We recently moved to Ubuntu 23.10 from 20.04 as our main image for benchmarking SSDs. In doing so, we noticed that many devices showed a significant drop in random I/O tests, with the drop more pronounced the faster the device was. We determined the root cause of the drop was the mitigation for the recent AMD SRSO vulnerability: https://docs.kernel.org/admin-guide/hw-vuln/srso.html A notable example is random reads on an m.2 Optane device: sudo fio --filename=/dev/nvme0n1 --name=job --ioengine=libaio --direct=1 --rw=randread --bs=4K --iodepth=8 --runtime=15 --time_based=1 With AMD SRSO mitigated: 1249 MiB/s, 320k IOPS With AMD SRSO mitigation disabled: 1682 MiB/s, 430k IOPS I was looking for feedback on best practices for deciding to disable mitigations when benchmarking SSDs. In general I'm hesitant to disable mitigations, and all devices are benchmarked with the same settings. It can be argued that yes, bandwidth is reduced, but the mitigation is similar to using benchmark hardware with a slower CPU, etc. As long as all SSDs are benchmarked with the same hardware and settings, all that is happening is a compression of the results - faster devices should still be faster, just not by as much. But this particular mitigation has such a big effect that it makes me consider disabling it. Thoughts?
