Re: Performance variation in traditional HDDs?

On 4/10/25 3:41 AM, Nick Neumann wrote:
> We recently benchmarked some identical HDDs (same part number,
> purchase date, and firmware) with fio and were surprised by speed
> differences between them. Intrigued, we bought a few more sets of
> identical drives and benchmarked them. It makes sense that there would
> be some variation - after all, these are rotating platters and
> physical rotation speed may vary a bit.

Very unlikely. You are talking about some very precise control in the drive here.

> All sets of identical drives showed some variation, with some showing
> significantly more than others. Note that when the exact same drive is
> rebenched, it generally shows the same results. So different instances
> of the same drive really do perform differently.
> 
> For example, we benched 6 of the same 2TB 7200 RPM drive and did a
> full drive write with fio. The slowest drive averaged 158MB/s over the
> run. The fastest averaged 174MB/s. And the average of all drives was
> 169MB/s.
> 
> I was wondering if anyone had pointers/thoughts on what kind of
> variation should be expected in performance with HDDs, or
> articles/research about this behavior?

Did you mount all drives exactly the same way ? Do the HDD mounts you use have
proper vibration dampening ?

Generally, HDD performance variations come from different vibration
characteristics, due to external factors, including noise.
E.g. a drive in an enclosure that is close to a fan may not perform as well as
another drive that is mounted further away from the fan.

Given the differences you are showing, I would say that it is likely that you
have HDD mounting problems. Of course, testing with the drive resting on a
table with no mounting is a non-starter...

Another thing: what fio command line did you use ? If you did not use direct
IOs at high queue depth, you can have problems getting the maximum speed.
Generally, something like this is good:

fio --name=test --filename=/dev/sdX --ioengine=libaio --iodepth=32 \
	--rw=read --direct=1 --bs=1M

-- 
Damien Le Moal
Western Digital Research