RPI3 swap experiments, was Re: GPT vs MBR for swap devices
bob prohaska
fbsd at www.zefox.net
Tue Jun 26 22:28:26 UTC 2018
On Tue, Jun 26, 2018 at 01:15:54PM -0700, Mark Millard wrote:
> On 2018-Jun-26, at 8:18 AM, bob prohaska <fbsd at www.zefox.net> wrote:
>
> > On Tue, Jun 26, 2018 at 07:37:59AM -0700, Mark Millard wrote:
> >>
> >>
> >> . . .
> >>
> >> As I remember, Bob P. Did reproduce drive errors even without
> >> the problem drive being used for swapping. This too suggests
> >> (A) as separate activity.
> >>
> > Indeed, it is a requirement. If the suspect device is used for swapping
> > OOMA kills prevent the test from progressing to the point of failure.
> >
>
> Looking back at http://www.zefox.net/~fbsd/rpi3/swaptests/
> and information about /dev/da0 rive errors it does not
> appear that a combination with:
>
> A) sufficient swap (> 1.5 GiByte total?) but no use of swap on
> any partition on /dev/da0
> and:
> B) use of /dev/da0 for /usr/ and /var/
> and:
> C) Records from the console showing errors (or notes
> indicating lack of such errors).
>
> exists. So I was remembering incorrectly.
>
> I'm not claiming such a combination is the best direction for
> the next tests, but absent such tests there is no
> compare/contrast to know if /dev/da0 would still get errors
> despite the system having sufficient swap present on other
> drives. Thus, I would not go so far as "is a requirement" on
> the evidence available.
>
I just didn't bother to record successful runs. I'm logging one now.
> We do have evidence for the system having insufficient swap
> space: this context seems to have the current status "is
> sufficient but might not be necessary" for /dev/da0
> getting drive errors.
>
Not sure I understand here. Basically there seem to be three cases:
Enough swap not on da0, -j4 buildworld completes.
Any swap on da0, -j4 buildworld is killed by OOMA
Not enough swap not on da0, -j4 buildworld crashes the machine eventually.
Are there other combinations I've overlooked? The first two don't seem
worth repeating, at least not often.
>
> As for simpler contexts, one that would swap but would be
> far simpler a context than buildworld buildkernel might be
> something like using the stress port via options like:
>
> stress -d 2 -m 3 --vm-keep
>
> (The option values likely could need adjustment from context
> to context to match available resources. The above is not
> carefully tailored to your context or a modern context.
> It dates back to 2016-Jan-22 for showing vnode based swap
> failures in a 1 GiByte RAM + 1 GiByte swap-file context
> [inside virtualbox on amd64 hardware]: see comment 3 of
> https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=206048 .)
>
>
This is new to me and entirely separate from Peter Holm's stress2.
It compiled without a hitch and seems worth a try.
Thank you!
bob prohaska
More information about the freebsd-arm
mailing list