Please enable JavaScript.
Coggle requires JavaScript to display documents.
One 7TB NVMe is slower on rhel8 than rhel7 - Coggle Diagram
One 7TB NVMe is slower on rhel8 than rhel7
observations
tests
hdparm -t /dev/nvme1n1
1/5 throughput of rhel7
hdparm -t --direct /dev/nvme0n1
same with rhel7 basically
only seen on one 7TB NVMe
hdparm -t
rhel8 takes ~3ms for each read(2MB)
rhel7 takes ~0.6ms for each read(2MB)
rhel8 vs. rhel7
readahead window size
rhel7
128K
rhel8
big as max_sectors_kb
doesn't make difference after chaning max_sectors_kb to 128k
how to investigate
ftrace on read()
read 'read ahead' code
run 'fio buffered randread', and see if there is regression