Re: Performance variation in traditional HDDs?

In the end, my goal is to understand enough about how HDDs work and
perform to know what we can reasonably expect to accomplish/learn with
using fio to benchmark them, and how to appropriately do the
benchmarking.

This last CMR drive has additional interesting performance
characteristics besides the differences between two of the same drive.
Zooming in on the beginning of a full drive write is surprising:

https://cdna.pcpartpicker.com/static/forever/images/content/c4c6d8413eef6effa17b2220b9414b6d.png

There is a very periodic behavior to bandwidth, with a period on the
order of a few minutes. The period of Drive 1 is actually a bit
shorter than that of drive 2. Drive 2, the slower drive, is faster
than drive 1 for the first 30+ seconds of writing because drive 2
starts at the high point of its periodic oscillation, while drive 1
starts at the low point. For both drives, the difference between high
and low bandwidth over a span of a few minutes is surprisingly large
(initially >25MB/s for drive 1 and > 50MB/s for drive 2), and this
difference gradually decreases over time.

On Fri, May 2, 2025 at 2:33 PM Nick Neumann <[email protected]> wrote:
>
> Here's another CMR drive (https://pcpartpicker.com/product/6FQcCJ).
> Here, two identical drives with same model number and firmware show
> about a 10% difference in performance.
>
> Looking at the average full drive write speed we saw the following
> results in MB/s:
>
> Drive 1: 153.86, 153.86, 153.86
> Drive 2: 140.63, 140.64, 140.65
>
> Again, each individual drive is incredibly consistent in its
> performance. But drive 1 is over 9% faster than drive 2. And the
> graphs of bandwidth over time for a full write are markedly different:
>
> https://cdna.pcpartpicker.com/static/forever/images/content/684ce1fea2858681575fb7666b8b987f.png
>
> Reporting benchmarked performance for HDDs is problematic when there
> is variation like this between two of the same drive.
>
> On Sun, Apr 27, 2025 at 12:23 PM Nick Neumann <[email protected]> wrote:
> >
> > On Wed, Apr 23, 2025 at 10:00 PM Damien Le Moal <[email protected]> wrote:
> > > So these are drive-managed SMR disks, right ?
> >
> > Yes. I chose these first because they show the largest performance
> > difference between two "identical" drives. We do an ATA SECURITY ERASE
> > to reset their state/LBA mapping - it takes longer than a full drive
> > write.
> >
> > Here is another set of identical drives
> > (https://pcpartpicker.com/product/XtjG3C) that just completed and also
> > show a significant performance difference. These are CMR drives with
> > the same model number and firmware.
> >
> > We did three full drive writes on each drive, using the same proper
> > horizontal mounting and position.
> >
> > Looking at the average full drive write speed we saw the following
> > results in MB/s:
> >
> > Drive 1: 162.72, 162.73, 162.73
> > Drive 2: 155.69, 155.70, 155.72
> >
> > Each individual drive is incredibly consistent in its performance. But
> > drive 1 is over 4% faster than drive 2.
> >
> > The difference can be seen in a graph comparing one full write run of each:
> > https://cdna.pcpartpicker.com/static/forever/images/content/221396115ee6bae8fbd35135fbc6441e.png