement for SLOG devices is that it has a 'super capacitor' to ensure all writes are finished in event of a power outage. Otherwise you could use just about anything for it.

 

[Spearfoot]

Hi,

I’m looking to replace an aging Intel 320 80GB drive as a slog device with something more capable and less likely to give up the ghost. I’m running home lab VMs locally requiring synchronous writes, and even the 320 makes quite a difference to performance. Budget not high - hopefully around the £250 mark (or ideally even less, if possible).

I know the standard advice is to go NVMe given latency benefits, but the sole accessible PCIe slot on the riser is taken up with my HBA and my budget precludes that in any case. I also doubt my pool would keep up as it currently stands, or even when extended with another mirrored vdev, so it would probably be a waste even if I could do it. As a result, I've been looking at cheaper enterprise SATA SSDs. I’ve been looking out for a UK/European 200GB Intel DC3700 (preferably new, or little used) on eBay, but there seem to be limited options that aren’t secondhand with no information on how hard they’ve been hammered or how long they've been in use, which is a bit of a worry given how old these drives are. I’ve already bought one that was claimed to be a hot spare with no write workload, but it was DOA so I’m returning - will hopefully get my money back without too much trouble … at least it’s a UK seller so easy to post.

While researching alternatives, I’ve come across the Samsung SM863a (http://www.samsung.com/semiconductor/minisite/ssd/downloads/document/PM863a_and_SM863a_Brochure.pdf), which looks at least superficially attractive. I can get a 240GB drive new for £235 from a reputable UK VAR (https://www.ballicom.co.uk/samsung-...WT20J69hfD_8kDUkLbJINaJPPRQgiABR28aAmp48P8HAQ). It’s a write-orientated enterprise SSD with PLP, rated for 3.6 DWPD over 5 years, and claims 450MB/s sequential write performance and ‘extremely low, consistent latency’ - although I can’t, unfortunately, find any actual figures on the latter. Or any posts regarding use as a slog device.

That looks promising enough that I'm tempted to abandon my search for a DC3700, even though the latter is frequently recommended for low-end slog. Any thoughts/experience would be very much appreciated.

Martin

—

Chenbro RM23212-O12C 12-bay 2u chassis/backplane
Tyan S7012 motherboard
2 x E5645 6-core Xeon
128GB RAM
LSI 9211-16i HBA
4 x WD RED 4TB, mirrored/striped zpool with Intel 320 80GB slog
2 x Kingston SSDNow UV400 120GB mirrored boot zpool

From the looks of it, the Samsung may be a good ZIL SLOG device and the specs claim that it has 'power loss protection', which is something your Intel 320 device lacks. How will it stack up against an Intel S3700? Hard to say. Intel seems to be the industry leader when it comes to producing durable SSDs; especially the types needed for a SLOG device. As you pointed out, there's a dearth of information on the interwebz about real-world experience using the Samsung so you'll pretty much be a pioneer.

If you're just running a lab, you can always turn off synchronous writes on your datasets. Performance will fly! I ran my lab system this way for a year before adding a SLOG. The caveat, of course, is that you won't have any of the protections afforded by using a SLOG.

 

[Martin Maisey]

Thanks very much for the advice.

From the looks of it, the Samsung may be a good ZIL SLOG device and the specs claim that it has 'power loss protection', which is something your Intel 320 device lacks.

Actually, the Intel 320 is one of the very few cheap drives that *does* have PLP. It's not a supercapacitor, but an array of lower-powered capacitors. Moreover, there's even been some third-party analysis done that shows it actually appears to work under torture-test conditions - http://lkcl.net/reports/ssd_analysis.html - unlike every other consumer drive PLP tested. There was a reason I used it in the first place - I am aware there wouldn't be any significant benefits (and plenty of downsides) in using a slog which dropped writes on the floor under power loss ;-)

There is no doubt that it's not well suited to being a slog in almost every other regard, notably the appalling write endurance (10-60TB depending on workload), high latency (90 microseconds) and low sequential write bandwidth (90MB). https://www.intel.com/content/dam/w...torage-application-specification-addendum.pdf .

But one came spare from an old desktop system, I found I got some benefit, and left it there on the basis that a double-failure (drive fail then power fail under write load) is necessary to cause data loss, and even then integrity should be OK, so the risk was low. The Proxmox pfeperf tool reports ~2000 fsyncs/s, which is about an order of magnitude higher than I was getting without it IIRC.

However, the media wearout indicator is now 96%, it's a very aged consumer-orientated drive, and I now have a much more capable server I'm moving some more important workloads to. Time for a change.

How will it stack up against an Intel S3700? Hard to say. Intel seems to be the industry leader when it comes to producing durable SSDs; especially the types needed for a SLOG device. As you pointed out, there's a dearth of information on the interwebz about real-world experience using the Samsung so you'll pretty much be a pioneer.

Suspected that would be the case, and it is salutary that the study above concludes with saying that if you want data consistency under power failure conditions, use an Intel drive. OTOH, the study was scoped to exclude enterprise write-orientated drives, Samsung aren't exactly a fly-by-night operator, and this is a drive explicitly targeted at OLTP workloads, where people care about data integrity. And buying a new drive, under warranty, from a proper supplier - as opposed to scrabbling around on eBay - definitely has its attractions. Hmm.

If you're just running a lab, you can always turn off synchronous writes on your datasets. Performance will fly! I ran my lab system this way for a year before adding a SLOG. The caveat, of course, is that you won't have any of the protections afforded by using a SLOG.

Sorry, I was unclear on that - it's mostly lab VMs, but also some 'production' home workloads where I do care about my data and which benefit from slog. What I haven't done - and you're right, I should do - is separate out the truly throwaway lab stuff and set sync=disabled on that. Higher performance and no SSD wear - what's not to like? :)

—

Chenbro RM23212-O12C 12-bay 2u chassis/backplane
2 x E5645 6-core Xeon
128GB RAM
LSI 9211-16i HBA
4 x WD RED 4TB, mirrored in zpool with Intel 320 80GB slog
2 x Kingston SSDNov UV400 120GB in mirrored boot zpool

 

[Martin Maisey]

Follow-up: sync=disabled gets >14k fsyncs/second with pveperf

Will be interested to see what a proper slog device gets me.

 

[Spearfoot]

Actually, the Intel 320 is one of the very few cheap drives that *does* have PLP. It's not a supercapacitor, but an array of lower-powered capacitors. Moreover, there's even been some third-party analysis done that shows it actually appears to work under torture-test conditions - http://lkcl.net/reports/ssd_analysis.html - unlike every other consumer drive PLP tested. There was a reason I used it in the first place - I am aware there wouldn't be any significant benefits (and plenty of downsides) in using a slog which dropped writes on the floor under power loss ;-)

I stand corrected... thanks!

There is no doubt that it's not well suited to being a slog in almost every other regard, notably the appalling write endurance (10-60TB depending on workload), high latency (90 microseconds) and low sequential write bandwidth (90MB). https://www.intel.com/content/dam/w...torage-application-specification-addendum.pdf .

But one came spare from an old desktop system, I found I got some benefit, and left it there on the basis that a double-failure (drive fail then power fail under write load) is necessary to cause data loss, and even then integrity should be OK, so the risk was low. The Proxmox pfeperf tool reports ~2000 fsyncs/s, which is about an order of magnitude higher than I was getting without it IIRC.

However, the media wearout indicator is now 96%, it's a very aged consumer-orientated drive, and I now have a much more capable server I'm moving some more important workloads to. Time for a change.



Suspected that would be the case, and it is salutary that the study above concludes with saying that if you want data consistency under power failure conditions, use an Intel drive. OTOH, the study was scoped to exclude enterprise write-orientated drives, Samsung aren't exactly a fly-by-night operator, and this is a drive explicitly targeted at OLTP workloads, where people care about data integrity. And buying a new drive, under warranty, from a proper supplier - as opposed to scrabbling around on eBay - definitely has its attractions. Hmm.



Sorry, I was unclear on that - it's mostly lab VMs, but also some 'production' home workloads where I do care about my data and which benefit from slog. What I haven't done - and you're right, I should do - is separate out the truly throwaway lab stuff and set sync=disabled on that. Higher performance and no SSD wear - what's not to like? :)

—

Chenbro RM23212-O12C 12-bay 2u chassis/backplane
2 x E5645 6-core Xeon
128GB RAM
LSI 9211-16i HBA
4 x WD RED 4TB, mirrored in zpool with Intel 320 80GB slog
2 x Kingston SSDNov UV400 120GB in mirrored boot zpool

Yes, that 320 is indeed nearly worn out! If you decide to use the Samsung, please let us know how it works out: the users here are always on the lookout for good SLOG devices.

Good luck!

 

[Martin Maisey]

Yes, that 320 is indeed nearly worn out! If you decide to use the Samsung, please let us know how it works out: the users here are always on the lookout for good SLOG devices.

To be perfectly honest I should have been checking the SMART data a little more often, so it came as a bit of a surprise. Oops. It seems that FreeNAS doesn't mail you unless there's an actual error. Anyhow, I doubt I will be able to squeeze as much "value for money" out of a 3 DWPD 240GB drive, no matter how hard I try ;-)

One further question: are IOPS considered at all important on a slog (I don't *think* they are, but just wanted to check)? Although on paper it's better than the 3700 in sequential write, the random write IOPS are significantly lower. Which I guess shows just how good a drive the 3700 was in its day.

I probably will order the drive unless someone pops up in the next day saying "don't do that", and if so will report back. Are there any particular tests that would be good to run to help others assess performance?


Sent from my iPad using Tapatalk

 

[Spearfoot]

To be perfectly honest I should have been checking the SMART data a little more often, so it came as a bit of a surprise. Oops. It seems that FreeNAS doesn't mail you unless there's an actual error. Anyhow, I doubt I will be able to squeeze as much "value for money" out of a 3 DWPD 240GB drive, no matter how hard I try ;-)

You ought to install a script to regularly email you the SMART test results. Mine is available in the Resources section ("Github repository for FreeNAS scripts, including disk burnin").

One further question: are IOPS considered at all important on a slog (I don't *think* they are, but just wanted to check)? Although on paper it's better than the 3700 in sequential write, the random write IOPS are significantly lower. Which I guess shows just how good a drive the 3700 was in its day.

I believe sequential writes are most relevant to a SLOG device but if I'm wrong, someone will surely come along and correct me.

I probably will order the drive unless someone pops in the next day saying "don't do that", and if so will report back. Are there any particular tests that would be good to run to help others assess performance?

It would be great if someone really knowledgable and with more real-world experience (@jgreco or @cyberjock, for example) would stop by and enlighten us.

 

[Martin Maisey]

I had itchy fingers and have now ordered it anyway :)

Thanks for the pointer to the script, will install it.

 

[Martin Maisey]

I've got the drive now. Interestingly, pveperf reports 2000 fsyncs/sec, basically the same as with the Intel 320. I suspect it's an extremely simple microbenchmark, though, and I haven't really researched what it's doing under the covers.

Anyway, regardless of the numbers, I'm happy that I'm now using a drive that isn't nearly dead and where the design appears broadly right for a slog workload.

As I mentioned earlier, if anyone's got better benchmarks they'd like me to execute, I'm happy to. Also happy to receive an Intel s3700 as a gift, if someone wants a head to head ;-)