Announcement

Collapse
No announcement yet.

Samsung 970 EVO NVMe SSD Benchmarks On Ubuntu Linux

Collapse
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • Samsung 970 EVO NVMe SSD Benchmarks On Ubuntu Linux

    Phoronix: Samsung 970 EVO NVMe SSD Benchmarks On Ubuntu Linux

    Last month Samsung introduced the 970 Series solid-state drives with the mainstream 970 EVO models and 970 PRO models for professionals/enthusiasts. The 970 Series moves to a 64-layer flash and uses a five-core Phoenix controller. For those curious about the Samsung 970 EVO performance under Linux, I have carried out some quick benchmarks to show off its potential under Ubuntu.

    http://www.phoronix.com/vr.php?view=26447

  • #2
    I noticed that using 3 MiB partitions alignment instead of common 1 MiB produces better results with Evo 970. See here.

    I was testing random read / write performance though, which you didn't include.
    Last edited by shmerl; 06-12-2018, 03:39 PM.

    Comment


    • #3
      Originally posted by shmerl View Post
      I noticed that using 3 MiB partitions alignment instead of common 1 MiB produces better results with Evo 970. See here.

      I was testing random read / write performance though, which you didn't include.
      I noticed this as well with my 960 Evo in raid0, when I created mdraid / f2fs filesystem on whole blockdevice, sequential read write performance was ~30% higher than when I had raid data partition + mirrored boot partition (1 MiB aligned).

      At the moment I simply use my old SSD for boot partition and mount raid from builtin initramfs inside kernel image.

      Comment


      • #4
        What I understood from various scattered info, it's because TLC NAND uses 3 bits per cell (that's T in Triple Level Cell). So erase block size for such SSD should be divisible by 3 and some suggested that for Evo-s, it's 1536 KiB. And it's generally beneficial to align partitions according to the erase block size. Samsung though are notorious for dodging any questions about their erase block size.
        Last edited by shmerl; 06-12-2018, 05:43 PM.

        Comment


        • #5
          Originally posted by shmerl View Post
          What I understood from various scattered info, it's because TLC NAND uses 3 bits per cell (that's T in Triple Level Cell). So erase block size for such SSD should be divisible by 3 and some suggested that for Evo-s, it's 1536 KiB. And it's generally beneficial to align partitions according to the erase block size. Samsung though are notorious for dodging any questions about their erase block size.
          From what I understand, this _shouldn't_ matter because IC on SSD _should_ handle all this crap behind the scenes, trasparently to the user and even OS.
          https://github.com/bradfa/flashbench

          flashbench -a /dev/nvme0n1 --blocksize=1024
          flashbench -O --erasesize=$[4 * 1024 * 1024] --blocksize=$[256 * 1024] /dev/nvme0n1 --open-au-nr=3


          Variance seems quite a bit higher than I would like, but I don't see any pattern or clear winners here (also tried with bs=4096/16/256k, /dev/md0, results seems to be basically uniform no matter what).
          Last edited by tpruzina; 06-12-2018, 07:16 PM.

          Comment


          • #6
            What's up with the standard deviation in some of the tests? (Or am I reading them wrong?)

            Comment


            • #7
              Never heard of flashbench, interesting tool.

              Comment


              • #8
                You forgot about Optane SSD in the benchmark.

                Comment


                • #9
                  Not just Optane, where's the 960 EVO & 960 PRO?

                  Comment


                  • #10
                    Not just Optane, where's the Samsung 960 EVO & 960 PRO in your tests.

                    Comment

                    Working...
                    X