On 02.09.26 11:41, Abdelkader Boudih wrote:
Hi all,
I am about to land the D59156 stack in about 24 hours:
https://reviews.freebsd.org/D59156
It is a stack of 6 commits, D59151 --> D59156.
This stack allows NVMe drives to run at full speed and use all
available lanes, instead of being quirked to use only one.
Older NVMe drives will probably see no performance improvement. People
with modern NVMe drives can see improvements between 2x and 13x. (from
tests)
On Apple machines, I measured up to 7.2x. Entreprise drive went from
800MB/s to 12GB/s in a PCIe-5 slot.
The system itself doesnt become 7x faster, but file transfers and
disk-heavy workloads become snappier.
I'd like people to test the stack, or at least be aware that the next
update will have it.
I have tested it with over 40 different NVMe models. The only
problematic one was the Apple T1 S3X NVMe, which wedged during
testing. It already has a quirk.
So, to avoid anyone being unhappy that such a blessing landed without
warning, here is the warning. :)
I fail to see how the six patches are supposed to change the negotiated
link width of NVMe drivers attached via PCIe. Does the claimed
improvement in throughput and "snappiness" come purely from queuing the
right number of requests in flight? Which hardware did you use for
testing? Do you have a recommended test regime e.g. just dd, mount + tar
to /dev/null, a specific fio invocation?
-- Jan Bramkamp
