Jump to content


Photo

Samsung SSD 840 Pro Enterprise SSD Review Discussion


  • You cannot start a new topic
  • Please log in to reply
42 replies to this topic

#1 Kevin OBrien

Kevin OBrien

    StorageReview Editor

  • Admin
  • 1,396 posts

Posted 15 April 2014 - 11:00 AM

The Samsung SSD 840 Pro isn't a new SSD, it was released 18 months ago and targeted toward the client market. At the time, we did a complete review of the SSD 840 Pro and found it to be a very capable SSD for its targeted use cases. A funny thing has happened though over time. Web hosts and many others looking for a good SSD with moderate write endurance have been turning to the SSD 840 Pro. It's a lower cost alternative than duty-built drives designed for enterprise users and it still offers a good mix of performance and write endurance. This phenomenon isn't new, even hybrid and all flash arrays have been "guilty" of using client SSDs as their mainstay for many years. Despite the best intentions of the SSD vendors who would prefer enterprise users buy their enterprise drives, the cost equation sometimes wins out, hence the 840 Pro wiggling itself into more and more server deployments, despite its relative age and client focus.

 

Samsung SSD 840 Pro Enterprise SSD Review


#2 pharaoh

pharaoh

    Member

  • Member
  • 15 posts

Posted 15 April 2014 - 11:23 AM

Brian, HUGE thank you! you are a god! :D

So it seems the S3700, when it comes to latency, especially for MySQL, is a better/faster choice than the 840, or am I reading the numbers wrong? 


#3 Brian

Brian

    SR Admin

  • Admin
  • 5,076 posts

Posted 15 April 2014 - 12:26 PM

You have it right. We'll also be adding SQL performance at some point, just don't have all the competitive data. 


Brian

Publisher- StorageReview.com
Twitter - @StorageReview

 

#4 unityole

unityole

    Member

  • Member
  • 43 posts

Posted 15 April 2014 - 11:55 PM

I want to correct something Brian.  Samsung 840 SSD line up is a joke when there are files filled into the drive especially doing heavy read/write work load.  I understand the tests here and how it's used but this goes against Anandtech review and especially Tweaktown's new consistency test.

 

840 pro 256gb proven to be a joke even in raid 0 was out performed by a single sandisk extreme II.  this is especially true in high end workload with load of queue depth and I have proven with my samsung 840 pro runs MUCH MUCH slower when i load multiple vmware from it instead from my OCZ Vector and sandisk extreme II.

 

also, 840 pro doesn't have OP, it'll get destroyed by any of the enterprise drive with heavy OP. please correct some of the information in the review in future, either stated light workload chart, or drive was tested empty.

 

btw, the charts and color switches which is pretty confusing..


Edited by unityole, 15 April 2014 - 11:57 PM.

#5 Rob0

Rob0

    Member

  • Member
  • 2 posts

Posted 16 April 2014 - 05:23 AM

Hahahaha I was doing this same benchmarking myself over the weekend. 2x128GB Raid 0 with 4k/4k stripe and allocation vs 256GB with Rapid enabled. I was comparing load data infile times for InnoDB with a 20GB flat file that produced 300 plus columns and 120M rows...I'll post back my results when I get to work.

#6 Rob0

Rob0

    Member

  • Member
  • 2 posts

Posted 16 April 2014 - 06:16 AM

I want to correct something Brian.  Samsung 840 SSD line up is a joke when there are files filled into the drive especially doing heavy read/write work load.  I understand the tests here and how it's used but this goes against Anandtech review and especially Tweaktown's new consistency test.
 
840 pro 256gb proven to be a joke even in raid 0 was out performed by a single sandisk extreme II.  this is especially true in high end workload with load of queue depth and I have proven with my samsung 840 pro runs MUCH MUCH slower when i load multiple vmware from it instead from my OCZ Vector and sandisk extreme II.
 
also, 840 pro doesn't have OP, it'll get destroyed by any of the enterprise drive with heavy OP. please correct some of the information in the review in future, either stated light workload chart, or drive was tested empty.
 
btw, the charts and color switches which is pretty confusing..




I didn't notice any significant latency while running HBase on a small Hadoop VM cluster off an 840 Pro Raid 0. It's all about stripe to allocation size. If you want responsiveness you'd better get it right. You can experience a 10x variation in 4kQ1 random writes across different settings. The 840 Pro for the price is no slouch, certainly not a joke. And with it's low price and 5 year warranty that's why it's in the "big data" data center.

And SR got it right about over provisioning. 840 Pro has OP via the magician software, but it doesn't support OP in Raid.

What high end workloads are you running?

#7 Kevin OBrien

Kevin OBrien

    StorageReview Editor

  • Admin
  • 1,396 posts

Posted 16 April 2014 - 08:49 AM

On our workloads, everything is brought into steady-state with heavy drive workloads, TPC-C in MySQL or NoSQL in our MarkLogic test are by no means "light enterprise". Each drive is secure erased and in the case of the MarkLogic test hammered for upwards of 72 hours.

 

Also you can always OP the drive using something like hdparm, which makes that change semi-permanent. That way it would carry its smaller size over into RAID environments.


#8 Brian

Brian

    SR Admin

  • Admin
  • 5,076 posts

Posted 16 April 2014 - 12:47 PM

I want to correct something Brian.  Samsung 840 SSD line up is a joke when there are files filled into the drive especially doing heavy read/write work load.  I understand the tests here and how it's used but this goes against Anandtech review and especially Tweaktown's new consistency test.

 

840 pro 256gb proven to be a joke even in raid 0 was out performed by a single sandisk extreme II.  this is especially true in high end workload with load of queue depth and I have proven with my samsung 840 pro runs MUCH MUCH slower when i load multiple vmware from it instead from my OCZ Vector and sandisk extreme II.

 

also, 840 pro doesn't have OP, it'll get destroyed by any of the enterprise drive with heavy OP. please correct some of the information in the review in future, either stated light workload chart, or drive was tested empty.

 

btw, the charts and color switches which is pretty confusing..

 

Yeah, we know we need to work on the charts to make them less confusing to read in a row, thanks for that pointer. 

 

The drive isn't intended to be in heavy write use cases, per our disclaimers in the review. That said, I'd take our real application tests as a better indicator of performance than synthetics. 


One other comment I'd add is that Pure Storage used to use SSD 830s in their AFAs. Not sure what's in there today but we've seen a lot of high-end client drives show up in AFAs or Hybrid Arrays. 


Brian

Publisher- StorageReview.com
Twitter - @StorageReview

 

#9 Kevin OBrien

Kevin OBrien

    StorageReview Editor

  • Admin
  • 1,396 posts

Posted 16 April 2014 - 01:48 PM

I want to correct something Brian.  Samsung 840 SSD line up is a joke when there are files filled into the drive especially doing heavy read/write work load.  I understand the tests here and how it's used but this goes against Anandtech review and especially Tweaktown's new consistency test.

 

Missed this comment but wanted touch back in it now. Synthetic consistency testing has very little value in the whole scheme of things when you look at actual enterprise workloads. The reason it goes against their testing is synthetic workloads try to replicate real-world testing by matching transfer sizes, but actual application testing looks at how the drive performs in harmony with the entire system around it. If those two don't correlate then the synthetic testing of it should be called into question. We had this same problem early on and have since moved to limit as much synthetic testing as possible and replace it with application tests.

 

Consistency is great, but show it in something that is meaningful that helps a company's bottom line. Noone buys SSDs for the sole purpose of hammering them with IOMeter, vdBench or FIO. They buy them to drive more value out of a compute platform that has an I/O bottleneck.


#10 unityole

unityole

    Member

  • Member
  • 43 posts

Posted 16 April 2014 - 10:40 PM

 

Missed this comment but wanted touch back in it now. Synthetic consistency testing has very little value in the whole scheme of things when you look at actual enterprise workloads. The reason it goes against their testing is synthetic workloads try to replicate real-world testing by matching transfer sizes, but actual application testing looks at how the drive performs in harmony with the entire system around it. If those two don't correlate then the synthetic testing of it should be called into question. We had this same problem early on and have since moved to limit as much synthetic testing as possible and replace it with application tests.

 

Consistency is great, but show it in something that is meaningful that helps a company's bottom line. Noone buys SSDs for the sole purpose of hammering them with IOMeter, vdBench or FIO. They buy them to drive more value out of a compute platform that has an I/O bottleneck.

 

thanks for the quick reply.

 

I want to mention that samsung is such a company to target these synthetic tests and benchmark software to get better score.  I am not trying to bring in their tampering with their galaxy phone on GPU benchmark but this is indeed the case, especially with rapid mode, overclocked controller to obtain higher numbers.

 

now as for storagereview's test method, it would be nice to see some real software running/encoding etc.  I understand samsung SSD isn't meant for heavy workload but the tests shown in review were meant for heavy workload or server type workload is it not? please correct me if I'm wrong.

 

I have done some basic tests myself such as running multiple vmware (starting up to shutting them down and repeat for a few minutes) to encoding video from RAM disk to targeted single SSD in order to bring out obvious bottleneck in the entire process and Samsung 840 pro and evo didn't do well.  compare to OCZ vector, corsair neutron and other SSDs like intel s3700.


#11 Brian

Brian

    SR Admin

  • Admin
  • 5,076 posts

Posted 17 April 2014 - 07:02 AM

What do you mean by "real software?" You don't find things like MarkLogic real? I'm sure they'd argue the point since they make money selling their NoSQL.

Brian

Publisher- StorageReview.com
Twitter - @StorageReview

 

#12 Kevin OBrien

Kevin OBrien

    StorageReview Editor

  • Admin
  • 1,396 posts

Posted 17 April 2014 - 08:14 AM

I'm as perplexed as Brian is on the comment towards real software for testing. What is more heavy-duty than actual Microsoft SQL Server, MySQL, or MarkLogic NoSQL databases running actual workloads on a given storage device? This isn't a PCMark or video encode level of test, this is very large scale database testing. If Samsung has found a way to make their SSD perform better in any of those areas, I'll applaud them since that performance measured is how those drives would hold up in a production setting. We poll latency from the database itself when running those tests, not from any sort of synthetic loadgen and extrapolate what it could perform like.

 

On the VMware comment when we have sufficient samples we run VMmark on SSDs, although it is very capacity intensive. We would need 16 512GB samples from Samsung.

 

Currently only Samsung has stepped up to the plate on the consumer/prosumer side to have us test their 840 Pro in an enterprise setting with application tests. We have reached out to OCZ and others for samples of (4) 480GB+ samples, but so far none have sent them in. We wrapped up testing on the Intel S3500 recently across these workloads since we had (4) 480GB samples, although for the S3700 we don't have samples of high enough capacity to run the tests. That said the S3700 even in a sample of one had no trouble running these workloads and has already proven itself handily.


#13 unityole

unityole

    Member

  • Member
  • 43 posts

Posted 17 April 2014 - 07:23 PM

if that is the case then is there a reason why anandtech and tweaktown both show samsung 840 pro to be a weak drive when doing heavy workload?  it makes little sense to me when both these sites also tested seagate 600 pro which is heavy in OP, same goes for s3700 but both doing worse than 840 pro which is a not enterprise drive.  I believe Anand runs an intel test while Tweaktown have their own test mythology as well as few other well known benchmark software.

 

also thanks for mentioning s3700 being 200gb only so slower performance does make some sense.  however I would be more incline say a company such as samsung wouldn't give enterprise storage performance in consumer level while they still sell enterprise ssds.

 

now I don't know if MySQL is the heaviest real software test but what I meant by real software is, daily used program, not benchmark software, something that we used everyday which still imposes heavy read/writes on the SSD.  something like running multiple VMware (10+) off a single SSD while installing microsoft offices on all of them at the same time and time it.

 

I'm just saying benchmark software can easily be manipulated and also be paid to work better for different types of controller/firmware.  intel/amd chipset is one of those, intel work better for window, an example.

 

i mean, if you guys are able to explain why a drive such as 600 pro heavy OP would lose out to samsung with NO OP then I would definitely come around to accept sammy ssd as a faster one however  OP is known to provide best enterprise stability and performance in write workload so this review is contradict to that.


Edited by unityole, 17 April 2014 - 07:26 PM.

#14 Kevin OBrien

Kevin OBrien

    StorageReview Editor

  • Admin
  • 1,396 posts

Posted 17 April 2014 - 08:26 PM

if that is the case then is there a reason why anandtech and tweaktown both show samsung 840 pro to be a weak drive when doing heavy workload?  it makes little sense to me when both these sites also tested seagate 600 pro which is heavy in OP, same goes for s3700 but both doing worse than 840 pro which is a not enterprise drive.  I believe Anand runs an intel test while Tweaktown have their own test mythology as well as few other well known benchmark software.

 

also thanks for mentioning s3700 being 200gb only so slower performance does make some sense.  however I would be more incline say a company such as samsung wouldn't give enterprise storage performance in consumer level while they still sell enterprise ssds.

 

now I don't know if MySQL is the heaviest real software test but what I meant by real software is, daily used program, not benchmark software, something that we used everyday which still imposes heavy read/writes on the SSD.  something like running multiple VMware (10+) off a single SSD while installing microsoft offices on all of them at the same time and time it.

 

I'm just saying benchmark software can easily be manipulated and also be paid to work better for different types of controller/firmware.  intel/amd chipset is one of those, intel work better for window, an example.

 

i mean, if you guys are able to explain why a drive such as 600 pro heavy OP would lose out to samsung with NO OP then I would definitely come around to accept sammy ssd as a faster one however  OP is known to provide best enterprise stability and performance in write workload so this review is contradict to that.

 

I'm not sure what you are trying to ask about real software. The tests we run are *actual* instances of enterprise applications. They are not fake by any means. This isn't a consumer daily use platform... these are enterprise production settings. If a product does well in them, that product is said to be very good at X workload. Many companies try to optimize for different database workloads, this is not like trying to game a benchmark. If you design to be really good in SQL Server, you end up with a product that is in fact really good in SQL Server.

 

Also not all drives with any amount of increased OP perform well. Some drives are inherently strong in certain workloads, others are not.

 

Also those sites you mention don't do application testing. We do and that is drastically different. I'm not sure I can explain that any better. We have close ties with Dell, HP, Lenovo, NetApp and many other enterprise vendors specifically because of the application testing we perform. It shows the true picture of performance.

 

Take our MarkLogic NoSQL test. It is a 8-CPU license version of MarkLogic ingesting Wikipedia and the test measures how long it takes the database takes to ingest that data. If a drive does well in that, MarkLogic customers buy drives or solutions based on that information because it directly reflects real-world performance.


#15 JesperA

JesperA

    Member

  • Member
  • 1 posts

Posted 18 April 2014 - 01:08 PM

Replacing SQL and NoSQL enterprise test with benchmark of installing Microsoft Office? The internet sure is i weird place...


#16 Kevin OBrien

Kevin OBrien

    StorageReview Editor

  • Admin
  • 1,396 posts

Posted 18 April 2014 - 01:32 PM

My feelings entirely. :huh:


#17 unityole

unityole

    Member

  • Member
  • 43 posts

Posted 18 April 2014 - 05:50 PM

Replacing SQL and NoSQL enterprise test with benchmark of installing Microsoft Office? The internet sure is i weird place...

 

heh, what I mentioned was just an example lol.  do you run Mysql at home 24/7 now?  well I sure don't and I am sure a lot of people don't as well, while they run bunch of other software almost all the time.  vmware for fun or not I would say it is much more realistic for me for sure.

 

so far I only get a solution on how tests should be performed and no real reason as to why.. anyway, are you able to answer my questions?

 

question 1: your reviews showing samsung 840 pro with no OP perform better than s3700 and seagate 600 pro enterprise SSDs with heavy OP.  I define "perform better" as simple as, better overall consistency, speed in 4k, sequential, read/write.

 

question 2: if sammy 840 pro is indeed better, the real reason I'd see it is long life on it's flash.  is flash life the real reason for them to use it for enterprise work?  or is the reason because of my 1st question the overall performance is just better?

 

 

 

Take our MarkLogic NoSQL test. It is a 8-CPU license version of MarkLogic ingesting Wikipedia and the test measures how long it takes the database takes to ingest that data. If a drive does well in that, MarkLogic customers buy drives or solutions based on that information because it directly reflects real-world performance.

 

I would like to mention that having a great benchmark score, or vantage score doesn't reflect to real performance.  any test will have flaws and MarkLogic will be no exception.  while I am only seeking the truth for real performance, testing storage devices over and over with the same test measure will only end up being targeted by manufacturer.  they can alter in design, change in firmware, change in controller etc to get better performance out of your test results which is one thing I weigh a lot.

 

just curious, whens the last time you guys have replaced test measure with MarkLogic?


Edited by unityole, 18 April 2014 - 06:07 PM.

#18 Brian

Brian

    SR Admin

  • Admin
  • 5,076 posts

Posted 18 April 2014 - 07:06 PM

What do you mean by replacing test measure?

As to your other comments, we ran the pro in enterprise tests at the request of our forums members. If you don't like the results, ignore them. Complaining that a manufacturer may tune their drive for SQL in response is ludicrous. If in fact they do, then anyone with a SQL database will rejoice. That's why OCZ, Fusion-io and other are offering duty specific products for core apps like SQL and MySQL amongst others. You're failure to understand the enterprise storage space and arguing in favor of synthetic tests that don't show real activity is absurd.

Brian

Publisher- StorageReview.com
Twitter - @StorageReview

 

#19 Kevin OBrien

Kevin OBrien

    StorageReview Editor

  • Admin
  • 1,396 posts

Posted 18 April 2014 - 07:16 PM

I'm honestly giving up at trying to reply again, you just don't get it. Our core audience are enterprise buyers looking for products that will work well in a business environment, not home buyers buying very expensive hardware to use at home. That is the disconnect here.

 

Our MarkLogic "test" is the actual MarkLogic environment running in real life, and we just measure the time a known set of data takes to get ingested into it. If Samsung tweaked or screwed with their SSD to be really good at it, then yay for them because they designed a really good drive for NoSQL. They aren't gaming a test, they would literally have a really good drive for any workload they drived to get a good "score" from us in, because we use and test their product with exactly that software.

 

You are missing the bigger picture in all of this. Our tests are leveraging software that enterprise customers use in revenue generating production environments. Those customers use this stuff 24/7/365, where any down time equals money lost.


#20 Kdawgca

Kdawgca

    Modinator

  • Mod
  • 632 posts

Posted 19 April 2014 - 02:15 AM

Unityole,

 

Have you looked that the Samsung 840 pro NON-enterprise review?

 

.... While the SSD 840 Pro has a few weak points which may limit its appeal for certain high-write applications, the SSD 840 Pro maintained average performance even in benchmarks when it did not finish in the first three positions

 

.

http://www.storagere...rise_ssd_review

 

That seems in line with your observations.  As mentioned in the this review, enterprise customers may be looking for augment their storage space  with cheap(er) and faster solution. 840pro is a great option for that role. This review gives a good idea at what a  robust client SSD can do in enterprise environments, but with all purchases, it should be vetted to make sure it is a right option.


Storagereview.com on Twitter
SSD vs. HDD Guide

  • MSI H81M-E33 V2 - Haswell Intel Pentium G3220+ 60GB Vertex 2 / 500 GB HDD/6GB/EVGA 9600GSO/ Windows 7 + VMs//Dell Latitude D410 -1.6ghz Pentium M, 120GB IDE HDD, 2 GB DDR ram.

#21 sam4ya

sam4ya

    Member

  • Member
  • 2 posts

Posted 24 April 2014 - 09:55 AM

Two cents from a newbie here (hi to all, btw), with no experience on high end setups or enterprise situations like the rest of you, just a first time impression on running our own server and having put an Samsung 840 Pro as the drive.

 

We have a small i7 quad-core, 6 VM osx server running 24/7 with 4 Linux and 2 Win VMs. The VMs only take up about 8GB RAM from the 16 total and we recently upgraded to a Samsung 840 Pro 256GB, left about 32GB for OP but the speed of this setup is extremely slow, going down to barely 22MB/s write and 150MB/s read on the highest workload times in the host machine! Never mind the VMs themselves. We've never had more than 125BMB/s write and 190MB/s read at any point in time, even though the advertised speeds are 500+. TRIM is enabled, VMs optimised as per vmware docs and nothing else running on the system or taking up CPU/IO. This machine is fully dedicated to running the VMs.

 

Personally I find this extremely frustrating and wish there were "real world" benchmarks that would have given us a good idea of how this drive performs in production situations (specifically to VMs), as we have another machine with an intel SSD that keeps its speed even when full and performs equally when empty or at full workload. 

 

We are installing an OWC drive today to compare and see if the issue is with our setup or it really is the disk that just cant keep up for what we need. I don't feel qualified to discuss the quality of the drive itself, just want to second that for what we use it for (which as a small agency, its running VMs), its less than ideal, and compared to the intel drive, its not up to the task.

 

My two cents...other than that, great forum! Glad to be here :)


#22 Brian

Brian

    SR Admin

  • Admin
  • 5,076 posts

Posted 24 April 2014 - 10:35 AM

Thanks for jumping in. There are a lot of variables here but generally speaking a single SSD will get thrashed pretty hard with all that random activity, so I'm not surprised you're seeing issues. Might be worth a new thread to dive into your config specifically to see what can be tuned for better performance. 


Brian

Publisher- StorageReview.com
Twitter - @StorageReview

 

#23 sam4ya

sam4ya

    Member

  • Member
  • 2 posts

Posted 24 April 2014 - 06:08 PM

UPDATE: We installed the OWC drive and freed up more space on the 840, now the speeds have gone up to a staggering 467 Write, 511 Read. Amazing, simply amazing.

It already had 32GB free which is a lot of free space, but I guess it needs more to keep up the speed. Maybe its safe to say that unless you buy an enterprise disk the consumer SSD have to be half full to perform well?

 

Anyways, conclusion is we had it too full, now its running with the same workload, and I'm happy to say CPU load has also reduced dramatically and everything works smoothly. 


#24 Kevin OBrien

Kevin OBrien

    StorageReview Editor

  • Admin
  • 1,396 posts

Posted 24 April 2014 - 09:12 PM

Well if you want it to mimic an enterprise SSD, chop its usable space down to 400GB. 


#25 ChrisCT

ChrisCT

    Member

  • Member
  • 7 posts

Posted 01 May 2014 - 12:04 PM

Great discussion, thanks for all the great follow-up.

 

I'm going to ask this question and duck- would RAPID mode plus 112GB over-provision down to 400GB let the Samsung drive compete with the Intel S3700 200GB?  From a money standpoint, both drives are comparable, regardless of the capacity. I suspect RAPID in a enterprise deployment is a gimmick, but the over-provisioning would offer noticeable benefits.

 

This is for my SMB of 20 users with a 6GB SQL file, QBooks, very light web server, and some file serving.  





1 user(s) are reading this topic

0 members, 1 guests, 0 anonymous users