MicroSD Card performance variance on XO-1.5

Tomeu Vizoso tomeu at sugarlabs.org
Sat Aug 7 05:27:27 EDT 2010


On Sat, Aug 7, 2010 at 02:52, Bernie Innocenti <bernie at codewiz.org> wrote:
> El Fri, 06-08-2010 a las 18:50 +0200, Tomeu Vizoso escribió:
>> > I also think our vm.dirty_* settings are wrong and likely causing our
>> > current fill-buffer-and-stutter behaviour. We are using the defaults
>> > and those are for spinning harddrives, with a significant cost to
>> > spinning up the disk on a laptop -- hence long expire_centisecs and
>> > writeback_centisecs and the assumption that once you've started
>> > writing, it'll be written out fast. We have zero seek costs, zero
>> > "spin disk up" costs, but slowish writes -- we have to start writing
>> > buffers out ASAP.
>>
>> I had also had the impression for some time that some knobs in the
>> kernel could have been better tuned for the XO-1 and the actual
>> workload.
>
> That was also my gut feeling.
>
> Swappiness also seems to be set too high. We should probably let the OOM
> killer kick in before the machine becomes totally unresponsive.

Btw, have read that some notifications about available memory have
landed in cgroups in recent kernels. The Sugar shell could listen to
those and give a chance to background activities to save their state
before killing them, thus avoiding OOM in some (most?) cases.

Regards,

Tomeu

> Tuning the VM knobs is no joke. We could proceed by applying some
> reasonable changes early in the next development cycle, then sit back
> and wait for users to tell us their overall impression of using the
> system as usual.
>
> --
>   // Bernie Innocenti - http://codewiz.org/
>  \X/  Sugar Labs       - http://sugarlabs.org/
>
>



More information about the Devel mailing list