Why does fio's performance seem relatively low when running IO with PI?

[MorningWeicheng]

I used fio to test the IO performance with PI on several different NVME SSD drives, and found that the performance in scenarios with PI configured is lower than that without PI. Has anyone encountered this phenomenon? Could you help confirm whether it's due to incorrect command configuration of fio with PI, or if there are some limitations in fio's PI configuration? Thanks a lot.
Here comes my fio cmd:
sudo fio -filename=/dev/ng0n1 -iodepth=128 -rw=write -bs=131328 -numjobs=1 -name=test -direct=0 -ioengine=io_uring_cmd -size=100% -pi_act=0 -pi_chk=GUARD,APPTAG,REFTAG --runtime=30

[vincentkfu]

I used fio to test the IO performance with PI on several different NVME SSD drives, and found that the performance in scenarios with PI configured is lower than that without PI. Has anyone encountered this phenomenon? Could you help confirm whether it's due to incorrect command configuration of fio with PI, or if there are some limitations in fio's PI configuration? Thanks a lot.
Here comes my fio cmd:
sudo fio -filename=/dev/ng0n1 -iodepth=128 -rw=write -bs=131328 -numjobs=1 -name=test -direct=0 -ioengine=io_uring_cmd -size=100% -pi_act=0 -pi_chk=GUARD,APPTAG,REFTAG --runtime=30

Fio needs to do more work when PI is enabled. Also make sure that fio is built linked to isa-l which considerably speeds up checksum calculations.

[MorningWeicheng]

Thanks! After check fio built with isa-l I found that IO performance seems much normal.

[MorningWeicheng]

I used fio to test the IO performance with PI on several different NVME SSD drives, and found that the performance in scenarios with PI configured is lower than that without PI. Has anyone encountered this phenomenon? Could you help confirm whether it's due to incorrect command configuration of fio with PI, or if there are some limitations in fio's PI configuration? Thanks a lot.
Here comes my fio cmd:
sudo fio -filename=/dev/ng0n1 -iodepth=128 -rw=write -bs=131328 -numjobs=1 -name=test -direct=0 -ioengine=io_uring_cmd -size=100% -pi_act=0 -pi_chk=GUARD,APPTAG,REFTAG --runtime=30

Fio needs to do more work when PI is enabled. Also make sure that fio is built linked to isa-l which considerably speeds up checksum calculations.

Hi,vincentkfu,I still got a problem: It seems that in the current fio code, the isal library is only used when calculating 16-bit CRC, while the calculation of 32-bit and 64-bit CRC still does not call isal. Is there any way to improve the performance for 32-bit and 64-bit cases at present? If not for the time being, are there any plans for optimization in the future? Tks！

[vincentkfu]

If isa-l supports the required 64-bit CRC calculation then it would be straightforward to have fio use the library.

Fio currently does not support 32-bit CRCs. I know of no current plans to add this feature.

[vincentkfu]

isa-l gained support for NVMe CRC64 calculations in May 2023.

intel/isa-l@2bbce31

Many platforms are still using an older isa-l version without support for this feature.

Please test this branch which uses isa-l if support is found for CRC64:

https://github.com/vincentkfu/fio/tree/isa-l-crc64

[MorningWeicheng]

Thanks, I tested with this branch, the performance are close to normal values under 64bit CRC check.