On Sat, 1 Nov 2025, at 11:31, Pádraig Brady wrote:
>> Interesting. Do you happen to have any performance benchmarks from when
>> fadvise was added to 'cp'? I assumed that the impact of it would be
>> minimal, but that report proves otherwise at least for ZFS.
>
> We saw a 5% speedup when reading from SSD:
> https://github.com/coreutils/coreutils/commit/15f1d0c9a
> I measured a more significant speedup with spinning disks.

Reading from SSD (or spinning disk, or hardware RAID controller) will do better 
with well-tuned readahead, vs waiting for the read(2) call from userspace. The 
queue depth of SSDs is especially beneficial when we can obtain simultaneous 
outstanding read opcodes.

Perhaps POSIX_FADV_SEQUENTIAL causes readahead to kick in sooner? I'd hope the 
kernel heuristics aren't too complex for sequential read, but haven't poked 
around in Linux.


Cheers,
Phil

Reply via email to