If I'm going to micro-optimize cold-cache access, I'll need to have some way to flush or bypass all of the caches between fastdupes and the drive platters.
I haven't yet found a way to force-bypass the hardware caches. I'll keep looking but I may simply have to do something like reading 256MiB of irrelevant data in between flushing the other caches and starting the test.
I still need to verify that there are no caches I've missed.
If I'm going to micro-optimize cold-cache access, I'll need to have some way to flush or bypass all of the caches between fastdupes and the drive platters.
/proc/sys/vm/drop_cacheswould cover all of the in-kernel caches.