NVMe drive performance

Expecting the same results as for SSD? Then you’re mistaken…

You might think a separate test for NVMe drives would be superfluous, but it’s not (otherwise we wouldn’t have bothered with it). Our five filesystems displayed quite different performance on the blazingly fast Samsung EVO 970 drive with non-volatile memory, so let’s see exactly how it differed from the legacy SSD drive performance.

First of all, Ext4 and XFS yielded the best results, while Btrfs dropped by one-third and looked more like Reiser5 in terms of speed in the Postmark test. We checked that deviation using dd , which enabled us to observe the reading speed of a large file, but it only confirmed our initial findings. Notice that this time all three tests indicated that Btrfs did not perform as well on an NVMe drive as it did on SSD. That said, Btrfs on the root partition of an NVMe drive will eventually result in slower application startup times and should probably not be used that way.

Another difference from the SSD performance is that the much faster nature of NVMe effectively smoothed out the differences between Btrfs, EXT4 and XFS. That means that on smaller I/O tasks (such as our test archive untar routine), there’s virtually no sense in considering one filesystem over any other – they all deliver practically the same performance. Reiser5 gained entry to the speedy folks club when it came to data unpacking, but it lagged behind by around 30 to 40 per cent in the two other tests. Nevertheless, Reiser5 showcased a decent daily driver level of performance on NVMe, and that’s definitely good news. By the way, NTFS is still a big no in Postmark, but a firm yes in Sysbench. It’s just a matter of the type of things you do with an NTFS partitions in Linux that makes it feel either sluggish or snappy.