Known Linux issues re: Lifetime of low-power USB hard drives

[sourcejedi]

EDIT: this post contains inline links. At least on my computer, they're not showing up unless you hover the cursor over them. Or press TAB to cycle through them. So I've marked them with $dollar signs$ as well.

A notice in particular for anyone with an always-on server, Debian Linux on ARM, and a low-power USB hard drive. Maybe people it's just me, maybe people already know exactly what I'm going to say, but I didn't find anything about this.

Here's what my drive looked like when I realized I needed to check it. It's been spun up for maybe half a year:

Code: Select all

$ MY_DRIVE_OPTS="-d sat /dev/sda"
$ sudo smartctl -a $MY_DRIVE_OPTS | grep Load_Cycle_Count
193 Load_Cycle_Count        0x0032   109   109   000    Old_age   Always       -       273664

For those not familiar with smartctl, some explanation may be needed .

I believe the normalized SMART value (109 in this case) are limited to the range 0-255. I recorded the SMART attributes when I bought the drive, and the initial value was 200. Lower values are worse. "000" is the "threshold", the value at which this attribute is considered to indicate "Old_age". From this we can guess that the drive has been exercised almost half-way to old age in six months - not good.

273664 is the raw value, literally how many times the drive heads have been unloaded and unloaded. And online sources confirm that many drives are rated for a lifetime of 600000 load cycles. (This is distinct from spinning the drive up/down - that's something that happens less often, and is represented by a different SMART attribute).

This issue was widely responsible for reducing the lifetime of hard drives of laptops when running Linux. As you'd hope, it was fixed quite nicely - but it seems only for laptop drives on x86.

From personal experience on Debian, and from a quick look at the $umbrella Launchpad bug$, the fix for excessive rates of increase in Load_Cycle_Count was implemented in the acpi_support package. Currently no Linux ARM platforms use ACPI, and it looks like the acpi_support package can't be installed on ARM.

$Note that configuring and running smartd isn't enough to warn you about an excessive rate of increase in Load_Cycle_Count.$ You need to run smartctl and look at how fast it's changing over time.

What would be a reasonable rate? Well, the $kernel wiki$ suggests I shouldn't complain. If I saw the equivalent rate on a laptop being using 12 hours a day, it might last 2 years before exceeding the rated figure. That may be a useful starting point, but I'm not happy with the idea of a 1 year rated life for my 24h drive.

So now I editted Debian's /etc/hdparm.conf (this should work well if you're booting off the drive, otherwise you need to worry about boot order, i.e. does the hdparm script run before the drive has been detected).

Code: Select all

/dev/disk/by-id/usb-WD_My_Passport_0740_575838314141315837363139-0:0 {
        apm = 254
}

I think 254 is supposed to work for most drives. 255 might be necessary for a few. And I think the launchpad bug implies some might be ok with 128, but I haven't tried that yet. Some drives may be hopeless and not take notice of any setting.

Then rebooting it would have been annoying, so I took a shortcut

$ sudo /etc/init.d/hdparm restart

and now it's been spinning for several hours with the raw value of Load_Cycle_Count holding steady at 273683. So I'm much happier about it (though I haven't compared the power consumption yet .

[caldimerda]

many thanks for this info sourcejedi

I've just set up a Raspi-based fileserver/webserver with a nice low-power portable hard disk and this would certainly have caught me out. As it was I was googling for smartctl and picked up your post as I was starting the config.

Right now I have a Load Cycle Count of 49 and it's creeping up in low single figures per day. Whatever kills the disk, it doesn't look like it'll be that

[sourcejedi]

Glad to hear someone found it useful .

[poing]

For Western Digital HDs this might be of interest: http://idle3-tools.sourceforge.net/