#Slow Buffered Read Performance in mdadm NVME raid0

17 messages · Page 1 of 1 (latest)

random hawk
#

I created a Raid 0 array of 2 identical 2TB SSDs and I am getting pretty slow buffered read performance on them. The top result is my other single boot NVME drive which is a better model drive. I am wondering if I am doing something wrong.

image.png
#

Here is the results if I just run hdparm on 1 drive, vs running it on the md0 raid 0 array

image.png
#

OK. It appears there is something up with the 2nd drive in the array...

image.png
#

I don't know why that would would be preforming so poorly

random hawk
#

/dev/nvme2 has passed bote a short and extended SMART test

#

I think I am just going to secure erase both drives and test them again. I can move the data off of them and back to them easily enough.

bright lodge
#

I'm curious to see results with fio with and without the direct=true option ?

#

I am using btrfs tho, but I get vastly worst performance (~500MiB/s) without O_DIRECT vs ~7GiB/s on a Raid1 that total ~14GiB/s without a filesystem.
After using perf I've find out my poor Ryzen3600 can't keep up and spend most of the time just memcpying between the different cache and buffers. O_DIRECT skip that *but that require specific application support, there is a new patch set to do that while being compatible but idk if it were merged yet

random hawk
#

I haven't destroyed the array yet, I just got done copying all the data off. I will look into fio right now

#

Here is the results without changing the direct setting.

#

The results using --direct=0

#

I didn't know how to test them independently using fio

#

But here is the result for the boot drive for comparision

image.png
#

Here is the results with --direct=1

image.png
#

It appears according to that the nvme2 is bottlenecking the array for some reason

#

At least according to that utilization %