Assistance - Calling all Samsung EVO owners

[Aurhinius]

Hi all,

I am looking to collate some information off Samsung EVO owners please.

I seem to have detected some pretty bad read performance degradation on 2 Samsung EVO 256GB drives that only manifests itself if you delve a bit deeper in to your drives performance.

I have tested on 2 different drives on 2 completely different spec pc's bought at completely different times. Both drives are used as an OS disk and have had rapid mode enabled for some time.

I am seeing read transfer speeds on areas of the disk that have been written too as low as 50MB/s!! When tested with HDTune read benchmark. Once the benchmark moves in to empty space the drive returns to your standard 450MB/s.

So if you have 2 mins to spare download hdtune free edition and run a read benchmark on your EVO. Please only run it on a in use drive with some data on as running on a secured erase drive will only show perfect performance of course.

Please download and use the Trial version of HD tune 5.50 - The free version has proven to give odd results - This doesn't appear to have effected the original drives issues so this is still a problem

If you can post a screenshot all the better.

I have seen one other user post on tomshardware with this issue and wonder if something perhaps in the latest firmware is to blame.

We shall see!

 

[Aurhinius]

For reference I do not see this behaviour on a heavily used Sandisk SSD on the same PC.

 

[Aurhinius]

Thanks for this guys the more people that can test and report back we can build up a case here.

It does seem like this is a drive specific issue because as said my sandisk does not exhibit this behaviour with the same version of HDtune so it can't be the HDtune benchmark.

This investigation by me was prompted by a strangely slow Acronis backup.

I have used Acronis to image my C drive(Samsung EVO) to D drive (Sandisk).

When I do this the copy size is approx 50GB and estimated completion time was 40 mins. When I did the calculation of how many MB/s this was I was seeing approx 50MB/s copy speed.

For an SSD that is VERY bad. When I fired up perfmon and resource mon I could see that the drives were manage exactly that an average of 50MB/s copy speed and queue depth's looked fine.

Digging further I tested in acronis a copy of D drive (sandisk) to C Drive (EVO) this produced a backup time of 10 mins for 70GB. Checking perfmon and resource mon I was seeing copy speeds at approx 250-300MB/s which is what I would expect for an SSD to SSD copy.

Testing further if you do a straight windows file copy again you see the same read performance from the EVO capped around the 50MB/s and an order of magnitude slower than the Sandisk to EVO copy.

So it seems the issue is the reading of data from the EVO drive in sections of the drive that has data on it. Any benchmark I tried i.e AS-SSD bench and copy test, samsung bench etc use spare drive section so report normal performance. With rapid enabled you get artificially inflated scores.

It's only HDTune that shows this well - haven't tried ATTO or HDtach.

It's not a benchmark only issue though because as I have said it shows up in Acronis drive backups - Both 11, 2011 and 2014 versions. I tried them all and standard windows copy.

I have tried with and without samsung magician installed as well and same behaviour.

So what is going on with this drive... firmware bug?

Definitely something wrong here.


Update -

For my drive it does seem to be the first 75GB of data on the drive. For my second drive which is a bit newer but the same size the problem is limited to the first 30-40GB but spreads out across the drive a bit more with slightly better peaks and troughs.

Not sure if it is an age related issue which gets progressively worse over time as more data gets written to the same section of the drive but you would expect wear level algoritims to prevent this kind of slow down. It's abnormally slow for an SSD drive read worse than a spinning HDD pretty much.

TRIM is and always has been enabled.

Have also checked with rapid enabled/disabled and write cache flushing and no difference.

 

[Aurhinius]

Thanks for the info Rotor I was merely corallating HDtune results with what I am seeing on the drive.

It seems mysterious that we are all seeing really low read performance from these drives when in the first section of the disk first half I would say. How that physically manifests itself in reality I don't know. It would make no sense as per your data distribution for the first part of the disk to provide such bad read performance but for this to correct deeper in to the drive.

If data is randomly written to individual chips and cells as I agree it actually is you would expect consistant read performance across the drive in the most part?

It's certainly not the case though.

Interestingly the write performance of the drive seems solid this is definitely a read performance issue and I encourage everyone to please test and post your results it takes 2 mins to run the test and you can stop it part the way through.

All we need to see is the initial sections of the drive.

Have you got any you can test Abarrass at OCuk see if you can replicate the issue?

I'd like to see this escalated to Samsung to see if we can get an answer. It might be by design and there's something not showing up in the reviews of all these drives but 50MB/s read from a drive that should produce 400-500MB/s isn't good and it's slower than a spinning HDD unit. That's a big problem for me and isn't just limited to benchmarks. I'm seeing this effect backup speeds and file copy speeds.

Frustratingly I can't test with the old firmware before they added the drive encryption to it.

So get testing folks and please post your results.

 

[Aurhinius]

More tests done by me -

HD Tune Pro: Samsung SSD 840 EVO 250G Extra Tests

Test capacity: 40 gB

Sequential outer 1354 IOPS 0.739 ms 84.615 MB/s
Sequential middle 497 IOPS 2.011 ms 31.080 MB/s
Sequential inner 605 IOPS 1.652 ms 37.840 MB/s

Cache test -

Cache
0.5 MB 307.540 MB/s
1.0 MB 374.739 MB/s
1.5 MB 140.910 MB/s
2.0 MB 133.268 MB/s
2.5 MB 105.347 MB/s
3.0 MB 94.682 MB/s
3.5 MB 94.810 MB/s
4.0 MB 83.930 MB/s
4.5 MB 84.106 MB/s
5.0 MB 76.919 MB/s
5.5 MB 78.549 MB/s
6.0 MB 77.862 MB/s
6.5 MB 76.549 MB/s
7.0 MB 74.698 MB/s
7.5 MB 76.576 MB/s
8.0 MB 79.839 MB/s
8.5 MB 78.502 MB/s
9.0 MB 77.004 MB/s
9.5 MB 77.374 MB/s
10.0 MB 76.680 MB/s
10.5 MB 76.512 MB/s
-----------------------------------------------------------------------

HD Tune Pro: SanDisk SDSSDHP256G Extra Tests

Test capacity: 40 gB

Sequential outer 7432 IOPS 0.135 ms 464.516 MB/s
Sequential middle 7415 IOPS 0.135 ms 463.436 MB/s
Sequential inner 7443 IOPS 0.134 ms 465.206 MB/s

Cache test -

0.5 MB 345.252 MB/s
1.0 MB 370.073 MB/s
1.5 MB 401.771 MB/s
2.0 MB 345.148 MB/s
2.5 MB 406.401 MB/s
3.0 MB 383.858 MB/s
3.5 MB 424.054 MB/s
4.0 MB 415.929 MB/s
4.5 MB 435.463 MB/s
5.0 MB 429.364 MB/s
5.5 MB 436.983 MB/s
6.0 MB 437.196 MB/s
6.5 MB 440.953 MB/s
7.0 MB 442.590 MB/s
7.5 MB 443.842 MB/s
8.0 MB 446.483 MB/s
8.5 MB 447.313 MB/s
9.0 MB 447.493 MB/s
9.5 MB 450.973 MB/s
10.0 MB 450.525 MB/s
10.5 MB 451.576 MB/s

This is me doing the HDTune short stroking the drives to the 40GB limit.

Both drive with a 250GB capacity

Startling really isn't it looking at the response time and speed of the Samsung compared to the sandisk. Both drives are about 50% full. Page file is on the sandisk interestingly enough.

 

[Aurhinius]

A couple of more tests at 2GB drive size

HD Tune Pro: Samsung SSD 840 EVO 250G Benchmark

Test capacity: 2 gB

Read transfer rate
Transfer Rate Minimum : 23.2 MB/s
Transfer Rate Maximum : 480.5 MB/s
Transfer Rate Average : 62.6 MB/s
Access Time : 0.217 ms
Burst Rate : 75.2 MB/s
CPU Usage : 0.3%

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

HD Tune Pro: SanDisk SDSSDHP256G Benchmark

Test capacity: 2 gB

Read transfer rate
Transfer Rate Minimum : 321.3 MB/s
Transfer Rate Maximum : 474.3 MB/s
Transfer Rate Average : 433.5 MB/s
Access Time : 0.080 ms
Burst Rate : 142.8 MB/s
CPU Usage : 1.2%

 

[Aurhinius]

Another good test John.

I think there is enough here to ask for samsung to check to see if they can replicate.

I'm seeing it on two used EVo 250GB, John has a EVO 120GB as does a few other people posting.

You won't see this on a clean secure erased drive obviously and I am sure secure erasing the drive will restore it's performance but that's not the point is it.

My EVO is about 50% full with the OS installed and programs/games on both drives I have and two different pc's.

Keep posting though owners if you can take the 2 mins to do it the more people we can show with the issue the more chance we have of narrowing down what causes this.

 

[Aurhinius]

Got pretty similar results as per you John using HD speed but that's what we would have expected.

e-mailed Samsung support. Suspect to get back scripted rubbish from tier 1 support fingers crossed though!

 

[Aurhinius]

I don't deny that performance seems fine on the drive.

Mine boots fine, loads things fine and plays fine but I'm using rapid which would mask the issue anyway to some degree.

It's not just a benchmark issue though as I have said I see the performance issue when attempting an acronis backup of the drive.

So there is an issue there just not entirely sure what.

It may well be that the drive hand off between dram cache and drive is bad with small files.

It stands to reason that the first section of my drive will have the OS installed on it which will be packed full of tiny files.

Which may well be the issue that the drive read speed for lots of small files is really bad (by design or bug) and that larger files don't exhibit this issue.

Throwing idea's out there.

Please test your drive regardless of whether you think it is fine though. It's not a witch hunt it's trying to understand an issue we have identified and judging how wide an issue it is.

Takes 2 mins to download and run the HDTune benchmark and it's free.

Ultimately a drive's read performance should not be effected by age nor file location on the drive. The drive should be able to read a data block on any nand cell at the same speed regardless of age but as this proves it just isn't happening. i.e reading cell 200 on chip 1 should be equally as fast as cell 1 on chip 4.

Locating the nand cell and then reading the data out should be the same.

As john and I have show we've tested sequential read's here so we are asking it to read cell1,2,3,4 in order which should be faster than asking for random cells but again it's not. It may well be how the data is distributed over cell boundaries or chip boundaries I don't know.

So although the file is sequential on the drive virtually it's not physically sequential and data could be distributed across chips I believe this is how the internal firmware file map would work. You would expect it to write a file across multiple chips to improve read access time as the controller has multiple channels. So faster to read all channels in one cycle rather than having to cycle the same chip to read out of it multiple times.

That may well explain the issue it may well be that samsung's attempt to limit drive wear is forcing small amounts of data to be less well distributed across the chips.

The more I type the more I think this makes sense with my limited understanding of nand and SSD technology.

 

[Aurhinius]

Had a response from Samsung to fill in a form etc so lets see what they come back with.

 

[Aurhinius]

Yeah Selekt0r you certainly are seeing the nasty sub 100MB for a good part of your drive.

The 2.7MB trough looks a bit dramatic as well.

Definitely a pattern here and I think it's perhaps a dirty reality of TLC nand but only shows up in certain benchmarks.

No word from Samsung so far.

 

[Aurhinius]

Toytown that is pretty dramatic and intersting given your detected issues with DB performance.

Once again showing this isn't a benchmark only issue.

Please keep posting folks we need more people with more EVO drives to run this very quick benchmark.

 

[Aurhinius]

Your 1 TB drive is looking fairly healthy there surveyor?

How old is it? recently secure erased?

It's still showing signs of degradation compared to the 830 though but no where near yet as dramatic as others.

So it's on it's way to the lows of under 100MB performance but not there yet.

 

[Aurhinius]

You probably don't know you have the problem Luke because you don't see it unless you look for it.

Doesn't mean it's not there.

Try running HDTune.

Ignorance is bliss in this case.

 

[Aurhinius]

Samsung have come back to me and said they are double checking the issue and will come back to me

That's something though

 

[Aurhinius]

Now that is interesting especially after a format.

That pulsing/step effect was something I was seeing and thought it rather odd.

Seems almost cyclic like buffer filling and being flushed over and over.

Unfortunately I can't secure erase my drive without major hassle so not going to do it.

Keep the tests coming. It's quite clear these drives are inconsistant.

The intresting thing is I doubt you would see this performance on a similarly cheap drive now like the MX100. Which makes the samsung drive pretty poor.

I think people are not reporting this issue because rapid mode helpds to mask it along with the rapid access time of an SSD even one that isn't working at peak performance.

 

[Aurhinius]

I got the form e-mailed to me once I e-mailed their support and they have been really genuine and helpful so far giving me the impression they will look at it.

So I would just e-mail them because they ask for a few other things e-mailed. I also attached some HD tune comparrison runs against my equally full Sandisk to prove the issue.

The end result remains to be seen.

Excellent news Abarrass, please let us know

 

[Aurhinius]

Certainly unless they can fix this I would go with the MX100.

You get a proven marvell controller and proven MLC nand rather than the TLC nand in the evo which could be the source of this issue if it is not a firmware bug.

 

[Aurhinius]

Could well be that small block read performance is really bad for some reason.

 

[Aurhinius]

Yes the full format results are the most interesting. I do wonder if a secure erase would have produced the same though.