Samsung 960 EVO NVMe SSD Benchmarks On Linux

[caligula]

These speeds are insane! I will not upgrade because the only applications I have that require drive speed are games and virtual machines. Both games and VMs requires too much capacity and it's just not practical to put it on SSD at this stage. I also don't have M.2 on my motherboard. My Crucial MX200 keeps me happy for now.

Plenty of people still have old style hard drives. They will eventually buy new computers, they'l want to avoid the hassle with cables (thus M.2) and indeed want a faster system than the old generation SATA SSDs provided. So, for them this is a natural upgrade step. For desktop Linux users upgrading up to date systems, the speedups are getting pretty small. You might save 1-2 seconds in boot time, at most. When compiling, you should anyways be using RAM unless the project is huge.

[edwaleni]

I have the 950 Pro 256Gb model. It is the slowest performing of the different sizes available in the M.2 NVMe line, though still incredibly fast. I believe testing in other press showed the 1Tb version from Samsung has shown much better performance overall than the smaller sizes. Not sure if that is a controller issue, caching or firmware preferences set by Samsung to push sales. I will look up any reviews on the 1Tb 960 Pro and see if they are seeing the same thing.

[Holograph]

I have the 950 Pro 256Gb model. It is the slowest performing of the different sizes available in the M.2 NVMe line, though still incredibly fast. I believe testing in other press showed the 1Tb version from Samsung has shown much better performance overall than the smaller sizes. Not sure if that is a controller issue, caching or firmware preferences set by Samsung to push sales. I will look up any reviews on the 1Tb 960 Pro and see if they are seeing the same thing.

It is always the case that large SSDs are faster than small SSDs when the small SSD doesn't use all of the flash channels on the controller and everything else is equal. Larger SSDs may also have more DRAM cache in some cases.

Also, every single model of the 950 is as fast or faster than any 850 pro in every single benchmark in Windows-land, so like I said a few posts ago, there is definitely a performance bottleneck in the Linux kernel somewhere (or a flaw in the benchmarking methods used). I don't know enough about the 960 series to claim the same, but I'd strongly lean towards the same being true.

[darkbasic]

Michael Did you align the partitions correctly? Since it is (Samsung) TLC I think you should align to 1536 KiB or just to be on the safe side to 6144 KiB (6 MiB). I did this to my 840 EVO and it got much faster (from 310 MB/s to 510 MB/s on sequential buffered reads (I think, whatever hdparm -t does anyway, so a very poor benchmark!), but of course some of that could be explained by the fact that all data were re-written and that SSD is known to slow down when data has been laying on disk. So, it could be that this doesn't change anything, but it might.
http://www.tech-g.com/2015/10/03/ali...-disk-problem/

Source? I knew aligning to 1MB was enough for all SSDs.

[Sonadow]

I'm questioning whether Michael is really using native NVMe or if it is falling back to AHCI mode.


Because if it is the latter, I'm not interested in reading anything about NVMe drives running in AHCI mode on Linux.

[GrayShade]

Source? I knew aligning to 1MB was enough for all SSDs.

Because this disk has a 1.5M erase block 1536 KiB and to be sure we want it to also align with 2048 KiB (Just in case the erase block is not the whole story), you can set the sector alignment value to 12288 (6144 KiB), which is a multiple of 1536 KiB and 2048 KiB.

So basically, 12288 is 3*4k, the three comes from the fact that it is a three level cell (TLC)

Hmm...

[mlau]

What are the current plans to improve the performance here? Clearly this is a kernel issue, because these Samsung NVMe drives don't lose benchmarks to the 850 series in Windows.... that's surprising to see in Linux to me. In Windows, sometimes Samsung's AHCI drives have beaten their NVMe counterparts*, but the high-end drives never get outperformed by the significantly lower-end drives.

*Edit to note: Not all of Sammy's NVMe drives have had AHCI equivalents but some like the XP941 have.

I read somewhere that Samsungs windows drivers don't use forced unit access (FUA) like the vanilla windows nvme and ahci drivers do. FUA does kill performance because it forces the disk to write data back to the medium (instead of keeping it in a volatile cache) before it can report the request as completed, but it also increases reliability.

[Michael]

I'm questioning whether Michael is really using native NVMe or if it is falling back to AHCI mode.


Because if it is the latter, I'm not interested in reading anything about NVMe drives running in AHCI mode on Linux.

They were in NVMe mode.

[darkbasic]

Hmm...

Yeah, but do you have any source?

[Tomin]

Source? I knew aligning to 1MB was enough for all SSDs.

I provided a link in the post, that's the source. That's all I know and the blogger seems to be certain that some Samsung SSDs have erase block size of 1536 KiB. It would be nice if the manufacturers did tell these things.

Usually the 1 MiB alignment is enough, but it seems that Samsung TLC (newer EVO drives) need something better.