[circuitboard]

But let's be honest. I'm not on this forum because I want practical computing.

Single NVMe

3 very fast (but by no means identical) NVMes in a RAID 0 array

Stripe size is max 128k.

 

[circuitboard]

its not mb/s. its latency. as I said above, for some reason they switched units.

I am not sure why my first result looks like that. It's sort of impossible to replicate now that I am RAIDed. My first screenshot doesn't have a dropdown for units which is sort of bothering me. I tried to leave everything on default so it would be the same. The text files should have all the information.

The first drive is a 2 TB Acer Predator 2280 NVMe.

------------------------------------------------------------------------------
CrystalDiskMark 8.0.4 Shizuku Edition x86 (C) 2007-2021 hiyohiyo
                                  Crystal Dew World: https://crystalmark.info/
------------------------------------------------------------------------------
* MB/s = 1,000,000 bytes/s [SATA/600 = 600,000,000 bytes/s]
* KB = 1000 bytes, KiB = 1024 bytes

[Read]
  SEQ    1MiB (Q=  8, T= 1):  7104.573 MB/s [   6775.4 IOPS] <  1179.49 us>
  RND    4KiB (Q= 32, T=16):  4553.659 MB/s [1111733.2 IOPS] <   459.76 us>

[Write]
  SEQ    1MiB (Q=  8, T= 1):  6660.250 MB/s [   6351.7 IOPS] <  1255.83 us>
  RND    4KiB (Q= 32, T=16):  4245.036 MB/s [1036385.7 IOPS] <   492.95 us>

Profile: Peak
   Test: 1 GiB (x5) [C: 20% (380/1907GiB)]
   Mode: [Admin]
   Time: Measure 5 sec / Interval 5 sec
   Date: 2023/03/14 23:20:26
     OS: Windows 11 Professional [10.0 Build 22621] (x64)

The array is composed the Predator and brand-new Samsung 980 pro along with a gently used WD Black. It is controlled by the CPU.

------------------------------------------------------------------------------
CrystalDiskMark 8.0.4 Shizuku Edition x64 (C) 2007-2021 hiyohiyo
                                  Crystal Dew World: https://crystalmark.info/
------------------------------------------------------------------------------
* MB/s = 1,000,000 bytes/s [SATA/600 = 600,000,000 bytes/s]
* KB = 1000 bytes, KiB = 1024 bytes

[Read]
  SEQ    1MiB (Q=  8, T= 1): 19365.184 MB/s [  18468.1 IOPS] <   432.43 us>
  SEQ  128KiB (Q= 32, T= 1): 18480.312 MB/s [ 140993.6 IOPS] <   223.02 us>
  RND    4KiB (Q= 32, T=16):  1355.159 MB/s [ 330849.4 IOPS] <  1546.10 us>
  RND    4KiB (Q=  1, T= 1):    71.940 MB/s [  17563.5 IOPS] <    56.86 us>

[Write]
  SEQ    1MiB (Q=  8, T= 1): 15852.066 MB/s [  15117.7 IOPS] <   528.45 us>
  SEQ  128KiB (Q= 32, T= 1): 14944.341 MB/s [ 114016.3 IOPS] <   279.50 us>
  RND    4KiB (Q= 32, T=16):  1205.821 MB/s [ 294389.9 IOPS] <  1738.20 us>
  RND    4KiB (Q=  1, T= 1):   176.593 MB/s [  43113.5 IOPS] <    23.11 us>

Profile: Default
   Test: 1 GiB (x5) [C: 7% (379/5588GiB)]
   Mode: [Admin]
   Time: Measure 5 sec / Interval 5 sec
   Date: 2023/03/16 0:04:03
     OS: Windows 11 Professional [10.0 Build 22621] (x64)

Q32, T16
The Acer had over a million IOPS in half a millisecond. The array had less than a third of that over more than 1.5 milliseconds. It looks like the naysayers are vindicated. I sacrificed real-world performance for a synthetic sequential read/write benchmark.

I am not done messing around with RAID. This configuration seems to be inviting drive failure. I am going to put RAID 1/0 on a discrete controller. I am expecting read improvements but a drop in write speed. I also want to try dropping my stripe size down to 4K and see how that affects it.