Re: [gentoo-dev] Re: Last rites: www-client/chromium-bin

[Rich Freeman]

On Sat, Mar 5, 2011 at 5:00 AM, Nikos Chantziaras <realnc@×××××.de> wrote:
> On 03/05/2011 04:41 AM, Rich Freeman wrote:
>> I need to
>> make sure I have /var/tmp/portage symlinked back to a non-tmpfs
>> location whenever I build it or else the system pretty-much dies from
>> a lack of RAM.
>
> Then I'd say you have your tmpfs mount options configured wrongly :-) You
> should specify a maximum amount of RAM it should use.  When it fill up, it
> then uses swap.
>

So, your later redaction aside, I should say that the kernel devs have
their swap behavior wrong in any case.

Think about it - if you write to a disk-based filesystem every byte
that you write ends up on disk, subject to cache (with little
discretion - all data needs to be written out within 30 seconds or
so).

If you have very little RAM free the tmpfs should really be no worse
since there too every byte that gets written gets swapped to disk.
However, the kernel has complete discretion about when this happens,
whether this happens at all, and it gets to write it in a partition
optimized for this purpose without regard for long-term
preservation/etc.  Yet, for whatever reason all this swapping seems
slower.

The likely reason is that the kernel isn't that good at figuring out
what to swap and what not to swap.  It might be prioritizing keeping
some object files around in RAM on a tmpfs while trying to swap in/out
some other process that keeps getting woken up or whatever.

Better swap behavior should let the tmpfs perform better than
ext4/etc, since it gives the kernel more discretion about what to
write out and when.