Great Job Waddlesplash - NVME Related


#1

file

CPU Usage was 50% over 1 Core and latest Nightly was used

I’d love to see the CPU usage come down and speed go up (it’s a Samsung PM981 and under Windows does 3.5GB/s with less than 10% CPU usage on this system)

But still NVME support that’s Fast Love it


#2

Are you using the same block size as the benchmark on windows? That has a significant impact on sequential speed. As the larger block sizes mitigate overhead.


#3

:slight_smile:

Yes, most of the overhead is probably the allocator calls, we could certainly trim those down, but for now it’s not a priority – most of this overhead occurs on dd only, as for in-kernel read requests an integer multiple of the device block size, which the kernel file cache and thus BFS always do, we use the passed buffers directly.


#4

Well I did some research and apparently CrystalDisk uses 1MB for it’s sequential tests
screenshot2


#5

well then you can rerun you test with bs=1M


#6

screenshot3


#7

I think the issue is that crystaldisk mark uses a VERY small part of the disk to test and that messes the numbers up bad

I would not be surprised to learn that a PM981 is really only 1.8GBs over the whole disk which is what Haiku is reporting

What I prolly SHOULD do is install Linux on this hardware and run the same test


#8

https://www.anandtech.com/show/12082/the-samsung-pm981-ssd-review-512gb-1tb-phoenix-3d-nand/6

Pretty sure Anandtech is not using crystaldisk.

https://ssd.userbenchmark.com/SpeedTest/413203/PM981-NVMe-Samsung-1024GB

Try running the test with only 1GB of data… it will probably be much faster as the SSD is probably thermal throttling, also I wouldn’t copy over my SSD like that unless I wanted it to fail sooner. The PM981 is rated for 150TBW (though most drives will exceed this amount before uncorrectable errors prevent it from working).