I doubt that FS alignment/closeness to end of the device matters. One obscure possibility is that ext4 has a bug if metadata/fs writes complete asynchronously, but I highly doubt it. NVMe is multi-core multi-threaded.
I'm still of the opinion that there is some latent corruption in your FS exposed by parallel access somehow.
Aside: I've found out that we've not been declaring current limits correctly for SD VDD1 - it defaults to some pointlessly small value which means the card defaults to 0.72W max power draw, if it pays attention to such limits. Pi 5 can deliver 2.88W (peak) to an SD card easily. In testing it seems to make a slight improvement to sequential write throughput for the fastest cards - but still tops out at about 63MB/s.
I'm still of the opinion that there is some latent corruption in your FS exposed by parallel access somehow.
Aside: I've found out that we've not been declaring current limits correctly for SD VDD1 - it defaults to some pointlessly small value which means the card defaults to 0.72W max power draw, if it pays attention to such limits. Pi 5 can deliver 2.88W (peak) to an SD card easily. In testing it seems to make a slight improvement to sequential write throughput for the fastest cards - but still tops out at about 63MB/s.
Statistics: Posted by jdb — Wed Apr 10, 2024 5:05 pm